# SyntaxRadar: full article text > SyntaxRadar (https://syntaxradar.com) explains AI, chips and developer tools in plain language. Each article below is complete and based on a named primary source. Figures are attributed in the text to the source that reported them. Index: https://syntaxradar.com/llms.txt Generated from the live site. 48 articles, newest first. ================================================================================ - URL: https://syntaxradar.com/radar/ai-bubble-pop-warning-signs - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-31T12:00:00.000Z - Last modified: 2026-10-31T12:00:00.000Z - Primary source: https://techcrunch.com/2026/10/30/cracks-are-emerging-in-the-ai-spending-boom/ # AI Bubble Pop? 5 Warning Signs Behind Big Tech's $740B Bet **Short answer:** Alphabet fell 12% on Oct 29, its worst day on record, after a second negative free cash flow quarter. The Nasdaq fell 2.9% the next day, S&P cut Meta's outlook to negative, and the WSJ reported OpenAI raised less than half of its $30B target. **Takeaway:** This week's AI bubble warnings came from cash burn and record borrowing, while the revenue from AI and cloud customers kept growing. **Key figures (as reported by the source):** - Alphabet one-day drop (Oct 29): 12%, about $500B in value - Big Tech debt raised in 2026: $326B, triple last year - Amazon trailing 12-month free cash flow: -$42.4B Fear of an AI bubble pop hit hard this week. On Thursday, October 29, Alphabet lost 12% in a single session, its worst day on record, wiping out about $500 billion in market value. On Friday the Nasdaq slid 2.9%, with Nvidia, Broadcom and Qualcomm all falling hard. Is the AI bubble popping right now? The warning lights this week are about money going out the door faster than it comes in, and about who is lending to cover the gap. The demand side, the part where customers actually pay for AI and cloud services, is still growing. That distinction decides whether this is a crash in the making or a very expensive growing pain. Here is what the earnings reports and ratings agencies showed this week, what they did not show, and what it would mean for ordinary investors and workers if the boom does crack. ### What is the AI bubble, and why did fears spike this week? The AI bubble is the worry that spending and stock prices tied to AI have run ahead of the money AI actually earns. The trigger was a run of third-quarter earnings from the biggest spenders on AI infrastructure. Alphabet, Amazon, Meta and Microsoft have outlined plans to spend more than $740 billion on capital expenditures this year, [according to TechCrunch](https://techcrunch.com/2026/10/30/cracks-are-emerging-in-the-ai-spending-boom/). Capital expenditure, or capex, is money spent on long-lived things like data centers, chips and other equipment. We explained the term in detail in [what AI capex is and why Alphabet's bill spooked Wall Street](/radar/what-is-ai-capex-alphabet), so here we will focus on what changed. What changed is that investors stopped giving the spending a free pass. Alphabet's 12% fall came after it reported another quarter of heavy building. A day later the sell-off spread to the chipmakers that supply the boom: Nvidia dropped 4%, Broadcom more than 5% and Qualcomm 11%. When the companies buying the chips and the companies selling them fall together, the market is questioning the whole supply chain at once. ### Is there an AI bubble? Five warning signs in the numbers Nobody can prove a bubble until it pops, but you can check the classic symptoms. Five of them showed up this week. #### 1. Free cash flow turned negative Free cash flow is the cash a company has left after paying for operations and capex. It is the money that funds dividends, buybacks and a cushion for bad years. Three of the four giants reported negative free cash flow. Alphabet reported negative free cash flow of $9.1 billion in Q3, its second straight negative quarter, on quarterly capex of $51.4 billion. Meta posted negative free cash flow of $2.1 billion. Amazon's trailing 12-month free cash flow was negative $42.4 billion, compared with positive $14.8 billion a year earlier. #### 2. The spending keeps climbing Alphabet expects to spend up to $205 billion on capex this year and says spending will keep rising in 2027. Amazon expects about $220 billion. Meta guided to as much as $145 billion. #### 3. Borrowing hit a record Microsoft, Alphabet, Amazon, Meta and Oracle raised a combined $326 billion in debt in 2026, triple last year's total, according to S&P Global Market Intelligence data cited by TechCrunch. Meta also said on October 26 it plans to raise $95 billion: $35 billion through underwritten public stock offerings and $60 billion through an at-the-market program, which sells shares gradually into the market. #### 4. A ratings agency blinked On October 30, S&P Global Ratings revised Meta's outlook from stable to negative, [as CNBC reported](https://www.cnbc.com/2026/10/30/meta-outlook-to-revised-to-negative-by-sp-after-cash-flow-negative.html). The rating itself stays at AA-, which is still very strong, but S&P cited uncertainty about elevated capex and negative free cash flow. A Meta spokesperson said the company "respectfully disagree[s]" with S&P's assessment of how it will make money from the spending. Meta stock fell 15% in October, its worst month in four years. #### 5. Private AI money got harder to raise OpenAI raised less than half of its $30 billion target in its latest round, according to a [Wall Street Journal report](https://www.wsj.com/tech/ai/openai-raises-less-than-half-of-30-billion-target-in-latest-round-91aecdc5). Anthropic filed for its IPO the same week, and the Financial Times reported on October 30 that Blackstone warned it may not qualify for an investment-grade rating. We covered the filing in [the five numbers that matter in Anthropic's IPO prospectus](/radar/anthropic-ipo-prospectus-five-numbers). ![Cards showing AI bubble free cash flow: Alphabet negative $9.1B, Meta negative $2.1B, Amazon negative $42.4B, Microsoft positive $18.6B](/uploads/blog/ai-bubble-pop-warning-signs/img1.webp) ### What is not breaking: the demand behind the spending The bear case runs into trouble here, because the revenue lines still look healthy. Amazon's cloud unit, AWS, grew revenue 39%. Meta's ad revenue grew 26%. Microsoft posted $18.6 billion in positive free cash flow, and its capex came in at $45.2 billion, below the $50 billion it had projected. Microsoft blamed the shortfall on delays in third-party capacity. That detail matters. In a classic bubble, companies build capacity nobody wants. Here, one of the biggest buyers blames late third-party capacity for spending less than planned. There is one soft spot on the demand side. Google Search ad growth slowed to 12%, its weakest pace in five quarters, and Search is roughly half of Google's revenue. If AI answers start eating into the ads that pay for Alphabet's build-out, that would be a real problem. Recheck the number in three months to see whether one quarter becomes a trend. ### Are we in an AI bubble? Weighing both sides Put the two lists side by side and a clearer picture forms. The cracks this week are in financing, while demand holds up.
SignalPoints toward a bubblePoints away from one
CashAlphabet, Meta and Amazon free cash flow negativeMicrosoft $18.6B positive free cash flow
SpendingOver $740B planned capex across four companiesMicrosoft could not spend as fast as planned
Funding$326B in debt, Meta raising $95B, OpenAI short of targetMeta's AA- rating unchanged
RevenueSearch ad growth slowed to 12%AWS +39%, Meta ads +26%
The bull reading: these are the richest companies on earth, borrowing heavily (Meta still holds an AA- rating) to build something customers are already paying for. Negative cash flow during a build-out is normal, the same way a factory burns cash before it opens. The bear reading: the build-out has no visible end date, the cash cushion is gone at three of the four, and the companies are now relying on debt markets and stock sales to keep going. A boom that needs outside money to continue is more fragile than one paid for from profits. S&P's own stated worry was elevated capex and negative free cash flow. Both readings fit the same numbers. The difference is whether AI revenue grows fast enough to cover spending that, by Alphabet's own guidance, will be even higher in 2027. ![Cards showing demand behind the AI bubble fears: AWS revenue up 39%, Meta ad revenue up 26%, Microsoft capex $45.2B, Google Search ad growth 12%](/uploads/blog/ai-bubble-pop-warning-signs/img2.webp) ### Will the AI bubble pop? Nobody can time it. This week's data shows demand holding and the margin for error narrowing. Three things would turn an expensive build-out into an AI bubble pop. First, revenue growth stalling while capex keeps rising, for example if AWS growth slowed sharply. Second, lenders demanding much higher rates or refusing to lend, which Blackstone's warning on Anthropic and OpenAI's short round hint at. Third, a big buyer cutting its spending plans, which would ripple straight down to Nvidia, Broadcom and the rest of the chip supply chain. None of those has happened yet. All three are now being watched by investors who, until recently, were not watching closely at all. ### What happens if the AI bubble bursts? What happens if the AI bubble bursts depends on who you are. For investors, this week's slide is a small preview: Alphabet alone lost about $500 billion in value in a day, and Nvidia, Broadcom and Qualcomm fell with it. For tech workers, a spending pullback would likely hit the jobs closest to the build-out first: data center construction, chip supply chains and the startups whose business depends on cheap computing. For everyday users, the effect would be slower and quieter. AI features that cost a lot to run could be priced higher, limited or dropped. The tools would not vanish, because the data centers already built do not disappear, but the pace of new free features would likely slow. ### What to watch next Whether this week becomes the start of an AI bubble pop or a scare that fades will show up in a few specific places: - **Next quarter's free cash flow** at Alphabet, Meta and Amazon. A third negative quarter for Alphabet would harden the bear case. - **Google Search ad growth.** A second slowdown below 12% would suggest AI is cannibalizing the business that pays for it. - **Debt and equity raises.** Watch whether Meta completes its $95 billion plan and whether other hyperscalers follow. - **Credit ratings.** S&P changed its outlook on Meta, and the rating itself is unchanged. A real downgrade would raise borrowing costs across the sector. - **The Anthropic IPO.** How public investors price it will show whether private AI valuations hold up in the open market. For now, the evidence points to a boom that is straining its wallet while its customers keep paying. ## Frequently asked questions ### Is there an AI bubble right now? There are bubble-like warning signs. Alphabet, Meta and Amazon reported negative free cash flow, and five big tech firms raised $326 billion in debt in 2026, triple last year. But demand is still growing: AWS revenue rose 39% and Meta's ad revenue rose 26%. The strain is in financing the build-out, while customers keep paying. ### What is the AI bubble? The AI bubble is the worry that spending and share prices tied to AI have run ahead of the revenue AI brings in. This week's evidence: Alphabet, Meta and Amazon reported negative free cash flow, and five big tech firms raised $326 billion in debt in 2026. Against that, AWS revenue rose 39% and Meta's ad revenue rose 26%. ### Why did Alphabet stock fall 12%? Alphabet fell 12% on October 29, its worst day on record, after third-quarter results showed negative free cash flow of $9.1 billion, its second negative quarter in a row, on $51.4 billion of quarterly capex. Search ad growth also slowed to 12%, its weakest pace in five quarters, and the company expects spending to keep rising in 2027. ### Why did S&P change Meta's outlook to negative? S&P Global Ratings moved Meta's outlook from stable to negative on October 30, citing uncertainty about its elevated capex and negative free cash flow. Meta's AA- rating did not change. Meta posted negative free cash flow of $2.1 billion in the third quarter and plans to raise $95 billion through stock offerings. Meta said it respectfully disagrees with the assessment. ### What happens if the AI bubble bursts? Investors would feel it first: this week Alphabet lost about $500 billion in value in one day. Jobs tied to data center construction and chip supply would likely be next. Everyday users could see fewer free AI features or higher prices, but the data centers already built would remain. This is a scenario, and this week's data shows strain in how the build-out is financed while demand keeps growing. ### Will the AI bubble pop? Nobody can time it, and this week's data shows demand holding up. Three things would turn the build-out into a real crash: revenue growth stalling while capex keeps rising, lenders demanding much higher rates or refusing to lend, or a big buyer cutting its spending plans. None of those has happened yet. ## Source Based on Q3 earnings reported by TechCrunch, CNBC and the WSJ. https://techcrunch.com/2026/10/30/cracks-are-emerging-in-the-ai-spending-boom/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/ai-bubble-pop-warning-signs. Last modified 2026-10-31T12:00:00.000Z. ================================================================================ - URL: https://syntaxradar.com/radar/xreal-aura-price-specs-who-should-buy - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-31T05:00:00.000Z - Last modified: 2026-10-31T05:00:00.000Z - Primary source: https://roadtovr.com/?p=233285 # Xreal Aura: Should You Buy the $1,099 Android XR Glasses? **Short answer:** Xreal launched the Aura, the first glasses built on Google's Android XR, on October 28, 2026. It costs $1,099 (regular $1,279) through December and ships mid-November to the US, Canada, Japan and South Korea. **Takeaway:** The Xreal Aura is a see-through Android XR computer for early adopters who want Gemini and spatial apps on the move, while anyone who just wants a big virtual screen should wait or buy simpler display glasses. **Key figures (as reported by the source):** - Launch price (12GB/256GB): $1,099 - Field of view: 70 degrees - Glasses weight: Under 95 g The Xreal Aura is the first pair of Android XR glasses, built around Google's operating system for headsets and glasses, and since October 28 you can actually order one. If you came here for the Xreal Aura price and release date, here is the short version: $1,099 for the base model until the end of December (regular price $1,279), with the first units shipping in mid-November. That is a lot of money for something you wear on your face, so the question is who it is worth it for. The Aura is a self-contained computer split in two: light glasses on your face and a pocket-sized puck that does the thinking. ### What is Xreal Aura? Xreal Aura is a pair of see-through smart glasses that run Android XR, Google's operating system for headsets and glasses, with Gemini built in. Xreal launched it at the Android Developer Summit on October 28, 2026, after announcing prices on October 7. The design splits the work in two. The glasses themselves weigh under 95 grams, according to [Road to VR's launch report](https://roadtovr.com/?p=233285), because the heavy lifting happens in a separate puck that sits in your pocket. That puck holds a Qualcomm Snapdragon Reality Elite chip, the battery, two USB-C ports and a trackpad. Think of it as a laptop where the screen is on your face and the computer is in your pocket. The glasses show you apps floating in the room, track your head and hands, and let you keep seeing the real world through the lenses. Xreal says more than 150 XR experiences are available at launch. If you want the bigger picture of where this category is heading, our explainer on [why smart glasses could replace smartphones](/radar/smart-glasses-replace-smartphones-meta-orion-ar-future) covers the long game. The Aura is one of the first products that lets you test that idea for real. ### How much is the Xreal Aura? The Aura price is $1,099 for the 12GB RAM and 256GB storage model, and $1,299 for the 16GB and 512GB model, during the launch window that runs through December 2026. After that, the regular prices return: $1,279 and $1,499. Here is the full ladder, with the launch prices taken from Road to VR and the regular prices from [9to5Google's price report](https://9to5google.com/2026/10/07/xreal-aura-with-android-xr-start-at-1279/):
ModelLaunch price (through Dec 2026)Regular price
12GB RAM / 256GB$1,099$1,279
16GB RAM / 512GB$1,299$1,499
So the discount is $180 on the base model and $200 on the bigger one. There is no extra charge for picking the Medium or Large frame size. Some context on the number. Xreal had promised a price "under $1,500" before it revealed the final figure, so $1,279 landed well below that ceiling. Early buyers who put down the $199 Priority Deposit had that amount counted toward the final price. ### Aura release date: when does it ship? The first batch ships in mid-November 2026 to the US, Canada, Japan and South Korea, according to [iDevice](https://idevice.com/xreal-s-aura-glasses-start-at-1-279-and-ship-in-mid-november). More countries follow in 2027. In the US you can buy it at XREAL.com and Best Buy, and Google Store locations in the US and Japan will carry it in 2027. ![Xreal Aura price ladder diagram: $1,099 and $1,299 launch prices, $1,279 regular price for the base model after December, first units shipping mid-November](/uploads/blog/xreal-aura-price-specs-who-should-buy/img1.webp) ### Xreal Aura specs: the display, the FOV and the puck The spec sheet reads like a headset's, squeezed into glasses. Here is what each number means for you. #### The display Each eye gets its own Sony Micro-OLED panel at 1920 x 1200, which works out to 32 pixels per degree (PPD). PPD is the number that matters in glasses: it tells you how sharp text looks at the size you actually see it. The panels run at up to 120Hz and reach 1,050 nits of brightness, a high ceiling for brightness. #### Field of view The field of view is 70 degrees, delivered through prism-style see-through optics. Field of view is how much of your vision the virtual image can fill. At 70 degrees, a virtual screen feels large in front of you, but it does not wrap around your whole view the way a closed VR headset does. The lenses also have five-level electrochromic dimming, meaning they can darken on command and block up to 99.9% of outside light. Dim them fully and the room mostly disappears. Leave them clear and you can see it. #### Tracking Two cameras on the glasses provide 6DoF inside-out tracking. In plain English, 6DoF (six degrees of freedom) means the glasses know both where you are looking and where you are moving, so a virtual window can stay pinned to a wall while you walk around it. The glasses also support hand tracking, so you are not limited to the trackpad. Xreal adds its own X1S Spatial Coprocessor to handle display and sensor processing, and the spec sheet lists 3ms motion-to-photon latency. That is the delay between moving your head and the image catching up. Lower is better, because lag is what makes people feel queasy. #### The puck The puck is the brain. It runs Android XR with Gemini on a Snapdragon Reality Elite chip rated at up to 48 TOPS (trillions of operations per second, a common measure of AI processing power). It carries a 4,455 mAh battery, rated at 34.84Wh. Putting the battery and chip in your pocket instead of on your nose is the main reason the glasses stay under 95 grams. #### Sizes There are two frame sizes based on IPD, the distance between your pupils: Medium fits 60 to 66 mm and Large fits 66 to 72 mm. If you do not know your IPD, an optician can measure it in a minute, and picking the right size matters for a sharp image. ### Xreal Aura vs Xreal One Pro The main difference is where the computing happens. The One Pro is Xreal's display-glasses line: glasses that show another device's screen as a big virtual display, with the connected device doing the computing. Our launch sources focus on the Aura and say little about the One Pro, so treat this as a comparison of the two approaches. The Aura brings its own computer. The puck runs Android XR and Gemini on its own chip, and the glasses add camera-based 6DoF tracking and hand tracking, so apps can live in your room rather than just floating in front of you. That is the jump from "a monitor on your face" to "a headset you can see through". The trade-off is cost and complexity. Our sources do not list One Pro pricing, so check Xreal's store before you compare, but display-only glasses have no separate puck to carry or charge. If what you want is a giant screen for movies or a second monitor, that kind of product is built for the job. If you want Android XR apps, Gemini and spatial tracking, only the Aura on this list offers them. ![Xreal Aura vs One Pro diagram: display glasses that show another device screen versus Aura glasses plus a puck running Android XR with Gemini](/uploads/blog/xreal-aura-price-specs-who-should-buy/img2.webp) ### Xreal Aura vs Meta glasses This comparison is closer than you might expect on price. At its regular $1,279, the Aura is $20 cheaper than Meta's VR Glasses, according to iDevice's report on the launch. The bigger difference is how each one feels to use. iDevice cites WIRED's hands-on testing, which found Meta's glasses better for sitting down and typing, while the Aura's see-through lenses work better when you are walking around. That lines up with the hardware: the Aura's prism optics let you see the real world clearly, with dimming when you want to shut it out. So the split is roughly this. If your plan is mostly seated productivity, Meta has the edge in WIRED's testing. If you want something you can wear while moving through a room, glancing at apps without losing track of what is around you, the Aura is the better fit. There is also the platform question. The Aura runs Android XR, which means Gemini and the Android developer world. Choosing between them is partly choosing which company's ecosystem you want to bet on for the next few years. ### Who should buy the Aura, and who should wait? Here is our verdict, based on the specs and early testing reported so far. **Buy now if:** - You want to try Android XR and Gemini on your face before almost anyone else, and $1,099 is money you can spend on a first-generation device. - You care more about seeing the real world while you use apps than about a fully immersive view. - You already live in Google's ecosystem and want the launch discount before it ends in December. **Wait if:** - You mainly want a big virtual screen for films or a laptop. Display glasses like the One Pro are built for that job. - You need it outside the US, Canada, Japan or South Korea. Those are the only first-batch countries, and others come in 2027. - You want to see how the 150+ launch experiences grow before paying a first-generation price. One practical note: measure your IPD before you order. Getting the wrong frame size on a device this expensive is an avoidable mistake. ### What to watch next The Aura is the first real test of whether Android XR can work in glasses people would wear outside the house. Here is what to track. First, the full Aura reviews once units ship. Units arrive in mid-November, and full reviews will show whether the 70-degree FOV, 32 PPD sharpness and puck battery hold up in daily use. Second, the price after December. If the $1,279 regular price returns on schedule, the launch window is the cheapest way in. Third, the app count. Launch day brings more than 150 XR experiences, and [Android Central's launch coverage](https://www.androidcentral.com/gaming/virtual-reality/launch-day-xreal-aura-is-here) is a good place to track what arrives next. Reviews should start landing when the first units ship in mid-November, and the launch price window closes at the end of December. ## Frequently asked questions ### What is Xreal Aura? It is a pair of see-through smart glasses that run Google's Android XR with Gemini. The glasses weigh under 95 grams because a pocket puck with a Snapdragon Reality Elite chip, battery and trackpad does the computing. It tracks your head and hands with two cameras and launched on October 28, 2026. ### How much is the Xreal Aura? Through December 2026 it costs $1,099 for the 12GB RAM and 256GB model and $1,299 for the 16GB and 512GB model. After the launch window, regular prices are $1,279 and $1,499. Medium and Large frame sizes cost the same, and it is sold at XREAL.com and Best Buy in the US. ### What is the Xreal Aura release date? The glasses went on sale on October 28, 2026, and the first batch ships in mid-November 2026 to the US, Canada, Japan and South Korea. Xreal says more countries will follow in 2027, when Google Store locations in the US and Japan will also start selling the glasses. ### Is the Xreal Aura better than Meta glasses? It depends on how you plan to use it. At its regular $1,279 price the Aura is $20 cheaper than Meta's VR Glasses. WIRED's testing, as reported by iDevice, found Meta better for sitting and typing, while the Aura's see-through lenses worked better for walking around and staying aware of the room. ### What are the Aura specs and field of view? Each eye gets a Sony Micro-OLED panel at 1920 x 1200 (32 PPD, up to 120Hz, 1,050 nits) with a 70-degree field of view. A pocket puck runs Android XR with Gemini on a Snapdragon Reality Elite chip, up to 48 TOPS, with a 4,455 mAh battery. The glasses weigh under 95 grams. ## Source Based on Road to VR, 9to5Google and iDevice launch reports. https://roadtovr.com/?p=233285 --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/xreal-aura-price-specs-who-should-buy. Last modified 2026-10-31T05:00:00.000Z. ================================================================================ - URL: https://syntaxradar.com/radar/claude-code-extension-vscode-backdoor - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-31T00:00:00.000Z - Last modified: 2026-10-31T00:00:00.000Z - Primary source: https://phorion.io/blog/offical-claude-code-vsx-compromise/ # Claude Code Extension VSCode Backdoor: How to Check and Clean Up **Short answer:** Phorion reports that Claude Code for VS Code 2.1.313 on the Cursor Marketplace hid a backdoor that planted an SSH key, opened a reverse tunnel and beaconed system details. Cursor pulled the release, but uninstalling does not remove the backdoor. **Takeaway:** If Claude Code for VS Code 2.1.313 ever ran on your machine, removing it is not enough: check for the svc-telemetry SSH key and persistence files, clean them up and rotate exposed credentials. **Key figures (as reported by the source):** - Affected version: 2.1.313 - Reconnect interval: Every 60 seconds - Secret env var names sent: Up to 40 Anyone with a Claude Code extension VSCode install, or its Cursor build, should check one thing first: the version number. According to [an advisory from security firm Phorion](https://phorion.io/blog/offical-claude-code-vsx-compromise/), release 2.1.313 of Claude Code for VS Code, served through the Cursor Marketplace, carried a hidden backdoor that ran the moment the extension loaded. Nothing asked for permission. Cursor has pulled the release. That helps people who never installed it, and does nothing for those who did. Phorion's analysis says the code planted an SSH key, opened a reverse tunnel to a remote server and set itself up to come back after every reboot. Removing the extension does not undo any of that. This guide explains what the Claude Code extension backdoor did in plain English, how to tell if your machine is affected, how to clean it up, and how to vet AI coding extensions before the next one lands. One caveat up front: Phorion is the only primary source so far, so every technical detail below is Phorion's finding. ### What happened to the Claude Code extension for VS Code? The Claude Code extension for VS Code is the add-on that puts Anthropic's coding assistant inside your editor. On October 30, 2026, Phorion researcher Calum Hall published an advisory saying version 2.1.313 of that extension, as distributed on the Cursor Marketplace, had been tampered with. According to Phorion, Cursor removed the release, and the NPM and VS Code versions of 2.1.313 were also pulled. The advisory does not change what you need to do: if 2.1.313 ever ran on your machine, treat that machine as compromised until you have checked it. The story spread fast among developers. The [Hacker News thread about the backdoor](https://news.ycombinator.com/item?id=50300683) reached 531 points. At the time of writing, a web search found no second independent report confirming the details, so keep that in mind as you read. ### How the Claude Code extension VSCode backdoor worked, in plain English Every VS Code extension has a main file, usually called `extension.js`. Normally the editor loads that file and then calls a function named `activate` when the extension is needed. Phorion found that someone had inserted an extra block of code at the very top of the file, written as an IIFE (an immediately invoked function expression, code that runs the instant it is read). That placement is the clever and nasty part. Because the code sat above everything else, it ran as soon as Node loaded the file, before `activate` and without any action from you. Opening your editor was enough. From there, Phorion describes three moves. #### Stage 1: a spare key to your machine The code appended an Ed25519 SSH key to `~/.ssh/authorized_keys`, labelled with the comment `svc-telemetry`. That file is the list of keys allowed to log in to your account over SSH. The label is chosen to look like boring housekeeping. #### Stage 2: a tunnel that dials out Most home and office networks block strangers from connecting in. So the code connected out instead. Phorion says it opened an SSH connection to `134.209.82[.]225` on port 22 as a user named `tunnel`, with a reverse tunnel option (`-R *:0:localhost:22`). In effect, your machine handed the attacker's server a door back to your own SSH service. The key it used was stored at `$TMPDIR/.stk_` followed by 10 hex characters. #### Stage 3: a beacon with your details The code also sent one HTTP POST to `/beacon` on the same address, port 8099. Per Phorion, it included your hostname, username, platform, CPU architecture, Git email, public IP and the names of up to 40 environment variables matching KEY, TOKEN, SECRET, PASS, AWS or GCP. That is a shopping list: it tells the attacker which machines hold cloud credentials worth going after. ![How the Claude Code extension VSCode backdoor worked: a spare login key is added, the machine dials out through a tunnel and a beacon sends the hostname, username and secret names, per the Phorion advisory](/uploads/blog/claude-code-extension-vscode-backdoor/img1.webp) ### Why uninstalling the extension is not enough The backdoor was built to outlive the extension. According to Phorion, it installed persistence on every major desktop system: - **macOS:** a LaunchAgent at `~/Library/LaunchAgents/com.apple.sysmond.plist`, plus files at `~/.config/.smt` and `.smtk`. The name imitates a real Apple service. - **Linux:** an `@reboot` cron entry. - **Windows:** a scheduled task called `SysMon`, or a `sysmon.cmd` file in the Startup folder. Once set up, it tried to reconnect every 60 seconds and used `autossh` if that tool was installed. The extension was only the delivery truck. Delete it and the cargo stays behind. There is one twist on Macs. Phorion found a logic bug: without admin rights, the command `systemsetup -getremotelogin` errors but still reports success, so the step that would switch on Remote Login was skipped. If Remote Login was already off, the inbound route may not have worked. The outbound beacon is a separate matter, and so is anything the attacker learned from it, so do not treat the bug as a free pass. ### How to check for malicious VSCode extensions on your machine Here is how to check for malicious VSCode extensions, using the indicators Phorion published. Run these on any machine that had version 2.1.313 of Claude Code for VS Code installed in Cursor or VS Code. In order, they look for the planted key, the macOS LaunchAgent and helper files, a Linux `@reboot` cron entry, the stored tunnel key and any live SSH or `autossh` processes you did not start. ```bash grep svc-telemetry ~/.ssh/authorized_keys ls -la ~/Library/LaunchAgents/com.apple.sysmond.plist ~/.config/.smt* 2>/dev/null crontab -l | grep @reboot ls -la "$TMPDIR"/.stk_* /tmp/.stk_* 2>/dev/null ps aux | grep -E "autossh|ssh .*tunnel@" | grep -v grep ``` On Windows, open Task Scheduler and look for a task named `SysMon`, then check your Startup folder for `sysmon.cmd`. Two more checks help if you want certainty. Phorion lists the SHA-256 hash of the backdoored `extension.js` as `ecb58cb29078d9646ad1dcdbe51526f2d9cfea06be5fb2ab183e8ecc3ab6bf15`, and the fingerprint of the planted key as `SHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk`. If either matches, you were hit. Your SSH logs are the other clue: Phorion says a login from `127.0.0.1` or `::1` using the `svc-telemetry` key is the sign that someone came back through the tunnel. #### How to clean it up Phorion's cleanup steps, in order: 1. Stop the retry loop and kill any SSH or `autossh` processes it started. 2. Remove the LaunchAgent, `~/.config/.smt` and `.smtk` (or the cron entry, scheduled task or `sysmon.cmd` on Linux and Windows). 3. Delete the `svc-telemetry` line from `~/.ssh/authorized_keys` and remove `/tmp/.stk_*`. 4. Confirm Remote Login is off on macOS. 5. Rotate your credentials if SSH was enabled or you suspect an inbound login. One addition to step 5: if the beacon could have listed token or cloud key names from your environment, rotate those keys anyway. Rotation is cheap compared with finding out later. ![How to clean up after the Claude Code extension backdoor: stop the SSH processes, remove the startup entry, delete the planted key, confirm Remote Login is off and rotate credentials](/uploads/blog/claude-code-extension-vscode-backdoor/img2.webp) ### Is the Claude Code extension safe? Is the Claude Code extension safe now? Based on Phorion's advisory, the problem was a single release, 2.1.313, and that release has been pulled from the Cursor Marketplace, NPM and VS Code. The advisory names only 2.1.313 as affected. That is a narrower statement than "safe". No account of how the tampered build reached the marketplace has been published in the sources we reviewed, and there is only one primary source. Until more is known, the sensible position is: confirm you are not on 2.1.313, run the checks above if you ever were, and watch for follow-up reports. The wider lesson is that AI coding tools hold an unusual amount of trust. They read your code, run commands and sit next to your SSH keys and cloud tokens. As we covered in our piece on [computer-use AI agents](/radar/computer-use-ai-agents-operating-systems-claude-operator), the more an assistant can do on your machine, the more a poisoned copy can do too. ### Is there a Claude Code extension for Cursor? Yes. Cursor is built on VS Code, so it runs VS Code extensions, and the Claude Code extension in Cursor is installed from the Cursor Marketplace. That is exactly the channel where Phorion found the backdoored 2.1.313 release. This is why the incident matters beyond one bad version. Phorion found the backdoor in a release served through the Cursor Marketplace itself, so the attack did not depend on a fake name. #### How to vet AI coding extensions from now on None of these steps would have made you immune, but each one shrinks the window: - **Pin and read versions.** Turn off automatic extension updates for tools that can run commands, and look at the changelog before you accept a new release. - **Wait a day on big tools.** A short delay before updating lets early reports, like this one, surface first. - **Watch your `authorized_keys` file.** Few developers ever look at it. A new line you did not add is an alarm. - **Keep secrets out of your shell.** The beacon only saw variable names, but those names told an attacker where to look. A secrets manager beats a long list of tokens in your environment. - **Treat an editor like a server.** If you would not run an unknown binary on a production box, be careful about what runs inside the editor that holds your keys. For a sense of how fast attackers are moving with AI tooling, see our report on [AI for hacking and CrowdStrike's ARTEX findings](/radar/ai-for-hacking-artex-crowdstrike). ### What to watch next Three things will tell us how big this really is. First, whether Anthropic or Cursor publish their own account of how a backdoored build of the Claude Code extension for VSCode reached the marketplace. Second, whether other security researchers confirm Phorion's indicators, since [Phorion's advisory](https://phorion.io/blog/offical-claude-code-vsx-compromise/) is still the only primary source. Third, whether marketplaces add stronger signing or review for extensions that can run code the moment they load. Until then, the to-do list for VS Code and Cursor users is short: check your version, run the checks, clean up if needed and rotate anything that might have leaked. ## Frequently asked questions ### Is the Claude Code extension safe? According to Phorion, the backdoor was in one release, Claude Code for VS Code 2.1.313 on the Cursor Marketplace, and that release has been pulled. The advisory does not report other affected versions. Confirm you are not on 2.1.313, and if you ever were, check your machine for the planted SSH key and persistence files, because uninstalling alone does not remove them. ### How to check for malicious VSCode extensions like this one? For this backdoor, Phorion's indicators are the clearest test. Search ~/.ssh/authorized_keys for a key labelled svc-telemetry, look for the com.apple.sysmond LaunchAgent on macOS, an @reboot cron entry on Linux or a SysMon task on Windows, and check SSH logs for logins from 127.0.0.1 or ::1 using that key. ### Is there a Claude Code extension for Cursor? Yes. Cursor is built on VS Code and installs the Claude Code extension from the Cursor Marketplace. That marketplace is where Phorion found the backdoored 2.1.313 release, which Cursor has since removed. If you use Claude Code in Cursor, check which version you have installed and run the cleanup checks if it was ever 2.1.313. ### Does uninstalling the backdoored extension remove the malware? No. Phorion says the code installed persistence that survives removal: a LaunchAgent and hidden config files on macOS, a reboot cron job on Linux and a scheduled task or Startup script on Windows. It retries its tunnel every 60 seconds. You need to kill those processes, delete the files and the svc-telemetry key, and rotate credentials if needed. ## Source Based on Phorion's security advisory (single primary source). https://phorion.io/blog/offical-claude-code-vsx-compromise/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/claude-code-extension-vscode-backdoor. Last modified 2026-10-31T00:00:00.000Z. ================================================================================ - URL: https://syntaxradar.com/radar/ao3-outage-why-is-ao3-down - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-30T17:00:00.000Z - Last modified: 2026-10-30T17:00:00.000Z - Primary source: https://www.otwstatus.org/incidents/g12jnxck9g6h # AO3 Outage: Why Is AO3 Down, and Is Your Account Safe? **Short answer:** AO3 went offline on Oct 29, 2026 under a maintenance notice, and OTW estimated 2 to 4 weeks to restore it. On Oct 30 OTW said it had detected evidence of a security incident, that no AO3 data was deleted, and advised changing reused passwords. **Takeaway:** OTW says no AO3 data was deleted, but until it finishes investigating, anyone who reused their AO3 password on another site should change it there now. **Key figures (as reported by the source):** - Outage first logged: Oct 29, 07:13 UTC - Security incident disclosed: Oct 30, 01:27 UTC - OTW restoration estimate: 2 to 4 weeks The AO3 outage began on October 29, 2026, when Archive of Our Own went dark under a status notice titled "AO3 Down for Maintenance." If you have spent the past day asking "is AO3 down?" or wondering why AO3 is down for this long, the story changed overnight. In an update posted at 01:27 UTC on October 30, the Organization for Transformative Works (OTW), which runs the archive, said it had detected evidence of a security incident. As of October 30 the site is still offline. OTW's estimate for bringing it back is around two weeks, possibly up to four, and it says the scope of the incident is still being determined and its response is in the early stages. Here is the short version for worried readers. OTW says no AO3 data was deleted, so works, bookmarks and kudos are safe. It says it does not currently believe the user database was stolen but cannot rule it out. And if you used your AO3 password anywhere else, OTW's advice is to change it on those sites. ### Why is AO3 down? AO3 has been offline since October 29, and OTW has not given a technical cause. It first logged the outage under an AO3 maintenance notice and later said it had detected evidence of a security incident. The [OTW incident log](https://www.otwstatus.org/incidents/g12jnxck9g6h) shows the order of events. The first entry, at 07:13 UTC on October 29, called it "AO3 Down for Maintenance" with a status of Investigating. At 09:03 UTC, OTW said maintenance was in progress and expected to take at least 12 hours, possibly longer. At 17:41 UTC the estimate jumped to around two weeks, perhaps up to four, with staff hoping to bring back a degraded version of the site sooner. Writing that day, [Books of Brilliance](https://booksofbrilliance.com/2026/10/29/ao3-down-extended-outage/) pointed out that the public update still gave no technical cause and no reopening date. The words "security incident" first appear on the status page in the 01:27 UTC update on October 30. What OTW has not said matters as much as what it has. In the status page updates, it has not described what kind of incident this is, how it happened, who is responsible, or whether any data was actually accessed or copied. It also has not said whether the original maintenance and the security incident are the same event. [TechCrunch](https://techcrunch.com/2026/10/29/fan-fiction-site-archive-of-our-own-could-be-offline-for-up-to-4-weeks/) added one useful piece of context. It notes that AO3 runs on Ruby on Rails, so the PHP vulnerabilities reported on October 21 are unlikely to be the cause. That is TechCrunch's reading; OTW has said nothing about cause. ### Is AO3 still down? Yes, AO3 is still down as of October 30, 2026. The main [AO3 status page](https://www.otwstatus.org/), run by OTW, lists the archive as a major outage, so if you want to know whether AO3 is down right now, check that page or the @AO3_Status account before trusting a screenshot passed around on social media. You are far from the only person refreshing. TechCrunch reports that OTW's status posts reached millions of views within two hours. The same status page lists OTW's other services, including Fanlore, the journal Transformative Works and Cultures and the OTW website, as operational. Two earlier entries on that page are easy to mix up with this one. Fanlore went offline on October 21 while it waited for PHP patches and was back the same day. AO3 emails, history and search updates were delayed from October 23 to 24, a problem marked resolved early on October 24. No source connects either one to the current outage. ![AO3 outage timeline from the OTW status page: site down on Oct 29 at 07:13 UTC, an estimate of about two weeks at 17:41 UTC, a security incident disclosed on Oct 30 at 01:27 UTC, and an estimate of two to four weeks](/uploads/blog/ao3-outage-why-is-ao3-down/img1.webp) ### Is AO3 safe? What OTW says about your data Your AO3 data is safe from deletion, according to OTW: it says no AO3 data was deleted, which means works, bookmarks and kudos are intact. Whether the user database was copied is a separate question, and OTW's answer is careful. It says it does not currently believe the user database was stolen, but it cannot rule that out. OTW also listed what that database may include: - usernames - email addresses - IP addresses - hashed passwords Some things sit outside it. AO3 does not collect real names, OTW says, and donor information is stored separately and was not exposed. If you have donated to OTW, your donor details fall under that statement. In general, it is hard to prove that something was never copied: it means reviewing logs and systems one by one. OTW has said the scope is still being determined. ### Should you change your AO3 password while AO3 is down? OTW's advice is direct: if you reuse your AO3 password elsewhere, change it on those sites. That makes the first job, while AO3 is down, finding every other site where you used your AO3 password. "Hashed" sounds reassuring, and hashed passwords are far safer than plain text. A hash is a one-way scramble. The site stores only the scrambled result, and when you log in it scrambles what you typed and checks whether the two match. Nobody can simply run the scramble backwards. The weak spot is guessing. Someone holding a list of hashes can take common passwords, scramble each one the same way and look for matches. A short password, or one that already appeared in an old leak, falls quickly. A long, unique one can take impractically long. [Bleeding Cool](https://bleedingcool.com/pop-culture/archive-of-our-own-could-be-down-for-the-next-2-4-weeks/) reports that OTW's notice makes this same point: weak or reused passwords can be cracked even when hashed. Reuse is what turns one leak into many. If the same email and password open your AO3 account, your inbox and a shopping site, a cracked AO3 hash becomes a key to all three. Attackers routinely try leaked email and password pairs on other popular sites, a technique called credential stuffing. OTW's own advice is the first step below. The rest is general good practice, not an OTW instruction. - Change the password anywhere you reused your AO3 one, starting with your email account, because email is how every other password gets reset. - Turn on two-factor authentication on those sites where it is offered. - Use a password manager so each site gets its own long password. - Be wary of unexpected emails claiming to be from AO3, since email addresses may be in the database. Do not click links in them; go to the official channels instead. OTW says staff may not answer individual inquiries and will post updates on its official channels. For a bigger example of how one weak link can open a much larger door, see our [Binance hacked](/radar/binance-hacked-engineer-laptop-theft) explainer. ![Is AO3 safe? What OTW says about your data during the AO3 outage: no AO3 data deleted, user database not believed stolen but cannot be ruled out, and reused passwords should be changed on other sites](/uploads/blog/ao3-outage-why-is-ao3-down/img2.webp) ### How long is AO3 down? AO3 could be down for about two weeks and possibly up to four, according to OTW's estimate. It first gave that range at 17:41 UTC on October 29 and repeated it after disclosing the security incident, now calling two weeks a minimum. That is a long time for a website people use every day. OTW's status updates do not explain why its estimate runs to weeks, so what follows is how recovery after a security incident generally works, not a description of AO3's plan. A routine outage can often be fixed by restarting a service or rolling back a bad update. After a security incident, teams typically have to work out how it happened, close that path, check what was touched and confirm nothing was left behind. That often means rebuilding servers from a known-clean state and restoring data carefully, because reopening too early risks letting the same problem back in. Each step has to finish before the next one can be trusted, which is how incident recoveries stretch into weeks. ### When will AO3 be back? OTW has not given a firm date for when AO3 will be back. Its range is two to four weeks, and it has said staff hope to restore a degraded service before a full one. OTW also warns that the site may return in a degraded or read-only state and could go down unexpectedly during recovery. In general, read-only means you can browse and read while anything that writes new data, such as posting, commenting or leaving kudos, stays switched off. Degraded generally means some features are slow or missing. OTW's status updates do not say which features would come back first. So if the archive flickers back and then drops again, that would fit what OTW warned could happen. ### AO3 status updates: what to watch next Four things will tell you the most over the coming days: - **The scope.** OTW says it is still determining the extent of the incident. Any update on what was affected is the most important one. - **The database question.** Watch for OTW to confirm or rule out whether the user database was copied. Until that answer arrives, the password advice above is the safe default. - **The timeline.** The two to four week range is an estimate, and OTW may revise it as the investigation moves. - **The first version back.** OTW has said the site may return degraded or read-only, so watch for which of those arrives first. For live AO3 status updates, use the [OTW status page](https://www.otwstatus.org/), the @AO3_Status account and OTW's own [update on the AO3 downtime](https://www.transformativeworks.org/update-on-ao3-downtime-october-30-2026/). Until OTW says more, the AO3 outage leaves fans with one useful job: retire any password they have used in more than one place. ## Frequently asked questions ### Why is AO3 down? AO3 went offline on October 29, 2026 under a maintenance notice. In a status update at 01:27 UTC on October 30, the Organization for Transformative Works said it had detected evidence of a security incident. OTW has not said what kind of incident it was, how it happened or who is responsible, and says its response is still in the early stages. ### Is AO3 down right now? As of October 30, 2026, yes. The OTW status page lists AO3 as a major outage, while Fanlore, Transformative Works and Cultures and the OTW website are listed as operational. Status can change quickly, so check otwstatus.org or the @AO3_Status account for the latest word rather than relying on screenshots shared elsewhere. ### When will AO3 be back? OTW has not given a firm date. It estimates a minimum of two weeks and possibly up to four, and says staff hope to restore a degraded service sooner. The site may first return in a degraded or read-only state and could go down unexpectedly during recovery, so a brief return followed by another drop would not be unusual. ### Is my AO3 data safe? OTW says no AO3 data was deleted, so works, bookmarks and kudos are safe. It does not currently believe the user database was stolen but cannot rule it out; that database may hold usernames, email addresses, IP addresses and hashed passwords. OTW advises changing your AO3 password on any site where you reused it. Donor information is stored separately and was not exposed. ### Was AO3 hacked? OTW's status updates do not say whether AO3 was hacked. It says it detected evidence of a security incident and that the scope is still being determined. It has not said what kind of incident it was, how it happened, who is responsible or whether any data was accessed or copied. It does not currently believe the user database was stolen but cannot rule it out. ### What is Archive of Our Own (AO3)? Archive of Our Own, or AO3, is the fan-fiction archive run by the Organization for Transformative Works. Bleeding Cool, a secondary source, puts the archive at more than 18,820,000 works across over 83,250 fandoms and says it was founded in 2008. As of October 30, 2026 it is offline, and OTW estimates two to four weeks to restore it. ## Source Based on OTW status updates, Oct 29 to 30. https://www.otwstatus.org/incidents/g12jnxck9g6h --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/ao3-outage-why-is-ao3-down. Last modified 2026-10-30T17:00:00.000Z. ================================================================================ - URL: https://syntaxradar.com/radar/ai-for-hacking-artex-crowdstrike - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-30T12:00:00.000Z - Last modified: 2026-10-30T12:00:00.000Z - Primary source: https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/ # AI for Hacking: CrowdStrike Says ARTEX Targeted 200+ Orgs **Short answer:** On Oct 29, 2026, CrowdStrike reported that VAPOR PANDA used ARTEX, an AI agent framework running mainly on DeepSeek, to target 200+ organizations in at least 14 countries over 11 days. CrowdStrike confirmed post-exploitation at nine. **Takeaway:** CrowdStrike's VAPOR PANDA report shows AI agents doing an intrusion's legwork while a human steers the key decisions, and its five defenses are standard identity, cloud and patching work. **Key figures (as reported by the source):** - Organizations targeted (CrowdStrike): 200+ - Confirmed post-exploitation (CrowdStrike): 9 - Campaign window from Oct 17 (CrowdStrike): 11 days Using AI for hacking now has a detailed, real-world case file. In a report published October 29, 2026, [CrowdStrike said](https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/) a group it tracks as VAPOR PANDA used ARTEX, an AI agent framework running mainly on the DeepSeek model, to target more than 200 organizations in at least 14 countries over 11 days starting October 17. The report describes an AI cyber attack in unusual detail, including how the AI hacking tools were organized and where a human stepped in. That headline number needs a careful reading. CrowdStrike's 200-plus figure counts organizations targeted. CrowdStrike confirmed post-exploitation activity at nine of them, and it saw password-spraying (guessing passwords across many accounts) and credential-scanning against at least 12 more. In CrowdStrike's account, then, the software went after hundreds of organizations, a much smaller set saw confirmed follow-on activity, and a human operator steered the agents at the key decision points. NBC News [reported the findings](https://www.nbcnews.com/tech/security/china-linked-hackers-used-ai-target-200-organizations-crowdstrike-says-rcna606343) the same evening; it attributed the figures to CrowdStrike and did not say it had verified them independently. ### What is ARTEX AI? ARTEX AI is an agentic penetration-testing tool, a kind of hacking AI: software that uses large language models to plan and carry out the stages of a network intrusion that a human security tester would normally work through by hand. Penetration testing itself is a legitimate trade. Companies hire testers to break into their own systems with permission so they can fix the holes before someone hostile finds them. An AI hacker tool built for that job can also be run against targets that never agreed to be tested, which is the situation CrowdStrike describes. ARTEX was first unveiled in July, according to Korea JoongAng Daily, and CrowdStrike's earlier report describes it as a recently released open-source tool developed in China. In the VAPOR PANDA campaign, CrowdStrike says the primary model behind ARTEX was DeepSeek, a Chinese-origin large language model. A second, subscription-based commercial LLM was also used, and CrowdStrike Intelligence could not confirm which one. #### The earlier South Korea case is a separate story ARTEX first made headlines in a different investigation. On October 7, [CrowdStrike reported](https://www.crowdstrike.com/en-us/blog/unknown-threat-actor-uses-artex-to-target-south-korean-finance/) that an unknown threat actor used ARTEX against South Korean financial organizations from late September to early October 2026. In that case CrowdStrike assessed, with moderate confidence, that the actor was financially motivated, and it noted that the bank breach details came from industry reporting rather than its own verification. In its words, "the number of organizations affected remains unconfirmed." After those attacks, the developer, who posts on GitHub as "Autumn-27", said ARTEX would become closed-source with no further public releases, citing "misuse of the tool", and [Korea JoongAng Daily](https://www.koreajoongangdaily.com/business/chinese-ai-tool-linked-to-korean-bank-hacks-halts-public-updates-as-suspect-denies-involvement/12913376) relayed Reuters' report that the GitHub page was taken down. Closing the code did not recall the copies already out there. [BleepingComputer reported](https://www.bleepingcomputer.com/news/security/hacker-used-artex-ai-and-claude-agents-to-target-south-korean-banks/amp/) that English- and Korean-language derivatives built from the code remain available as they were. The October 7 case involves an actor CrowdStrike does not name. The October 29 report covers VAPOR PANDA, and every number in the rest of this article comes from that second report. ### What is AI hacking? How ARTEX splits an intrusion between agents AI hacking, in ARTEX's case, means splitting an intrusion into jobs and handing each job to a different AI agent. According to CrowdStrike, a main agent coordinates subordinate planner agents and worker agents, and all of them share a graph of facts, findings and goals. Think of it as a small crew with one project manager. The main agent is the manager. Planner agents decide what should happen next. Worker agents do the hands-on steps. The fact graph is the shared whiteboard: when one worker learns something about a network, that finding is written down where every other agent can read it, and it can become the next goal. ARTEX calls each engagement a "task", and CrowdStrike says the agents move a task from initial access through host reconnaissance and lateral movement. In plain terms, that means getting a first foothold, looking around the machine that was breached, and then reaching other machines on the same network. Those are the same stages a human intruder works through. Two figures from the report give a sense of scale. CrowdStrike says a single ARTEX controller was used against more than 200 distinct organizations, and that planning agents in a single intrusion consumed roughly 1.4 billion input tokens. A token is a small chunk of text that a language model reads, so that is a very large volume of text fed to the planners for one intrusion. ![Diagram of how AI hacking works inside ARTEX per CrowdStrike: a main agent coordinates planner agents and worker agents that share a fact graph](/uploads/blog/ai-for-hacking-artex-crowdstrike/img1.webp) ### Can AI hack by itself? AI did not hack by itself in the campaign CrowdStrike describes. CrowdStrike found that a human operator steered the agents at decision points, with instructions written in Chinese, labeled "human instruction" and timestamped in China Standard Time. In CrowdStrike's account, the agents handled the legwork: tracking findings, planning next moves and carrying out the steps. The human stepped in at the decision points. The operator is still needed, but the job shifts from doing every step to steering. What CrowdStrike describes is supervised automation, with a person making the key calls. ### AI for hacking at scale: what CrowdStrike verified Here is what CrowdStrike says it saw in the VAPOR PANDA campaign, figure by figure.
Finding (all per CrowdStrike)Figure
Organizations targeted by a single ARTEX controllerMore than 200
Countries with targeted entitiesAt least 14
Time window, starting October 17, 202611 days
Organizations with confirmed post-exploitation (BLOODALCHEMY deployed, follow-on activity)9
Further organizations hit by password-spraying and credential-scanningAt least 12
U.S.-based organizations targeted in one engagement19
CrowdStrike lists the targets as foreign ministries, government bodies, think tanks, and organizations in pharmaceuticals, semiconductors, defense, nuclear, finance and telecoms. The single engagement against 19 U.S.-based organizations, CrowdStrike says, covered pharmaceutical, semiconductor, defense, nuclear and financial sectors. The confirmed intrusions are the nine where CrowdStrike says the adversary deployed a tool called BLOODALCHEMY and carried out follow-on activity. The password-spraying and credential-scanning against at least 12 more organizations are login attacks, and CrowdStrike lists them separately from the nine confirmed cases. Everything else in the 200-plus count is targeting. ![AI for hacking by the numbers per CrowdStrike: 200+ organizations targeted in at least 14 countries over 11 days, 9 confirmed post-exploitation and at least 12 more hit by credential attacks](/uploads/blog/ai-for-hacking-artex-crowdstrike/img2.webp) ### Can AI hack websites? The CrowdStrike findings covered here do not single out website hacking. They describe ARTEX going after whole organizations, with agents that move from initial access through host reconnaissance and lateral movement. One of CrowdStrike's five recommendations, patching internet-facing appliances, does cover systems exposed directly to the internet. Most of CrowdStrike's list deals with what happens after a first foothold: credentials, cloud permissions and monitoring. On the defensive side of the same debate, our explainer on [Gemini 4 Carbon](/radar/gemini-4-carbon-explained) covers why Google's new AI goes to defenders first. For the agent idea outside security, see how [computer-use agents are learning to click and type like a human](/radar/computer-use-ai-agents-operating-systems-claude-operator). ### How do you defend against an AI cyber attack? You defend against an AI cyber attack like this one with the five steps CrowdStrike lists in its report. None of them is exotic. 1. **Rotate secrets after suspected exposure.** CrowdStrike says to rotate signing keys, preauthentication keys and service credentials. A key that may have been seen should stop working. 2. **Tighten identity and cloud settings.** Enforce SAML signature validation, require IMDSv2 and limit the permissions cloud roles have to execute commands. SAML handles single sign-on, and IMDSv2 is the newer, harder-to-abuse version of the cloud metadata service that tells a server about itself. 3. **Hunt for ARTEX markers.** CrowdStrike lists "artx" or "artex" in command lines, cloud roles with an artx prefix, "art18-" comments, and the folder `C:\ProgramData\artx5v\`. 4. **Patch internet-facing appliances.** These are devices exposed directly to the internet, such as VPN gateways and firewalls. 5. **Deploy a real-time EDR agent across the environment.** EDR (endpoint detection and response) software watches each machine for suspicious behavior as it happens, so coverage gaps are blind spots. The list is short on AI-specific measures. Apart from the ARTEX markers, it is ordinary identity, cloud and patching hygiene. ### AI cyber attacks in 2026: what to watch next CrowdStrike expects the trend to continue. The company assesses it is "highly likely" that China-nexus adversaries will keep adopting agentic frameworks like ARTEX over the next 12 months. Three open questions remain. First, whether CrowdStrike or other researchers identify the second commercial LLM in the VAPOR PANDA campaign. Second, whether CrowdStrike updates the nine confirmed cases. Third, whether the ARTEX derivatives that BleepingComputer says remain available turn up in new campaigns. For defenders, the lesson of this AI for hacking case is practical: CrowdStrike's five steps, from rotating keys to full EDR coverage, are where to start. ## Frequently asked questions ### What is ARTEX AI? ARTEX is an agentic penetration-testing tool that uses large language models to plan and run the stages of a network intrusion. CrowdStrike's earlier report describes it as a recently released open-source tool developed in China, and Korea JoongAng Daily says it was first unveiled in July. Its developer later said it would become closed-source, but BleepingComputer reports that derivatives remain available. ### Can AI hack by itself? Not in the VAPOR PANDA campaign. CrowdStrike found that a human operator steered the ARTEX agents at decision points, with instructions labeled "human instruction" and timestamped in China Standard Time. The agents tracked findings, planned next moves and carried out steps, while the human made the key calls. ### Can AI hack websites? The CrowdStrike findings covered here do not single out website hacking. They describe ARTEX agents going after whole organizations, moving from initial access to host reconnaissance and lateral movement. One of CrowdStrike's five recommendations is patching internet-facing appliances, the devices exposed directly to the internet. Most of its list covers credentials, cloud permissions and monitoring. ### How do you defend against an AI cyber attack? Against an ARTEX-style AI cyber attack, CrowdStrike lists five steps: rotate signing keys, preauthentication keys and service credentials after suspected exposure; enforce SAML signature validation, require IMDSv2 and limit cloud role command-execution permissions; hunt for ARTEX markers; patch internet-facing appliances; and deploy a real-time EDR agent across the environment. ## Source Based on CrowdStrike's Oct 29 and Oct 7 reports. https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/ai-for-hacking-artex-crowdstrike. Last modified 2026-10-30T12:00:00.000Z. ================================================================================ - URL: https://syntaxradar.com/radar/binance-hacked-engineer-laptop-theft - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-30T07:38:48.000Z - Last modified: 2026-10-30T07:38:48.000Z - Primary source: https://www.trmlabs.com/resources/blog/north-korea-steals-usd-6-36-billion-from-binance-in-the-largest-hack-in-history # Binance Hacked: How One Laptop May Have Led to a $6.36B Theft **Short answer:** On Oct 29, 2026, attackers drained Binance hot wallets across ten blockchains. TRM Labs counts about $6.36B stolen and links it to North Korea's TraderTraitor; Binance blames a compromised engineer laptop and says it will cover all losses. **Takeaway:** The record Binance theft did not need a leaked private key: by Binance's early account, one compromised laptop reached the system that submits withdrawals, and only part of the loot, mostly stablecoins, could be frozen. **Key figures (as reported by the source):** - Stolen, per TRM Labs (theft-time prices): $6.36B - Frozen by 00:30 UTC Oct 30, per TRM: ~$2.7B - Blockchains hit: 10 According to blockchain analytics firm TRM Labs and its [analysis of the Binance hack](https://www.trmlabs.com/resources/blog/north-korea-steals-usd-6-36-billion-from-binance-in-the-largest-hack-in-history), money started leaving Binance's hot wallets at 20:05 UTC on October 29, 2026, and most of the outflows were done about 20 minutes later. By October 30, TRM had counted about $6.36 billion stolen across ten blockchains, roughly four times the $1.5 billion Bybit hack of February 2025. Is Binance hacked? Yes, and by TRM's count it is the largest crypto theft on record. The surprising part is the entry point. Binance's early explanation points to one compromised laptop belonging to a DevOps engineer. ### What happened with Binance? On October 29, Binance was hacked: users began reporting funds draining from Binance wallets, and Binance suspended deposits and withdrawals. [TechCrunch's report that Binance paused withdrawals](https://techcrunch.com/2026/10/29/binance-pauses-withdrawals-as-hackers-suspected-of-stealing-billions-in-crypto/), published at 3:27 PM Pacific time, quoted Binance spokesperson Ross O'Leary saying Binance withdrawals were suspended on multiple networks as a precaution and that user funds were "completely safe." At that point nobody had a firm number. TechCrunch relayed that crypto watchers estimated at least $2 billion had left Binance's hot wallets, and that CoinDesk put it at upward of $5.7 billion, unconfirmed and possibly still growing. By October 30 the picture had sharpened. In a [statement on the Binance hack reported by CoinLaw](https://coinlaw.io/binance-halts-withdrawals-6-3-billion-hack/), Binance said about $6.3 billion in assets were affected and about $3 billion had been frozen on-chain. It said it would cover all losses from its SAFU fund (its emergency reserve for users) and company funds, that it had revoked the affected address authorization, and that it had brought in law enforcement. The same day, TRM Labs published its own analysis and attributed the theft, with high confidence, to North Korea's TraderTraitor group. Binance.US and Coinbase both said they were not affected. Binance said the breach involves Binance.com only. ### How was Binance hacked? By Binance's early account, the Binance hack began on one DevOps engineer's laptop and ended in a hot wallet. An exchange keeps most customer crypto offline but needs a pool of coins online to pay out withdrawals quickly. That online pool is the hot wallet, and a piece of software decides when coins leave it. Binance said its early checks suggest the attacker compromised the laptop and used it to reach a service container: a small, isolated piece of Binance's infrastructure whose job is to submit withdrawals. From there, the attacker sent withdrawal requests that the system treated as legitimate. Binance also said it found no evidence that a wallet private key leaked. That detail matters. Think of a bank where nobody stole the vault key. Instead, someone got onto the computer that sends instructions to the vault, and the vault obeyed because the instructions came from the right place. The chain data shows how fast and how organized the payout was. On Ethereum, TRM reports that the attacker used EIP-7702, an Ethereum feature that lets an ordinary wallet hand its powers to a piece of smart contract code, along with 50 pre-signed authorizations. A single transaction at 20:20 UTC moved 167,464 ETH and 63 different tokens, including 1.615 billion USDT. TRM puts Ethereum's share at about $3.79 billion. Bitcoin went at the same minute. TRM says 18,270 BTC, worth about $1.47 billion, left a Binance Bitcoin hot wallet at 20:20 UTC and was split through a peel chain, a laundering pattern where a large balance moves through a long series of wallets, shaving off smaller amounts at each hop. TRM's breakdown of the rest of the Binance attack lists Tron (about $628 million), the XRP Ledger (about $296 million), Solana (about $96 million), Arbitrum, Optimism and Base (about $39 million combined), Avalanche (about $28 million) and Polygon (about $10 million). TRM reports that withdrawals from Tornado Cash, a crypto mixer, funded the attacker's operator wallet starting October 28, at least 36 hours before the theft. ![Timeline of the Binance hack: Tornado Cash funding on Oct 28, first transfer at 20:05 UTC, 167,464 ETH and 18,270 BTC moved at 20:20 UTC, most funds gone by 20:25 UTC and about 2.7B USD frozen at 00:30 UTC on Oct 30, per TRM Labs](/uploads/blog/binance-hacked-engineer-laptop-theft/img1.webp) ### Why do Binance hack numbers range from $2 billion to $6.36 billion? The Binance hack totals in different headlines differ mostly by timing and wording, so nobody is necessarily wrong. The table lines up each figure with its source.
FigureWho said itWhat it measures
At least $2 billionCrypto watchers, via TechCrunch (Oct 29)First partial estimate, now superseded
Upward of $5.7 billionCoinDesk, via TechCrunch (Oct 29)Early estimate, unconfirmed at the time
About $6.3 billionBinance (Oct 30)Assets "affected"
About $6.36 billionTRM Labs (Oct 30)Assets stolen, at prices when taken (about $6.23 billion at current prices)
About $3 billionBinanceFrozen on-chain so far
About $2.7 billionTRM LabsFrozen as of 00:30 UTC Oct 30
About $2.4 billionTether and Circle CEOsFrozen by those two issuers only
The $2 billion figure came in the first hours, before anyone had traced all ten chains. The $6.3 billion and $6.36 billion figures are the full count, with Binance saying "affected" and TRM saying "stolen." The frozen totals differ because they were measured at different times and include different issuers. One more comparison you will see: TRM and Reuters put the Bybit hack at about $1.5 billion, while TechCrunch says $1.4 billion. Either way, by TRM's count this Binance exchange hack is roughly four times larger. ### Can stolen Binance crypto be frozen? Only part of it can. Stablecoins such as USDT and USDC are issued by companies, Tether and Circle, and their token contracts let the issuer block an address. Tether CEO Paolo Ardoino said Tether froze about $1.6 billion, and Circle CEO Jeremy Allaire said Circle froze more than $800 million in USDC, according to CoinLaw. TRM's tally as of 00:30 UTC on October 30 was about $2.7 billion frozen, including about 1.77 billion USDT, 825 million USDC, 40.5 million RLUSD and 27.8 million USD1. There was one more lever. On Arbitrum, a network that runs on top of Ethereum, TRM says the Arbitrum Security Council forced transfers of 8,426 ETH and 74.6 million ARB at 00:08 UTC. Bitcoin and ETH on Ethereum's main network have no issuer and no council with an off switch. Once they leave, nobody can claw them back on-chain. TRM says the attacker still holds more than 420,000 ETH plus the 18,270 BTC, all beyond any issuer's reach. ![Bar chart of why Binance hack numbers differ: loss estimates from at least 2B USD to about 6.36B USD next to frozen totals of about 3B, 2.7B and 2.4B USD](/uploads/blog/binance-hacked-engineer-laptop-theft/img2.webp) ### Who is TraderTraitor, the North Korea group blamed for the Binance hack? TraderTraitor is a hacking group that TRM Labs links to North Korea, and TRM attributes the Binance theft to it with high confidence. The FBI named the same group as responsible for the Bybit hack in February 2025. Attribution here is TRM's assessment; Binance has said it is working with law enforcement, and an official finding may come later. TRM's chain analysis of this North Korea crypto hack describes a planned operation. The Ethereum withdrawals were bundled using 50 pre-signed authorizations. Most of the money moved inside about 20 minutes, across ten blockchains. Then came a peel chain for the Bitcoin and a swap of stablecoins into ETH. Both thefts, tied to North Korean crypto hackers by TRM in this case and by the FBI for Bybit, hit major exchanges. If Binance's early account holds, this one started with one engineer's laptop. ### Is Binance safe? Binance says user funds are safe: the company said customer deposits were untouched and that it would cover all losses from the SAFU fund and company funds. Co-CEO Yi He said Binance would absorb all hacking losses, according to [Reuters, as syndicated by globetv.app](https://globetv.app/news/north-korean-hackers-steal-63-billion-from-binance-in-largest-crypto-heist-firms-say-6ae43869ad113). Those are Binance's statements. Outsiders cannot yet verify its balance sheet, so treat them as promises to watch rather than settled facts. On whether Binance is insolvent, none of the sources used here say so. They say Binance has committed to covering the loss itself, and the real test is whether withdrawals reopen and process normally. On timing, Binance said its suspended withdrawals should resume within 48 hours and that a detailed technical post-mortem will follow. Yi He also said withdrawals stay disabled until Binance confirms its environment is secure, so the second statement can override the first. Whether Binance is safe from hackers in future is a separate matter, and no exchange can promise that. By Binance's early account, the entry point was an engineer's laptop and the systems it could reach, and that kind of risk exists at any custodial platform. #### Was my Binance account hacked? No report so far says a Binance account was hacked or that individual customer accounts were breached. Binance and TRM describe a drain of Binance's own hot wallets through its internal systems, and Binance said customer deposits were untouched. Still, events like this attract scammers. A few sensible steps: - Get updates only from Binance's official app, website and verified social accounts. - Ignore messages offering "refunds," "wallet verification" or help recovering funds. Real support never asks for your password, 2FA codes or seed phrase. - Use an authenticator app or hardware security key for 2FA rather than SMS. - Turn on a withdrawal address allowlist if your exchange offers one, so new addresses need extra approval. - If you use Binance.US or Coinbase, both said they were not affected. ### Binance hacked today: what happens next? Four things are still open after the Binance hack. **Withdrawals.** Binance said they should be back within 48 hours. Whether they reopen smoothly will be the clearest early test of Binance's promise to cover losses. **The post-mortem.** Binance has promised a detailed technical write-up. The key question is how one laptop could reach a container that submits withdrawals, and which checks a single transaction moving 167,464 ETH passed on its way out. **The untouchable coins.** TRM says more than 420,000 ETH and 18,270 BTC sit outside any freeze. Watch for how much more gets traced, and whether more of the laundering trail can be blocked. **Not just Binance.** On the same day, Polymarket briefly halted trading and paused withdrawals after a third-party company was compromised; researchers estimate about $2 million in user funds was stolen via DebugBear software. Different scale, similar weak point: a person or a third-party tool. ## Frequently asked questions ### Is Binance hacked? Yes. On October 29, 2026, attackers drained Binance.com hot wallets across ten blockchains. TRM Labs counts about $6.36 billion stolen, the largest crypto theft on record, and Binance says about $6.3 billion in assets were affected. Binance's early checks point to a compromised DevOps engineer's laptop, with no evidence of a private key leak. ### Is Binance safe from hackers? No exchange can promise that. By Binance's early account, the weak point was a compromised engineer laptop and the systems it could reach, and that kind of risk exists at any custodial platform. Binance says customer deposits were untouched and that it will cover all losses, but those are its own statements and outsiders cannot yet verify them. ### How was Binance hacked? Binance's early checks suggest a compromised DevOps engineer's laptop gave an attacker access to a service container used to submit withdrawals, which the system then treated as legitimate. Binance said it found no evidence that a wallet private key leaked. TRM Labs attributes the theft to North Korea's TraderTraitor group with high confidence. ### Is Binance insolvent? None of the sources reviewed say so. Binance says it will cover all losses from its SAFU emergency fund and company funds, and that customer deposits were untouched. Those are Binance's own statements, and outsiders cannot yet check its finances, so the practical test is whether withdrawals reopen and process normally. ### When will Binance withdrawals resume? Binance said withdrawals should resume within 48 hours, according to Reuters, and that a detailed technical post-mortem will follow. Co-CEO Yi He also said withdrawals stay disabled until Binance confirms its environment is secure, so the timing depends on that check rather than on a fixed date. ### Is Binance US safe? Binance.US and Coinbase both said they are not affected. Binance has said the breach involves Binance.com only, the global exchange. If you hold funds on Binance.US or Coinbase, there is no reported loss there. Wherever you hold funds, the main risk this week is scam messages pretending to offer refunds or account recovery, so rely only on official channels. ## Source Based on TRM Labs' analysis and Binance's statements. https://www.trmlabs.com/resources/blog/north-korea-steals-usd-6-36-billion-from-binance-in-the-largest-hack-in-history --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/binance-hacked-engineer-laptop-theft. Last modified 2026-10-30T07:38:48.000Z. ================================================================================ - URL: https://syntaxradar.com/radar/what-is-ai-capex-alphabet - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-30T05:00:00.000Z - Last modified: 2026-10-30T05:00:00.000Z - Primary source: https://thetechportal.com/2026/10/29/google-parent-alphabet-sees-nearly-500bn-ai-led-market-cap-wipeout-largest-in-company-history/ # What Is AI Capex? Why Alphabet's $51.4B Bill Spooked Wall Street **Short answer:** Alphabet's Q3 2026 capex hit $51.4B, more than double a year earlier, pushing free cash flow to negative $9.1B for a second straight quarter. Shares fell about 12% on Oct. 29, its worst trading day on record. **Takeaway:** AI capex is the money spent on chips, datacenters and power, and Alphabet's quarter shows that bill can outrun even Google's cash machine while the payoff is still arriving. **Key figures (as reported by the source):** - Q3 2026 capex: $51.4B (vs $24B a year earlier) - Free cash flow: -$9.1B - 2026 capex guidance: $195B-$205B On October 29, Alphabet had its worst single trading day on record. The company behind Google lost roughly $500 billion in market value in one session, after a quarterly report showing $51.4 billion of capital spending in three months. That spending has a name you will hear a lot this year: AI capex, short for AI capital expenditure. AI capex is the money tech giants pour into the physical stuff that runs artificial intelligence: datacenters, the specialized chips inside them, and the power and cooling that keep them alive. Alphabet's Q3 2026 numbers are one of the clearest real-world lessons in how AI capex spending works, and why it can make even a fast-growing company look stretched. ### What happened in Alphabet's Q3 2026 earnings Alphabet reported third-quarter results on October 28. On the surface, plenty went right. Total revenue reached $127.6 billion, up 25% from a year earlier, according to [The Tech Portal's breakdown of the results](https://thetechportal.com/2026/10/29/google-parent-alphabet-sees-nearly-500bn-ai-led-market-cap-wipeout-largest-in-company-history/). Google Cloud more than doubled its revenue to $30.7 billion. Three things spoiled the mood. Revenue came in about $640 million short of Wall Street's expectations. Search, Google's oldest money machine, grew 12% to $63.35 billion, below the $64.96 billion analysts had pencilled in. And the spending line was enormous: capital expenditures of $51.4 billion, more than double the $24 billion Alphabet spent in the same quarter of 2025. Shares fell more than 10% after hours. The next day they were down about 12%, and [NBC News reported](https://www.nbcnews.com/business/markets/alphabet-google-stock-ai-advertising-rcna606322) it was the company's worst trading day on record. Even after the drop, Alphabet was worth about $3.42 trillion at midday, the fourth-largest US company. ### What is AI capex? AI capex is capital spending aimed at artificial intelligence. To see what that means, start with plain capex. #### What does capex mean? Capex, short for capital expenditure, is money a company spends on assets it expects to use for years. Typical capex examples are buildings, land, servers and network gear. A datacenter keeps working for a long time, so accountants spread its cost across its useful life instead of charging it all at once. #### Capex vs opex: what is the difference? The difference between capex and opex comes down to timing. Opex, or operational expenditure, covers ordinary running costs such as salaries, electricity bills and marketing, which are used up in the period they are paid. Capex buys things that last, so its cost is spread over the years the asset is used, and that split decides how an expense shows up in the accounts. #### What AI capex buys AI capex is that same capital spending aimed at artificial intelligence. For a company like Alphabet, the shopping list looks roughly like this: - **Chips.** AI models are trained and run on specialized processors. - **Datacenters.** The warehouses that hold those chips, along with the land, construction, racks and networking that tie thousands of machines together. - **Power and cooling.** AI hardware draws a lot of electricity and produces a lot of heat. Securing reliable power and keeping machines cool is part of the bill. Every Gemini answer, every AI Mode search and every cloud customer renting AI compute runs on this hardware. Executives say the Gemini app and AI Mode have reached 1 billion users. Serving that many people with large models takes a lot of physical infrastructure, and that is what the spending buys. It also ripples outward. When hyperscalers (the handful of companies that run giant cloud networks) buy hardware at this scale, they compete with everyone else for the same parts. Hyperscaler capex, the combined capital spending of those cloud giants, therefore reaches well beyond their own balance sheets. We covered one example of that in our explainer on [why AI datacenters are making SSD storage cost more](/radar/ssd-prices-ai-nand-shortage). ![Diagram of capex vs opex: capex is spread over years, opex is paid this period, and AI capex buys chips, datacenters and power](/uploads/blog/what-is-ai-capex-alphabet/img1.webp) ### Why is Alphabet free cash flow negative? Free cash flow is the cash a business has left after paying to run itself and paying for its capital spending. The free cash flow formula is simple: take the cash the business brings in from operations, then subtract what it spent on buildings and equipment. Whatever remains is "free" to pay dividends, buy back shares, pay down debt or sit in the bank. In Q3, that number went below zero. Alphabet's free cash flow was negative $9.1 billion, according to [CNBC's earnings report](https://www.cnbc.com/2026/10/28/alphabet-googl-earnings-q3-2026.html), which put the precise figure at negative $9.07 billion. It was the second straight negative quarter. So what does negative free cash flow mean? It means the business spent more cash than it collected. A company can have negative free cash flow and still be profitable, because capex is paid in cash up front while the accounts spread its cost over years. Alphabet's adjusted earnings per share of $2.67 shows the profit is there. Alphabet still makes a lot of money from ads and cloud. The capex bill is simply larger than the cash those businesses throw off right now. That gap has to be filled somehow, and part of the answer shows up on the balance sheet. Alphabet's long-term debt rose to $125.1 billion, up from $46.5 billion at the end of 2025. The coverage does not say what the borrowing paid for, but the jump arrived alongside the spending surge. Two negative quarters in a row and a debt load that more than doubled in under a year show how much the AI race is costing, even for one of the richest companies on earth. ### How much do big tech companies and hyperscalers spend on AI capex in 2026? Start with Alphabet's capex guidance, which doubles as its AI capex forecast for 2026. The company kept its full-year capex at $195 billion to $205 billion, and it flagged a significant further increase for 2027. To put that in perspective, the $51.4 billion Alphabet spent in Q3 alone is more than twice what it spent in the same quarter a year earlier. And the company is telling investors that next year will be bigger still. Alphabet is one hyperscaler among several, so treat its numbers as one company's AI capex outlook for 2026. If you follow AI capex, the useful habit is to look at each company's capex line next to its free cash flow, because that pairing tells you whether the spending is being paid for by the business or by borrowing. Why is AI capex so high? One piece of evidence is demand. Google Cloud's backlog, the value of contracted work customers have committed to but Google has yet to deliver, crossed $514 billion. ![Bar chart of Alphabet AI capex spending: $24B in Q3 2025, $51.4B in Q3 2026 and $195B to $205B 2026 capex guidance, with free cash flow of -$9.1B](/uploads/blog/what-is-ai-capex-alphabet/img2.webp) ### If Google Cloud doubled, why did investors still react so badly? On paper, Cloud was the star of the quarter. Revenue more than doubled to $30.7 billion, and its operating margin jumped to 35.6% from 20.7%. That means Cloud is keeping far more of each dollar it earns, which is exactly what a business selling AI compute hopes to show. The trouble is that Cloud's success sits next to two less comfortable facts, and both bear on whether AI capex is paying off. The first is Search. Growth slowed from 19% in Q1 to 17% in Q2 to 12% in Q3. Schindler attributed part of that slowdown to AI Mode, Google's AI-powered search experience. JPMorgan analysts called the search slowdown "steeper than expected," according to NBC News. The second is the profit picture. Adjusted earnings per share came in at $2.67 against a consensus of $3.02. Put that next to a doubling capex bill and negative free cash flow, and the market's question becomes simple: when does all this spending start paying for itself? The bill arrives first and the payoff still has to be proven, which is where AI capex ROI concerns come from. #### Is AI capex paying off? A booming Cloud unit and a $514 billion backlog are evidence that the bet could work. Slower Search growth, a profit miss and rising debt are evidence of what it costs while everyone waits to find out. Anyone asking whether there is an AI capex bubble is asking this same ROI question, and the numbers so far point both ways. ### What AI capex means for you You use the output of a datacenter every day, even if you never see one, and this spending reaches you in four ways. **AI features are the payoff.** A company spending $195 billion or more in a year needs people to use what it built, and usage is how that hardware earns its keep. **Search is changing shape.** Alphabet linked part of its Search slowdown to AI Mode, so how you find information, and how websites get visits, is already part of this story. **AI pricing is tied to the bill.** Cheap AI access and a huge infrastructure bill pull in opposite directions, a tension we touched on in [why LLM API pricing fell sharply in 2026](/radar/llm-api-pricing-cuts-2026). How providers balance the two is an open question. **Hardware competition touches consumer gear.** When giant buyers soak up chips, memory and storage, those parts can get scarcer or pricier for everyone else. ### What to watch next If you follow hyperscaler capex news, these signals tell you whether the spending is paying off: - **Alphabet's 2027 capex number.** The company has already flagged a significant increase. The size of that jump will show how far it plans to push the build-out. - **Free cash flow.** After the second straight negative quarter, watch whether it returns to positive free cash flow or stays below zero. - **Search growth.** The slide from 19% to 12% is the trend to follow, along with any detail on how AI Mode affects ads. - **Cloud backlog and margins.** If the $514 billion backlog keeps converting into revenue at strong margins, the capex case gets stronger. - **Debt.** Long-term debt of $125.1 billion is a new feature of Alphabet's finances. Watch how fast it grows. Alphabet's quarter showed what AI capex looks like when the bill arrives before the payoff: $51.4 billion spent in three months, free cash flow below zero for a second straight quarter, and a cloud backlog of more than $514 billion waiting to be delivered. ## Frequently asked questions ### What is AI capex in simple terms? AI capex is capital expenditure on the physical infrastructure behind artificial intelligence: datacenters, specialized chips, networking, power and cooling. Companies spread the cost of these long-lived assets over years instead of treating them like everyday expenses. For Alphabet, that spending reached $51.4 billion in Q3 2026, more than double the $24 billion it spent a year earlier. ### Why is Alphabet's free cash flow negative, and what does it mean? Free cash flow is cash from operations minus capital spending. In Q3 2026 Alphabet's capex was so large that it exceeded the cash its businesses generated, leaving free cash flow at negative $9.1 billion, according to CNBC. It was the second straight negative quarter, and long-term debt rose to $125.1 billion from $46.5 billion at the end of 2025. ### How much is Alphabet spending on AI capex in 2026? Alphabet kept its full-year 2026 capital expenditure guidance at $195 billion to $205 billion, according to The Tech Portal. The company also flagged a significant further increase for 2027. In the third quarter alone it spent $51.4 billion, more than double the $24 billion from the same quarter a year earlier. ### Why did Alphabet stock fall if Google Cloud doubled? Google Cloud revenue more than doubled to $30.7 billion with a 35.6% operating margin, but other numbers worried investors. Search growth slowed to 12%, total revenue missed expectations by about $640 million, adjusted earnings per share missed, and capex more than doubled. Shares fell about 12% on October 29, Alphabet's worst trading day on record. ### Is AI capex paying off yet for Alphabet? It is too early to say, and the evidence points both ways. Google Cloud revenue more than doubled to $30.7 billion and its backlog crossed $514 billion. Against that, Search growth slowed to 12%, adjusted earnings per share missed, free cash flow was negative $9.1 billion and long-term debt rose to $125.1 billion. ### What does capex mean, and what is the difference between capex and opex? Capex, short for capital expenditure, is money spent on long-lived assets such as buildings, land, servers and network gear, with the cost spread over the asset's useful life. Opex covers everyday costs such as salaries and electricity, which are used up when paid. Alphabet's capex reached $51.4 billion in Q3 2026. ## Source Based on Alphabet's Q3 2026 results as reported by CNBC, NBC News and The Tech Portal. https://thetechportal.com/2026/10/29/google-parent-alphabet-sees-nearly-500bn-ai-led-market-cap-wipeout-largest-in-company-history/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/what-is-ai-capex-alphabet. Last modified 2026-10-30T05:00:00.000Z. ================================================================================ - URL: https://syntaxradar.com/radar/mistral-large-4-open-weights-explained - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-30T00:00:00.000Z - Last modified: 2026-10-30T00:00:00.000Z - Primary source: https://mistral.ai/news/mistral-large-4-oss/ # Mistral Large 4: What a 1-Trillion-Parameter Open Model Means **Short answer:** Mistral published the weights of Mistral Large 4, a 1-trillion-parameter mixture-of-experts model, on October 29 under a modified MIT license that lets anyone use and fine-tune it but requires a separate license for Model-as-a-Service providers. **Takeaway:** Mistral Large 4 gives anyone a downloadable frontier-scale model with commercial rights, but its 2 TB-class memory footprint and hosting-service license clause mean most people will use it through an API rather than at home. **Key figures (as reported by the source):** - Total parameters: 1 trillion - API price (input / output): $0.70 / $2.10 per 1M tokens - Coding Agent Index (Mistral-reported): 48% Mistral just put a trillion-parameter AI model on the internet for anyone to download. On October 29, Mistral published the weights of Mistral Large 4, three weeks after opening a public preview through its API, [according to Mistral's announcement](https://mistral.ai/news/mistral-large-4-oss/). That means anyone can download it, inspect it and fine-tune it themselves. But "open" is doing a lot of work in that sentence. The license has a catch for one kind of business, the trillion parameters are not all used at once, and every impressive benchmark in the launch post was measured by Mistral itself. Below: what the model is, what the license lets you do, what it costs per token, and whether your own computer can run it. ### What is Mistral Large 4? It is a general-purpose large language model that Mistral describes as natively multimodal, meaning it was built to handle more than plain text from the start rather than having image support bolted on later. It is the successor in the line that includes the Mistral Large 3 model. Its scale is what makes this release stand out, and the release date for the weights, October 29, followed a public preview that opened on October 6. The headline figure for the model's parameters is 1 trillion. Parameters are the numbers a model learns during training; more of them usually means more knowledge and skill packed inside, at the cost of more memory and compute to run. The model has three reasoning modes: Low, Medium and High, with High as the default. In practice that is a dial between speed and care. Low answers quickly and spends fewer tokens thinking; High works through a problem more carefully and costs more per answer. For developers, Mistral says the model can be served with the popular open serving tools vLLM and SGLang, and it mentions an NVIDIA NIM container and TensorRT LLM for high-throughput deployments. ### How can a 1-trillion-parameter model use only about 50 billion at a time? Large 4 is a mixture-of-experts model. Instead of one giant network where every parameter fires for every word, it contains many smaller specialist sub-networks, called experts, plus a router that picks a few of them for each token it processes. So there are two numbers that matter: - **Total parameters:** everything stored in the model, about 1 trillion. This decides how much memory you need to hold it. - **Active parameters:** the slice actually used to produce each token. This decides how much computing each word costs. Mistral's own launch post does not spell out the active count. The clue is in the name of the Hugging Face repository, Mistral-Large-4-1T-A52B, which by the usual naming convention reads as 1 trillion total and 52 billion active. Some earlier reports cited 49 billion, so the safe summary is roughly 50 billion active parameters per token. A useful analogy is a large hospital. It employs thousands of specialists (total parameters), but any single patient sees only a handful of them (active parameters). The hospital still needs a building big enough for everyone, which is why the memory bill stays huge even when the work per patient is modest. The payoff is that you get the knowledge of a very large model while paying the compute cost of a much smaller one on each token, which helps explain the API prices below. ![Mixture-of-experts diagram showing a router sending each token to 3 of 32 expert blocks, so about 50 billion of the open weight model's 1 trillion parameters are active per token](/uploads/blog/mistral-large-4-open-weights-explained/img1.webp) ### Is Mistral Large 4 open source? Not in the strict sense, and the difference matters if you plan to build a business on it. Mistral released the model under what it calls the Mistral Ecosystem Open-Weight license, a modified MIT license. MIT is one of the most permissive licenses in software, and this version keeps most of that freedom: according to Mistral, you may use, modify and fine-tune the model, including for commercial purposes. The modification is the catch. Businesses that run a Model-as-a-Service offering, meaning they host the model and sell access to it to others, must get a separate commercial license from Mistral. On that reading, a startup that builds a legal research tool on top of the model is fine, while a cloud company that wants to sell access to the model itself needs to talk to Mistral first. That is why the accurate label is "open weights" rather than "open source," a distinction that applies to most open weight models today. Classic open source licenses do not restrict who may offer the software as a service. When people search for the best open source LLM, they often mean "a model I can download and run," and on that practical test Large 4 qualifies. On the stricter definition it does not. If the distinction feels fuzzy, our explainer on [what open weight models are and why they matter](/radar/open-weight-models-explained-ibm) walks through it in more detail. The short version: you get the finished model, and the license still limits how you can compete with its maker. That holds for open weight AI models in general, so read the license before you read the benchmark table. ### Mistral Large 4 vs GLM-5.3: what do the benchmarks say? Every number in this section comes from Mistral's announcement and has not been independently verified. Treat it as the company's own scorecard. The comparison Mistral leans on hardest is with GLM-5.3, another large open-weight model. On the Coding Agent Index, which tests how well a model works through programming tasks on its own, Mistral reports a score of 48%, effectively tied with GLM-5.3. The interesting claim is efficiency: Mistral says its model uses 47% fewer tokens per task on average. Fewer tokens means faster answers and smaller bills for the same result.
Benchmark (Mistral-reported)Mistral Large 4Context
Coding Agent Index48%Effectively tied with GLM-5.3, 47% fewer tokens per task
Terminal-Bench 4.041%Up from 28% in the public preview, 30% fewer output tokens
BrowseComp84.5%Up from 75% for an earlier checkpoint
FinWorkBench73.3%Second among open models, behind Kimi K3
B3 Agent Security97.1%Attack resistance, up from 93.3% in preview
A few of these deserve a plain-English gloss. Terminal-Bench measures whether a model can get real work done in a command-line environment, and the jump from 28% to 41% between preview and final release is the biggest improvement Mistral reports. BrowseComp tests whether a model can hunt down hard-to-find facts on the web. FinWorkBench focuses on financial work tasks, where Mistral places itself second among open models behind Kimi K3. On safety, Mistral cites a 51% score on Artificial Analysis's Cyber Index, a 96.6% average refusal rate on malicious cyber requests, and the 97.1% attack-resistance figure in the table. What this adds up to: on Mistral's numbers, Large 4 is competitive with the strongest open-weight models rather than clearly ahead of them, and its main selling point is doing similar work with fewer tokens. Independent testers such as [Vals AI, which lists the model](https://vals.ai/models/mistralai_mistral-large-4), are the place to check whether that holds up. ### What does Mistral Large 4 cost to use? If you do not want to host anything, you can call the model through Mistral's API. According to the announcement, pricing is: - **$0.70** per million input tokens (what you send in) - **$0.07** per million cached input tokens (repeated input the system has already seen, such as a long instruction prompt reused across requests) - **$2.10** per million output tokens (what the model writes back) A token is roughly a short word or a chunk of one. As a simple illustration of our own, sending a million tokens in and getting a million tokens back would cost $2.80 in total. The cached rate is a tenth of the normal input price, which rewards apps that reuse the same long system prompt over and over. Output costs three times as much as input, so a chatty model gets expensive. That is why Mistral's claims about using fewer tokens per task matter beyond bragging rights: if they hold up, they translate directly into a lower bill. For the wider picture on how these prices have been falling, see our breakdown of [LLM API pricing cuts in 2026](/radar/llm-api-pricing-cuts-2026). ### Can you run Mistral Large 4 locally? For almost everyone, no. Mistral does not publish hardware requirements, so here is our own back-of-envelope arithmetic, not a Mistral figure. Every parameter has to sit in memory. At 16-bit precision, each parameter takes 2 bytes, so 1 trillion parameters works out to about 2 terabytes of weights. A high-end gaming graphics card has a small fraction of that, and so does a well-equipped laptop. Even compressed versions (quantized to 8 or 4 bits per parameter, which shrinks the file at some cost in quality) would by the same arithmetic still need roughly 1 TB or 500 GB. Remember the mixture-of-experts point: only about 50 billion parameters are active per token, which keeps the compute per word manageable. But the router can pick any expert for any token, so the whole model still has to be loaded and ready. Each token uses only the active parameters, but the full set still has to sit in memory. Realistically, this is a model for companies and research groups with multi-GPU servers, which fits with Mistral pointing to vLLM, SGLang and NVIDIA's serving stack. For an individual, the practical routes are the API or a hosting provider that has sorted out the license. If you want something on your own machine, smaller open source LLMs remain the better fit. The real winners from the weight release are mid-sized companies that want a frontier-scale model on their own infrastructure, for privacy, compliance or customization, without sending data to someone else's API. ![Bar chart of the memory needed to run Mistral Large 4 locally: about 2 TB at 16-bit, 1 TB at 8-bit and 500 GB at 4-bit](/uploads/blog/mistral-large-4-open-weights-explained/img2.webp) ### What to watch next - **Independent benchmarks.** Every score so far is Mistral's own. Third-party results will show whether the token-efficiency claims survive outside the company's tests. - **The context window.** Sources disagree: Mistral's documentation lists 1 million tokens, while Vals AI lists 512,000. Expect clarification as people test it. - **License fights.** Watch how hosting companies respond to the Model-as-a-Service clause, and whether Mistral licenses resellers or keeps the API business to itself. - **Fine-tunes.** With weights public, specialized versions built on it should start appearing. [Early coverage from The New Stack](https://thenewstack.io/mistral-large-4-weights/) tracked the preview; the full release is where the community takes over. ## Frequently asked questions ### What is Mistral Large 4? Mistral Large 4 is a natively multimodal large language model from Mistral with about 1 trillion total parameters. It uses a mixture-of-experts design, so roughly 50 billion parameters are active for each token. It offers Low, Medium and High reasoning modes and is available through Mistral's API and as downloadable weights released on October 29. ### Is Mistral Large 4 open source? It is open weights rather than classic open source. The Mistral Ecosystem Open-Weight license, a modified MIT license, allows use, modification, fine-tuning and commercial use. However, businesses that run a Model-as-a-Service offering, hosting the model and selling access to others, need a separate commercial license from Mistral. ### Can you run Mistral Large 4 locally? Not on normal consumer hardware. Mistral does not publish requirements, but by our own arithmetic 1 trillion parameters at 16-bit precision is about 2 TB of weights, and even 4-bit compression would leave roughly 500 GB. It is built for multi-GPU servers running tools like vLLM or SGLang, so individuals are better off using the API. ### How does Mistral Large 4 compare with GLM-5.3? According to Mistral, Large 4 scores 48% on the Coding Agent Index, effectively tied with GLM-5.3, while using 47% fewer tokens per task on average. These figures come from Mistral itself and have not been independently verified, so check third-party benchmark sites before treating the comparison as settled. ## Source Based on Mistral's open-weights announcement. https://mistral.ai/news/mistral-large-4-oss/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/mistral-large-4-open-weights-explained. Last modified 2026-10-30T00:00:00.000Z. ================================================================================ - URL: https://syntaxradar.com/radar/ssd-prices-ai-nand-shortage - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-29T18:42:38.817Z - Last modified: 2026-10-29T18:42:38.817Z - Primary source: https://www.notebookcheck.net/SSD-prices-explode-by-303-in-a-year-as-AI-fuels-NAND-flash-shortage.1433261.0.html # SSD Prices: Why AI Datacenters Are Making Storage Cost More **Short answer:** The contract price of a 256GB TLC SSD hit about $134.20 in Q3 2026, up 303% from $33.30 a year earlier, as AI cloud buildouts absorb NAND flash. TrendForce expects NAND contract prices to rise another 15-20% in Q4 2026. **Takeaway:** AI datacenters are buying up the same NAND flash chips that go into laptops, phones and PCs, so storage is getting pricier and relief probably waits until new factory capacity arrives in mid-to-late 2027. **Key figures (as reported by the source):** - 256GB TLC SSD contract price, Q3 2026: $134.20 (from $33.30, +303% YoY) - TrendForce 4Q26 NAND contract price forecast: +15-20% QoQ - Enterprise SSD bit demand growth, 2026: More than 80% YoY If you have priced a new laptop, a gaming PC upgrade or a spare drive lately, you have probably noticed that SSD prices no longer behave the way they used to. Drives that were easy to shrug at a year ago now cost noticeably more. The reason sits in giant server halls. Cloud companies building AI services are buying enormous amounts of NAND flash memory, the storage chip inside every SSD and phone, and the factories that make it have already promised much of their output to those customers. Your next laptop is competing for the same chips. ### Why are SSD prices going up? The reason SSDs are so expensive right now starts in the wholesale market, and the sharpest number comes from there. According to [Notebookcheck's report on the NAND shortage](https://www.notebookcheck.net/SSD-prices-explode-by-303-in-a-year-as-AI-fuels-NAND-flash-shortage.1433261.0.html), the contract price of a 256GB TLC SSD reached about $134.20 in the third quarter of 2026. A year earlier it was about $33.30. That is a 303% rise in twelve months, and roughly 20% above the previous quarter alone. Bigger drives followed the same path. The 512GB equivalent reportedly went from about $62.40 to $252 over the same period. Further down the supply chain, the price of a 1-terabit TLC NAND flash chip, one of the raw parts drive makers solder onto a circuit board, reportedly rose 123% in the six months to October 21. The driver is demand from AI. Market research firm TrendForce projects that enterprise SSD bit demand, meaning the total amount of storage capacity that datacenter drives need, will grow more than 80% year over year in 2026. Cloud providers expanding their AI inference infrastructure, the servers that answer your chatbot prompts, are behind that growth. ![Column chart of SSD contract prices showing a 256GB TLC SSD rising from $33.30 a year ago to $134.20 in Q3 2026, a 303% increase, and a 512GB SSD from $62.40 to $252](/uploads/blog/ssd-prices-ai-nand-shortage/img1-v2.webp) ### What is NAND flash memory? In plain terms, NAND flash storage is the type of memory that keeps your files when the power is off. Your laptop's SSD, your phone's storage, a USB stick and a memory card all use it. Inside each chip is a dense grid of tiny cells that trap an electrical charge, and the level of that charge encodes your data. Manufacturers squeeze more data into each cell to cut costs. TLC (triple-level cell) flash stores three bits per cell and is the mainstream choice for consumer SSDs. QLC (quad-level cell) stores one more bit per cell, which makes it cheaper per gigabyte but slower to write and less durable. That trade-off suits huge datasets that are read far more often than they are rewritten. Phones use the same NAND flash, just packaged differently. Instead of a removable drive, it ships as eMMC or UFS storage soldered to the phone's board. So when NAND gets expensive, laptops, desktops and phones all feel it, even if each product shows the increase at a different time. ### Contract prices vs shelf prices: what the 303% figure means for SSD pricing That 303% number describes contract prices: what big buyers such as PC makers and drive brands pay manufacturers for bulk orders, usually negotiated every quarter. Notebookcheck is explicit that these are contract figures and do not mean every consumer SSD rose by the same amount. The shelf price of an SSD moves later and less evenly, because stores and brands sell from inventory bought at older prices and set margins drive by drive. Shelf prices are rising too, though unevenly, and for some models the jump is steep. The 1TB drive is the number most buyers feel: Notebookcheck reports that a 1TB Gen4 NVMe SSD averaged about $144 in mid-October. One popular model, the Samsung 990 Evo Plus 1TB, was listed at $237, against a previous low of about $59. The average tells you what a typical 1TB fast drive costs today. The Samsung listing shows how far an individual model can drift above its previous low when stock runs short. Your own experience will depend on which drive you want, where you shop and how much old inventory that retailer still holds.
Price typeWho pays itExample from the data
Contract pricePC makers and drive brands buying in bulk256GB TLC SSD: about $33.30 a year ago, about $134.20 in Q3 2026
Component priceDrive makers buying raw NAND chips1-terabit TLC NAND: up 123% in six months to October 21
Retail priceYou, at a store1TB Gen4 NVMe average: about $144 in mid-October
### How AI datacenters ended up competing with your laptop for NAND flash AI models need fast access to vast amounts of data, and much of it has to live on flash so it can be fetched in milliseconds. According to [TrendForce's fourth-quarter memory outlook](https://www.trendforce.com/presscenter/news/20260930-13258.html), agentic AI, software that takes multi-step actions on its own, is driving data growth for real-time retrieval and caching. Rising use of QLC drives in vector databases, the search systems that help AI find relevant information, adds to the pull. Those buyers also have deep pockets and long planning horizons. Notebookcheck reports that Micron has committed more than 75% of its expected fiscal 2027 output, and that SK hynix has allocated its NAND production through the end of 2026. When supply is spoken for that far ahead, anyone shopping on a shorter timeline, from a laptop brand to a home builder, pays whatever is left over. TrendForce says enterprise SSD price growth is expected to accelerate in the fourth quarter, making it the only memory category with a bigger increase than in the prior quarter. Consumer device makers are responding in a different way. PC brands are cutting SSD capacities in mainstream models to hold down costs, and they are leaning on inventory built in the first half of 2026. Smartphone brands, the main buyers of eMMC and UFS storage, are also relying on existing stock, so their new purchases are very limited. That may be why shelf prices have not fully matched the wholesale numbers yet: device makers are drawing on stock bought earlier. ![Diagram showing AI datacenters and laptop and phone makers both drawing on the NAND flash supply, with Micron committing 75% of fiscal 2027 output, SK hynix sold out through 2026 and contract prices forecast to rise 15-20% next quarter](/uploads/blog/ssd-prices-ai-nand-shortage/img2-v2.webp) ### Will SSD prices drop? The forecasts say not soon. TrendForce projects NAND flash contract prices will rise another 15-20% quarter over quarter in the fourth quarter of 2026. That is on top of the increases already baked into this year's contracts. The bigger constraint is factory capacity. Adding flash production is slow, and Notebookcheck reports that new capacity may not meaningfully add supply until mid-to-late 2027. Until then, a drop would need AI demand to cool, and the current forecasts point the other way. Nothing in the current data says when prices will turn, so plan around rising SSD prices through 2026 and treat any dip before new capacity arrives as a bonus. ### What buyers should do now: laptops, phones and PCs If you need storage soon, waiting is unlikely to pay off. With contract prices still forecast to rise and new supply a long way off, a drive you need this year will probably cost more later in the year. Buy what you need, but there is little reason to stockpile. #### Building or upgrading a PC Check a model's price history before you pay, because finding a cheap SSD right now means comparing options. The gap between the 990 Evo Plus listing and its previous low shows how much individual drives can swing, so compare several brands of similar speed and size. A slightly slower drive at a normal price often beats a premium model at a shortage price, especially for gaming and everyday work where the difference is hard to notice. #### Buying a laptop Read the storage line on the spec sheet carefully. Because PC brands are trimming SSD capacities in mainstream models, this year's version of a laptop may ship with less storage than last year's at a similar price. If you keep lots of photos, games or video locally, paying for the bigger configuration upfront may save you buying more storage later. #### Choosing a phone Phone makers are still drawing on existing storage inventory, according to TrendForce, which may delay price changes. Storage is usually fixed on a phone, so pick the capacity you will need for the whole time you plan to own it, and lean on cloud backup for older photos and videos rather than paying for a top tier you will not fill. #### Already own a drive? If your current SSD works, keep it. Freeing up space and moving old files to an external drive or cloud storage is the cheapest upgrade available while prices are high. ### What to watch next Watch three signals. First, TrendForce's next quarterly outlook, which will show whether contract increases are speeding up or slowing after the forecast 15-20% fourth-quarter rise. Second, laptop and phone launches: smaller default storage or higher prices for the same capacity will show that cheaper inventory has run out. Third, updates from flash makers such as Micron and SK hynix on when new capacity starts shipping, since [Notebookcheck's reporting](https://www.notebookcheck.net/SSD-prices-explode-by-303-in-a-year-as-AI-fuels-NAND-flash-shortage.1433261.0.html) points to mid-to-late 2027 before supply catches up. Until then, AI datacenters set the pace for SSD prices. Every new wave of AI services needs more fast storage, and [TrendForce's demand projections](https://www.trendforce.com/presscenter/news/20260930-13258.html) suggest that appetite is still growing. ## Frequently asked questions ### Why are SSD prices going up? Cloud providers expanding AI inference infrastructure are buying huge amounts of NAND flash, the chip inside every SSD. TrendForce projects enterprise SSD bit demand will grow more than 80% in 2026, and manufacturers such as Micron and SK hynix have committed much of their production in advance, leaving less supply for laptops, PCs and phones. ### Did consumer SSDs really rise 303%? No. The 303% figure is a contract price, what large buyers pay manufacturers in bulk, for a 256GB TLC SSD, which went from about $33.30 to $134.20 in a year. Retail prices move later and unevenly. A 1TB Gen4 NVMe SSD averaged about $144 in mid-October, though some individual models are listed far above their previous lows. ### When will SSD costs come down? Not in the near term on current forecasts. TrendForce expects NAND flash contract prices to rise another 15-20% quarter over quarter in the fourth quarter of 2026, and new factory capacity may not meaningfully add supply until mid-to-late 2027. A drop before then would need AI demand to cool, which the forecasts do not show. ### What is NAND flash memory? NAND flash is the memory that keeps data when the power is off. It sits inside SSDs, phones, USB sticks and memory cards. Each chip holds a dense grid of cells that store data as trapped electrical charge. TLC flash is the mainstream consumer type, while QLC packs more data per cell at lower cost but with slower writes and less durability. ## Source Based on TrendForce's 4Q26 outlook and Notebookcheck's price data. https://www.notebookcheck.net/SSD-prices-explode-by-303-in-a-year-as-AI-fuels-NAND-flash-shortage.1433261.0.html --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/ssd-prices-ai-nand-shortage. Last modified 2026-10-29T18:42:38.817Z. ================================================================================ - URL: https://syntaxradar.com/radar/is-a-computer-science-degree-worth-it - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-29T12:02:02.365Z - Last modified: 2026-10-29T12:02:02.365Z - Primary source: https://www.studentclearinghouse.org/nscblog/computer-science-enrollment-is-cooling/ # Is a Computer Science Degree Worth It? What the 8.1% Drop Means **Short answer:** US undergraduate computer science enrollment at four-year colleges fell 8.1% in fall 2025 and 8.4% in spring 2026, per the National Student Clearinghouse, as overall enrollment grew. A Hoover study found CS's share of enrollment fell for the first time in about 20 years. **Takeaway:** CS enrollment is falling after a decade of growth, for several reasons including AI, so the degree still pays off for students who build deep fundamentals and real experience, and less for those who only wanted a guaranteed job. **Key figures (as reported by the source):** - Four-year CS enrollment, fall 2025: -8.1% - Four-year CS enrollment, spring 2026: -8.4% - CS bachelor's degrees awarded, 2024: 122,000 Is a computer science degree worth it? That question sits under a fresh wave of headlines about students walking away from the major, and the numbers are smaller, and more interesting, than the panic suggests. According to the [National Student Clearinghouse](https://www.studentclearinghouse.org/nscblog/computer-science-enrollment-is-cooling/), undergraduate computer science enrollment at four-year US colleges fell 8.1% in fall 2025 and 8.4% in spring 2026, while overall enrollment grew. That is a real drop. By one long-running measure, it is the first time in about two decades that computer science has lost ground as a share of what students take. Whether AI caused it is a separate question, and the economist behind that measure says his data cannot answer it. For a student comparing majors, a parent paying for one, or a career changer eyeing a return to school, the useful question is narrower than the headlines: is computer science worth it in 2026, and which parts of a CS education still pay off, and for whom. ### What the computer science enrollment decline actually shows In fall 2025, four-year colleges had about 606,100 undergraduates in computer science, down from about 659,700. That is the 8.1% figure. The drop was not even. Depending on the type of institution, fall 2025 declines ranged from 3.6% at primarily associate-granting colleges that also award bachelor's degrees to 14.0% at the graduate level. Over the same period, total US postsecondary enrollment grew 1.0% while computer science shrank. The spring 2026 follow-up showed the trend holding:
Spring 2026, computer scienceChange
Four-year institutions-8.4%
Primarily associate-granting baccalaureate institutions-9.3%
Two-year colleges-11.2%
All undergraduate enrollment (every major)+1.3%
A second dataset points the same way. Stanford Hoover Institution economist Jacob Light analysed more than 50 million course sections across 1,019 US colleges going back to 1996, [as reported by The Next Web](https://thenextweb.com/news/computer-science-enrollment-falls-ai-first-time). In the 2025-26 school year, computer science's share of enrollment fell for the first time in about 20 years, and average CS class enrollment at a school dropped about 4.6%. National Science Foundation data cited in the same report shows CS bachelor's degrees more than doubling, from about 56,000 in 2014 to 122,000 in 2024. Enrollment remains above its 2022 level. The field has a long history of booms followed by busts, and an 8% dip is still a lot of students. ![Bar chart comparing the 8.1% fall 2025 and 8.4% spring 2026 drop in computer science enrollment against small gains in overall enrollment, from National Student Clearinghouse data](/uploads/blog/is-a-computer-science-degree-worth-it/img1-v2.webp) ### Why is computer science enrollment dropping? Anyone wondering why fewer students pick the major should start with Light's list. He names several likely drivers, and only one of them involves AI directly. They are a weaker job market for junior developers, growing doubts about the value of the degree, weaker maths preparation among incoming students, and the rise of data science as a separate major that pulls in people who once would have picked CS. Where the students are going is clearer than why. The Clearinghouse reports significant spring 2026 gains in engineering and health professions, though it gives no figures. [Moneywise, via Yahoo Finance](https://finance.yahoo.com/sectors/technology/articles/computer-science-once-golden-ticket-140500823.html), reports mechanical and electrical engineering enrollment growing about 11% and 14%. At Princeton, the Class of 2028 had 74 computer science B.S.E. declarations, down from 117 and 150 in the prior years. Students who start in CS are also less likely than peers in other big majors to still be enrolled a year later. Among the ten most popular bachelor's majors, computer science had the lowest second-fall persistence, at 85.0%, which is 1.3 points below the fall 2022 cohort. Engineering had the highest, at 93.1%. The job market backdrop explains some of the nerves. New college graduates have had an unemployment rate above the overall US rate for five straight years, according to the New York Times as cited by Moneywise. That is a problem for every major, but CS had long been seen as the safe one, so the disappointment lands harder. ### Is computer science enrollment really down because of AI? Not provably. Light calls his numbers descriptive, not causal, and makes no claim that AI caused the decline. AI is the easy headline, but the other drivers on his list (a tighter junior hiring market, maths preparation, data science) would push numbers down with or without chatbots. What AI plausibly does is sharpen the doubt. When the boss of JPMorgan says AI has erased a third of jobs in some units, as TNW reports, an 18-year-old choosing a major hears that every programming job is next. The enrollment data cannot tell you whether that fear is accurate. It can only tell you that fewer students are acting on the old assumption that CS is a sure thing. The boom-and-bust history is the other reason to be careful. CS enrollment has risen and fallen before as the job market swung. This drop follows a decade of growth, and totals are still above 2022. It could be a cycle turning or the start of a longer slide. One year of data cannot settle which. ### Is a CS degree worth it with AI? The argument for the degree still holds with AI, on new grounds. Google DeepMind's chief argues that a STEM degree still makes people far more effective with AI, according to TNW. AI tools write code quickly, but someone has to decide what to build, check whether the output is right, and fix it when it fails in ways the tool did not predict. That work leans on the parts of a CS education that are hardest to shortcut. Algorithms, data structures, operating systems, networking, probability and linear algebra are what let you read AI-generated code and spot the subtle bug, the security hole or the design that will not scale. Those fundamentals keep their value as the tools change. The pressure is real at the entry level. Light points to a weaker junior developer market, and that is where entry-level graduates feel it first. So whether a CS degree is worth it now turns on one question: will you leave school able to do work a tool cannot do alone? ### Is a computer science degree worth it? A practical way to decide Are computer science degrees worth it for everyone? No, and they never were. If you are deciding whether to major in computer science, five questions about your own situation do most of the work. ![Five numbered question cards for deciding whether a computer science degree is worth it: love the problems, neighbouring major, real experience, career changer, and skills that last](/uploads/blog/is-a-computer-science-degree-worth-it/img2-v2.webp) #### Do you like the problems, or the old paycheck story? If you mostly want the salary, the return on a computer science degree depends on finishing it and on landing real work afterwards. The 85.0% persistence figure means about 15 in every 100 CS bachelor's starters were not still enrolled by their second fall, the worst rate among the big majors. If the attraction was a guaranteed job rather than the work itself, that story is weaker now. If you enjoy logic, maths and building things, a computer science major is worth it, because the degree still trains exactly that. #### Would a neighbouring major open the same doors? Students are already moving to electrical engineering, mechanical engineering and data science. Many careers draw on all of them, and computing is rarely the only route in. Electrical engineering sits close to hardware and chips, and data science sits close to AI work. Pick the field whose hard problems you want to own, then add programming on top. #### Can you get real experience before you graduate? Does a computer science degree guarantee a job? It never did, and with entry-level hiring tighter the degree alone carries less weight. Internships, research with a professor, open-source contributions and shipped side projects are what separate graduates. If your school makes those hard to get, factor that in. #### Are you a career changer? Is computer science worth it if you already have a degree and a job? Sometimes. The full degree makes most sense if you want roles that lean on deep fundamentals, such as systems, security or machine learning research. If you want to apply software to a field you already know, targeted coursework plus a portfolio may get you further, faster. #### Which skills hold their value either way? Maths and statistics, systems thinking, the ability to review and test code you did not write, and clear writing about technical trade-offs. None of these get easier for a chatbot to replace just because it can produce code. All of them make you better at using one. ### What to watch next The Clearinghouse's fall 2026 numbers will show whether this was a one-year correction or the start of a longer slide. Watch three other signals alongside them: whether the engineering and data science gains continue, whether junior developer hiring recovers, and whether new graduate unemployment keeps running above the national rate. For students who build fundamentals and real experience, the evidence still says a computer science degree is worth it, with less certainty than before. If fewer students enter CS this year, fewer will graduate into the market four years from now. The field has a history of booms and busts, though nobody can promise that pattern repeats. ## Frequently asked questions ### Is a computer science degree worth it in 2026? For students who enjoy maths, logic and building software, yes, as long as they pair the degree with internships, research or shipped projects. Enrollment fell 8.1% in fall 2025, but CS bachelor's degrees still more than doubled between 2014 and 2024. The degree pays off most as training in fundamentals AI tools cannot replace. ### How much has computer science enrollment dropped? According to the National Student Clearinghouse, undergraduate CS enrollment at four-year US colleges fell 8.1% in fall 2025, from about 659,700 to 606,100 students, and 8.4% in spring 2026. Two-year colleges saw an 11.2% spring drop. Overall undergraduate enrollment grew 1.3% over the same spring period, so the decline is specific to computing. ### Is AI the reason fewer students study computer science? The data cannot prove it. Hoover economist Jacob Light, who analysed more than 50 million course sections, calls his numbers descriptive, not causal. He lists a weaker junior developer job market, doubts about the degree's value, weaker maths preparation and the rise of data science as likely drivers. AI may sharpen those doubts, but it is not the only factor. ### What are students studying instead of computer science? Engineering and health professions posted significant gains in spring 2026, according to the Clearinghouse. Moneywise reports mechanical and electrical engineering enrollment growing about 11% and 14%. Data science is also drawing students who might once have chosen CS. Many of these fields still involve heavy programming, so the skills overlap more than the major names suggest. ### Is computer science still a good major with AI? It can be, for students who build deep fundamentals. Google DeepMind's chief argues a STEM degree still makes people far more effective with AI, according to TNW. Entry-level hiring is tighter, with a weaker junior developer market, so internships, research and shipped projects matter as much as the degree itself. ## Source Based on National Student Clearinghouse data. https://www.studentclearinghouse.org/nscblog/computer-science-enrollment-is-cooling/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/is-a-computer-science-degree-worth-it. Last modified 2026-10-29T12:02:02.365Z. ================================================================================ - URL: https://syntaxradar.com/radar/macbook-touch-screen-new-macbook-pro-explained - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-29T06:56:52.882Z - Last modified: 2026-10-29T06:56:52.882Z - Primary source: https://www.apple.com/newsroom/2026/10/apple-unveils-groundbreaking-macbook-pro-with-proled-xdr-display/ # MacBook Touch Screen: Which New MacBook Pro Finally Has One **Short answer:** On October 28, 2026, Apple announced the MacBook Pro with PROLED XDR Display, 14- and 16-inch models with M5 Pro or M5 Max and the first touch screen on a MacBook Pro. It ships November 6. The M6 14-inch MacBook Pro announced a day earlier has no touch screen. **Takeaway:** Only the new M5 Pro and M5 Max MacBook Pro with the PROLED XDR OLED display has touch, it lists $400 above the previous models at each chip tier, and every performance claim is Apple's until reviews arrive after November 6. **Key figures (as reported by the source):** - Starting price (14-inch, M5 Pro): $2,899 - Peak brightness (Apple's claim): 2000 nits - Availability: November 6, 2026 Yes, there is finally a MacBook touch screen. On October 28, 2026, Apple announced the [MacBook Pro with PROLED XDR Display](https://www.apple.com/newsroom/2026/10/apple-unveils-groundbreaking-macbook-pro-with-proled-xdr-display/), a 14-inch and 16-inch laptop that Apple says is the first MacBook Pro with Touch. Pre-orders opened the same day, and the machines arrive on Friday, November 6. One day earlier, Apple announced a different laptop, the 14-inch MacBook Pro with M6, and that one does not have a touch screen. Both are called MacBook Pro and both ship November 6, so it is easy to buy the wrong one and find you cannot tap the screen. Nobody has independently tested these laptops yet. Every battery and display number below is Apple's claim until reviews land. ### Does MacBook have touch screen? Only one family does, and it is the new high-end one. The touch screen MacBook Pro comes in 14-inch and 16-inch sizes, uses Apple's new PROLED XDR display, and runs M5 Pro or M5 Max chips. Not M6. That surprises people, because the M6 is the newer chip. The M6 model is the cheaper 14-inch MacBook Pro. [MacRumors describes it](https://www.macrumors.com/2026/10/28/apple-new-macbook-pro/) as the low-end 14-inch option, sitting under the touch models, while Apple keeps selling the existing mini-LED MacBook Pro models alongside both. The two models side by side:
MacBook Pro with PROLED XDR (touch)14-inch MacBook Pro with M6
AnnouncedOctober 28, 2026October 27, 2026
Touch screenYesNo
Sizes14-inch and 16-inch14-inch
ChipsM5 Pro or M5 MaxM6 (12-core CPU, 12-core GPU)
DisplayTandem OLED, up to 2000 nits peak HDR, 10,000,000:1 contrastLiquid Retina XDR, up to 1600 nits peak HDR, 1,000,000:1 contrast
Starting price$2,899 (14-inch, M5 Pro)$1,999 (16GB, 1TB)
Battery (Apple's claim)Up to 24 hoursUp to 24 hours
ShipsNovember 6November 6
Sources: [Apple's touch MacBook Pro announcement](https://www.apple.com/newsroom/2026/10/apple-unveils-groundbreaking-macbook-pro-with-proled-xdr-display/), [Apple's M6 announcement](https://www.apple.com/newsroom/2026/10/apple-introduces-the-new-14-inch-macbook-pro-with-m6/), [MacRumors pricing](https://www.macrumors.com/2026/10/28/new-macbook-pro-maxed-out-cost/). If you want to know whether the M6 chip itself is worth the upgrade over M5, we covered that in [MacBook Pro M6 vs M5: who should upgrade](/radar/macbook-pro-m6-vs-m5-upgrade). This article sticks to the touch model. ![Side by side comparison of the MacBook Pro with PROLED XDR display and M5 Pro or M5 Max chips, touch yes, announced Oct 28, versus the 14-inch MacBook Pro with M6, touch no, announced Oct 27](/uploads/blog/macbook-touch-screen-new-macbook-pro-explained/img1-v2.webp) ### What the MacBook touch screen actually does Think of touch as a third way to control the Mac, next to the keyboard and trackpad, not a replacement for them. Apple says touch works alongside keyboard shortcuts and trackpad gestures, so in principle you can tap when it is handy and use the keyboard and trackpad the rest of the time. The gestures Apple lists are the ones you already know from a phone. Swipe to scroll a webpage. Pinch to zoom in on a photo. Drag slides from one presentation into another. Double-tap a window to fill the screen, or flick it to tile it into place. On its [product page](https://www.apple.com/macbook-pro/proled-xdr-display/), Apple adds drawing and annotating onscreen, rotating 3D objects and playing virtual instruments. According to Apple, many existing apps work with touch right away because macOS automatically translates touches into clicks. A tap is a click. For many apps, that means basic tapping works without an update. Developers who want more than that can use updated system gesture APIs to build controls designed for fingers. Apple names a long list of third-party apps that will take advantage of touch: Notion, Canva, Zoom, Microsoft Word, Excel, PowerPoint and Outlook, Adobe Lightroom, DaVinci Resolve, Blender, Maxon Cinema 4D, Maxon ZBrush, Shapr3D and Algoriddim djay Pro. That list leans heavily toward creative and 3D work, which tells you who Apple has in mind. Three smaller details matter in daily use: - **A sturdier hinge.** Apple says a newly designed hinge provides greater stability while you use touch. - **Steadier video calls.** Video stabilization smooths the camera feed in real time when you use touch on a video call. - **A test bench for app makers.** Developers can test how touch performs in their iOS and iPadOS apps directly on the MacBook Pro display. One thing we cannot tell you yet is whether touch can be switched off. Apple's materials do not say, and the laptops are not in reviewers' hands. Apple also says the touch sensor is built directly into the display panel rather than layered on top of it, which the company says reduces latency (the delay between your finger moving and the screen responding). Whether that feels instant in practice is exactly the kind of claim reviews will need to check. ### Why Apple says it added touch now Apple's own executives gave the reasons in an interview with Wired, [reported by MacRumors](https://www.macrumors.com/2026/10/28/apple-executives-touch-screen-explanation/). Apple CEO John Ternus did not oversell it. He said: "The Mac is first and foremost an indirect input device, built around the trackpad and keyboard. But I think there's times when it's comfortable and handy to have touch." His main condition was cost to the product. Ternus said Apple needed to reach a point where adding touch "didn't make the product lesser in any way", meaning not heavier and not thicker, and that this came together with the OLED screen. Two other executives filled in the rest, per the same report. Marketing chief Greg Joswiak said pro users were waiting for touch and that a generation of young creators expect it. Hardware chief Johny Srouji said display technology, Apple silicon and software had evolved enough to meet Apple's bar. Apple frames touch as a handy extra that pro users and young creators were waiting for, added once it would not make the laptop heavier or thicker. ### Is MacBook Pro OLED? The new touch models are, and that is the other half of this launch. The MacBook Pro OLED panel is what Apple calls PROLED XDR, a tandem OLED display (two light-emitting layers stacked to push brightness higher). According to [Apple's specs](https://www.apple.com/macbook-pro/proled-xdr-display/), it reaches up to 2000 nits (a measure of screen brightness) peak brightness for both HDR and SDR content, with up to 1000 nits sustained full-screen brightness. Apple quotes a contrast ratio of 10,000,000:1, against 1,000,000:1 for the Liquid Retina XDR screen on the M6 model. OLED pixels can switch fully off, so dark scenes look darker and bright highlights stand out more. The panel also runs at up to 120Hz with ProMotion, with refresh rates starting from 10Hz, covers P3 and Adobe RGB color, and has an antireflective coating, with a nano-texture glass option. The screens are slightly bigger too. Apple says the displays grow to 14.2 and 16.3 inches thanks to slimmer borders. The notch is gone, replaced by the Dynamic Island, which comes to the Mac for the first time and becomes the home for Spotlight, Live Activities and Siri AI. And the bodies got lighter, which lines up with Ternus's "not heavier" condition. According to Apple: - The 14-inch model weighs 3 pounds, half a pound less than before, and is half a millimeter thinner. - The 16-inch model weighs 4 pounds, 0.7 pounds lighter, and is nearly a millimeter thinner. ### What the touch screen MacBook Pro costs The MacBook Pro with touch screen costs $400 more than the previous model at each chip tier. That is our arithmetic from the prices below, and it works out to roughly 9 to 16 percent depending on the configuration. [MacRumors lists](https://www.macrumors.com/2026/10/28/new-macbook-pro-maxed-out-cost/) the starting points: - **14-inch, M5 Pro:** $2,899 (15-core CPU, 16-core GPU, 24GB memory, 1TB storage), up from $2,499. - **16-inch, M5 Pro:** $3,399 (18-core CPU, 20-core GPU, 24GB, 1TB), up from $2,999. - **14-inch, M5 Max:** $4,499 (36GB, 2TB), up from $4,099. - **16-inch, M5 Max:** $4,799 (36GB, 2TB), up from $4,399. Fully loaded with M5 Max, 128GB of memory, 16TB of storage and nano-texture glass, the 14-inch reaches $14,249 and the 16-inch $14,549. Memory tops out at 64GB with M5 Pro and 128GB with M5 Max. Battery life, at least on paper, did not suffer. Apple claims up to 24 hours, and the per-model estimates [listed by MacRumors](https://www.macrumors.com/2026/10/28/macbook-pro-touch-battery-life/) match the existing mini-LED models' best figures. The 16-inch M5 Pro is rated for up to 24 hours of video streaming and 17 hours of web browsing; the 14-inch M5 Pro for 22 and 14 hours. The M5 Max versions are rated slightly lower: 22 and 16 hours on the 16-inch, 20 and 13 hours on the 14-inch. All of these are Apple's figures, and OLED battery claims are worth checking against real-world tests. ![Column chart of the MacBook touch screen price jump: 14-inch M5 Pro from $2,499 to $2,899 and 14-inch M5 Max from $4,099 to $4,499, each plus $400](/uploads/blog/macbook-touch-screen-new-macbook-pro-explained/img2-v2.webp) ### Who should buy a touch screen MacBook, and who should wait **Consider the touch model if** you edit photos or video, work in 3D, draw, mark up documents or make music, and you already planned to spend Pro or Max money. You get touch, the brighter OLED screen, the Dynamic Island and a lighter body for $400 more than last generation's equivalent. **Pick the M6 14-inch if** you want a new MacBook Pro and do not care about tapping the screen. At $1,999 it is $900 cheaper than the cheapest touch model ($2,899 minus $1,999, our arithmetic), ships the same day and carries the same up to 24 hours battery claim. **Wait if** you are unsure whether you would actually use touch. The laptops ship November 6, and independent reviews should answer the real questions: how the hinge holds up, whether touch latency feels as low as Apple claims, and whether the battery numbers survive normal use. Waiting for them costs you little. If you are curious how other hardware makers are rethinking screens, our [iPhone Duo vs Galaxy Z Fold 8 teardown](/radar/iphone-duo-vs-galaxy-z-fold-8-teardown) is a good companion read. ### What to watch next - **Reviews after November 6.** Independent testing of brightness, battery and touch response will confirm or correct Apple's claims. - **App updates.** Watch whether the named apps (Lightroom, DaVinci Resolve, Blender, Microsoft Office and others) ship touch-specific features, or rely on macOS turning taps into clicks. - **Developer uptake.** The updated gesture APIs only matter if developers use them. Apple's named apps lean toward creative and 3D tools, so start there. - **The lineup split.** Apple now sells three kinds of MacBook Pro at once: M6, mini-LED and the touch OLED models. Clear labelling at checkout will matter, because the names alone will not tell buyers which one has a touch screen. ## Frequently asked questions ### Does the MacBook Pro have a touch screen? Yes, but only the MacBook Pro with PROLED XDR Display that Apple announced on October 28, 2026. It comes in 14-inch and 16-inch sizes with M5 Pro or M5 Max chips. The cheaper 14-inch MacBook Pro with M6, announced a day earlier, does not have a touch screen, and neither do the existing mini-LED models. ### Is the touch screen MacBook Pro OLED? Yes. Apple calls it a PROLED XDR display, a tandem OLED panel that Apple says reaches up to 2000 nits peak brightness, 1000 nits sustained full-screen brightness and a 10,000,000:1 contrast ratio. It also runs at up to 120Hz with ProMotion and offers an optional nano-texture glass finish. These are Apple's figures, not independent measurements. ### How much does the touch screen MacBook Pro cost? According to MacRumors, the 14-inch model with M5 Pro starts at $2,899 and the 16-inch with M5 Pro at $3,399. The M5 Max versions start at $4,499 and $4,799. Each is $400 more than the previous model at the same chip tier. A fully loaded 16-inch configuration reaches $14,549. ### Do Mac apps work with touch without an update? Many do, according to Apple, because macOS automatically translates touches into clicks, so a tap acts like a click. Developers can use updated gesture APIs to build controls designed for fingers. Apple names Notion, Canva, Zoom, Microsoft Office apps, Adobe Lightroom, DaVinci Resolve and Blender among the apps that will take advantage of touch. ## Source Based on Apple's October 28 announcement and MacRumors reporting. https://www.apple.com/newsroom/2026/10/apple-unveils-groundbreaking-macbook-pro-with-proled-xdr-display/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/macbook-touch-screen-new-macbook-pro-explained. Last modified 2026-10-29T06:56:52.882Z. ================================================================================ - URL: https://syntaxradar.com/radar/what-is-gemini-4-carbon - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-29T06:17:34.000Z - Last modified: 2026-10-29T06:17:34.000Z - Primary source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ # What Is Gemini 4 Carbon? Google's Bug-Hunting AI, Explained **Short answer:** On October 28, 2026, Google DeepMind announced Gemini 4 Carbon, a frontier model built to find, exploit and patch software vulnerabilities. It goes first to a small group of cyber defenders through the Fairwind Program, with wider access promised but undated. **Takeaway:** Gemini 4 Carbon is Google's strongest model on paper, but it is a gated security tool for vetted defenders first, so most developers can read about it today and only plan for it later. **Key figures (as reported by the source):** - Artificial Analysis Intelligence Index v5.0: 59 (ranked #1) - SEC-bench Pro (finding real vulnerabilities): 88.2% - Introductory price per million tokens: $2 input / $10 output Google DeepMind's newest model is not a chatbot you can open in a browser tab tonight. Gemini 4 Carbon, announced on October 28, 2026, is a model aimed first at one job: finding holes in software before attackers do, and then fixing them. So what is Gemini 4 Carbon, and should you care if you are not a security engineer? Short answer: it tops a widely watched intelligence ranking, it posts state-of-the-art scores on finding and exploiting software flaws, and almost nobody outside a vetted circle of defenders can use it yet. The price is already public, though, and it doubles after an introductory period. Below: what the benchmarks actually measure, how Carbon differs from Gemini 4 Argon, what it will cost, and which pre-launch rumors to set aside. ### What is Gemini 4 Carbon? Carbon follows Gemini 4 Argon and was introduced in a [blog post by Koray Kavukcuoglu](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/), Google DeepMind's SVP and Chief AI Architect. It is a frontier model, and Google is pitching and releasing it as a security tool first. Think of it as a very fast, very patient code auditor. You point it at software, and it looks for weaknesses, checks whether those weaknesses can actually be abused, and writes a fix that addresses the cause rather than the symptom. That focus helps explain the cautious rollout. A model that is good at spotting exploitable bugs is useful to the people defending a bank's systems, and could be just as useful to the people trying to break into them. Google is handing it to defenders first. It also ranks well on general intelligence. According to Google, Carbon scores 59 on the Artificial Analysis Intelligence Index v5.0, which puts it at #1 on that leaderboard. ### Gemini 4 Carbon benchmark scores, in plain English Benchmarks are standardized exams for AI models. The names are opaque, so here is what each Gemini 4 Carbon benchmark is actually testing, using the figures from Google's announcement.
BenchmarkWhat it testsCarbon score
SEC-bench ProFinding real vulnerabilities in real code88.2% (state of the art, per Google)
ExploitGymProving a flaw can actually be exploited54.8% (new state of the art)
CWE-bench (Collinear)Patching a flaw at its root cause71.0%
WebDev ArenaBuilding web apps, judged by human votesRank #4, 1768 Elo (up 89)
The pattern matters more than any single number. Carbon scores close to nine in ten at finding bugs, a little over half at exploiting them, and 71% at fixing them at the root. These are different tests, so the scores are not directly comparable, but the gap suggests exploiting is the hardest of the three. Google also reports figures on behavior, which may matter more than raw skill for a tool like this. On Terminal-Bench 4.0, successful cheating fell to 0.0%, with cheating attempts at 0.4%. High-severity misalignment alerts in live coding environments dropped by almost 80%. On the Gray Swan prompt-injection test, where hidden instructions try to hijack the model, the attack success rate was 0.0% for Carbon against 0.7% for Argon. One caveat: these are company-reported numbers. Independent testers have not yet had broad access to check them. ### Gemini 4 Carbon vs Gemini 4 Argon: what changed? Argon came first. Google announced it on September 30, 2026, also [initially for selected cyber defenders](https://techenclave.com/news/items/google-introduces-gemini-4-argon-initially-for-selected-cyber-defenders). Carbon arrived four weeks later with a sharper focus on writing safe code and staying honest. Here is what Google says improved, measured against Argon on its internal benchmarks: - **Secure code generation:** success rate more than 2x Argon's. - **Unsupported claims:** cut by roughly half, meaning fewer confident statements with nothing behind them. - **Overreach:** down 26%. That is the model doing more than it was asked. - **Under-service:** down 29%. That is the opposite failure, stopping short of what was asked. On security tests the comparison is murkier. According to the [Fairwind Program page](https://deepmind.google/fairwind-program/), Argon scored 68.0% on CWE-bench v1, tied for first with Grok 4.7 and GPT-6 Astra. Carbon's 71.0% is on CWE-bench as run by Collinear. The labels differ, so treat the gap as a hint, not a verdict. Argon was already strong on its own terms: 85.8% on real-world vulnerability discovery versus 71.0% for Gemini 3.8 Flash Cyber, and 70.9% versus 58.2% on Wiz's penetration-testing benchmark. Carbon's pitch is less "finds more bugs" and more "writes safer code and behaves more predictably while doing it." ### Gemini 4 Carbon release date: who can use it today? The announcement went out on October 28, 2026. But the Gemini 4 Carbon release date depends on who you are. Access starts the week of the announcement for a subset of cyber defenders. More trusted partners join gradually through the Fairwind Program. Broader availability to developers, businesses and consumers is described only as "coming soon," with no date attached. #### What is the Fairwind Program? Fairwind is Google DeepMind's vetting scheme for its most capable security models. It works with over 650 partners worldwide, grouped into three kinds: governments and national cyber authorities, critical infrastructure operators, and core technology platforms. The rules are tight. Fairwind partners may give Argon access only to their internal cybersecurity, incident response or penetration testing teams. A company in the program cannot simply hand the model to its whole engineering department. If your organization is not eligible, Google points to CodeMender, a code security agent that runs on its public models. It is the consolation prize, but it exists today. ![Diagram showing who gets Gemini 4 Carbon first: Fairwind Program partners in governments, critical infrastructure and core tech platforms, with cyber defenders first and everyone else coming soon](/uploads/blog/what-is-gemini-4-carbon/img1-v2.webp) ### Gemini 4 pricing vs Opus 5.5: what we actually know Google has published Carbon's price even though most people cannot buy it yet, which helps anyone planning a budget.
Gemini 4 CarbonInput (per 1M tokens)Output (per 1M tokens)
Introductory price$2$10
Standard price afterwards$4$20
Cached input, meaning text the model has already seen in a recent request, is 95% off. That discount is the one to design around: if your tool sends the same large codebase over and over, caching it cuts that part of the bill sharply. Note the 2x step. Every line doubles when the introductory period ends, and Google has not said when that is. If you build a cost model on $2 and $10, build a second one on $4 and $20. For how vendors have been moving prices lately, see our breakdown of [why LLM API pricing fell up to 80%](/radar/llm-api-pricing-cuts-2026). Now the Opus 5.5 question, which many readers are searching for. The comparison started before launch. On October 9, 2026, Business Insider reported that Carbon was a Gemini 4 test build used on Jetski, Google's internal coding platform, and that one unnamed employee compared it to Claude Opus 5.5 at coding, as [Let's Data Science summarized](https://letsdatascience.com/news/gemini-4-carbon-checkpoint-reportedly-appears-on-chatbot-are-ee165ebf). That claim was secondhand, from a single source, and the official announcement has now replaced it. Google's post does not benchmark Carbon against Opus 5.5 and our sources do not include a matching price sheet, so any head-to-head pricing table you see right now is guesswork. Wait for independent tests on the same tasks. ![Column chart of Gemini 4 Carbon pricing per 1M tokens: introductory $2 input and $10 output rising to $4 input and $20 output, with cached input 95% off](/uploads/blog/what-is-gemini-4-carbon/img2-v2.webp) ### What it means for developers and everyone else If you run security at a government agency, a utility or a major platform, Carbon may already be on your desk, or about to be. The behavior numbers (no successful cheating on Terminal-Bench, no successful prompt injections on Gray Swan) are the ones to verify in your own environment first. If you are an everyday developer, nothing changes this week. You cannot call Carbon from your code yet. What you can do: - Plan around the published pricing, including the doubling. - Try CodeMender on public models if code security is your concern now. - Treat "#1 on the Intelligence Index" as a reason to test Carbon when it opens up, not a reason to rip out what you use today. If you are not technical at all, the story is still worth knowing. Google has now released two models in a row, Argon and Carbon, to defenders first. Here the gate is part of the product. (For a very different kind of DeepMind model, see our explainer on [Genie 4's playable AI worlds](/radar/google-genie-4-world-model-explained).) One gap to note: Google's post does not state Carbon's context window, the amount of text it can read at once. For code auditing, that number matters, and it is missing for now. ### What to watch next - **A date for general access.** "Coming soon" is the only promise so far. - **When the introductory price ends.** That is when $2 and $10 become $4 and $20. - **Independent benchmarks.** Every figure above is company-reported. Outside tests on SEC-bench Pro and ExploitGym will show whether the lead holds. - **The context window.** Google has not published it yet. - **A real Opus 5.5 comparison.** The leak-era claim is retired. A fair test on identical coding tasks would settle it. ## Frequently asked questions ### What is Gemini 4 Carbon? Gemini 4 Carbon is a frontier AI model from Google DeepMind, announced on October 28, 2026. Google is releasing it first as a security tool that finds, exploits and patches software vulnerabilities. It scores 59 on the Artificial Analysis Intelligence Index v5.0, ranked first, and goes to vetted cyber defenders before anyone else. ### When can developers use Gemini 4 Carbon? Not yet for most people. Access began the week of the October 28, 2026 announcement for a subset of cyber defenders, with more trusted partners added gradually through the Fairwind Program. Google says broader availability for developers, businesses and consumers is coming soon, but it has not given a date. ### How much does Gemini 4 Carbon cost? Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input 95% off. After the introductory period, prices double to $4 per million input tokens and $20 per million output tokens. Google has not said when the introductory period ends, so plan budgets for both levels. ### Is Gemini 4 Carbon better than Claude Opus 5.5? There is no official answer. Before launch, Business Insider reported that one unnamed Google employee compared Carbon to Claude Opus 5.5 at coding, but that was a single secondhand claim. Google's announcement does not benchmark Carbon against Opus 5.5, so a fair comparison needs independent tests on identical tasks. ## Source Based on Google DeepMind's Carbon announcement and Fairwind Program page. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/what-is-gemini-4-carbon. Last modified 2026-10-29T06:17:34.000Z. ================================================================================ - URL: https://syntaxradar.com/radar/gemini-4-carbon-explained - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-29T06:17:22.641Z - Last modified: 2026-10-29T06:17:22.641Z - Primary source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ # Gemini 4 Carbon: Why Google's New AI Goes to Defenders First **Short answer:** Google announced Carbon on Oct 28, 2026, a security-focused Gemini 4 model that scores 88.2% on SEC-bench Pro and 0.0% prompt-injection success. It goes to a subset of cyber defenders first; wider access is 'coming soon'. **Takeaway:** Carbon is Google's strongest model on security and safe-agent tests, but it goes to vetted defenders first, and its prices double after an introductory period of unannounced length. **Key figures (as reported by the source):** - SEC-bench Pro (finding vulnerabilities): 88.2% - Gray Swan prompt-injection success: 0.0% (Argon 0.7%) - Intro price per 1M input / output tokens: $2 / $10 Google has released a second Gemini 4 model, and this one was built for security work first. Gemini 4 Carbon, announced on October 28, 2026, is tuned to find software vulnerabilities, write safer code and resist the tricks attackers use against AI agents. It arrived about two weeks after Gemini 4 Argon reached developers. The catch: most people cannot use it yet. Google is handing Carbon to a small group of cyber defenders first and widening access slowly. Several API-reseller blogs ranking for the name describe it as a leak or an internal build. It is neither. Google announced it [on its official blog](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/), with benchmarks and prices. By the end of this piece you will know what Carbon is, how it differs from Argon, what the security scores actually measure, when you might get access and what it will cost. ### What is Gemini 4 Carbon? Carbon is a Gemini 4 model aimed at defensive security and safe coding. Koray Kavukcuoglu, SVP of Google DeepMind and Google's Chief AI Architect, introduced it. Think of Argon as the general model for developers and Carbon as a sibling trained to be a better security engineer and a more careful agent. It is not a narrow tool, though. According to Google's announcement, Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59, a broad measure of general capability. It also gained 89 Elo points on WebDev Arena, reaching fourth place with a rating of 1768. So it is a strong all-round model that happens to be pointed at security first. A model that is very good at finding and exploiting bugs is useful to defenders and attackers alike, so the staged rollout is worth understanding. Google is choosing who gets it first. ### The Gemini 4 Carbon benchmark scores, in plain English Google published a [long list of numbers](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/). They fall into three groups: finding bugs, fixing them, and behaving safely while working. #### Finding and exploiting vulnerabilities - **SEC-bench Pro: 88.2%.** This test checks whether a model can find real vulnerabilities in code. Google calls the score state-of-the-art. - **ExploitGym: 54.8%.** This measures whether the model can turn a bug into a working exploit, which is how security teams prove a flaw is real and urgent. Also a new state-of-the-art, per Google. #### Fixing what it finds - **CWE-bench: 71.0%.** CWE is the standard catalogue of software weakness types, and this benchmark tests patching them. - **Secure code generation:** on Google's internal benchmarks, Carbon's success rate is more than 2x higher than Argon's. - **Chrome's verdict:** the Chrome team reported more than 10x better patch quality than another lab's large frontier model. Google does not name that lab. #### Behaving safely This group matters most if you plan to let an AI agent work on your code or inbox. - **Gray Swan prompt injection: 0.0% attack success rate,** versus 0.7% for Argon. Prompt injection is when hidden text in a web page, document or email tells the AI to do something its user never asked for. - **Terminal-Bench 4.0 cheating:** successful cheating fell to 0.0% and cheating attempts to 0.4%. "Cheating" here means gaming a task, for example making tests pass without fixing the real problem. - **Misalignment alerts:** high-severity alerts in live coding dropped by almost 80%. - **Honesty and judgement:** unsupported claims were cut about 2x, overreach (doing more than asked) fell 26%, and under-service (doing less than asked) fell 29%. One caveat: these are Google's own reported results, and several, like the secure-code and patch-quality claims, come from internal tests. Independent testers will need time with the model before anyone can confirm them. ![Bar chart of Gemini 4 Carbon benchmark scores reported by Google: 88.2% on SEC bench Pro, 71.0% on CWE bench and 54.8% on ExploitGym](/uploads/blog/gemini-4-carbon-explained/img1-v2.webp) ### Gemini 4 Carbon vs Gemini 4 Argon: what actually changed? Both belong to the same Gemini 4 generation. The difference is focus and availability.
ArgonCarbon
FocusGeneral-purpose model for developersSecurity work and safer agent behaviour
Who can use itDevelopers via the Gemini APIA subset of cyber defenders, then trusted partners
Prompt injection (Gray Swan)0.7% attack success0.0% attack success
Secure code generationBaselineMore than 2x higher success rate
Argon is the one you can build with today. Its [Gemini API model page](https://ai.google.dev/gemini-api/docs/models/gemini-4-argon) lists the model ID `gemini-4-argon`, a context window of 1,048,576 tokens and 65,536 default max output tokens, rising to 1,048,576 with long decode. In plain terms, Argon can read an enormous codebase or document set in one go. Google has not published equivalent context-window specs for Carbon in the material covered here, so treat any reseller table that lists them as unconfirmed. If you are choosing today, the answer is easy: Gemini 4 Argon is what you can actually call. Carbon matters when your work involves security reviews, patching, or agents that read untrusted content. ![Two panel chart comparing Gemini 4 Carbon with Gemini 4 Argon: prompt injection success of 0.0% for Carbon versus 0.7% for Argon, and secure code generation success more than 2X higher for Carbon](/uploads/blog/gemini-4-carbon-explained/img2-v2.webp) ### When is Gemini 4 Carbon released, and who can use it? The Carbon release date depends on who you are. Google's rollout runs in three stages: 1. **Now:** from the week of the October 28 announcement, a subset of cyber defenders. 2. **Next:** trusted partners, added gradually through what Google calls the Fairwind Program. 3. **Later:** broader availability for developers, businesses and consumers, which Google describes only as "coming soon." There is no date. So if you searched for a Gemini 4 release date expecting a public switch-on, the honest answer for Carbon is: not yet, and Google has not said when. Argon, by contrast, has been in developers' hands since mid-October. This staged approach is the story's real tension. The same skills that make Carbon good at spotting a flaw in a browser could help someone attack that browser. Google's stated approach is to give defenders access first and widen it gradually. ### How much does Gemini 4 Carbon cost? Google has already set prices, which is a sign the wider launch is planned rather than hypothetical. Per the [pricing in Google's Carbon announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/):
PeriodInput (per 1M tokens)Output (per 1M tokens)
Introductory$2$10
After introductory period$4$20
Cached input, meaning text the model has already seen and you send again, gets a 95% discount. That matters for security work, where you often feed the same large codebase to the model over and over. The key detail: prices double once the introductory window ends, and Google has not said how long that window lasts. If you are budgeting a project, plan around the higher rate. For a wider view of where model prices are heading, see our breakdown of [why LLM API prices fell up to 80%](/radar/llm-api-pricing-cuts-2026). ### Who wins, who loses, and what it means for you **Security teams** are the clear early winners. If the reported numbers hold up in real codebases, a model that finds bugs at 88.2% on SEC-bench Pro and patches them well could shorten the gap between a flaw existing and a fix shipping. **Developers building AI agents** should watch the safety numbers more than the hacking ones. A 0.0% prompt-injection success rate on Gray Swan, if it survives outside testing, removes one of the biggest worries about letting an agent read email or browse the web. **Everyone else** may benefit indirectly. Google says its Chrome team already tested Carbon on patch quality, so browser fixes are one place the gains could show up. The losers, for now, are developers who want Carbon today and cannot get it, plus anyone relying on reseller blogs that promise access Google itself is not yet offering. The interest is real: the announcement reached 766 points on the Hacker News front page the day after launch. ### What to watch next - **A date for broad access.** "Coming soon" is the most important missing detail. - **The end of introductory pricing.** Rates double after it, so its length shapes real costs. - **Independent benchmarks.** Outside testers rerunning SEC-bench Pro and the prompt-injection tests will show whether Google's numbers travel. - **Fairwind Program partners.** Who joins will hint at where Carbon is used first. Google DeepMind is shipping fast across its lineup, from Gemini to world models like [Genie 4](/radar/google-genie-4-world-model-explained). Carbon shows that safety results, not just raw capability, are now part of how Google pitches a model. ## Frequently asked questions ### Is Google's Carbon model a leak or an internal build? No. Google officially announced Carbon on its blog on October 28, 2026, with benchmark results and pricing. Some API-reseller sites describe it as a leaked or internal model, but it is a real Gemini 4 release. Access is simply limited for now, starting with a subset of cyber defenders before it widens to trusted partners. ### Can I use Carbon through the Gemini API today? Probably not. Google is giving Carbon first to a subset of cyber defenders, then gradually to trusted partners through its Fairwind Program. Broader availability for developers, businesses and consumers is described as coming soon, with no date. If you need a Gemini 4 model now, Gemini 4 Argon is available to developers through the Gemini API. ### How is Carbon different from Gemini 4 Argon? Argon is the general-purpose Gemini 4 model for developers, with a context window of about one million tokens. Carbon is tuned for security and safer agent behaviour. Google reports a 0.0% prompt-injection attack success rate for Carbon versus 0.7% for Argon, and more than twice Argon's success rate on internal secure code generation tests. ### How much will Carbon cost after the introductory period? Google's announcement lists introductory pricing of $2 per million input tokens and $10 per million output tokens, with cached input 95% off. After the introductory period, prices rise to $4 per million input tokens and $20 per million output tokens. Google has not said how long the introductory period will last, so budget around the higher rate. ## Source Based on Google's Gemini 4 Carbon announcement. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/gemini-4-carbon-explained. Last modified 2026-10-29T06:17:22.641Z. ================================================================================ - URL: https://syntaxradar.com/radar/tmux-ghostty-guide-should-you-switch - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-28T17:34:54.807Z - Last modified: 2026-10-28T17:34:54.807Z - Primary source: https://ghostty.org/docs/install/release-notes/1-4-0 # Tmux Ghostty Guide: What 1.4 Changes and Who Should Switch **Short answer:** Ghostty 1.4.0 shipped on October 27, 2026 after seven months of work by 172 contributors. It parses plain text up to 5.7x faster, cuts idle scrollback memory by 70% to 90%, adds full Kitty graphics support and a new Linux renderer, and fixes 10 security advisories. **Takeaway:** Ghostty 1.4 makes a fast, native terminal lighter on memory, so heavy terminal users have a strong reason to try it, while tmux stays useful mainly for persistent remote sessions. **Key figures (as reported by the source):** - Plain-text parsing on M4 Max: 128 MB/s to 725 MB/s - Idle scrollback memory saved: 70% to 90% - Empty terminal memory (macOS): 128 KiB to 16 KiB Ghostty 1.4 shipped on October 27, 2026, and by the next day it was sitting on the Hacker News front page with 1,239 points. That is a lot of attention for a terminal app. The headline numbers explain some of it: on an M4 Max Mac, plain text now parses at 725 MB/s instead of 128 MB/s, and a fresh, empty terminal uses about an eighth of the memory it used to. If you have searched "tmux ghostty" because you are wondering whether the two belong together, or whether Ghostty should replace the Terminal app you already have, this guide is for you. The release covers seven months of work from 172 contributors across 2,889 commits, and most of it is invisible until you know where to look. By the end you will know what a terminal emulator actually does, what the 1.4 numbers mean in daily use, how Ghostty compares with macOS Terminal and tmux, and whether switching is worth your afternoon. ### What is Ghostty? A terminal emulator is the window where you type commands. Decades ago, a terminal was a physical screen and keyboard wired to a big shared computer. Today it is an app that imitates one: it shows text, passes your keystrokes to a shell (the program that actually runs your commands, such as zsh or bash), and decodes a stream of hidden control codes that say where the cursor goes, which colour to use and when to clear the screen. Decoding those codes quickly and correctly is most of the job. Ghostty is one of these apps. According to its [about page](https://ghostty.org/docs/about), Mitchell Hashimoto started it as a passion project in his free time, with a simple goal: a terminal that is native, feature-rich and fast at once. Many terminals give up one of those three to get the other two. "Native" is meant literally. On macOS, Ghostty is written in Swift with AppKit and SwiftUI, so its windows, tabs and menus behave like any other Mac app. On Linux it is written in Zig on top of the GTK4 toolkit, so it fits in with other Linux desktop apps. Underneath both apps sits libghostty, a shared core that handles terminal emulation, fonts and rendering. That core may turn out to be the most important part of the project, and we come back to it at the end. ### What changed in Ghostty 1.4, in plain numbers The [1.4.0 release notes](https://ghostty.org/docs/install/release-notes/1-4-0) are long. Here is what the main figures mean for someone who just wants their terminal to feel quick. #### Speed you feel when output floods in On an M4 Max Mac, plain ASCII text parsing went from 128 MB/s to 725 MB/s, about 5.7 times faster. Throughput through the pty, the pipe that connects the terminal window to your shell, rose 25% to 55% over the 1.3.x releases. You notice this when a build spits out thousands of lines or you `cat` a huge log file. A slow terminal stutters and catches up in bursts, and can even hold up the program feeding it. A fast one just keeps scrolling. Resizing a window that holds 10,000 lines of history is also about six times faster. #### Memory you stop thinking about This is the change most people will benefit from without ever noticing. Ghostty now compresses idle scrollback (the history you scroll back through), which cuts physical memory use by 70% to 90% for typical text. Because history got cheaper to keep, the default scrollback limit per terminal went up from 10 MB to 50 MB. A brand-new empty terminal on macOS now uses about 16 KiB of page memory, down from 128 KiB in 1.3.1. If you keep a dozen tabs open all day, that adds up. How much memory software spends holding on to past context is a recurring theme across computing; AI models have their own version of the problem, which we covered in [KV cache explained](/radar/kv-cache-explained-cloudflare-prompt-leak). #### Images, Linux and security Ghostty now supports 100% of the published [Kitty graphics protocol](https://sw.kovidgoyal.net/kitty/graphics-protocol/), the standard that lets command-line tools draw real images inside a terminal, now including animation. The release notes say it is the first terminal other than Kitty itself to get there, and parsing a large 64 MiB image transfer is about 16 times faster in their x86_64 Linux benchmark. On Linux and FreeBSD, a new EGL-based renderer adds triple buffering and true fractional scaling, so a display set to 125% or 150% looks crisp instead of soft. The release also closes 10 published GitHub security advisories, covering paste protection, Kitty graphics, tmux control mode and macOS link handling. It adds macOS 27 support (macOS 13 Ventura is still the minimum) and four new interface languages: Basque, Belarusian, Danish and Serbian. ![Two panel chart of Ghostty 1.4 vs 1.3: text parsing speed rising from 128 MB/s to 725 MB/s, 5.7X faster, and empty terminal memory falling from 128 KiB to 16 KiB](/uploads/blog/tmux-ghostty-guide-should-you-switch/img1-v2.webp) ### Ghostty vs Terminal: is the built-in app good enough? Every Mac ships with Terminal, and every Linux desktop comes with a terminal of its own. For running a script now and then or checking a git status, they are perfectly fine. The ghostty vs terminal question comes down to how many hours a day you spend in that window. Where Ghostty pulls ahead, going by the 1.4 release notes: faster parsing when output floods in, lower memory use and full Kitty image support. Where the stock app wins: it is already installed, it updates with the operating system, and there is nothing new to learn. The usual reason people stay on Terminal is that third-party terminals feel slightly foreign on a Mac. Ghostty removes most of that objection, because it is built with the same Apple frameworks. If you rarely open a terminal, keep the one you have. If you live in it, Ghostty is worth a trial week. ### Warp terminal vs Ghostty: which fits you? Warp is a separate terminal app that we did not test for this article, and the 1.4 release notes do not compare the two, so there is no speed verdict to give here. What the sources do tell us is what Ghostty is aiming for: a terminal that is native, feature-rich and fast at once, built around the classic terminal model. So the warp terminal vs ghostty choice is less about a benchmark and more about what you want from the tool. Look at Warp's own documentation for what it offers, then compare it against the Ghostty numbers above and decide which trade-offs suit your day. Searches for cmux vs ghostty point to a third option: cmux is a separate project that readers mention alongside Ghostty. We have not tested it either, so check its own page before choosing. If you want a plain, general-purpose terminal, Ghostty is the one this release is about. ### Tmux Ghostty setups: do you still need tmux? [tmux](https://github.com/tmux/tmux) is not a terminal. It is a terminal multiplexer, a program that runs inside any terminal and lets one window hold many shell sessions, split into panes. Its killer feature is that those sessions keep running after you close the window or lose your SSH connection. Start a long job on a remote server inside tmux, shut your laptop, reattach tomorrow, and it is still there. Ghostty is a native app, so check whether its own window handling covers how you use tmux locally. If you rely on tmux for persistent sessions, especially on servers, keep it. Running tmux inside Ghostty works like running it in any other terminal. The one common snag is that a remote machine may not recognise Ghostty's terminal type, which can garble colours or keys inside tmux and other full-screen programs. The new `ghostty +ssh` command in 1.4 handles this: it wraps ssh, installs Ghostty's terminfo (the file that tells programs what a terminal can do) on the remote host, and forwards your environment variables.
If you...Best fit
Open a terminal a few times a weekKeep macOS Terminal or your Linux default
Spend hours a day in the terminal and want speed with a native feelGhostty
Need sessions that survive disconnects on remote serverstmux, running inside whichever terminal you pick
Want a different interface philosophyCompare Warp's own docs against Ghostty's numbers
![Diagram of tmux running inside Ghostty with three tmux sessions, connected by the ghostty +ssh command to a remote server, showing that tmux sessions survive a closed window](/uploads/blog/tmux-ghostty-guide-should-you-switch/img2-v2.webp) ### How to use Ghostty: install, set up, connect Ghostty install takes a couple of minutes. On a Mac, download it from the [official install page](https://ghostty.org/docs/install). You need macOS 13 Ventura or newer. For Ghostty on Linux, the install docs list packages for many distributions. One catch with 1.4: GTK builds now require GTK 4.18 and libadwaita 1.7, so older long-term-support distributions may not have a ready package yet. Once it is installed, here is how to use Ghostty without a learning curve: 1. **Open it and work as usual.** Ghostty hosts the shell you already have, so your prompt, aliases and history come with you. 2. **Check the docs for settings.** Ghostty's documentation lists every option, so start there before changing anything. 3. **Compare before you commit.** Use it for a week next to your current terminal and see whether the speed and memory gains show up in your own work. 4. **Keep tmux for sessions that must outlive the window.** 5. **Use `ghostty +ssh user@server` for remote work,** so programs on the server know what your terminal supports. ### What to watch next The bigger story is libghostty. The about page describes it as a shared, C-ABI-compatible core for terminal emulation, fonts and rendering. Today it is not a stable API and has not been released as a standalone library. Once it is, other apps could in principle embed the same terminal core instead of writing their own. Ghostty 1.5 is planned for March 2027, and the project is keeping its six-month release cycle. Three things to watch until then: whether libghostty gets a stable release, how quickly Linux distributions catch up with the new GTK requirements, and whether other apps start building on Ghostty's engine. For now, if you spend real time in a terminal, 1.4 is the best reason yet to give it a week. ## Frequently asked questions ### What is Ghostty? Ghostty is a terminal emulator, the app you type commands into, started by Mitchell Hashimoto as a passion project. It is a native app on each platform: Swift with AppKit and SwiftUI on macOS, Zig with GTK4 on Linux. Its stated goal is to be native, feature-rich and fast at the same time, rather than trading one away. ### Does Ghostty work with tmux? Yes. tmux runs inside Ghostty the same way it runs inside any terminal. You may only need tmux for sessions that must survive a closed window or dropped SSH connection. The new ghostty +ssh command in 1.4 installs Ghostty's terminfo on remote hosts, which avoids garbled colours or keys inside tmux on servers. ### How do I install Ghostty on Mac or Linux? On macOS, download Ghostty from the official install page; it needs macOS 13 Ventura or newer. On Linux, the install docs list packages for many distributions. GTK builds of version 1.4 require GTK 4.18 and libadwaita 1.7, so some older distributions may need to wait for a package or build from source. ### What is libghostty? libghostty is the shared core inside Ghostty that handles terminal emulation, fonts and rendering, with a C-ABI-compatible interface so other programs could embed it. According to Ghostty's about page, it is not yet a stable API and has not been released as a standalone library. ## Source Based on Ghostty's 1.4.0 release notes. https://ghostty.org/docs/install/release-notes/1-4-0 --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/tmux-ghostty-guide-should-you-switch. Last modified 2026-10-28T17:34:54.807Z. ================================================================================ - URL: https://syntaxradar.com/radar/macbook-pro-m6-vs-m5-upgrade - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-28T12:55:21.422Z - Last modified: 2026-10-28T12:55:21.422Z - Primary source: https://www.apple.com/newsroom/2026/10/apple-introduces-the-new-14-inch-macbook-pro-with-m6/ # MacBook Pro M6 vs M5: Who Should Upgrade, and Who Can Skip It **Short answer:** Apple announced the 14-inch MacBook Pro with M6 on October 27, 2026. It claims up to 25% faster on-device AI than M5 and a Neural Engine up to 1.9x faster. Pre-orders opened October 28; it ships November 6 from $1,999. **Takeaway:** The M6 MacBook Pro is a real step up for AI and rendering on Apple's numbers, but M5 owners can skip it while M1-era owners have a strong reason to upgrade. **Key figures (as reported by the source):** - US starting price: $1,999 - Neural Engine vs M5 (Apple claim): Up to 1.9x faster - Availability: November 6, 2026 Apple's new 14-inch MacBook Pro with M6 went up for pre-order on October 28, one day after the [official announcement](https://www.apple.com/newsroom/2026/10/apple-introduces-the-new-14-inch-macbook-pro-with-m6/), and it reaches buyers on November 6. The pitch is AI. Apple says the M6 runs on-device AI up to 25% faster than the M5, and that its Neural Engine is up to 1.9 times quicker. The catch is the sticker: the base model starts at $1,999, which is $400 more than the M5 model cost when it launched in October 2025. So the MacBook Pro M6 vs M5 question is not really "is it faster?" It is. The real question is whether the gains land on the work you actually do, and whether they justify the gap. By the end you will know what changed inside the chip, what those AI multipliers mean if you run models on your own machine, what the laptop costs, and whose current Mac should retire. One caveat runs through all of it: every performance figure below is Apple's own claim. Independent benchmarks are not out yet. ### MacBook Pro M6 vs M5: what actually changed? The headline change is the chip. Apple says the M6 has a 12-core CPU and a 12-core GPU, each with two more cores than the previous generation. That is the plainest way to read the Apple M6 chip vs M5 story: more of everything, plus a much bigger bet on AI hardware.
PartM6Change vs M5 (Apple's claim)
CPU12 coresTwo more cores
GPU12 coresTwo more cores, up to 30% faster graphics
Neural EngineDual 16-coreUp to 1.9x faster
On-device AI overallNot given as a single specUp to 25% faster
US starting price$1,999M5 model launched at $1,599
The rest of the Apple M6 chip specs and the laptop around it read like a polished version of a familiar machine: 16GB of unified memory as standard, configurable to 32GB, with up to 170GB/s of memory bandwidth. You get a Liquid Retina XDR display that peaks at 1600 nits in HDR, three Thunderbolt 4 ports, HDMI, MagSafe 3, an SDXC slot, a headphone jack, Wi-Fi 7 and Bluetooth 6. What Apple did not say matters too. Its announcement does not name the M6's process node or architecture, so treat any claim you see elsewhere about how the chip is built as unconfirmed until Apple or a teardown says so. ### Apple M6 chip performance: what 1.9x and 1.3x mean for local AI If you run language models on your own laptop, two of Apple's numbers jump out. The Neural Engine is "up to 1.9x faster than M5". And in Apple's app tests, the M6 is up to 1.3x faster at LLM "time to first token". Time to first token is the pause between pressing Enter and seeing the model start to answer. During that pause the chip reads and processes your whole prompt. That is why Apple pairs its 25% on-device AI claim with "LLM prompt processing" as the example. If you paste long documents or large chunks of code into a local model, this is the wait you feel most, and a 1.3x improvement would cut it by roughly a quarter. Notice the gap between 1.9x and 1.3x. The Neural Engine is only one of the places AI work can run, and real tasks also depend on the GPU, memory and software. A faster block of silicon only helps the apps written to use it. Apple does not explain the gap, but that is one plausible reason its app-level figure is smaller than its Neural Engine figure. Memory is the other half of the story. On a laptop like this, the model has to fit in the unified memory it shares with macOS and your other apps. The M6 starts at 16GB and tops out at 32GB, so this is a machine for small and mid-sized local models, not the largest open models. Once the first token appears, how fast the rest streams out depends heavily on memory bandwidth, which Apple puts at up to 170GB/s. If private, offline AI is why you are shopping, our explainer on [why devices are moving AI off the cloud](/radar/on-device-phone-ai-local-models-apple-android-privacy) covers the trade-offs. ![Bar chart of MacBook Pro M6 vs M5 AI speed: M5 as the 1X baseline, up to 1.3X faster time to first token and up to 1.9X for the Neural Engine, marked as Apple's own claims and not yet tested](/uploads/blog/macbook-pro-m6-vs-m5-upgrade/img1-v2.webp) ### Apple M6 chip benchmark claims for coding, creative work and battery Apple's comparisons against the M5 go beyond AI. According to the company, the M6 is up to 1.3x faster at Xcode builds and Blender ray tracing, and up to 1.2x faster at image processing in Affinity. Graphics overall are up to 30% faster. For a developer, a 1.3x faster build is real but modest. It shaves time off every compile without changing how you work. For 3D artists the Blender figure matters more, because ray-traced renders can run for long stretches where a 1.3x speedup adds up across a day. Read every "up to" carefully. It describes the best result Apple measured, usually on a task picked to show the chip at its strongest. Until independent Apple M6 chip benchmark results arrive, the honest summary is: probably a solid step, size to be confirmed. Battery may be the upgrade more people actually notice. Apple rates the M6 model at up to 24 hours, which it calls the longest battery life ever in a Mac. A compatible 96W USB-C adapter charges it to 50% in 30 minutes. The 12MP Center Stage camera and six-speaker system round out the everyday package. ### How much does the MacBook Pro M6 cost? The MacBook Pro M6 price in the US starts at $1,999, or $1,899 with education savings, according to [Apple's announcement](https://www.apple.com/newsroom/2026/10/apple-introduces-the-new-14-inch-macbook-pro-with-m6/). If you would rather pay monthly, Apple's Upgrade lease runs $38.99 a month over 36 months or $53.99 a month over 24 months, before taxes and any trade-in credit. In India, Apple's [India newsroom](https://www.apple.com/in/newsroom/2026/10/apple-introduces-the-new-14-inch-macbook-pro-with-m6/) lists the same model at INR 239,900, or INR 227,900 with education savings. Now the comparison that matters. When the base 14-inch M5 MacBook Pro launched on October 15, 2025, it cost $1,599, according to [PhoneArena's launch coverage](https://www.phonearena.com/apple-macbook-pro-m5-release-date-price-features-news). The 1TB version was $1,799, and the 24GB, 1TB configuration was $1,999. That last line is the useful one. At launch, $1,999 bought an M5 with 24GB of memory. Today it buys an M6 with 16GB as standard. Our sources do not list the base storage, so check the configurator before you compare like for like. If you plan to run local models, the memory you can afford may matter more than the chip generation. ![Bar chart of the MacBook Pro base price: $1,599 for the M5 in October 2025 and $1,999 for the M6 now, a $400 increase, with 16GB at $1,999 on the M6 against 24GB at $1,999 on the M5](/uploads/blog/macbook-pro-m6-vs-m5-upgrade/img2-v2.webp) ### Is the MacBook Pro M6 worth upgrading from M5? For most M5 owners, no. The M5 is about a year old, and Apple's own claims put the M6 at 20% to 30% ahead on most tasks, with the larger Neural Engine jump only helping apps built to use it. When reviewers looked at the M5, they called it an incremental upgrade over the M4. On Apple's numbers, the M6 looks like a bigger step than that, but not one that makes a recent machine feel slow. The math changes the further back you go. Apple compares the M6 directly with the 13-inch M1 MacBook Pro and claims: - Up to 13x faster AI - Up to 11x faster graphics - Up to 11.7x faster Blender ray tracing - Up to 3.3x faster Affinity image processing - Up to 2.5x faster Xcode builds Here is how we would sort buyers: - **M5 owners:** skip it, unless long prompts to local models or heavy rendering are your daily bottleneck and the time saved pays for itself. - **M4 owners:** a gray zone. Apple's comparisons in our sources cover the M5 and the M1, not the M4, so wait for independent tests. - **M1 owners and older:** this is the generation where upgrading makes clear sense, especially if you want to try local AI. - **First-time buyers:** decide on memory first. A discounted M5 with 24GB may suit local AI work better than a base M6 with 16GB, if you can find one. ### MacBook Pro M6 release date, and the iMac that launched with it The MacBook Pro M6 release date is November 6, 2026. Apple announced it on October 27, and pre-orders opened on October 28. It did not arrive alone. Apple also unveiled an [iMac with M6](https://www.apple.com/imac/) the same day, with a 24-inch 4.5K display, up to 32GB of memory and Wi-Fi 7. iMac pre-orders opened at 10:10 a.m. PT on October 28, with availability also on November 6. Apple's iMac page shows no US price yet. On materials, Apple says the MacBook Pro enclosure is 100% recycled aluminum and that the laptop is 45% recycled content overall. ### What to watch next The first independent reviews should land around the November 6 release. Three things are worth checking in them. First, whether the 1.3x time-to-first-token claim holds in the local-model apps people actually use, and whether those apps start using the faster Neural Engine. Second, how the M6 behaves under long, sustained loads such as big builds and renders, where a laptop's cooling sets the real ceiling. Third, what Apple or a teardown reveals about how the chip is built, since the announcement left that out. And watch M5 prices. If retailers cut the older model, the $400 gap could widen further, and for many buyers the best MacBook Pro M6 vs M5 decision may turn out to be the cheaper laptop. For more buyer-first hardware calls, see our [iPhone Duo vs Galaxy Z Fold 8 teardown verdict](/radar/iphone-duo-vs-galaxy-z-fold-8-teardown). ## Frequently asked questions ### What is the difference between the MacBook Pro M6 vs M5? Apple says the M6 has a 12-core CPU and 12-core GPU, two more cores each than the M5, plus a dual 16-core Neural Engine up to 1.9x faster. It claims up to 25% faster on-device AI and up to 30% faster graphics. These are vendor figures, and independent benchmarks are not yet available. ### When is the MacBook Pro M6 release date? Apple announced the 14-inch MacBook Pro with M6 on October 27, 2026. Pre-orders opened on October 28, and the laptop is available from November 6, 2026. An iMac with M6 was announced the same day and follows the same schedule, with pre-orders from 10:10 a.m. PT on October 28. ### How much does the MacBook Pro M6 cost? In the US it starts at $1,999, or $1,899 with education savings. Apple's Upgrade lease costs $38.99 a month over 36 months or $53.99 a month over 24 months, excluding taxes and trade-in credit. In India the same model is listed at INR 239,900. The base M5 model launched at $1,599. ### Is the MacBook Pro M6 worth upgrading from M5? For most M5 owners, no. Apple's own comparisons put the M6 about 1.2x to 1.3x ahead in apps like Xcode, Blender and local AI models. Owners of an M1 MacBook Pro have a far stronger case, with Apple claiming up to 13x faster AI and 2.5x faster Xcode builds. ## Source Based on Apple's October 27, 2026 announcement. https://www.apple.com/newsroom/2026/10/apple-introduces-the-new-14-inch-macbook-pro-with-m6/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/macbook-pro-m6-vs-m5-upgrade. Last modified 2026-10-28T12:55:21.422Z. ================================================================================ - URL: https://syntaxradar.com/radar/google-genie-4-world-model-explained - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-28T06:12:34.222Z - Last modified: 2026-10-28T06:12:34.222Z - Primary source: https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/ # Google Genie 4: DeepMind's Playable AI Worlds Now Have Sound **Short answer:** On October 27, 2026, Google DeepMind announced Genie 4, a real-time world model it says is the first to generate audio natively. It is a research preview for trusted testers, coming first to Google AI Ultra subscribers in Project Genie, with no firm date. **Takeaway:** Genie 4 turns a text prompt into a world you can walk through and hear, but it is a trusted-tester preview with no published specs, and its real prize is training robots, not entertaining people. **Key figures (as reported by the source):** - Genie 1 speed (March 2024): 1 frame per second - Genie 3 real-time output (Aug 2025): 720p at 24 fps - Project Genie exploration cap: 60 seconds Type "a rainy harbour town at dusk" and moments later you are standing in it. You walk with the arrow keys, the camera follows, and now you can hear the rain on the cobblestones and gulls somewhere off to the left. Nothing here was filmed and nothing was built by a game studio. A model is making up every frame and every sound as you move. That is the pitch for **Google Genie 4**, which Google DeepMind [announced on October 27, 2026](https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/). DeepMind calls it the first real-time world model that generates audio natively: ambient sound, music and even spoken dialogue, made alongside the picture instead of pasted on afterwards. You probably cannot try it yet, and some of the most interesting parts have nothing to do with entertainment. By the end of this piece you will know what a world model actually is, how it differs from ChatGPT-style chatbots and video generators, what changed from Genie 3, and when ordinary people might get access. ### What Google Genie 4 is, in plain English Google Genie 4 is a research preview from Google DeepMind, credited to Raia Hadsell, Diego Rivas and Hang Qi. It is open only to trusted testers for now, not the public. The basic idea is the same as earlier versions. You describe a place, and the model generates an explorable scene you can steer with standard keyboard controls. While you are inside it, you can type natural-language prompts to change things on the fly: switch the weather, move the camera, or drop in new people and objects. What Genie 4 adds, according to [DeepMind's announcement](https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/), falls into three buckets: - **Native sound.** Ambient noise, music and interactive dialogue generated in real time with the visuals. - **A Gemini guide.** Gemini 3.8 Flash watches the frames as they are generated, breaks your goal into steps, checks whether each step worked, and adjusts its plan. A separate model, Gemini Omni, keeps the simulation anchored to how things look and behave in the real world. - **Robot rehearsal.** A demo pairs Genie 4 with Google's robotics models: the task is broken into steps and Genie 4 shows them in a simulated scene. Every frame of video and every second of audio carries SynthID, Google's invisible watermark for AI-generated media. ### What is a world model in AI? A world model is an AI system that learns how an environment behaves, so it can predict what happens next when something acts inside it. Think of how you catch a ball. You do not solve physics equations. Your brain has a rough internal model of the world: thrown things arc, then fall. You run that model forward in your head and put your hand where the ball will be. A world model AI tries to learn the same kind of intuition from huge amounts of video. The key word is *interactive*. A world model does not just produce a clip. It takes an action from you (step forward, turn left, open the door) and produces the next moment of the world in response. Then it does it again, many times per second. That loop is why people describe Genie as a "playable" world rather than a video. It is also why AI labs care about world models far beyond games. If a model can reliably imagine what happens when a robot arm reaches for a cup, the robot can practise in that imagined world thousands of times without breaking a single real cup. ### From 1 frame per second to real time: Genie 1 to Genie 3 The jump to Genie 4 makes more sense once you see how fast each Google Genie AI release moved. These figures come from the [Wikipedia summary of the Genie models](https://en.wikipedia.org/wiki/Genie_(world_model)):
VersionReleasedWhat it could do
Genie 1March 20242D only, trained on video game footage, 1 frame per second
Genie 2December 20242D and 3D at 360p, worlds lasting 10 to 20 seconds
Genie 3August 2025Real time at 720p and 24 fps, with one minute of memory
Genie 4October 2026Adds native audio, a Gemini guiding agent and robot planning demos; no resolution or frame rate published
Genie 1 was closer to a flipbook than a game. Genie 2 added depth, but each world lasted only 10 to 20 seconds. **Google Genie 3** was the moment it started to feel real: smooth enough to walk around in, and able to remember what was behind you for about a minute. That memory point matters more than resolution. If you look away from a red car and look back, a weak model may show you a blue one, or no car at all. Holding a world steady over time is the hard part. Genie 3 also had critics. Fast Company's Thomas Smith called 24 fps too choppy for real-world use. And the Genie 3 world model has already been put to work outside Google DeepMind's lab: the Waymo World Model, from February 2026, is built on Genie 3 and, per the same summary, outputs lidar (the laser-based depth data self-driving cars rely on) four times faster than Genie 3. ![Staircase timeline of Google Genie versions: Genie 1 in March 2024 at 1 FPS, Genie 2 in December 2024 at 360P, Genie 3 in August 2025 at 720P and 24 FPS, and Genie 4 in October 2026 with native audio](/uploads/blog/google-genie-4-world-model-explained/img1-v2.webp) ### What Genie 4 adds: sound, a co-pilot and robot rehearsals #### Sound that belongs to the scene DeepMind says Genie 4 generates ambient sound, music and interactive dialogue natively, alongside the picture. Anyone who has played a game with the sound off knows how much flatter it feels. Sound is a big part of why a place feels like a place. #### A Gemini agent steering the generation The more technical change is the guiding agent. Instead of the world model simply reacting to your keystrokes, Gemini 3.8 Flash observes the generated frames, splits a goal into smaller steps, checks each one succeeded, and changes course if not. In practice, that should make it easier to ask for something with several parts ("walk to the market, then have a vendor greet me") and get a coherent result. #### Rehearsing robot tasks The robotics demo is the most telling part of the announcement. DeepMind shows Gemini 4 Argon working with Gemini Robotics to break a task into steps, while Genie 4 renders each step as real-time text captions in the simulated scene. The example is Apptronik's Apollo humanoid playing checkers. That is the long-term business case. Robots are slow and expensive to train in the physical world, a theme we covered in our look at [humanoid robots moving into homes](/radar/humanoid-robots-daily-life-figure-02-tesla-optimus-homes). A world model that can imagine the next few seconds of a task gives the robot a place to think before it moves. #### What still breaks DeepMind is upfront about the limits. Visual details and object permanence can drift over long sessions, after several prompt edits, or while you move around. Readable text, spatial accuracy, audio accuracy and long-term consistency are all listed as open research areas. The announcement gives no resolution, frame rate or session length, so anyone claiming Genie 4 runs at a specific spec is guessing. ![Three cards showing what Genie 4 adds: native sound, a Gemini guide that plans steps and checks each one, and robot rehearsal shown in a simulated scene](/uploads/blog/google-genie-4-world-model-explained/img2-v2.webp) ### What is a world model vs an LLM? This is the comparison most people trip over, so here it is directly. A **large language model** (LLM), like the one behind ChatGPT or Gemini, learns patterns in text. It predicts the next word. It can describe a kitchen beautifully, but it has no picture of the kitchen and no sense of what happens if you push a glass off the counter, beyond what people have written about it. A **world model** learns patterns in how scenes change over time. It predicts the next frame given an action. It does not need words at all to know that a pushed glass falls and that walls stop you. So the world model AI vs LLM split is roughly *words* versus *worlds*. And Genie 4 shows they work better together: Gemini (the language side) plans and checks the steps, while Genie (the world side) renders what those steps look and sound like. What about AI video generators? They are cousins, not twins. A video generator produces a fixed clip from a prompt, and you watch it. As we explained in our piece on [AI video generation and Hollywood](/radar/ai-video-generation-hollywood-revolution-sora-kling-creators), that is great for making a scene. A world model lets you act inside the scene and keeps generating in response, which is a harder problem, because it has to stay consistent with choices it could not see coming. ### Can you use Genie 4 today? Not unless you are one of DeepMind's trusted testers. The plan, per the announcement, is a staged rollout: 1. First to Google AI Ultra subscribers inside **Project Genie**. 2. Then to early-access partners through the Gemini API and Google AI Studio. There is no firm date for either step. Project Genie is the consumer app that already exists for the previous generation. It launched on January 29, 2026 for AI Ultra subscribers through Google Labs, with each exploration capped at 60 seconds, and gained Street View integration on May 19, 2026. That Street View link hints at where this could go: stepping into a generated version of a real street rather than an invented one. If you are not on Google's top consumer plan, the realistic route to Google Genie 4 is the API stage, and that will matter more for developers building games, training tools or robot software than for casual users. ### What to watch next Three things will tell you whether Google Genie 4 is a demo or a platform. **Published specs.** DeepMind has not said how sharp, how fast or how long Genie 4 worlds run. Genie 3's 720p and 24 fps were criticised as choppy. If Genie 4 ships without beating that, the audio and agent features carry the whole story. **Consistency over minutes, not seconds.** DeepMind itself says details drift over long sessions. The first hands-on reports from Project Genie users will show whether a world holds together long enough to be useful. **Robotics partners.** The Apptronik checkers demo is a single example. Watch for more robot makers, or a Waymo-style product built on Genie 4, as the signal that world models are becoming training infrastructure and not just a party trick you can hear. ## Frequently asked questions ### Can I use Google Genie 4 right now? Not yet. Genie 4 is a research preview limited to trusted testers. Google DeepMind says it will come first to Google AI Ultra subscribers inside Project Genie, then to early-access partners through the Gemini API and Google AI Studio. The company has not given a firm date for either stage, so treat any launch date you see online with caution. ### What is new in Genie 4 compared with Genie 3? DeepMind describes Genie 4 as the first real-time world model to generate audio natively, including ambient sound, music and dialogue. It also uses Gemini 3.8 Flash as a guiding agent that plans and checks steps, and a robotics demo shows task steps inside a simulated scene. Genie 3 ran at 720p and 24 fps with one minute of memory; Genie 4's specs are unpublished. ### What is a world model in AI? A world model is an AI system that learns how an environment behaves so it can predict what happens next when something acts inside it. You give it an action, such as walking forward, and it generates the next moment of the scene. That interactive loop makes it useful for games, simulation and training robots safely before they act in the physical world. ### How is a world model different from an LLM? A large language model predicts the next word in text, so it can describe a scene but has no picture of it. A world model predicts the next frame of a scene in response to an action, learning how things move and collide. In Genie 4 the two work together: Gemini plans and checks the steps, and Genie renders what they look and sound like. ## Source Based on Google DeepMind's Genie 4 announcement. https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/google-genie-4-world-model-explained. Last modified 2026-10-28T06:12:34.222Z. ================================================================================ - URL: https://syntaxradar.com/radar/kv-cache-explained-cloudflare-prompt-leak - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-27T12:19:50.722Z - Last modified: 2026-10-27T12:19:50.722Z - Primary source: https://blog.cloudflare.com/kv-cache-exposure-remediation/ # KV Cache Explained: How a Cloudflare Bug Leaked Users' Prompts **Short answer:** Cloudflare disclosed that 155 responses from its GLM-5.3 model on Workers AI contained fragments of other customers' prompts. A 681,280-token KV cache pool behind a 1,048,576-token context, plus an SGLang bug, let the model read stale memory. **Takeaway:** A KV cache pool smaller than the advertised context, plus an SGLang admission bug, let GLM-5.3 read other users' leftover memory, so anyone running inference should check pool size against context length after every config change. **Key figures (as reported by the source):** - Affected responses: 155 - Prefill KV pool before fix: 681,280 tokens - Advertised context window: 1,048,576 tokens For nearly two weeks this summer, a small number of people using Cloudflare's GLM-5.3 model got back answers stitched partly from someone else's prompt. Not a hack, not a stolen password: a memory bookkeeping mistake inside the part of an AI model called the KV cache. Explained simply, that cache is the model's short-term notepad, and for 155 responses it was reading pages a previous customer had left behind. Cloudflare published a [coordinated disclosure of the GLM-5.3 KV cache exposure](https://blog.cloudflare.com/kv-cache-exposure-remediation/) on October 26, and it is one of the clearest real-world lessons in how large language models actually run. The leak window ran from August 27 to September 9, 2026, and only hit prompts longer than 681,280 tokens. By the end of this piece you will know what a KV cache is, why it eats so much GPU memory, how two numbers that did not match turned it into a privacy problem, and what to do about it whether you send prompts to a shared API or run your own servers. ### What happened to GLM-5.3 on Workers AI? GLM-5.3 is a large model Cloudflare hosts on Workers AI, its pay-per-use AI platform, under the name `@cf/zai-org/glm-5.3`. According to the [launch changelog](https://developers.cloudflare.com/changelog/post/2026-08-28-glm-5.3-workers-ai/), it shipped with a context window of 1M tokens, priced at $1.40 per million input tokens, $0.26 per million cached input tokens and $4.40 per million output tokens. That 1M figure was the problem. The advertised context was 1,048,576 tokens, but the memory pool that held the prefill KV cache could only fit 681,280 tokens. Any prompt bigger than the pool should have been refused. Instead, a bug in SGLang, the open-source serving engine underneath, let some of those requests through in a broken state. The result, per Cloudflare's write-up: 155 responses contained fragments of other customers' prompts and partial generations. Gadi Evron of the security firm Knostic reported it through HackerOne on September 8 at 23:11 UTC. Cloudflare reproduced it with two test accounts within 25 minutes of declaring an incident the next day. ![Bar chart comparing the advertised 1,048,576 token context with the 681,280 token KV cache pool, a red hatched zone marked should be refused and 155 responses affected, explaining the KV cache bug](/uploads/blog/kv-cache-explained-cloudflare-prompt-leak/img1-v2.webp) ### What is KV cache in LLM? Start with how a chatbot writes. A large language model produces text one token (roughly a word fragment) at a time. To pick each new token, it runs a step called attention, which looks back over every earlier token in the conversation and decides which ones matter right now. For each earlier token, attention needs two small sets of numbers: a key (what this token is about) and a value (what it contributes). Recomputing those for the whole conversation on every new word would be absurdly wasteful. So the model stores them. That store is the KV cache, short for key-value cache. That is the KV cache meaning in one line: saved work, so the model does not redo attention math over tokens it has already seen. As the [Towards AI explainer](https://pub.towardsai.net/the-kv-cache-explained-why-long-conversations-get-expensive-4bd8f77dd7e7) puts it, the cache grows with context length times the number of people being served at once, and it can end up bigger than the model's own weights. That growth is why long context is expensive. A million-token window is a promise that the server has room to keep a million tokens' worth of keys and values in GPU memory for you, alongside everyone else being served by the same machines. ### How does KV cache work? Serving engines like SGLang do not hand each request one giant block of memory. They carve GPU memory into a shared pool of fixed-size pages and lend pages out as needed. When your request finishes, its pages go back to the pool and get reused by the next request. Nobody wipes them first, because overwriting them is the next request's job. Cloudflare's setup was also disaggregated, meaning the work is split across two kinds of machine: - **Prefill nodes** read your whole prompt at once and build its KV cache. - **Decode nodes** take that cache and generate the reply token by token, reading the cache at every step. Between the two sits an admission step that decides whether a request is valid and ready to hand over. This design is common because the two jobs stress hardware differently. It also means the KV cache travels between machines, and both sides have to agree on exactly which pages hold real data. If you want the performance side of this story, our explainer on [RadixAttention and prefix caching in SGLang](/radar/radixattention-dynamic-prefix-caching-sglang-vllm-benchmarks) covers how the same engine reuses cached prefixes to cut latency for multi-turn agents. ### KV cache explained through the bug: seven stale pages Here is the chain of events from Cloudflare's root-cause analysis, in plain terms. 1. A prompt arrives that is longer than 681,280 tokens, so it cannot fit in the prefill pool. 2. SGLang marks it as rejected and swaps in a one-token placeholder instead of failing loudly. 3. The disaggregated admission check only asks one question: is this request finished? A rejected request counts as finished, so it passes through as if it were normal. 4. The decode node reserves 8 KV pages, expecting a full handoff. Prefill wrote only one page, the placeholder. 5. The decode node then runs attention over all 8 pages. Seven of them still hold keys and values from earlier requests to the same model, belonging to other users. The model did what models do: it wrote fluent text based on what was in its notepad. Some of that notepad was other people's prompts. The output came back coherent, with an HTTP 200 success code rather than a 400 error, so nothing looked wrong on either side. That is also why it slipped past testing. Cloudflare says its pre-launch tests checked status codes, not what the model actually said. A test that expected a rejection would have seen a cheerful 200 and moved on. ![Diagram of eight KV cache pages where page 1 is a placeholder and pages 2 to 8 are stale pages from other users, with the note status 200 no error](/uploads/blog/kv-cache-explained-cloudflare-prompt-leak/img2-v2.webp) ### Can an AI model leak other users' data? Yes, and this case shows the route is often plumbing rather than the model itself. The weights of GLM-5.3 did not memorize anyone's secrets. The leak happened because shared memory was reused without a hard guarantee that a request could only read its own pages. Every shared AI API serves many customers on the same GPUs, because that is what makes per-token pricing cheap. Isolation between those customers is enforced by software: page tables, admission checks, size limits. When one of those checks has a gap, data from one request can bleed into another. It is worth keeping the scale in proportion. Cloudflare counts 155 affected responses, all from very long prompts, on one model, and says it saw no evidence of deliberate exploitation. The bug was not unique to Cloudflare either: the same SGLang flaw had been reported upstream as issue #30233 on July 6, 2026, and a fix (PR #36988) was merged on September 6, three days before Cloudflare's incident, in SGLang v0.5.20. ### Is Cloudflare Workers AI safe? On the evidence published, Cloudflare handled the response quickly once it knew. The timeline on September 9, all UTC:
TimeStep
13:22Incident declared
14:00GLM-5.3 set to private
14:35Root cause identified
14:52All prefill nodes on corrected config
19:40Model restored to public
After the fix, the prefill pool holds about 2.04 million tokens, roughly twice the context length, so a maximum-size prompt now fits. Cloudflare says no other Workers AI model was affected, met the SGLang maintainers at RadixArk on September 10, and coordinated the public write-up for October 26. So it is reasonably safe in the sense that the hole is closed and was disclosed openly. The honest caveat is that it ran for nearly two weeks before an outside researcher reported it. #### If you send prompts to a shared AI API Treat any multi-tenant API as a place where a bug could, rarely, expose what you send. Keep secrets, credentials and regulated personal data out of prompts unless you have a contract and a risk review that cover it. Very long prompts deserve extra thought, since this bug only fired above a size threshold. #### If you run your own inference Cloudflare's advice is specific. After any change to model weights, quantization, parallelism or memory settings, compare the KV pool size with the advertised context length on every prefill and decode node. Upgrade SGLang to v0.5.22 or later; v0.5.21 added a follow-up fix (PR #38935). And, given how this bug hid, test what the model returns, not only the status code: send an over-length prompt and confirm it is refused. Most KV cache optimization work, from KV cache quantization to compression to offloading pages into CPU memory, exists to squeeze more tokens into a fixed pool. This incident is the reminder that the pool size is also a security boundary, and each of those tricks changes how many tokens fit in it. ### What to watch next Watch for other providers running SGLang with disaggregated prefill and decode to say whether they checked their own pool sizes. Watch SGLang itself: v0.5.21 on October 2 and v0.5.22 on October 19 both followed the original fix, a sign the maintainers are still hardening this path. And watch whether hosted AI platforms start publishing isolation tests the way cloud providers publish security audits. The bigger takeaway is simple. A KV cache is just memory, and memory that gets shared needs the same care as any other shared resource. ## Frequently asked questions ### What is a KV cache in simple terms? It is the model's short-term notepad. For every token already in the conversation, the model stores a key and a value so it does not recompute attention over earlier text each time it writes a new token. The cache grows with context length and the number of users served at once, and it can exceed the size of the model's weights. ### How did the Cloudflare GLM-5.3 leak happen? Prompts longer than the 681,280-token prefill pool were marked rejected by SGLang and replaced with a one-token placeholder, but the admission check only asked whether a request was finished, so they passed through. Decode reserved 8 KV pages while prefill wrote one, so the model read seven stale pages holding other users' data. ### Was my data exposed on Workers AI? Only if you used GLM-5.3 between August 27 and September 9, 2026, with very long prompts. Cloudflare says 155 responses were affected, all to prompts longer than 681,280 tokens, that no other Workers AI model was affected, and that it saw no evidence of deliberate exploitation. The model was fixed and restored on September 9. ### What should teams running SGLang do now? Upgrade to SGLang v0.5.22 or later. After any change to weights, quantization, parallelism or memory settings, compare the KV pool size with the context length on every prefill and decode node. Also check what the model actually returns for an over-length prompt, because this bug produced an HTTP 200 with coherent text instead of an error. ## Source Based on Cloudflare's coordinated disclosure. https://blog.cloudflare.com/kv-cache-exposure-remediation/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/kv-cache-explained-cloudflare-prompt-leak. Last modified 2026-10-27T12:19:50.722Z. ================================================================================ - URL: https://syntaxradar.com/radar/anthropic-ipo-prospectus-five-numbers - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-27T07:32:20.591Z - Last modified: 2026-10-27T07:32:20.591Z - Primary source: https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm # Anthropic IPO Prospectus: The 5 Numbers That Actually Matter **Short answer:** Anthropic's S-1 went public on SEC EDGAR and hit the Hacker News front page on October 27, 2026. It shows 2025 revenue of about $4.6 billion, a roughly $42 billion net loss and about $518 billion in planned compute commitments. **Takeaway:** Anthropic grew revenue roughly twelvefold in 2025, but its S-1 shows a business still losing billions in cash terms and betting roughly $518 billion of commitments on that growth continuing. **Key figures (as reported by the source):** - 2025 revenue: ~$4.6B (about 12x prior year) - 2025 net loss: ~$42B (~$34B non-cash) - Planned compute commitments: ~$518B The Anthropic IPO prospectus is public. The S-1, the document a company must file before it sells shares to the public, is now [readable on SEC EDGAR](https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm), and on October 27 it climbed to the front page of Hacker News with 1,573 points. The headline numbers pull in opposite directions. Anthropic, the company behind Claude, grew revenue roughly twelvefold in 2025, to about $4.6 billion. Over the same year it reported a net loss of roughly $42 billion. Both figures are true, and the gap between them is the single most useful thing to understand about this filing. This guide reads the prospectus the way a friend with an accounting habit would: five numbers, what each one actually means, and why they matter if you use or build on Claude. It is an explainer, not stock advice. ### What does the Anthropic S-1 say? An S-1 is the registration statement a US company files with the Securities and Exchange Commission before it lists on a stock exchange. It has to lay out audited financials, how the business makes money and everything management thinks could go wrong. Anthropic first filed a confidential draft in June 2026, [according to The Next Web](https://thenextweb.com/news/anthropic-ipo-2-trillion-october-revenue-multiple). The version now on EDGAR is the one anyone can read. The short version: revenue is soaring, costs are soaring too, and the company has signed up for an enormous computing bill. [The Decoder's breakdown of the Anthropic IPO filing](https://the-decoder.com/anthropics-ipo-filing-shows-soaring-revenue-mounting-costs-and-existential-risks/) pulls out the figures that matter most: - **Revenue:** about $4.6 billion in 2025, roughly twelve times the year before - **Operating loss:** $8.06 billion in 2025, up from $2.98 billion - **Net loss:** roughly $42 billion, about $34 billion of it a non-cash accounting charge - **Compute commitments:** about $518 billion planned over the coming years - **Customer concentration:** two customers made up nearly a quarter of 2025 revenue The tone stands out as much as the numbers. Nearly a third of the prospectus is risk factors, and they go well past the usual warnings about competition and regulation. Anthropic flags existential risks to humanity and the possibility that AI models could manipulate, blackmail or act unpredictably. Few companies would put that in a document meant to sell shares, but a prospectus exists to disclose risk, and Anthropic has chosen to disclose a lot of it. ### How much revenue does Anthropic make? Far more than a year ago, and the pace has accelerated since. Revenue for 2025 came in around $4.6 billion, roughly twelve times the prior year, per The Decoder. The 2026 figures are on a different scale. Revenue for the second quarter of 2026 alone was $11.5 billion, more than double all of 2025 in a single quarter. When Anthropic raised its Series H, [NPR reported](https://www.npr.org/2026/06/01/nx-s1-5843199/anthropic-ipo-filing-ai-large) that run-rate revenue had passed $47 billion in May 2026, up from $9 billion at the end of 2025. Run-rate deserves a sentence of explanation, because you will see it everywhere in coverage of the Anthropic IPO. It takes recent revenue and projects it across a full year, assuming the latest pace holds. For a company growing this fast it is a fair way to show momentum, and a flattering one, because it assumes nothing slows down. The $11.5 billion quarter is the sturdier number. That is money actually booked in three months, not a projection. ![Bar chart of Anthropic revenue on a dark background: $4.6B for the 2025 full year, $11.5B for Q2 2026 alone and a $47B May 2026 run rate, with a callout that 2025 was roughly 12X the year before, from the Anthropic IPO prospectus](/uploads/blog/anthropic-ipo-prospectus-five-numbers/img1-v2.webp) ### Is Anthropic profitable? Not by standard accounting. By one narrower, company-chosen measure, it recently has been. Both answers sit in the same filing. Start with the operating loss, which measures the cost of actually running the business against what it earns. That loss widened from $2.98 billion to $8.06 billion in 2025. Revenue grew twelvefold, and the operating loss still nearly tripled. Growth this fast is expensive. The roughly $42 billion net loss looks far worse, but most of it is not cash leaving the building. About $34 billion is a non-cash accounting charge tied to financing that could convert into stock. Charges of this kind typically arise when a company has raised money through instruments that can turn into shares: as the company's value climbs, those instruments are worth more on paper, and the increase is booked as a loss. Nobody wrote a $34 billion check. Take that charge out and what remains is in the same territory as the $8.06 billion operating loss. Still large, but a very different story from $42 billion. Now the better news. According to The Decoder, Anthropic is on track for a second consecutive quarter of operating profit in 2026, but only on an adjusted basis. Adjusted means the company decides which costs to leave out, often items like stock-based pay or one-off charges. Read it as a direction of travel, not as proof the business model is settled. ![Diagram of the roughly $42B Anthropic 2025 net loss split into a ~$34B charge that is not cash and ~$8B the rest, with operating loss bars of $2.98B in 2024 and $8.06B in 2025](/uploads/blog/anthropic-ipo-prospectus-five-numbers/img2-v2.webp) ### The compute bill inside the Anthropic IPO prospectus Here is where the money goes. Compute and infrastructure cost $7.33 billion in 2025, about three times the year before and more than half of all operating costs. That is the price of the chips and data centers that train Claude models and answer every prompt you send. The forward commitment is the number that makes people stop scrolling. Anthropic plans roughly $518 billion in cloud, compute and infrastructure commitments over the coming years. An analysis by Let's Data Science estimates about 80 percent of that is non-cancelable, meaning Anthropic owes it whether or not demand keeps climbing. For scale, $518 billion is more than ten times the $47 billion run rate. The bet is simple to state: lock in capacity now, at a size that only makes sense if revenue keeps compounding for years. Customer concentration is the second pressure point. Two customers accounted for nearly a quarter of 2025 revenue, and many large customers have no long-term contracts. If either of those two shifted spending to a rival model, a big slice of revenue would leave with them, and customers without long-term contracts can walk quickly. Pair that with a mostly non-cancelable compute bill and you see why investors read this section closely. ### When is the Anthropic IPO? There is no confirmed Anthropic IPO date yet. Reuters sources put the listing in November 2026, while The Next Web has reported October. Morgan Stanley, Goldman Sachs and JPMorgan reportedly lead the offering, according to Quartz as cited in [TNW's coverage](https://thenextweb.com/news/anthropic-ipo-2-trillion-october-revenue-multiple). Making the S-1 public is a strong signal the listing is close, since companies that file confidentially must publish the document before they start pitching the shares to investors. The Anthropic IPO valuation talk starts at $965 billion, the post-money valuation (what the company was worth right after the round) from the $65 billion Series H, per NPR. Six backers told the Financial Times they expect an IPO valuation of $2 trillion or more, TNW reports. That would be roughly 43 times the $47 billion run rate. Listed AI peers have traded at about 55 times revenue, so by that yardstick $2 trillion would be a discount. Anthropic has not confirmed any figure. The wider context matters. SpaceX listed in June at roughly $1.77 to $1.8 trillion, and OpenAI confidentially filed its own IPO paperwork the same month. A $2 trillion debut would put Anthropic above SpaceX's listing valuation. What about the Anthropic IPO ticker and Anthropic IPO share price? The reporting this guide relies on does not cite either. A price range usually arrives in an amended filing once the banks have gauged demand, so treat any site quoting a share price before that appears on EDGAR with suspicion. ### What the Anthropic IPO filing means if you build on Claude If you pay for Claude or ship products on its API, three things in the filing are worth your attention. **Pricing will be watched by shareholders.** With compute above half of operating costs, every pricing and rate-limit decision becomes a margin decision that gets reported every quarter. API prices across the industry have been falling fast, as we covered in [our look at LLM API pricing cuts](/radar/llm-api-pricing-cuts-2026). A listed Anthropic will have to balance that price race against investors who want losses to shrink. **More capacity, and more pressure to sell it.** The $518 billion commitment should mean more compute coming online, which helps availability. It also means Anthropic needs demand to grow into it. Given that many large customers have no long-term contracts, expect a harder push toward multi-year enterprise deals. **Far better visibility into your vendor.** A public company reports results every quarter and has to disclose material risks. If you are betting a product on one model provider, that is more information than you have ever had. It is also a reminder to keep your code able to switch providers, the same advice any team relying on a single supplier should follow. ### What to watch next - **The amended S-1.** This is where a price range, share count and ticker should appear. - **The next quarter.** Whether adjusted operating profit holds for a third quarter, and whether the gap to standard accounting profit narrows. - **The valuation gap.** Whether the market accepts roughly 43 times run-rate revenue, or prices Anthropic closer to its $965 billion private mark. - **OpenAI's filing.** Its confidential paperwork will eventually go public too, and it will give investors the first direct comparison between the two labs. - **Concentration.** Whether the share of revenue from the two biggest customers shrinks as Claude's customer base widens. ## Frequently asked questions ### When is the Anthropic IPO? Anthropic has not confirmed a date. Reuters sources point to November 2026, while The Next Web reported October. Morgan Stanley, Goldman Sachs and JPMorgan reportedly lead the offering. The S-1 becoming public on SEC EDGAR is a strong sign the listing is close, but the exact day depends on the investor roadshow (the pitch tour before shares go on sale) and an amended filing with a price range. ### How much revenue does Anthropic make? About $4.6 billion in 2025, roughly twelve times the year before. Growth then accelerated: second-quarter 2026 revenue alone was $11.5 billion, and NPR reported run-rate revenue passed $47 billion in May 2026, up from $9 billion at the end of 2025. Run-rate projects recent revenue over a year, so the quarterly figure is the firmer number. ### Is Anthropic profitable? Not under standard accounting. Its 2025 operating loss was $8.06 billion, and the net loss was roughly $42 billion, though about $34 billion of that is a non-cash charge tied to financing that could convert into stock. The Decoder reports Anthropic is on track for a second straight quarter of operating profit, but only on an adjusted basis. ### Where can I read the Anthropic S-1 filing? The full prospectus is public on the SEC's EDGAR database under the filing anthropicpbc-sx1.htm. Nearly a third of it is risk factors, including warnings about existential risks and models that could act unpredictably. If you want the numbers without reading the whole filing, the financial statements and the risk factors section are the two parts that matter most. ## Source Based on Anthropic's S-1 on SEC EDGAR and press reporting. https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/anthropic-ipo-prospectus-five-numbers. Last modified 2026-10-27T07:32:20.591Z. ================================================================================ - URL: https://syntaxradar.com/radar/open-weight-models-explained-ibm - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-27T07:23:52.900Z - Last modified: 2026-10-27T07:23:52.900Z - Primary source: https://www.theregister.com/systems/2026/10/22/ibm-ceo-says-open-weight-models-are-now-irresistibly-good-and-cheap-on-prem/5304640 # Open Weight Models Explained: IBM Says 90% as Good, Cheaper **Short answer:** On IBM's Q3 earnings call, reported 2026-10-22, CEO Arvind Krishna said open-weight AI models now cost 10 to 25 percent of frontier token prices and are "90 percent as effective," making on-prem AI practical for most business tasks. **Takeaway:** Open-weight models now trail the best closed models by months rather than years, so for routine, high-volume or privacy-sensitive work they are worth testing before you pay frontier prices. **Key figures (as reported by the source):** - Open model token price vs frontier (per IBM's CEO): 10-25% - Effectiveness vs frontier (Krishna's claim): 90% - Intelligence Index: best closed vs best open: 60 vs 51 On IBM's third-quarter earnings call, [reported by The Register](https://www.theregister.com/systems/2026/10/22/ibm-ceo-says-open-weight-models-are-now-irresistibly-good-and-cheap-on-prem/5304640) on October 22, 2026, CEO Arvind Krishna made a blunt argument: open weight models, the AI models anyone can download and run on their own hardware, are now good enough and cheap enough for most everyday business work. He put their price at 10 to 25 percent of frontier token prices and called them "90 percent as effective." What is an open weight model, and how does open weight vs open source differ? This piece answers both. Keep in mind who is talking. That 90 percent figure is a CEO's claim on an earnings call, not a lab result, and IBM sells the hardware and software to run these models. Still, independent benchmark data points the same way: the gap between the best closed models and the best open ones is real, but it is now measured in months, not years. By the end you will know why the label is more than a wording quibble, where closed models still win, and how to decide which kind fits your team. ### What Is an Open Weight Model? An AI language model is, at bottom, a very large file of numbers called weights (you will also hear "parameters"). During training, a lab feeds the model huge amounts of text and nudges those numbers over and over until the model gets good at predicting what comes next. The finished weights are the model. Everything it has learned lives in them. An open weight model is one whose finished weights the developer publishes for anyone to download. You can run it on your own servers, on a rented cloud GPU, or, for the smaller ones, on a decent laptop. You can also fine-tune it, which means training it a little further on your own data so it gets better at your contracts, your support tickets or your codebase. A closed model works the other way round. The weights never leave the company's servers. You reach the model through an API: you send text in, text comes back, and you pay per token (a token is roughly a fragment of a word). You never touch the model itself. #### What is actually in the download? When a lab releases an open weights model, you typically get the weight files, a description of the model's architecture so software knows how to load them, and a licence that spells out what you may do. Free tools such as vLLM or Hugging Face's libraries handle the job of actually running it, a step called inference. What you usually do not get is the recipe: the training data, the full training code and the many choices about filtering and tuning that shaped the model's behaviour. Think of it as receiving a finished cake and the right to slice it, re-ice it and serve it, without the recipe or a list of where the flour came from. ### Open Weight vs Open Source: Why the Label Matters Plenty of headlines call these models "open source." In most cases that is loose language. In software, open source means you get the source code and the freedom to study, change and share it. The closest AI equivalent would include the training data and training code, so anyone could rebuild the model from scratch and check how it was made. Most open-weight releases stop at the weights. Licences add a second wrinkle. Some open weight releases come with permissive terms, while others attach conditions on commercial use or on what the model may be used for. Where exactly "open" begins for AI is still argued over, so treat the word as a starting point and read the licence before you build a product on any model.
Closed modelOpen weight modelFully open source model
How you use itThrough the vendor's APIDownload and run anywhereDownload and run anywhere
Weights you can inspectNoYesYes
Training data and codeNoUsually notYes
Fine-tune on your own dataOnly through vendor tools, if offeredYesYes
Your data leaves the buildingYesNot if you self-hostNot if you self-host
Main costPer-token feesHardware and staff (or a host's per-token fee)Hardware and staff
![A developer in a small office server room downloading open weight AI models onto a rack server, lit by the glow of a laptop screen](/uploads/blog/open-weight-models-explained-ibm/img1.webp) ### What IBM's CEO Said About Open Weight Models Krishna framed his case as a story of a closing gap. In 2024, he said, closed frontier models were clearly superior: on a broad benchmark, the top model might score 70 while an open model managed 20 or 30. By late 2025, by his account, that had narrowed to roughly 70 versus 50. Today, he argued, open-weight models cost 10 to 25 percent of frontier token prices while being, in his words, "90 percent as effective." His point was less about benchmarks than about the work companies actually do. Summarizing emails, reviewing contracts and handling customer calls, he said, do not need the most expensive model on the market. In his view, many businesses can build useful on-prem AI in two or three racks of servers. He was also candid about the catch. According to [The Register's report](https://www.theregister.com/systems/2026/10/22/ibm-ceo-says-open-weight-models-are-now-irresistibly-good-and-cheap-on-prem/5304640), Krishna said enterprises struggle to run these systems because doing it well takes kernel engineers and network specialists, people most companies do not employ. IBM plans software to make deployment easier, and it expects on-prem demand to lift sales of its Spyre accelerator cards for mainframes. So the pitch is honest about the hard part, and IBM is positioned to sell the fix. ### Are Open Weight Models as Good as Closed Models? Short answer: close, not equal, and it depends on the job. The [FutureSearch analysis of mid-2026 leaderboards](https://futuresearch.ai/app/p/a/open-weight-llm-overtaking-closed-weight) gives the clearest picture. On the Artificial Analysis Intelligence Index, Claude Fable 5 scores 60 and GPT-5.6 Sol scores 59, while the best open-weight model, GLM-5.2 (max), scores 51, about 9 points behind. On LMArena's text leaderboard, where people vote on which answer they prefer, Claude Fable 5 leads with an Elo rating of 1507; the top open model, GLM-5.1, sits around 27th place at 1471. Time is the other way to measure it. By FutureSearch's count, open models have trailed the closed frontier by roughly four months on average since January 2026. A May 2026 evaluation by CAISI at NIST put the lag for DeepSeek V4 Pro closer to eight months on less public benchmarks. That difference matters: tests that are widely published can be studied and tuned for, so private tests may give a more sober reading. On some narrow tasks the race is a near tie. A [June 2026 comparison](https://braindetox.kr/en/posts/open_vs_closed_llm_gap_2026.html) reports DeepSeek V4-Pro at about 80.6 percent on SWE-bench Verified, a coding benchmark built from real software bugs, against roughly 80.8 percent for Claude Opus 4.6. That is a single blog source, so treat it as a direction rather than a verdict. #### Where the gap still shows According to [one blog write-up](https://mythos.one/me/brianswichkow/12b11e), closed models still tend to lead in three areas: following complex instructions with many constraints at once, handling multimodal input such as images alongside text, and staying reliable over long, multi-step agent tasks. Treat that as directional: it is a single blog-level source, not a benchmark result. There is also a strategic reason the very top may stay closed. FutureSearch notes that Meta is keeping frontier models such as Muse Spark 1.1 proprietary, and that OpenAI's open-weight gpt-oss-120b sits well below the frontier. So the strongest open models may keep sitting a step behind the best closed ones. #### Which open weight AI models lead right now? Based on the sources above, the names to know are the GLM family (GLM-5.2 tops the open side of the Intelligence Index), DeepSeek V4 Pro (strong at coding), and OpenAI's gpt-oss-120b. Nvidia has also shown how far tuning can push a free model, which we covered in [our look at Nemotron](/radar/how-nvidias-nemotron-outsmarted-gpt4-secret-sauce-ai-tuning). ![An IT manager in a quiet company data center checking a row of GPU servers used to run open-weight AI models on-prem](/uploads/blog/open-weight-models-explained-ibm/img2.webp) ### What Do Open-Weight AI Models Really Cost to Run? Start with Krishna's range. If his numbers hold, every dollar a team spends on frontier tokens would buy the same volume of work from an open model for 10 to 25 cents. For a company pushing millions of emails or call transcripts through a model, that is the whole argument. (Closed prices are falling too, as we covered in [our breakdown of LLM API pricing cuts](/radar/llm-api-pricing-cuts-2026), so check current rates before you do the math.) But per-token price only tells you part of the story when you run the model yourself. A self hosted LLM turns a usage bill into a fixed one: you buy or rent GPUs, pay for power and space, and pay people to keep it fast and secure. That trade pays off when your volume is high and steady, and it hurts when your servers sit idle. Krishna's own warning about kernel and network engineers is the hidden line item. #### Self-hosted LLM vs using an API - **Using a closed API:** no hardware, best raw quality, pay per token, and your data goes to the vendor. - **Using an open model through a hosting provider's API:** many cloud providers serve open weight models per token, which is where much of the price gap shows up. Little setup, but your data still leaves your network. - **Self-hosting an open model:** your data stays in your building, you can fine-tune freely, and costs become predictable. In exchange, you own every outage, upgrade and security patch. ### Who Should Use Open Weight Models (and Who Should Not) Here is a simple way to decide. **Good fit:** - **Regulated or privacy-sensitive businesses**, such as healthcare, finance, legal and government, where keeping data on your own machines matters more than squeezing out the last few benchmark points. - **High-volume, repetitive tasks** like the ones Krishna listed: summarizing, classifying, reviewing documents, routing support calls. - **Teams that want to fine-tune** a model on their own data and own the result. - **Students and hobbyists**, who can run smaller open models locally to learn how these systems work with no API bill at all. **Poor fit, for now:** - **Products that live or die on frontier quality**, such as complex agents, multi-step reasoning or rich image understanding, where closed models still hold an edge. - **Small teams with no infrastructure people.** If nobody can debug a GPU server at 2 a.m., start with an API, closed or open-hosted, and revisit when volume grows. A practical middle path for founders: build on an API first, measure what your tasks really need, then test an open model on the same prompts. If it is good enough on your work, not on a leaderboard, the savings are yours to take. ### What to Watch Next Watch whether IBM's promised deployment software makes on-prem AI realistic for companies without specialist engineers, and whether Spyre sales show real demand behind the CEO's pitch. Watch the lag numbers: if the average gap shrinks below four months, the case for paying frontier prices on routine work gets thin. Watch whether more labs follow Meta in keeping their best models closed. And watch independent evaluations like CAISI's, which test models on benchmarks they have not been tuned for. ## Frequently asked questions ### What is an open weight model in simple terms? It is an AI model whose trained weights, the large file of numbers that holds everything the model learned, are published for anyone to download. You can run it on your own hardware, fine-tune it on your own data and keep your information private. You usually do not get the training data or full training code, and a licence still sets the rules. ### What is an open weight model vs open source? Open weight means you can download and run the finished model. Open source, borrowed from software, implies you can also see how it was built, which for AI would mean the training data and training code. Most models sold as open release only the weights, so open weight is usually the more accurate label. Licence terms also vary between models. ### Are open weight AI models really cheaper? Per token, IBM CEO Arvind Krishna says they cost 10 to 25 percent of frontier model prices. Self-hosting changes the bill: you pay for GPUs, power and skilled engineers instead of per-token fees. That works out well for high, steady volume and poorly for light or unpredictable use, where a hosted API is often cheaper overall. ### Which open-weight models are closest to closed ones? In mid-2026 data compiled by FutureSearch, GLM-5.2 (max) was the top open-weight model on the Artificial Analysis Intelligence Index at 51, against 60 for Claude Fable 5. DeepSeek V4-Pro reportedly scored about 80.6 percent on SWE-bench Verified, close to Claude Opus 4.6. Closed models still lead on complex agents and multimodal tasks. ## Source Based on The Register's report of IBM's Q3 earnings call and FutureSearch leaderboard data. https://www.theregister.com/systems/2026/10/22/ibm-ceo-says-open-weight-models-are-now-irresistibly-good-and-cheap-on-prem/5304640 --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/open-weight-models-explained-ibm. Last modified 2026-10-27T07:23:52.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/iphone-duo-vs-galaxy-z-fold-8-teardown - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-26T07:56:00.497Z - Last modified: 2026-10-26T07:56:00.497Z - Primary source: https://www.ifixit.com/News/119895/iphone-duo-teardown # iPhone Fold Teardown: iPhone Duo vs Galaxy Z Fold 8 for Buyers **Short answer:** iFixit tore down Apple's first foldable, the $1,999 iPhone Duo, and gave it a provisional 5/10 repair score, the best for any foldable so far. Its hinge gathered less dust than the Galaxy Z Fold8's, but an out-of-warranty inner screen costs $999. **Takeaway:** The iPhone Duo is the most repairable foldable iFixit has scored and undercuts the Galaxy Z Fold8 Ultra by $100, but a $999 inner screen and a hinge that still crunches with dust make AppleCare close to mandatory. **Key figures (as reported by the source):** - iFixit provisional repair score: 5/10 - Inner display replacement (out of warranty): $999 - Starting price (Galaxy Z Fold8 Ultra: $2,099): $1,999 Apple's first foldable, the iPhone Duo (the phone most people have been searching for as the Apple iPhone Fold), has been opened up, and the most interesting part is not the screen. It is the glue. In its [iPhone Duo teardown](https://www.ifixit.com/News/119895/iphone-duo-teardown), iFixit gave the phone a provisional repairability score of 5 out of 10, the highest it has handed any foldable so far. That changes the iPhone Duo vs Galaxy Z Fold 8 question, because with a $1,999 phone, what happens after a bad drop matters as much as the spec sheet. The catch is a single number: replacing the inner folding display out of warranty costs $999, according to iFixit. Half the price of the phone, for one part. Below: how a folding iPhone works inside, what the teardown says about repair, how it compares with Samsung's and Google's foldables on the points actually tested, and who should buy now versus wait. ### iPhone Fold teardown: what did iFixit find inside the iPhone Duo? Apple unveiled the iPhone Duo on September 9, 2026 as its first foldable iPhone, [BGR reported](https://bgr.com/2254151/apple-iphone-duo-foldable-price-release-date-specs). For anyone tracking the iPhone Fold release, pre-orders opened on October 16 in more than 70 countries and sales began October 23. iFixit published its teardown two days later. The headline findings, in short: - **A complicated hinge.** Gears, springs and linked metal joints. Apple says it has more than 100 components. - **A layered folding screen.** Apple describes more than ten ultrathin layers and adhesives. - **Two batteries, not one.** Cells of 7.5 Wh and 11.6 Wh, for 19.1 Wh total. - **Glue you can switch off.** Both batteries sit on electrically releasable adhesive. - **A simpler logic board.** It is single-layer, unlike the double-layer boards Apple has used in iPhones since the iPhone 8 (the XR and SE were exceptions). Getting all of this apart took five different screwdriver types. That is not a friendly number for home repair, but the overall picture is better than any foldable iFixit has scored before. ### How does an iPhone that folds actually work inside? A foldable phone has to solve three problems that a normal phone does not: bend a screen thousands of times, split its electronics across two halves, and keep dirt out of a moving joint. Here is how the Duo handles each. #### The screen is a stack, not a sheet of glass Glass does not fold, so a foldable screen is a sandwich of thin, flexible layers glued together. Apple says the Duo's folding display has more than ten of them. The top surface is a nano-texture cover sitting on a flexible polymer layer, and that layer can be replaced without replacing the OLED panel underneath, according to iFixit. That detail matters for your wallet. The cover layer is the part your fingernails, keys and pocket grit attack every day. Swapping it costs $49 without AppleCare or $19 with it, versus $999 for the whole inner display. #### The hinge is a tiny gearbox The hinge is built from gears, springs and linked metal joints, more than 100 components by Apple's count, packed into the spine of the phone. In iFixit's dust tests it collected less visible dust than the Galaxy Z Fold8 and the Pixel 11 Pro Fold. The bad news: enough still got in to make it crunch. #### Two halves, one charger The Duo carries two battery cells. Only one side has charging hardware, so power and data cross the hinge on an accordion-pleated flex cable, a ribbon folded like a paper fan so it can stretch and compress as the phone opens and shuts. One of the two batteries charges through that cable. #### Batteries that let go on command Phone batteries are normally held in with stretch-release tape or stubborn glue, and prying them out risks a puncture. The Duo uses adhesive that releases when you run a small electric current through it, using iFixit's VoltClip with a FixHub, or a plain 9V battery. Both cells are also protected by metal shells. ![A repair technician at a workbench lifting the folded halves of a titanium smartphone apart with a plastic opening tool](/uploads/blog/iphone-duo-vs-galaxy-z-fold-8-teardown/img1.webp) ### How repairable is the iPhone Duo? Better than any foldable iFixit has scored, still not easy. The 5/10 score is provisional, and iFixit says it will finalize it once replacement parts are on sale. The good parts are real. Batteries come out with electricity instead of force. The cover layer on the folding screen can be replaced on its own. The single-layer logic board is simpler to work on than the stacked boards in recent iPhones. And when iFixit removed the outer display, it came off with its trim intact. The bad parts are also real. Five screwdriver types is a lot of tooling. The hinge is a 100-plus-part mechanism that already crunched with dust in testing. And the inner display, the part most exposed on any foldable, costs $999 to replace out of warranty.
RepairCost (out of warranty)
Inner folding display$999
Folding display cover layer, no AppleCare$49
Folding display cover layer, with AppleCare$19
Read that table as a buying instruction. The cover layer is built to be the cheap, replaceable part, and the OLED underneath is the expensive one. The $19 fix only helps if the damage stays in the top layer. ### iPhone Duo vs Galaxy Z Fold 8: which is better? On price, Apple undercut expectations. The Duo starts at $1,999, while IDC had expected $2,500, [The Next Web reported](https://thenextweb.com/news/apple-iphone-duo-foldable-1999-price). Samsung's Galaxy Z Fold8 Ultra costs $2,099, so the iPhone is $100 cheaper at the starting line. Storage runs from 256GB to 2TB. On the teardown bench, the Duo came out ahead on the two things iFixit compared directly: - **Dust.** The Duo's hinge gathered less visible dust than both the Galaxy Z Fold8 and the Pixel 11 Pro Fold, though it was not dust-proof. - **Screen removal.** The Fold8's inner-screen trim crumbled as iFixit took it off. The Duo's outer display and trim came off intact. Those are signs of build quality and repair friendliness, not proof that one phone will outlast the other. The sources we used do not include Samsung's full spec sheet, battery claims or repair pricing, so we are not going to guess at a spec-by-spec scorecard. What we can say is what Apple is bringing to the fight: - **Screens:** a 7.6-inch inner display, about 50% larger than the iPhone 18 Pro Max's, and a 5.4-inch outer display with about 90% of the iPhone 18 Pro's screen area. - **Chip:** the A20 Pro, the same chip as the $1,199 iPhone 18 Pro. Apple claims up to 35% better sustained performance than the iPhone 17 Pro. (For why phone chips now matter beyond raw speed, see our explainer on [on-device NPUs](/radar/next-gen-silicon-edge-computing-on-device-npus).) - **Battery:** up to 31 hours of video on the inner screen, 44 hours on the outer, 24 hours of mixed use, and about 50% charge in roughly 20 minutes wired or 30 minutes on MagSafe or Qi2. - **Build:** grade 5 titanium, IP68, Ceramic Shield 2 on the front. - **Cameras:** dual 48MP rear cameras (Fusion Main and Fusion Ultra Wide), with 4K120 Dolby Vision video. There are trade-offs Apple fans should know before ordering. The Duo has no Face ID; Touch ID lives in the side button. It is eSIM only, so there is no physical SIM tray. It comes in two colours, star white and night sky. What about the crease? Neither the teardown nor the launch coverage we drew on measures it, so this is one to judge with your own eyes in a store. ![A commuter on a train unfolding a large foldable phone to read on its inner screen](/uploads/blog/iphone-duo-vs-galaxy-z-fold-8-teardown/img2.webp) ### Is the iPhone Fold worth it? Whether Apple's new iPhone foldable is worth it depends on which kind of buyer you are. **Buy it if** you already live in Apple's ecosystem, want a tablet-sized screen in your pocket, and will pay for AppleCare. The teardown suggests this is the most repairable foldable iFixit has tested, the battery life claims are strong, and the $19 cover-layer swap with AppleCare makes the most common kind of screen wear cheap to fix. **Wait if** you are rough on phones, work in dusty places, or will not buy AppleCare. A hinge that crunches with dust and a $999 inner screen are a bad pairing for a phone that lives in a toolbox or a beach bag. Waiting also means the repair score gets finalized and real-world durability reports arrive. **Look at Samsung or Google if** you are already on Android, or the Duo's eSIM-only design is a dealbreaker. Switching platforms for a first-generation foldable is a big ask, and on the evidence here the Duo's lead is about repair and build details, not a feature you cannot get elsewhere. **Skip it if** you mostly want the A20 Pro chip and a great camera. The $1,199 iPhone 18 Pro has the same chip for $800 less, without a hinge to worry about. ### What to watch next Three things will settle whether the iPhone Fold is a smart buy or an expensive first draft. 1. **The final repair score.** iFixit will lock in its number once Apple sells replacement parts. The 5/10 is provisional, so it could move either way. 2. **Hinge durability over months.** The dust crunch showed up in a teardown lab. Owner reports from beaches, building sites and coat pockets will show whether it matters in daily life. 3. **Wider availability.** Sales expand to 28 more markets from October 30, which will bring more buyers, more reviews and a clearer picture of how the Duo holds up against Samsung's and Google's foldables. If you are curious what Apple's chips will do with all that screen, our look at [why phones are moving AI on-device](/radar/on-device-phone-ai-local-models-apple-android-privacy) is a good next read. ## Frequently asked questions ### Is the iPhone Duo worth it? It is worth it if you are already an iPhone user, want a 7.6-inch screen that folds into your pocket, and will pay for AppleCare. If you are rough on phones or mainly want Apple's fastest chip, the $1,199 iPhone 18 Pro has the same A20 Pro for $800 less, with no hinge or $999 folding screen to worry about. ### How repairable is the iPhone Duo? iFixit gave it a provisional 5 out of 10, the highest score it has given any foldable so far. The batteries release with an electric current, the folding screen's cover layer can be swapped on its own, and the logic board is single-layer. Against that, the teardown needed five screwdriver types, and the inner display costs $999 to replace out of warranty. ### Does the iPhone Duo have a crease? The sources we used do not measure the crease, so we cannot give you a verdict on how visible it is. What we know is how the screen is built: more than ten ultrathin layers and adhesives, topped by a nano-texture cover on a flexible polymer layer that can be replaced without replacing the OLED. Check it in person before you buy. ### iPhone Duo vs Galaxy Z Fold 8: which is better? On the evidence available, the Duo leads on repair and build details. Its hinge gathered less dust than the Fold8's in iFixit's tests, and its screen came off intact while the Fold8's inner-screen trim crumbled. It also starts at $1,999 versus $2,099 for the Fold8 Ultra. Android users should still look at Samsung. ## Source Based on iFixit's iPhone Duo teardown. https://www.ifixit.com/News/119895/iphone-duo-teardown --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/iphone-duo-vs-galaxy-z-fold-8-teardown. Last modified 2026-10-26T07:56:00.497Z. ================================================================================ - URL: https://syntaxradar.com/radar/llm-api-pricing-cuts-2026 - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T22:13:25.843Z - Last modified: 2026-10-25T22:13:25.843Z - Primary source: https://minimaxir.com/2026/10/cheap-llms/ # LLM API Pricing Fell Up to 80%: Why, and How to Pay Less **Short answer:** Between June and September 2026, OpenAI cut its small Luna model by 80% and both OpenAI and Anthropic launched cheaper flagships: GPT-6.1 Sol at $2/$10 and Claude Opus 5.5 at $4/$20 per million tokens. **Takeaway:** AI model prices are falling because serving got cheaper and open-weight rivals made switching easy, so builders should start on the cheapest tier, design for caching and re-check prices every quarter. **Key figures (as reported by the source):** - GPT-5.6 Luna price cut (July 30): 80% - GPT-6.1 Sol input/output per 1M tokens: $2 / $10 - Claude Opus 5.5 input/output per 1M tokens: $4 / $20 If you build on ChatGPT or Claude, or use apps that do, the meter just slowed down. LLM API pricing, the per-token rates developers pay to send text to these models, fell hard between June and late September 2026. OpenAI's smallest model got 80% cheaper in one cut, and both OpenAI and Anthropic shipped new flagship models that cost less than the ones they replaced. The catch is in the fine print. That headline 80% applied to one small model, not the whole lineup. The reasons behind the cuts also say a lot about who holds the power in this market now. By the end you will know what changed, what a dollar buys today, the four forces pushing prices down, and a short checklist for paying less without settling for a worse model. ### What actually got cheaper since June? First, how to read LLM API pricing. Providers quote prices in dollars per million tokens, in three parts. **Input** is what you send (your question, a document, some code). **Output** is what the model writes back. **Cache read** is a discounted rate for input the provider has already processed recently, such as a long set of instructions you resend with every request. A token is a small chunk of text, often a piece of a word. Here is the sequence, drawn from [a detailed price timeline published on minimaxir.com](https://minimaxir.com/2026/10/cheap-llms/) and [Reuters' report on OpenAI's July cut](https://www.thestar.com.my/tech/tech-news/2026/07/31/openai-cuts-prices-on-smaller-models-as-businesses-scrutinize-ai-spend): - **June 26:** OpenAI launched GPT-5.6 Sol at $5 input, $30 output and $0.50 cache read, the same as GPT-5.5. Its small sibling, GPT-5.6 Luna, came in at $1, $6 and $0.10. - **June 30:** Anthropic launched Claude Sonnet 5 at $2, $10 and $0.20, first described as temporary launch pricing. On August 10 it made that price permanent. - **July 30:** OpenAI cut Luna by 80%, to $0.20, $1.20 and $0.02. The mid-tier Terra model dropped less: input from $2.50 to $2, output from $15 to $12. The flagship Sol did not move. - **September 1:** Anthropic released Fable 5.1 at $10, $50 and $0.25, and cut its cache read price by 75%. - **September 22:** OpenAI released GPT-6 Sol at $2, $10 and $0.20, plus GPT-6 Luna at $0.10, $0.50 and $0.01. The same day Anthropic released Claude Opus 5.5 at $4, $20 and $0.20, down from Opus 5's $5, $25 and $0.50. - **September 29:** OpenAI followed with GPT-6.1 Sol at $2, $10 and $0.10, halving the cache read price again. So the flagship story is real too, just slower. OpenAI's top model went from $30 per million output tokens in June to $10 by the end of September. The timeline's author calls the Sol family 75% cheaper cumulatively, and puts Opus 5.5's total cost reduction against Opus 5 at 40% to 60%. ### LLM API price comparison: what does $1 buy now? Price lists are hard to feel. A simpler view is to divide a dollar by the rate. The table works as a quick LLM API cost comparison, collecting the list prices from the timeline above, all per million tokens.
ModelInputOutputCache read
GPT-5.6 Sol (June)$5$30$0.50
GPT-6.1 Sol (Sept 29)$2$10$0.10
GPT-5.6 Luna (June)$1$6$0.10
GPT-5.6 Luna (after July 30 cut)$0.20$1.20$0.02
GPT-6 Luna (Sept 22)$0.10$0.50$0.01
Claude Opus 5$5$25$0.50
Claude Opus 5.5 (Sept 22)$4$20$0.20
Claude Sonnet 5$2$10$0.20
Kimi K3.1 Pro$2$10$0.20
DeepSeek V4 Flash$0.14$0.28$0.0028
Now the division. In June, $1 of output from OpenAI's small model bought you about 167,000 tokens. At GPT-6 Luna's $0.50, the same dollar buys 2 million. That is twelve times more writing for the same money in under three months. At the top end, $1 of Opus 5.5 output buys 50,000 tokens. Same dollar on GPT-6.1 Sol buys 100,000. Neither is wrong to pay for. They are different tools, and the gap between "small" and "flagship" is now far wider than the gap between the two big labs. ![A software developer at a wooden desk by a window at dusk, studying a laptop while weighing which AI model to pay for](/uploads/blog/llm-api-pricing-cuts-2026/img1.webp) ### Why is LLM pricing dropping so fast? Four forces show up in the sources. None of them is charity. **1. Models got cheaper to run.** OpenAI told Reuters the lower prices were partly enabled by efficiency gains from GPT-5.6, including improved code and performance optimization during internal development. Anthropic was blunter about Opus 5.5, saying it "requires less compute to serve than Opus 5, and its pricing reflects that." When serving a request costs less in chips and electricity, a lab can cut the price and keep its margin. (For one concrete example of how that kind of saving happens, see [how DeepSeek cut AI server costs by 80%](/radar/deepseek-v3-mla).) **2. Open-weight rivals set a low ceiling.** [Reuters reports](https://www.thestar.com.my/tech/tech-news/2026/07/31/openai-cuts-prices-on-smaller-models-as-businesses-scrutinize-ai-spend) that open-source Chinese models such as Z.ai's GLM-5.2 nearly match US performance at lower cost. DeepSeek V4 Flash, released in April, lists at $0.14 input and $0.28 output. The timeline's author also claims Kimi K3.1 Pro charges about half of Opus 5.5's input and output price at similar benchmark scores. Every one of those is a price a buyer can point to in a negotiation. **3. Companies started reading their AI bills.** The same Reuters report says businesses facing high AI costs are scrutinizing what they spend. The first wave of AI projects was often funded as experiments. Once an app is running in production and sending millions of requests, the finance team notices, and a cheaper model becomes a line item, not a nice-to-have. **4. Switching is easy.** DeepSeek V4.1 Flash is hosted by 29 providers on OpenRouter, up from 13 for V4 Flash in May, [according to the timeline](https://minimaxir.com/2026/10/cheap-llms/). When the same model is sold by dozens of hosts, and most apps reach LLM APIs through a simple call, moving to a cheaper option can take an afternoon. What about the rumour that OpenAI is cutting prices to look good ahead of an IPO? The author mentions the IPO talk but argues price cuts would look counterproductive for IPO financials. No company has given that as a reason. ### Is Claude cheaper than GPT? At list price, mostly no, but it depends on which tier you compare. At the flagship level, GPT-6.1 Sol costs $2, $10 and $0.10. Claude Opus 5.5 costs $4, $20 and $0.20. On all three numbers, Opus is exactly double. Anthropic's pricier Fable 5.1, at $10 input and $50 output, sits in a bracket of its own. One tier down, the picture evens out. Claude Sonnet 5 at $2, $10 and $0.20 matches GPT-6 Sol's launch price on every line. Against GPT-6.1 Sol it matches on input and output, but pays double for cached reads. At the cheap end, the timeline does not report a new small Claude price, so the race for the cheapest LLM API is between OpenAI's Luna and open-weight models. GPT-6 Luna, at $0.10 input, now undercuts DeepSeek V4 Flash's $0.14 on input, while DeepSeek stays cheaper on output ($0.28 against $0.50) and on cache reads. Two cautions. List price is not cost per task: a model that solves your problem in one try can beat a cheaper one that needs three attempts. And benchmark claims like "similar performance" are averages, so test on your own work before you switch. ![A small team of founders around a cafe table reviewing monthly costs together on a tablet and paper notes](/uploads/blog/llm-api-pricing-cuts-2026/img2.webp) ### What is a cache read price and why does it matter? Many AI apps send the same long block of text over and over. A customer support bot resends its rulebook with every message. A coding assistant resends large chunks of your project. A cache lets the provider remember that repeated part, and charge you a much lower rate to read it again. This is where some of the quietest cuts in LLM API pricing landed. Anthropic cut Fable 5.1's cache read price by 75%. Opus 5.5's cache read fell from $0.50 to $0.20. GPT-6.1 Sol's dropped to $0.10, which is one twentieth of its $2 input price. For a chatbot that is mostly repeated instructions, the cache read price can matter more than the headline input rate. Two models with the same input price can produce very different bills once caching is counted. Opus 5.5 shows it: its cache read price fell 60% while input and output fell 20%, so a cache-heavy workload saves more. ### What this means for builders and everyday users If you pay for an API, the practical moves are simple: - **Start with the cheapest tier that could work.** Try a small model like Luna or an open-weight option first. Move up only when it fails on real examples from your app. - **Design for the cache.** Put the parts of your prompt that never change (instructions, reference material) at the start, and the parts that change at the end, so repeated text can be served at cache read rates. - **Watch output, not just input.** Output costs five times input on GPT-6.1 Sol and Opus 5.5. Asking for shorter answers is often the cheapest optimization you have. - **Re-check LLM API pricing every quarter.** These two companies announced new prices more than six times between June and September. Sonnet 5's "temporary" launch price became permanent. Last spring's comparison spreadsheet is already wrong. - **Keep switching cheap.** Route model calls through one place in your code, so moving to the cheapest LLM API provider is a config change instead of a rewrite. If you just use AI apps, the effect is indirect but real. The companies behind your writing tools, coding helpers and chat apps now pay far less per answer. Whether that shows up as lower subscription prices, bigger usage limits or fatter margins is each company's choice. ### What to watch next in LLM API pricing Three things will tell you whether this keeps going. First, whether Anthropic answers OpenAI's Luna cuts at the small-model tier. Second, whether flagship prices keep falling now that GPT-6.1 Sol and Opus 5.5 have set a new floor. Third, how many hosts pick up the next open-weight release, since that number is a good proxy for how hard the price pressure will be. Efficiency research keeps pushing in the same direction, including ideas like [1-bit models that swap hard math for simple addition](/radar/the-1-bit-llm-revolution-bitnet-b158-ternary-addition-vs-multiplication). Nobody can promise the cuts continue, so build as if LLM API pricing will keep moving. ## Frequently asked questions ### Why are LLM prices dropping so fast? Four forces are at work. Newer models are cheaper to serve, according to both OpenAI and Anthropic. Open-weight rivals like DeepSeek and Z.ai GLM-5.2 offer similar quality at lower prices. Businesses are scrutinizing their AI bills. And with popular open models hosted by dozens of providers, switching to a cheaper option takes little effort. ### How much does the OpenAI API cost per million tokens in 2026? As of late September 2026, the flagship GPT-6.1 Sol costs $2 per million input tokens, $10 per million output tokens and $0.10 for cached reads. The small GPT-6 Luna costs $0.10 input, $0.50 output and $0.01 for cached reads. Prices have changed several times since June, so check the official page before budgeting. ### Is Claude cheaper than GPT? At the flagship level, no. Claude Opus 5.5 costs $4 input and $20 output per million tokens, exactly double GPT-6.1 Sol. Claude Sonnet 5, at $2 input and $10 output, matches GPT-6.1 Sol on those two rates but costs twice as much for cached reads. Real cost also depends on how many attempts each model needs for your task. ### What is a cache read price and why does it matter? A cache read price is the discounted rate for input text the provider has recently processed, like instructions your app resends with every request. It is often far below the normal input rate: GPT-6.1 Sol charges $0.10 for cached reads versus $2 for fresh input. For apps that repeat long prompts, it can decide most of the bill. ## Source Based on minimaxir.com price timeline and Reuters. https://minimaxir.com/2026/10/cheap-llms/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/llm-api-pricing-cuts-2026. Last modified 2026-10-25T22:13:25.843Z. ================================================================================ - URL: https://syntaxradar.com/radar/ai-skin-cancer-detection-old-cnn-beats-llms - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T20:59:17.579Z - Last modified: 2026-10-25T20:59:17.579Z - Primary source: https://lynchlab.net/llm-skin-cancer # AI Skin Cancer Detection: Why a 10-Year-Old Model Beat GPT-6 **Short answer:** Dermatologist Magnus Lynch tested GPT-6 Astra, Claude Fable 5.1 and the 10-year-old Inception v3 CNN on 200 biopsy-confirmed skin lesions. The CNN got 62% of diagnoses right versus 42% for both chatbots, but missed 18 of 30 melanomas at its default setting. **Takeaway:** A small model fine-tuned on skin images out-diagnosed two frontier chatbots, but both kinds of AI failed in ways that make a dermatologist, not an app, the right final check. **Key figures (as reported by the source):** - Correct first diagnosis: CNN vs both chatbots: 62% vs 42% - Harmless lesions correctly discharged: CNN / GPT-6 Astra / Claude Fable 5.1: 76 / 32 / 22 of 100 - Melanomas the CNN discharged at its default setting: 18 of 30 A 10-year-old image-recognition network just beat two of the most powerful AI models of 2026 at a job that matters: telling skin cancer apart from harmless moles. In a dermatologist-run test of AI skin cancer detection, the old model picked the right diagnosis 62% of the time. GPT-6 Astra and Claude Fable 5.1 both managed 42%. That headline is true, and it is not the whole story. At its default setting, the old model also sent home 18 of the 30 melanomas it saw, the most serious mistake a screening tool can make. The chatbots made the opposite error: they told almost everyone to see a doctor, including people with perfectly harmless spots. Below: what the study measured, why a small specialist can outperform a giant generalist here, and what it means if you were hoping to check a mole with your phone. ### What did the AI skin cancer detection test actually find? The test comes from Dr Magnus Lynch, a consultant dermatologist and Mohs surgeon at Guy's Hospital and an honorary reader at King's College London. He [published the comparison on the Lynch Lab website](https://lynchlab.net/llm-skin-cancer) on 25 October 2026. It is a blog post, not a peer-reviewed paper, and that matters. It still landed on the Hacker News front page the same day, with 274 points and 133 comments. Lynch put three models through the same exam. Two were frontier chatbots that can read images: GPT-6 Astra and Anthropic's Claude Fable 5.1, which he describes as two of the most capable AI models available this month. The third was Inception v3, a convolutional neural network (CNN), a type of model built purely to recognise patterns in pictures. He calls it a 10-year-old baseline. Here is how they scored on 200 lesions, half dangerous and half harmless:
ResultGPT-6 AstraClaude Fable 5.1Inception v3 (CNN)
Correct first diagnosis (of 200)84 (42%)85 (42%)125 (62%)
Melanomas named correctly (of 30)8 (27%)11 (37%)13 (43%)
Harmless lesions correctly discharged (of 100)322276
Dangerous lesions correctly kept for follow-up (of 100)979677 at default setting
Melanomas wrongly discharged (of 30)3418 at default setting
Read across the rows and two different personalities appear. The CNN is much better at naming the actual condition and at reassuring people whose spots are harmless. The chatbots almost never let a dangerous lesion go, but they get there by flagging most of the harmless ones too. ### How do you test an AI on moles? "Discharge" is the key word in this study. In a skin clinic, discharging a patient means telling them no follow-up is needed. So the real question for each model was simple: would it send someone home who should have stayed? Lynch drew 200 lesions at random from MILK10k, a public dataset released through the International Skin Imaging Collaboration (ISIC) in 2025. Every lesion had been biopsied and confirmed under the microscope, so the right answer was known for certain. Each came with two images: an ordinary clinical photo and a dermoscopic image, the magnified, lit close-up a dermatologist takes with a handheld scope. The 100 lesions that should not be discharged were 30 melanomas, 25 basal cell carcinomas, 25 squamous cell carcinomas (including Bowen's disease) and 20 actinic keratoses. The 100 safe ones were 50 benign naevi, the ordinary moles most of us have, and 50 seborrhoeic keratoses, common harmless growths. Every model had to pick one of those six diagnoses as its first choice. A script then scored each answer against the lab result, so no human judgement crept into the marking. The CNN had homework first. It started life trained on everyday photographs, then Lynch fine-tuned it on 68,011 dermoscopic and clinical images from the ISIC Archive. The training ran on a single H100 GPU, one high-end AI chip. ![Dermoscopic close-up of a mole used to test AI skin cancer detection models](/uploads/blog/ai-skin-cancer-detection-old-cnn-beats-llms/img1.webp) ### Why would an old image model beat a newer chatbot? Think of a brilliant general practitioner who has read every textbook against a specialist who has examined tens of thousands of skin lesions and been told the correct answer each time. The generalist can talk fluently about melanoma. The specialist has simply seen more of them. That is roughly the gap here. A frontier chatbot is trained to handle almost anything you throw at it, from tax questions to poetry, and it reads images as one skill among many. The CNN does exactly one thing. Fine-tuning on 68,011 labelled skin images taught it which pixel patterns go with which diagnosis, and nothing else competes for its attention. There is a second, quieter advantage: the CNN's caution can be adjusted. The study shows the model's cut-off for what counts as "worth a closer look" being moved, with large effects on what it missed. The chatbots were each tested at whatever level of caution they showed by default. None of this is new. A [2017 Nature paper by Esteva and colleagues](https://www.Gwern.net/doc/ai/nn/2017-esteva.pdf) trained a CNN with the same Inception v3 architecture, also pretrained on general object recognition, on 129,450 clinical images covering 2,032 diseases. It performed on par with all 21 board-certified dermatologists it was tested against on biopsy-proven images. Lynch's result suggests the frontier chatbots of 2026 still have not caught up with that kind of focused training on this task. A detail from his caveats sharpens the point. Because MILK10k is public, Lynch notes the chatbots may have seen it during training, while the CNN had not. If that happened, it would have tilted the test toward the chatbots, and they still lost on diagnosis. ### Where did each AI go wrong? The chatbots failed by over-warning, and each had a favourite wrong answer. GPT-6 Astra labelled 75 lesions as basal cell carcinoma when only 25 were, and that pile included 14 of the 30 melanomas. Claude Fable 5.1 had a different habit, calling 53 lesions actinic keratosis (a rough, sun-damaged patch that can turn into cancer) when only 20 were. Those mistakes look strange but explain the table. Calling a melanoma a basal cell carcinoma is the wrong diagnosis, yet the patient still gets follow-up because both are on the "do not discharge" list. Calling a harmless mole an actinic keratosis keeps a healthy person in the system. The result is high safety on paper and a lot of unnecessary worry. The CNN failed in the opposite direction. At its preset threshold it kept only 77 of the 100 dangerous lesions and discharged 18 of the 30 melanomas. That is the number to remember before you trust any app's default setting. The cut-off matters here. Picture a smoke alarm with a sensitivity dial: turn it down and it stops going off for toast, but it may also miss a real fire. When Lynch turned the CNN's dial up to match the chatbots' caution, it kept 98 or 99 of the dangerous lesions and let only 2 or 1 melanomas through. Same model, very different safety record. The test set was also hard. The CNN's area under the ROC curve, a standard score for how well a model ranks risky cases above harmless ones (higher is better), was 0.86 on these lesions against 0.96 on ordinary archive images. Every case here had been worrying enough to biopsy, which is not what a typical mole looks like. ![Doctor reviewing a grid of mole photos on a laptop during AI skin cancer screening](/uploads/blog/ai-skin-cancer-detection-old-cnn-beats-llms/img2.webp) ### How much should you trust one blog study? Less than the headline suggests, and Lynch says so himself. His caveats are unusually candid: - Only 30 melanomas were tested, so the uncertainty around those numbers is wide. - Each chatbot saw each case once, and chatbot answers vary from run to run. - Every lesion was biopsied, so the set is harder than what a clinic sees day to day. - The dataset is public, so the chatbots may have met it in training. - Inception v3 is a baseline, not the best image model available. He also discloses that Claude Code, an Anthropic tool, helped with the analysis and first drafts, and that one of the tested chatbots is Anthropic's too. The scoring itself was done by script against the lab results. What makes the finding credible is that it fits a wider pattern. A JAMA Dermatology study published this year by Anriot and colleagues, [summarised by Practical Dermatology](https://practicaldermatology.com/news/analysis-ai-model-outperforms-early-career-physicians-in-skin-lesion-diagnosis/2487332/), used 1,117 cases and 652 physicians. A dermatology foundation model called PanDerm reached 72.2% multiclass accuracy, a multimodal model 66.3% and a first-generation CNN 56.7%. Physicians with more than 10 years of experience still scored highest, at 74.2%. Put the two together and the picture for skin cancer AI is consistent: purpose-built skin models keep improving, general chatbots trail them on this task, and experienced dermatologists remain the benchmark. ### Can ChatGPT or Claude diagnose skin cancer from a photo? Not in any useful sense, based on this test. Lynch ran a second, patient-style experiment using only the ordinary clinical photo and the question "do I need to see a doctor?" For a further 100 biopsy-proven melanomas, the two chatbots said yes 99 and 98 times. That sounds reassuring until you learn they also said yes for almost every harmless lesion. A tool that answers "see a doctor" to nearly everything is not doing AI skin diagnosis. It is giving you the safe default. That is not harmful advice, but it tells you nothing you could not have worked out yourself. So is an AI skin cancer detection app accurate enough to replace a dermatologist? This study did not test any commercial app, so it cannot rank them. What it does show is the trade-off every such tool faces. A specialist model can be far sharper than a chatbot, but its safety depends on where the cut-off is set, and at its preset setting it sent home 18 of 30 melanomas. The practical takeaway is short. Use AI as a nudge, never as permission. If a spot on your skin worries you, book an appointment whatever an app or chatbot says. Skin checks are a case where the human specialist still sets the standard. ### What to watch next The obvious next step for skin cancer AI detection is a bigger, repeated version of this test: more melanomas, several runs per chatbot, and a peer-reviewed write-up. Lynch's own caveats read like a to-do list for that study. Watch, too, whether general chatbots start handing medical images to specialist models instead of answering from their own broad vision skills, and how foundation models like PanDerm perform when they meet everyday cases rather than curated ones. Until then, the lesson from this AI skin cancer detection test travels well beyond dermatology: for narrow, high-stakes image tasks, a focused model trained on the right data can still beat the biggest AI in the room. ## Frequently asked questions ### Can AI detect melanoma? Sometimes, but not reliably enough to stand alone. In Dr Magnus Lynch's test of 30 biopsy-proven melanomas, a specialist image model named 13 correctly, Claude Fable 5.1 named 11 and GPT-6 Astra named 8. At its default setting the specialist model also discharged 18 of the 30, which is why a doctor should make the final call. ### Can ChatGPT or Claude diagnose skin cancer from a photo? Not usefully, based on this test. Shown only a clinical photo and asked whether to see a doctor, GPT-6 Astra and Claude Fable 5.1 said yes for 99 and 98 of 100 melanomas, but also for almost every harmless lesion. That is a safe default, not a diagnosis, so it adds little beyond telling you to book an appointment. ### Why would an old image-recognition model beat a newer chatbot at medical images? Because it is a specialist. Inception v3 was fine-tuned on 68,011 labelled skin images and does nothing but classify them, while chatbots spread their skills across every topic. Its cut-off can also be tuned to be more cautious. A 2017 Nature study found the same architecture matched 21 dermatologists on biopsy-proven images. ### Is an AI skin cancer checker app accurate enough to replace a dermatologist? No app was tested in this study, so it cannot rank products. It does show the core trade-off: a specialist model's safety depends on its cut-off setting. In a separate 2026 JAMA Dermatology study, physicians with over 10 years of experience scored 74.2%, above every AI model tested. Treat any app as a prompt to see a doctor. ## Source Based on Dr Magnus Lynch's Lynch Lab test. https://lynchlab.net/llm-skin-cancer --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/ai-skin-cancer-detection-old-cnn-beats-llms. Last modified 2026-10-25T20:59:17.579Z. ================================================================================ - URL: https://syntaxradar.com/radar/how-nvidias-nemotron-outsmarted-gpt4-secret-sauce-ai-tuning - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T09:14:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Instruct-HF # Nvidia Nemotron: What the Open Model Did and Did Not Prove **Short answer:** In October 2024 Nvidia released Llama-3.1-Nemotron-70B-Instruct, a Llama 3.1 model retuned with a reward model. Its model card reports 85.0 on Arena Hard, ahead of GPT-4o (79.3) and Claude 3.5 Sonnet (79.2), with caveats. **Takeaway:** A reward model tuned a mid-sized open model to top three AI-judged chat benchmarks, though knowledge benchmarks barely moved. **Key figures (as reported by the source):** - Arena Hard score (model card): 85.0, vs GPT-4o 79.3 - AlpacaEval 2 LC (model card): 57.6, vs GPT-4o 57.5 - RewardBench score (paper): 94.1 Nvidia Nemotron is a family of open AI models from Nvidia, and the one that made headlines in October 2024 was a retuned version of Meta's Llama 3.1 70B. On three chat-quality tests, it scored above OpenAI's GPT-4o and Anthropic's Claude 3.5 Sonnet. The gain came from better training feedback applied to the same 70B model. The tests measure how much a judge prefers a model's answers in open-ended chat, and Nvidia's own model card says the model was not tuned for specialized areas such as math. ### What is Nemotron? Nemotron is Nvidia's name for its own line of language models and the tools used to train them. The model in the 2024 headlines is Llama-3.1-Nemotron-70B-Instruct. Its model card on Hugging Face (a site where AI developers share models) says the starting point was Meta's Llama-3.1-70B-Instruct. The "70B" means roughly 70 billion adjustable settings, called parameters. Nvidia took an existing open model and ran an extra round of training on top. The model card says the license is the NVIDIA Open Model License, with the Llama 3.1 Community License also applying, and the page says the model is "Built with Llama." The name now covers more than that one model. In December 2025 Nvidia [announced the Nemotron 3 family](https://nvidianews.nvidia.com/news/nvidia-debuts-nemotron-3-family-of-open-models), which comes in three sizes. The later sections return to that. ### What is a reward model, and how did it tune Nemotron? A reward model is a second AI that scores answers. It reads a question and a reply, then outputs a number for how good the reply is. Developers use those numbers as a training signal. The training method is called reinforcement learning from human feedback, or RLHF. The main model writes answers, the reward model scores them, and the main model is nudged toward answers that score higher. The model card says Nvidia used a specific variant called REINFORCE, run in Nvidia's NeMo Aligner software. The reward model itself was trained on human preference data. According to the [HelpSteer2-Preference paper](https://arxiv.org/abs/2410.01257), Nvidia released new preference labels for its HelpSteer2 dataset, where people marked which of two answers was better and wrote justifications. The paper compares two common ways of building a reward model and proposes a method that combines them. The paper reports that a Llama-3.1-70B reward model built this way scored 94.1 on RewardBench, a public test of reward models. The authors say that put it first among more than 140 reward models as of 1 October 2024. That score describes the judge, and the chat benchmarks below describe the model it trained. ![A researcher compares two printed answer sheets side by side at a desk in a quiet office](/uploads/blog/how-nvidias-nemotron-outsmarted-gpt4-secret-sauce-ai-tuning/gemini-supporting.jpg) ### How did Nemotron score against GPT-4o and Claude? The model card lists three automatic benchmarks, with figures dated 1 October 2024. Arena Hard and AlpacaEval 2 are tests where another AI model judges which answer is better. MT-Bench scores answers on a ten-point scale, also with an AI judge (GPT-4-Turbo here). According to the model card, the results were:
ModelArena HardAlpacaEval 2 LCMT-Bench
Llama-3.1-Nemotron-70B-Instruct85.057.68.98
GPT-4o (2024-05-13)79.357.58.74
Claude 3.5 Sonnet (20240620)79.252.48.81
Llama-3.1-405B-Instruct69.339.38.49
Llama-3.1-70B-Instruct (starting point)55.738.18.22
The jump from the starting model is large: Arena Hard rose from 55.7 to 85.0. Nemotron also beat Meta's own 405B model, which is roughly six times larger by parameter count. On AlpacaEval 2 the gap to GPT-4o is 0.1 points, which is inside the margin of error the card lists (a standard error of 1.65 for Nemotron and 1.47 for GPT-4o). Calling that a win overstates it. The model card also reports a Chatbot Arena rating of 1267 and a rank of 9 as of 24 October 2024, which places it near the top but below the number one slot. ### Did Nvidia Nemotron really beat GPT-4o? On those three tests, yes, by the model card's numbers. Judge-based chat benchmarks reward answers that an AI judge prefers. Researchers who built the length-controlled version of AlpacaEval wrote that existing open models had exploited its length bias, which is why the "LC" in the table exists. The Arena Hard authors likewise describe adjusting scores for style features such as length, markdown headers, bold text and lists. A model trained against a reward model can learn what the judges like. There was also a public question about contamination, meaning training on test data. Latent Space [looked into it](https://latent.space/p/ainews-legacy-24-10-16-ainews-did-nvidias-nemotron-70b-train-on-test) and reported that two knowledge benchmarks, GPQA and MMLU Pro, came in unchanged or worse compared with the base Llama 3.1 70B. It concluded the explanation was likely benign, with the gains coming from the new preference dataset and no training on test. The tuning made the model's chat answers much more to the liking of AI judges, while the two knowledge benchmarks stayed level or dropped. Nvidia's card describes the model as a demonstration of techniques for improving helpfulness in general instruction following. ![A server technician in a data center aisle checks a rack of GPU servers](/uploads/blog/how-nvidias-nemotron-outsmarted-gpt4-secret-sauce-ai-tuning/gemini-supporting.jpg) ### Can you run Nemotron yourself? The weights are public, so developers can download them. The practical cost is hardware. The model card says the model needs two or more 80GB GPUs and at least 150GB of free disk space. Output is capped at 4,000 tokens (a token is a word fragment, roughly three quarters of a word in English). The card also warns that the Hugging Face version has been converted from the NeMo Aligner version, so results may differ slightly from the published scores. For most people, using the model through a hosted service is easier than running it. How those services price open models is covered in our look at [LLM API pricing cuts](/radar/llm-api-pricing-cuts-2026). Open here means the weights can be downloaded under a license with conditions. Our explainer on [open-weight models](/radar/open-weight-models-explained-ibm) covers how that differs from fully open source. ### What are Nemotron 3 Nano, Super and Ultra? Since the 2024 model, Nvidia has moved to a new generation. Its December 15, 2025 announcement describes Nemotron 3 in three sizes, using figures from Nvidia's release. Nano has 30 billion total parameters with up to 3 billion active at a time. Super has about 100 billion total with up to 10 billion active per token. Ultra has about 500 billion total with up to 50 billion active per token. "Active" parameters are the share of the model used for any single piece of text, a design that keeps running costs down. At launch only Nano was available, on Hugging Face and through providers including Baseten, DeepInfra, Fireworks, FriendliAI, OpenRouter and Together AI. Nvidia said Super and Ultra were expected in the first half of 2026. Nvidia's model pages are the place to check whether those two have shipped. Nvidia says Nano delivers up to 4x higher token throughput than Nemotron 2 Nano and cuts reasoning-token generation by up to 60%. The release also says Nvidia is publishing three trillion tokens of training datasets and open-source training libraries (NeMo Gym and NeMo RL), though it does not name a license in that announcement. ### What to watch next The 2024 Nemotron result matters mostly as a method. A strong reward model trained on carefully labeled human preferences can move a mid-sized open model a long way on judge-scored chat tests. Nvidia published the preference data and a paper, so others can test the claim. Independent evaluations that control for length and style will show whether the same gap to GPT-4o holds. The Nemotron 3 Super and Ultra releases, and how they score on knowledge and reasoning tests, will show whether the method carries over to the new generation. Nvidia's decision to keep publishing training data matters because that is what lets outsiders check the work. ## Frequently asked questions ### What is Nvidia Nemotron? Nvidia Nemotron is Nvidia's line of open language models and training tools. The 2024 headline model, Llama-3.1-Nemotron-70B-Instruct, is Meta's Llama 3.1 70B retrained with reinforcement learning from human feedback. In December 2025 Nvidia announced Nemotron 3, a newer family in Nano, Super and Ultra sizes that it also calls open. ### Did Nemotron really beat GPT-4o? On three AI-judged chat benchmarks, the model card reports it did: Arena Hard 85.0 against 79.3, and MT-Bench 8.98 against 8.74. On AlpacaEval 2 the two are 57.6 and 57.5, within the margin of error. The model card says it was not tuned for specialized domains such as math, so the wins are narrow. ### What is a reward model in AI? A reward model is an AI that scores another model's answers. Developers train it on human judgments of which answers are better, then use its scores as a training signal in reinforcement learning from human feedback. The main model is adjusted over many rounds toward replies the reward model rates higher, which is how Nvidia tuned its Llama-based model. ### Can I run Nvidia Nemotron for free? The 70B model's weights can be downloaded under the NVIDIA Open Model License and the Llama 3.1 Community License, so there is no per-token fee when you host it yourself. The model card says it needs two or more 80GB GPUs and at least 150GB of disk space, so the hardware is the real cost for most people. ## Source Based on Nvidia's model card and paper. https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Instruct-HF --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/how-nvidias-nemotron-outsmarted-gpt4-secret-sauce-ai-tuning. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/inside-grok-bot-xai-colossus-supercomputer-real-time-ai - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T09:12:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://nvidianews.nvidia.com/news/spectrum-x-ethernet-networking-xai-colossus # xAI Colossus: What the Memphis Supercomputer Really Is **Short answer:** xAI built Colossus in Memphis with 100,000 Nvidia Hopper GPUs in about 122 days to train Grok. It has since drawn permit disputes over its gas turbines, and reports say Anthropic now uses all of Colossus 1. **Takeaway:** Colossus is a real, fast-built training cluster, but the real-time Grok answers come from search tools, and most performance claims still come from the companies selling the hardware. **Key figures (as reported by the source):** - GPUs at launch: 100,000 Nvidia Hopper GPUs - Build time: About 122 days - Network throughput (NVIDIA claim): 95% with Spectrum-X xAI Colossus is the AI training supercomputer that xAI built in a former factory in Memphis, Tennessee, to train its Grok chatbots. [NVIDIA's announcement](https://nvidianews.nvidia.com/news/spectrum-x-ethernet-networking-xai-colossus) from October 2024 put the first version at 100,000 Nvidia Hopper GPUs (graphics processing units, the chips that do most of the math in AI training) and said it went up in about 122 days. Since then the story has grown past Grok. xAI merged into SpaceX, Colossus 1 has reportedly been rented out to Anthropic, and the site has drawn a long fight over the gas turbines that power it. ### What is xAI Colossus, and what does it do? Colossus is a data center, a building full of servers, built for one job: training large language models, the software behind chatbots like Grok. Training means running a model over enormous amounts of text again and again until it gets good at predicting the next word. That takes thousands of chips working on the same problem at once, wired together into a cluster, a group of computers that work as one machine. According to NVIDIA, Colossus was built to train xAI's Grok models, which are offered to X Premium subscribers. [Supermicro](https://www.supermicro.com/en/featured/xai-colossus), which supplied the servers, describes it as the largest liquid-cooled AI cluster in the world and says it uses 100,000 NVIDIA Hopper GPUs with direct-to-chip liquid cooling. Liquid cooling pipes coolant right to the chips, which is how servers this dense stay at a safe temperature. Colossus is where the model is made. NVIDIA's release describes it as a training system, and the chatbot you talk to is the finished product of that training. ### How was xAI Colossus built in 122 days? NVIDIA's release says the build took about 122 days from start to finish and that training began 19 days after the first server rack arrived. The building was an old Electrolux factory in South Memphis, so xAI did not have to pour a new shell. The companies have not published how xAI managed that pace or a construction timeline beyond the 122-day and 19-day figures. The documented result is a cluster of 100,000 Hopper GPUs, liquid-cooled and linked by one Ethernet network, running training within weeks of the first rack arriving. The 122-day figure is the number NVIDIA and xAI gave for the first 100,000 GPUs; larger and later expansions took different amounts of time. NVIDIA also said xAI planned to double the cluster to 200,000 Hopper GPUs. An Inside Deep Tech guide that cites xAI's own page says that doubling took 92 days. ![Rows of liquid-cooled AI servers inside a large data center hall with technicians walking the aisle](/uploads/blog/inside-grok-bot-xai-colossus-supercomputer-real-time-ai/gemini-supporting.jpg) ### How does the network keep 100,000 GPUs working together? The chips also have to swap data constantly, and if the network is slow, expensive GPUs sit idle waiting for each other. Colossus uses NVIDIA's Spectrum-X Ethernet, built on Spectrum SN5600 switches and BlueField-3 SuperNICs (network cards that offload traffic handling from the main processor). NVIDIA says the system saw no application latency degradation or packet loss from flow collisions, which happen when many data streams crowd the same network path. NVIDIA also reports 95% data throughput with its congestion control, compared with about 60% for standard Ethernet at this scale. That is a vendor claim from the company selling the network. The network is what lets 100,000 chips work together as one training system. For readers who want more on the physics of moving data between chips, the [co-packaged optics explainer](/radar/co-packaged-optics-silicon-photonics-gpu-cluster-bandwidth-wall) covers why networking has become the bottleneck for clusters like this. ### Does Grok really get real-time data from X? Partly. The training run in Memphis produces a fixed model with a knowledge cutoff, like every other chatbot. Fresh information comes from search tools that Grok calls when it answers. xAI's developer documentation describes an [X Search tool](https://docs.x.ai/developers/tools/x-search) that lets Grok run keyword search, semantic search, user search and thread fetch on X, then cite the posts it found. The docs list a price of $5 per 1,000 posts fetched, on top of normal token costs, and say developers switch it on by adding it to a request. Search at answer time is a separate step from training, so the quality of a live answer depends on what the search finds as well as on the model built in Memphis. So "Grok knows what happened thirty seconds ago" is only true when that search runs and returns something useful. ### Why is Colossus controversial in Memphis? An Inside Deep Tech guide citing Memphis Light, Gas and Water says the utility was supplying 150 megawatts to the site in May 2025, with another 150 requested. xAI also brought in on-site gas turbines, which residents and regulators have argued about. [Inside Climate News](https://insideclimatenews.org/news/17072025/elon-musk-xai-data-center-gas-turbines-memphis/) reported in July 2025 that residents and activists said Colossus ran turbines without the right permits. According to reporting summarized by Global Energy Monitor, the Shelby County Health Department then issued a permit for up to 15 turbines producing about 247.2 megawatts. Aerial imagery from April 2025 reportedly showed about 35 turbines on site, more than the permit covered at the time. A permit sets a legal ceiling on output, which is why the gap between the permitted 15 turbines and the roughly 35 seen in imagery drew attention. The Southern Environmental Law Center and the NAACP pushed back, and residents of nearby neighborhoods raised health concerns. xAI has described the turbines as backup power. Critics dispute that and have called for stricter permitting. Exact turbine counts and legal outcomes have kept changing, so check current local reporting before quoting any figure. ![Gas turbines and power equipment beside an industrial building at dusk with a residential street in the foreground](/uploads/blog/inside-grok-bot-xai-colossus-supercomputer-real-time-ai/gemini-supporting.jpg) ### What is Colossus 2, and who uses Colossus 1 now? Colossus 2 is the next-generation site, spread across the Tennessee and Mississippi border near Memphis and built around newer Nvidia GB200 and GB300 chips. Epoch AI, an independent research group, estimates it at about 1.11 million H100-equivalents on 946 megawatts of IT power. An H100-equivalent expresses newer chips in units of one Nvidia H100 so that different clusters can be compared. Colossus 1 has a new tenant. News coverage from CNBC, Axios and Data Center Dynamics reported on May 6, 2026 that Anthropic agreed to use all of Colossus 1's capacity, described as more than 300 megawatts and over 220,000 Nvidia GPUs. Anthropic said the capacity would help its paid Claude Pro and Max subscribers. One outlet reported a price of $1.25 billion per month, but that comes from a single source and is unconfirmed. ### What to watch next Whether Colossus 2 delivers a better Grok is the main question. Bigger clusters do not guarantee better models, and xAI has to show results. The Anthropic deal also raises the question of how much capacity xAI wants to keep for itself. The power fight continues. The same problem shows up across the industry, and the [Anthropic IPO numbers piece](/radar/anthropic-ipo-prospectus-five-numbers) shows how much money now rides on getting compute online. Court rulings and permit decisions in Tennessee and Mississippi could change how fast sites like this get built. Independent measurements would help. Almost every performance claim about Colossus comes from the companies that sell or run it. Third-party figures from groups like Epoch AI are worth more. Grok's search tools are also worth following. If real-time answers are what set Grok apart from rivals, the X Search tool and its pricing will matter more than the chip count. ## Frequently asked questions ### What is xAI Colossus? It is an AI training data center in Memphis, Tennessee, built in a former factory to train the Grok language models. NVIDIA's October 2024 announcement described the first version as 100,000 Hopper GPUs linked by Spectrum-X Ethernet, built in about 122 days. Supermicro supplied the liquid-cooled servers. ### Where is xAI Colossus located? The original site is in South Memphis, Tennessee, in a former Electrolux factory. Reporting places the second site, Colossus 2, across the Tennessee and Mississippi border near Memphis. The power setup at these sites, especially the on-site gas turbines, has been disputed by local residents and environmental groups. ### How many GPUs does xAI Colossus have? NVIDIA said the first cluster used 100,000 Hopper GPUs and that xAI planned to double it to 200,000. News coverage in May 2026 describes Colossus 1 as holding over 220,000 Nvidia GPUs. Epoch AI, a third-party group, estimates Colossus 2 at about 1.11 million H100-equivalents. ### Does Grok use real-time data from X? Grok can search X when a search tool is enabled. xAI's documentation describes an X Search tool that runs keyword search, semantic search, user search and thread fetch, and cites the posts it finds. The Colossus training run itself produces a fixed model, so live information comes from search at answer time. ## Source Based on NVIDIA's Colossus announcement. https://nvidianews.nvidia.com/news/spectrum-x-ethernet-networking-xai-colossus --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/inside-grok-bot-xai-colossus-supercomputer-real-time-ai. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/claude-duet-human-ai-pair-programming-artifacts-co-creation - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T09:10:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ # AI Pair Programming: What Studies Show About Claude and Copilot **Short answer:** AI pair programming now spans chat tools, editors and agents like Claude Code. Controlled studies disagree on speed: one found 55.8% faster, another found experts 19% slower. Newer data is promising but weak. **Takeaway:** AI pair programming can help, but the best evidence shows the gain depends on the task and the developer, so measure your own results and review every change. **Key figures (as reported by the source):** - Copilot trial: time saved: 55.8% faster (95% CI 21% to 89%) - METR 2025: experienced developers: 19% longer with AI - METR 2025: developers' own belief afterward: AI sped them up by 20% AI pair programming means working on code with an AI model the way you would with a human partner: you describe the goal, the AI proposes code, and you review, redirect and decide. Tools such as GitHub Copilot, Cursor and Claude Code all work this way, and Anthropic's Artifacts feature adds a visible side panel where the result appears while you talk. ### What Is AI Pair Programming? Pair programming is an old practice: two developers share one task, one writing ("driving") while the other reviews and thinks ahead. AI pair programming swaps the second person for a language model, which is software trained on large amounts of text and code to predict useful next words. In practice the model plays several roles. It can write a first draft from a plain-language request, explain code you did not write, or suggest the next few lines as you type. The human still owns the decisions: what to build, whether the result is right, and when to throw it away. Anthropic describes Claude Code as "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools." "Agentic" means the AI takes several steps on its own, such as editing files and running tests, instead of answering one question at a time. According to [Anthropic's Claude Code documentation](https://code.claude.com/docs/en/overview), it runs in a terminal, in code editors such as VS Code and JetBrains, in a desktop app, and in a web browser. In practice that means you can ask it to change a file, run the tests and report back, then review what it did. ### Pairing With AI vs Vibe Coding: What Is the Difference? "Vibe coding" is a casual label for describing an app in plain words and accepting whatever the AI produces, often without reading the code. The difference is review. In a pair, the human reads the proposed changes, runs the program and pushes back. In vibe coding, the human judges the result only by whether the app seems to work. Both use the same tools, so the line is about habits. A developer who reads each change, runs the tests and rejects weak suggestions is pairing, even with the same tool that someone else uses to generate a whole app unread. For a throwaway prototype, skipping review may be fine. For anything that handles money, personal data or other people's work, someone has to read the code, because the model can write plausible code that is wrong. ![A designer and a developer sitting side by side at a shared desk, pointing at a laptop screen while discussing a project](/uploads/blog/claude-duet-human-ai-pair-programming-artifacts-co-creation/gemini-supporting.jpg) ### What Are Claude Artifacts, and Where Do They Fit? Artifacts are the visible part of the pairing in Claude's chat app. According to [Anthropic's help center](https://support.claude.com/en/articles/9487310-what-are-artifacts-and-how-do-i-use-them), an artifact is "anything Claude makes for you that you'd put in front of someone," and it opens beside your conversation, where you can edit it, return to it and share it. The help center lists documents, code snippets, single-page websites, images, diagrams, dashboards and small interactive tools as artifact types. You can ask Claude in chat to change one, and in template artifacts such as documents you can edit directly. Artifacts start private to you. This matters for non-programmers because the output is something you can look at and click. You can say "make the button green" and see the change. Because the artifact sits next to the chat, each request and its result stay visible together, so it is easy to see what changed after every instruction. Availability depends on the plan. The help center says creating artifacts in a chat works on Free, Pro, Max, Team and Enterprise plans, and that the feature requires "Cloud code execution and file creation" to be switched on in settings. Artifacts also work inside Claude Code on every plan that includes it. ### Does Pairing With AI Actually Make Developers Faster? Two studies measure speed. One is a randomized experiment by researchers at Microsoft Research, GitHub and MIT. According to [the paper on GitHub Copilot](https://arxiv.org/abs/2302.06590), developers asked to write an HTTP server in JavaScript finished 55.8% faster with Copilot than a control group, with a 95% confidence interval of 21% to 89%. That is a wide range, and it covers one task. The other is from METR, a nonprofit that evaluates AI systems. In [its 2025 study](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/), 16 experienced open-source developers worked on 246 real issues in large codebases they knew well. With AI allowed, they took 19% longer. They had expected a 24% speedup, and afterward still believed AI had sped them up by 20%. The Copilot test used a fresh, self-contained task. The METR test used mature projects where the developers already knew the code. AI help seems to matter more when you are new to the task or the code, and less when you are the expert. Some of the gap between the studies may come from what counts as work. In the METR update, developers using newer agentic tools chose different kinds of tasks than before, so differences in time do not map cleanly onto differences in value. The Copilot paper measured one clearly defined task with a clear finish line, which is easier to time than open-ended project work. ### Has the Picture Changed Since Early 2025? METR followed up. In [its February 2026 update](https://metr.org/blog/2026-02-24-uplift-update/), it studied 57 developers, 143 repositories and more than 800 tasks with late-2025 tools. For the original developers it estimated an 18% speedup (confidence interval from 38% faster to 9% slower). For newly recruited developers it estimated a 4% speedup (confidence interval from 15% faster to 9% slower). Both estimated ranges include zero, so the data cannot rule out no change at all. METR itself called these unreliable signals. Many developers declined to take part because they did not want to work without AI, and some withheld tasks they did not want to do without it. METR says this likely biases the estimate downward, and calls the data "very weak evidence" for the size of any increase. The early-2025 slowdown probably no longer describes current tools, but nobody has yet measured the new speedup well. ![A developer reviewing printed code line by line with a pen at a quiet workspace](/uploads/blog/claude-duet-human-ai-pair-programming-artifacts-co-creation/gemini-supporting.jpg) ### How Do You Pair With AI Without Fooling Yourself? The METR result carries a warning: developers felt faster while measuring slower. A few habits help, and they come from the studies' own caveats. Read every change before accepting it. METR noted that developers reported differences in the quality of final work between conditions. Keep tasks small. A request you can verify in a minute is easier to trust than a request that rewrites a dozen files. Write down the project rules. Claude Code reads a file called CLAUDE.md at the start of every session, where you can set coding standards and preferred libraries, according to Anthropic's documentation. Other tools have similar settings. Measure your own work. Time yourself on a few tasks with and without the AI. That is the same method the studies used, at a small scale. Each of these habits puts a check between the model's output and your project, so a wrong answer is caught while it is still cheap to fix. For related background on how AI tools connect to other software, see our explainers on [computer-use AI agents](/radar/computer-use-ai-agents-operating-systems-claude-operator) and the Model Context Protocol, the open standard Anthropic's documentation says lets Claude Code read documents and update tickets in other tools. ### What to Watch Next METR said its late-2025 data was weak, so cleaner studies are the first thing to follow. Look for results on tasks of different sizes, since the two studies above disagree partly because they tested different work. On the Artifacts side, Anthropic's help center lists templates still marked beta, including Dashboards and Motion, and sharing and data-storage features are limited to paid plans. ## Frequently asked questions ### What is AI pair programming? AI pair programming means building software alongside an AI model that drafts, explains and edits code while you set the goal and review the results. Tools include GitHub Copilot, Cursor and Claude Code. The human stays responsible for decisions and for checking that the code is correct before using it. ### Is AI pair programming the same as vibe coding? No. Both use the same AI tools, but pairing means the human reads proposed changes, runs the program and pushes back. Vibe coding usually means accepting whatever the AI produces based on whether the app seems to work. The difference is the level of review, which matters more as projects handle real data. ### What are Claude Artifacts? According to Anthropic's help center, an artifact is anything Claude makes that you would put in front of someone, such as a document, a single-page website, a dashboard or a small interactive tool. It opens beside the chat so you can edit it and share it. Artifacts start private to you. ### Does AI make programmers faster? Evidence is mixed. A Microsoft Research, GitHub and MIT trial found developers finished one task 55.8% faster with Copilot. METR found experienced open-source developers took 19% longer in 2025. METR's 2026 update suggested a speedup with newer tools but called its own data very weak evidence. ## Source Based on METR and Microsoft Research studies. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/claude-duet-human-ai-pair-programming-artifacts-co-creation. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/death-of-slow-charging-solid-state-sodium-ion-batteries - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T09:05:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://liveinthefuture.org/stories/solid-state-battery-scorecard-2026 # Solid State Battery: Where the Technology Really Stands in 2026 **Short answer:** No automaker has shipped a fully solid-state battery in a customer car yet. Toyota, Samsung SDI and QuantumScape target 2027 to 2028, while CATL's sodium-ion pack heads into a mass-market car in China. **Takeaway:** Solid state batteries have promising lab numbers but no customer cars yet, so judge them by 2027 production results. **Key figures (as reported by the source):** - QuantumScape B-sample charge, 10 to 80%: 12.2 minutes - Samsung SDI claimed density vs its P5 cell: 900 Wh/L, about 40% higher - CATL Naxtra sodium-ion density: 175 Wh/kg A solid state battery swaps the liquid inside a lithium-ion cell for a solid material, and as of late October 2026 no automaker has put one in a customer's car. Toyota, Samsung SDI and QuantumScape all point to 2027 or 2028 for first production, according to a [2026 industry scorecard](https://liveinthefuture.org/stories/solid-state-battery-scorecard-2026). Meanwhile a different chemistry, sodium-ion, is already heading into a mass-market car. ### What is a solid state battery? Every battery has two electrodes and a material between them that lets charged atoms (ions) travel back and forth. That material is the electrolyte. In today's lithium-ion cells it is a flammable liquid. In a solid state battery it is a solid, such as a ceramic, a glass-like material or a sulfide compound. The solid layer is meant to do two jobs. It can allow a lithium metal electrode, which stores more energy than the graphite used today, and it can resist the thin metal spikes called dendrites that grow during charging and can short-circuit a cell. Researchers still disagree about how well solid electrolytes block dendrites. A University of Maryland team reported that two leading solid electrolyte materials, known as LLZO and LPS, can facilitate dendrite growth rather than stop it, according to a DOE-funded project announcement. The appeal is energy density, meaning how much energy fits in a given volume or weight. Samsung SDI, for example, claims 900 watt-hours per litre for its oxide-based cells, about 40% above its own current P5 lithium-ion cell, according to the scorecard. That is a company claim for prototype cells. QuantumScape's B-sample cells reach 844 watt-hours per litre in the same scorecard. Both numbers describe cells shipped to partners for testing. A finished battery pack adds casing, cooling and wiring, so pack-level figures come out lower than cell-level ones. ### How fast can a solid state battery charge? The headline numbers come from two sources. QuantumScape reports its B-sample cells charge from 10% to 80% in 12.2 minutes, and Toyota claims 10% to 80% in 10 minutes, per the same scorecard. Toyota's figure is a projection, since the cells are not in customer cars. Fast charging is hard on any battery because high current pushes lithium ions hard against the electrode. That is where dendrites tend to form. A solid layer that stays stable under high current would allow quicker charging, which is why the claims cluster around the ten-minute mark. A ten-minute charge in a test cell says little about years of daily charging in a real car, and that kind of track record does not exist yet. ### Is a solid state battery safer than lithium-ion? Often yes, with an important exception. Sandia National Laboratories researchers published a [study in Joule in March 2022](https://www.sciencedaily.com/releases/2022/03/220307113014.htm) comparing the heat released by lithium-ion cells and two solid state designs. The design containing a little liquid electrolyte released about one-fifth the heat of a comparable lithium-ion battery in fire exposure, and the fully solid version produced no heat in that test. The same study found that dendrite short circuits released similar heat in all three types, depending on how much lithium metal was present. It also reported that a short circuit in the all-solid cell could release a dangerous amount of heat. So the idea that solid-state cells carry no fire risk overstates the evidence. A solid electrolyte removes one fuel source, the flammable liquid, while leaving other failure routes in place. ![A technician in protective glasses holding a small prototype battery cell in a clean laboratory](/uploads/blog/death-of-slow-charging-solid-state-sodium-ion-batteries/gemini-supporting.jpg) ### When will a solid state battery reach cars? Company targets point to 2027 and 2028, and the track record of those targets is poor. The scorecard lists Toyota as targeting mass production in 2027 with first cars in 2027 to 2028, after earlier promises of 2020 and then 2025. It adds that Toyota received production approval in Japan in October 2025. Samsung SDI also targets 2027, from a roadmap that once showed prototypes in 2025. QuantumScape targets field tests in Volkswagen vehicles in 2026 and mass production in 2027 to 2028, versus a 2024 projection in a 2020 investor presentation. QuantumScape's PowerCo license covers up to 40 gigawatt-hours a year, the scorecard says. The scorecard counts zero all-solid-state cells in customer cars so far and says the step from lab to factory has added two to four years each time. Its listed hurdles include resistance where the electrolyte meets the electrodes, dendrites under high current and factory conditions with moisture below one part per million. ### Why is it so expensive? The scorecard cites analyst estimates (BloombergNEF and IDTechEx) of $400 to $600 per kilowatt-hour for solid state cells, against roughly $80 to $100 for lithium-ion. It reports no expected price parity before 2032 to 2035. The same piece says Toyota plans its first cars for Lexus flagships, at an estimated 800,000 to 1,000,000 yuan (about $110,000 to $137,000). Lithium-ion got cheap by being made in huge quantities, and solid state has not yet had that chance. ### Solid state battery vs sodium ion battery: which is closer? Sodium-ion is closer to the road. CATL and Changan announced the Changan Nevo A06 with CATL's Naxtra sodium-ion pack, described by [New Atlas](https://newatlas.com/automotive/catl-naxtra-sodium-ion-changan-nevo-a06/) as the first mass-production EV using sodium-ion batteries. It is due in mid-2026, initially in China. Sodium is a cheaper, more abundant material than lithium. According to the report, the chemistry avoids cobalt and nickel and uses aluminum foil instead of copper. The trade-off is energy density: Naxtra is reported at 175 Wh/kg, which CATL says is the highest in mass production for sodium, and which still trails lithium iron phosphate (LFP), a common lower-cost lithium chemistry. CATL also reports that the pack keeps over 90% of its capacity at -40 °C and that crush, puncture and saw tests produced no smoke or flames. The A06 is reported at over 400 km of range. These are CATL's claims and have not been independently verified. Sodium-ion targets cheaper, cold-tolerant packs, while solid state targets longer range and faster charging at higher cost. CATL's sodium grid-storage unit is rated at about 160 Wh/kg and over 15,000 cycles to 80% capacity, with deployment planned before the end of 2026. ![A row of electric cars plugged in at a winter charging station at dusk](/uploads/blog/death-of-slow-charging-solid-state-sodium-ion-batteries/gemini-supporting.jpg) ### What about semi-solid batteries? Semi-solid cells, which use a gel or reduced-liquid electrolyte, are the part of this story already on roads. NIO has shipped a 150 kWh semi-solid pack since 2024, with about 360 Wh/kg reported, and an ET7 drove 1,070 km on one in April 2024, according to the scorecard. It says "thousands" are in use, available through swap stations in China and select European markets. CATL targets semi-solid mass production for late 2026 and claims 500 Wh/kg for those cells. The scorecard lists no customer cars with that cell yet. Semi-solid is the likelier first taste of the technology for most drivers. It also muddies the labels, because marketing may call a gel cell "solid state" when the cell still contains liquid components. ### What to watch next Samsung SDI's mass production start in 2027 is the first marker. Toyota putting an all-solid-state car in customer hands, and at what price, is the second. QuantumScape's Volkswagen field tests will show how the cells behave over repeated charging. The Changan Nevo A06 will show whether the range and cold-weather behavior CATL claims hold up on the road. For readers shopping for a car now, the practical advice is to buy on today's lithium-ion or LFP specifications, because nothing sold today depends on solid state arriving. ## Frequently asked questions ### What is a solid state battery? A solid state battery is a rechargeable battery that uses a solid electrolyte, the material that carries ions between the electrodes, instead of the flammable liquid in today's lithium-ion cells. Makers hope this allows more stored energy, faster charging and better safety, but the technology is still at pilot stage and no customer car uses a fully solid cell yet. ### When will solid state batteries be in cars? Toyota, Samsung SDI and QuantumScape each target mass production in 2027, with first cars in 2027 to 2028, according to a 2026 industry scorecard. Earlier targets slipped repeatedly, and the scorecard counts no all-solid-state cells in customer cars so far. Treat 2027 as a company goal. ### Are solid state batteries safer than lithium-ion? Often, but not always. A 2022 Sandia National Laboratories study in Joule found a solid state design with a little liquid released about one-fifth the heat of a comparable lithium-ion battery in fire exposure. The same study found short circuits from lithium dendrites released similar heat in all types, and a fully solid cell could still release dangerous heat. ### Is a sodium-ion battery the same as a solid state battery? No. Sodium-ion batteries replace lithium with sodium, while solid state batteries replace the liquid electrolyte with a solid. CATL's Naxtra sodium-ion cell is reported at 175 Wh/kg and is headed into the Changan Nevo A06 in China, according to New Atlas. It targets low cost and cold weather. ## Source Based on a 2026 industry scorecard and Sandia's Joule study. https://liveinthefuture.org/stories/solid-state-battery-scorecard-2026 --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/death-of-slow-charging-solid-state-sodium-ion-batteries. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/ai-video-generation-hollywood-revolution-sora-kling-creators - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T09:00:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://techcrunch.com/2026/03/24/openais-sora-was-the-creepiest-app-on-your-phone-now-its-shutting-down/ # AI Video Generation in 2026: Sora Is Gone, Kling 4.0 Is Here **Short answer:** OpenAI announced on 24 March 2026 that it was shutting down Sora, and Kuaishou launched Kling 4.0 on 15 October 2026. Creators still stitch short AI clips together to make longer films. **Takeaway:** AI video tools can produce convincing short clips, but longer films still depend on stitching shots together and on human editing, and the market is shifting fast. **Key figures (as reported by the source):** - Sora shutdown announced: 24 March 2026 - Kling 4.0 single-pass length: Up to 30 seconds - Sora downloads, Nov to Feb (Appfigures): 3.33M to 1.13M AI video generation can now turn a written prompt into a short, realistic clip, but no tool makes a finished feature film in one go. The field also looks different from a year ago: OpenAI announced on 24 March 2026 that it was shutting down Sora, its video app, and Kuaishou's Kling launched a 4.0 model on 15 October 2026. Creators who want to make films with AI today stitch many short clips together and edit them like ordinary footage. ### Is Sora still available? Sora is no longer a consumer app. [TechCrunch reported](https://techcrunch.com/2026/03/24/openais-sora-was-the-creepiest-app-on-your-phone-now-its-shutting-down/) that OpenAI announced the shutdown of the Sora app on 24 March 2026, about six months after launch, and gave no reason. The company said it would share timelines for the app and its API later. According to TechCrunch, citing the analytics firm Appfigures, Sora downloads peaked at about 3.33 million in November and fell to about 1.13 million by February. Appfigures estimated that the app earned about $2.1 million from in-app credit purchases over its lifetime. A planned Disney partnership went away with it. [Euronews reported](https://www.euronews.com/2026/03/25/openai-to-abruptly-close-sora-video-app-following-backlash-over-deepfakes-and-ai-slop) that the three-year deal would have let Sora users make videos with more than 200 licensed Marvel, Pixar and Star Wars characters. Disney said it respected OpenAI's decision to exit the video generation business and would keep working with AI platforms. For developers, a secondary site reports that OpenAI's Sora API was scheduled to close on 24 September 2026, with no replacement named. OpenAI has said that ChatGPT's image generation is unaffected. ### How does AI video generation work? Most current video generators are built on a design called a diffusion transformer. A diffusion model starts from random noise and removes it step by step until a picture or clip emerges that matches the text prompt. A transformer is the type of neural network that lets the model weigh how every part of the clip relates to every other part. [Wikipedia's summary of Kling](https://en.wikipedia.org/wiki/Kling_AI) says Kling uses this kind of architecture, plus a self-developed 3D variational autoencoder. That component compresses video across both space and time, so the model works on a small summary of the clip instead of every raw pixel. The model does not film anything. It predicts what plausible footage looks like, based on patterns learned from huge amounts of training video. That is why results can look convincing and still break physical rules, such as a glass that does not shatter the way glass should. Companies sometimes describe this as learning a "world model", an internal sense of how objects and scenes behave. The label is an aspiration. No independent benchmark shows that any generator reliably simulates physics. ![A filmmaker reviewing a short film sequence on a monitor in a small home editing studio](/uploads/blog/ai-video-generation-hollywood-revolution-sora-kling-creators/gemini-supporting.jpg) ### What does Kling 4.0 add? Kling 4.0 is the newest model from the Chinese short-video company Kuaishou. According to an [S&P Capital IQ article republished by MarketScreener](https://www.marketscreener.com/news/kuaishou-technology-launches-next-generation-kling-4-0-video-generation-model-of-kling-ai-ce785cdade8dff24), it launched on 15 October 2026 and generates up to 30 seconds of video in a single pass. Longer extensions of up to two minutes were described as coming soon. The same report describes a feature called Omni Reference. It lets a creator supply up to 15 reference items, such as images, video clips and voices, so that a character or style stays consistent. It also accepts up to 10 keyframes, which are images that mark where a shot should pass through key moments. Kling 4.0 is also connected to Kling's MCP and command-line tools, which means AI agents can call it automatically. Kling is not new to this. Wikipedia lists its first public beta as 10 June 2024, with a 3.0 release in early 2026. Almost all Kling 4.0 specifications come from Kuaishou's own materials. ### Can AI make a whole movie? AI can contribute to a whole movie, but no documented tool produces a feature-length film in one generation. Comparison guides disagree about per-clip limits for Veo, Runway and Kling, and they range from a few seconds to roughly half a minute for typical single generations. Those guides are mostly vendor-comparison blogs, so the figures are best treated as rough. Making a longer film therefore means generating hundreds of separate clips, choosing the best, and keeping characters and settings looking the same from shot to shot. Consistency across shots is the main technical hurdle, which is why reference tools like Kling's Omni Reference exist. Studios are also experimenting, with humans still in charge. Lionsgate announced in September 2024 that it would build a custom model with Runway, trained on its own catalog of more than 20,000 titles and restricted to affiliated filmmakers. OpenAI backed an animated feature called Critterz, announced with a budget under $30 million, a nine-month production schedule and a planned Cannes debut in 2026. Reports describe it as combining AI tools with human voice acting and hand-drawn elements. Whether it premiered as planned is unconfirmed. ![A small crew of animators and a director discussing storyboards around a desk in a bright studio](/uploads/blog/ai-video-generation-hollywood-revolution-sora-kling-creators/gemini-supporting.jpg) ### Who actually benefits from AI video tools? The clearest winners are people who need short, inexpensive footage: advertisers testing ideas, teachers illustrating a lesson, and independent filmmakers drafting a scene before a real shoot. Lionsgate's stated interest in using generated clips to decide whether to shoot a scene points the same way. In those cases the clip stands in for a rough sketch. The risks are also documented. TechCrunch reported that Sora's guardrails against videos of public figures were easy to evade, and that deepfakes of Martin Luther King Jr. and Robin Williams appeared. The daughters of both men asked users to stop. Users also made videos with copyrighted characters. Euronews noted that critics warned open-ended prompts enabled nonconsensual images, realistic deepfakes and low-quality "AI slop." OpenAI limited content featuring public figures only after pushback from family estates and an actors' union. Control over content varies by provider. Wikipedia says Kling follows strict content rules aligned with Chinese government regulations, and that it blocks topics such as politics and protests. It also records that fake Kling websites and ads were used in May 2025 to spread infostealer malware, so creators should reach the tool only through its official site. ### How should a creator approach AI video tools today? The evidence supports a modest approach. Start with short clips for one purpose, such as a storyboard test, a social post or a lesson illustration, and keep the edit in ordinary editing software so no single vendor holds the project. Keep original files and prompts for every clip, since a service can change or close. Sora is the example: users were told to export their content because it was going away. Check the rules on likenesses and copyrighted characters before publishing anything. The Sora reporting shows that guardrails can fail, and a clip that breaks a family's wishes or a studio's rights lands on the creator. ### What to watch next Whether any tool can deliver reliable character consistency across a whole film is the open question. Watch for independent tests of Kling 4.0 once more people have used it, and for the two-minute extension that Kuaishou says is coming. Also watch how studios handle licensing. The Disney and OpenAI deal collapsed along with Sora, yet Disney said it will keep talking to AI platforms. Related world-building research is covered in [Google's Genie 4 explainer](/radar/google-genie-4-world-model-explained), which looks at models that generate interactive scenes instead of fixed clips. Finally, check each tool's current documentation before building a workflow around it, since Sora's exit shows how quickly an AI video service can disappear. ## Frequently asked questions ### Is Sora still available? No. OpenAI announced on 24 March 2026 that it was shutting down the Sora app and gave no reason. A secondary site reports the Sora API was scheduled to close on 24 September 2026. OpenAI says ChatGPT image generation is unaffected. Check OpenAI's own documentation for the current status of any Sora feature. ### What is Kling 4.0? Kling 4.0 is the latest video model from Kuaishou's Kling AI. According to an S&P Capital IQ report, it launched on 15 October 2026 and can generate up to 30 seconds of video in one pass. It adds an Omni Reference feature for keeping characters and styles consistent, using images, clips and voices as references. ### Can AI make a whole movie? Not in one generation. Creators generate many short clips, pick the best and edit them together, while working to keep characters consistent. Studios such as Lionsgate are testing custom models with Runway, and the OpenAI-backed film Critterz planned a 2026 Cannes debut, but both still involve people. ### How does AI video generation work? Most systems use a diffusion transformer. The model starts with random noise and removes it step by step until a clip matching the text prompt appears. Kling also uses a 3D variational autoencoder that compresses video across space and time. The model predicts plausible footage from training data, so it can still break physical rules. ## Source Based on TechCrunch and Kuaishou reporting. https://techcrunch.com/2026/03/24/openais-sora-was-the-creepiest-app-on-your-phone-now-its-shutting-down/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/ai-video-generation-hollywood-revolution-sora-kling-creators. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/smart-glasses-replace-smartphones-meta-orion-ar-future - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T08:55:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://about.fb.com/news/2024/09/introducing-orion-our-first-true-augmented-reality-glasses/ # Meta Orion Glasses: What the Prototype Can and Cannot Do **Short answer:** Meta's Orion AR glasses are a prototype that will not be sold. The Ray-Ban Display, a simpler product with one small screen, reached Germany on 13 October 2026 at 899 euros and still needs a phone. **Takeaway:** Orion shows what wide-view AR glasses can do, but cost, battery and phone dependence mean today's glasses extend your phone. **Key figures (as reported by the source):** - Orion weight (PPC Land): 98 grams - Orion field of view (PPC Land): About 70 degrees - Ray-Ban Display price in Germany: 899 euros Meta Orion glasses are a working prototype of true augmented reality (AR) eyewear, and Meta has said they will not be sold. The glasses weigh about 98 grams, according to the [PPC Land write-up of the specs](https://ppc.land/orion/), and show a wide display through see-through lenses. They are a research device, so they cannot replace your phone today. A different, much smaller Meta product, the Ray-Ban Display, is the one you can buy now. ### What are Meta Orion glasses? Orion is an AR prototype that Meta unveiled in September 2024. In its [announcement](https://about.fb.com/news/2024/09/introducing-orion-our-first-true-augmented-reality-glasses/), Meta calls it an internal product prototype and says it has the largest field of view in the smallest AR glasses form to date. Augmented reality means digital objects appear overlaid on the real world you see through the lenses, as opposed to virtual reality, which replaces your view entirely. Meta did not turn Orion into a retail product. According to [UploadVR](https://www.uploadvr.com/meta-to-reportedly-offer-orion-ar-glasses-to-developers-2026/), it is going to a small group of developers, and Meta reportedly plans a consumer glasses launch in 2027 under the codename Artemis. That report is from January 2025, so the 2027 date is a reported target. ### How do Meta Orion AR glasses work? Three parts make up the system, and the glasses are only one of them. The glasses carry the lenses, sensors and a battery in the arms. PPC Land reports a field of view of about 70 degrees diagonal, which it calls the widest yet in a glasses form. The lenses use thin plates made of silicon carbide, a material that bends light more strongly than ordinary glass, so one plate can carry a wider picture. A separate compute puck does the heavy processing. It is a pocketable unit with a large battery and a Meta-designed chip, and it streams finished frames to the glasses wirelessly. Keeping the processor out of the frame is how the glasses stay light. The cost is one more object to carry. The third part is a wristband. It uses electromyography (EMG), which means it reads the small electrical signals in your forearm muscles when your fingers move. PPC Land stresses that it reads muscle activity and not brain signals, whatever the "neural" label suggests. You pinch or swipe with your hand at your side to control what you see. PPC Land adds that Magic Leap 2, a headset, matches the 70 degree figure, while Snap's SPECS claims 51 degrees. Orion's achievement is fitting that width into something shaped like glasses. ![A person in a cafe wearing thin black-framed glasses and a fabric wristband while looking at a laptop](/uploads/blog/smart-glasses-replace-smartphones-meta-orion-ar-future/gemini-supporting.jpg) ### Is the Meta Orion price really $10,000? The figure most often quoted is about $10,000 per unit to make. PPC Land notes that it was first reported by The Verge and that sources disagree on whether Meta confirmed it. UploadVR, quoting Meta, says each unit would cost around $10,000 to make. Because Orion is not for sale, that is a manufacturing cost. PPC Land also cites a German outlet's estimate that the display accounts for roughly 90 percent of the manufacturing cost. That estimate suggests the screen technology is the expensive piece, and cheaper optics are one of the things Meta says it needs. Its announcement lists a sharper display, a smaller form factor and lower costs through large-scale production as the priorities for future devices. ### Can you buy smart glasses with a display today? Yes, in the form of the Meta Ray-Ban Display, which is a far simpler device than Orion. According to [Telekom](https://www.telekom.com/en/newsroom/latest-updates/media-information/2026/10/launch-telekom-offers-meta-ray-ban-display), which began selling it in Germany on 13 October 2026, it has a 600 by 600 pixel colour screen in the right lens only. That screen shows messages, photos, notifications, live captions, translations, map directions and answers from Meta AI. It comes with a Neural Band, the same EMG idea as Orion's wristband, and the band is not sold separately. Telekom lists the price as 899 euros without a mobile plan. [Neowin's launch coverage](https://www.neowin.net/amp/meta-launches-ray-ban-display-glasses-and-neural-band-for-799/) puts the US price at $799, including the band. PPC Land puts the Ray-Ban Display's field of view at about 20 degrees, against roughly 70 for Orion. One lens, a small window of information and no compute puck make a product you can sell today. Orion's wide, two-lens view is the part that is still too costly. Telekom adds some practical limits. Prescription lenses are fitted only at selected certified opticians, and only single-vision lenses from -4.00 to +4.00 dioptres are supported. People with stronger prescriptions or varifocals cannot use the glasses as sold. ### Will smart glasses replace smartphones? The Ray-Ban Display shows why not yet. Telekom's own page says the Ray-Ban Display needs a compatible smartphone, the Meta AI app, a Meta account and an internet connection. The glasses extend a phone, so the phone stays in your pocket. Battery life is the second limit. Telekom quotes up to six hours for the glasses, with up to 24 more hours from the folding case and up to 18 hours for the Neural Band, according to Meta. Orion, per a report cited by PPC Land, lasts about two hours, and the puck has to stay close. Meta's own roadmap hedges on timing. The company says it plans to begin shipping consumer AR glasses "in the near future" and to release new devices over the next few years. It gives no date and does not claim they will replace phones. ![A hand holding a smartphone beside a pair of thin wire-frame glasses on a wooden table](/uploads/blog/smart-glasses-replace-smartphones-meta-orion-ar-future/gemini-supporting.jpg) ### What would have to change first? Cost is the first problem between Orion and an everyday pair of glasses. A display that makes up most of a $10,000 build cost cannot sit in a mass-market product. Size and power come second. Orion offloads its processor to a puck, and its battery lasts about two hours. The third is the field of view. UploadVR reports that Artemis may be sleeker and lighter than Orion but may have a narrower field of view, so Meta is expected to trade some of Orion's width for a wearable design. The fourth is the phone dependency. Today's display glasses lean on a phone for connectivity and the Meta AI app. For glasses to replace the phone, they would need to carry their own connection, apps and storage in a frame people are willing to wear all day. Our piece on [on-device phone AI](/radar/on-device-phone-ai-local-models-apple-android-privacy) covers the phone side of that shift, and the [iPhone Duo vs Galaxy Z Fold 8 teardown](/radar/iphone-duo-vs-galaxy-z-fold-8-teardown) shows how much hardware engineering a phone still holds. ### What to watch next Watch for the consumer glasses Meta has codenamed Artemis. UploadVR's January 2025 report points to 2027, and Meta has confirmed only "the near future." If a date and price appear, check three numbers against Orion: field of view, battery life and whether a compute puck is still required. Watch the Ray-Ban Display rollout as well, since it is the closest thing to a mass-market test of the idea. It reached the US in September 2025 and Germany in October 2026, with a UK launch reported for September 2026, so the real test is how many people keep wearing it once the novelty fades. Meta has not published sales figures. Finally, watch the weight and the lenses. If Meta can keep Orion's wide picture while holding the frame under 100 grams, glasses start to look like a serious alternative for some tasks. ## Frequently asked questions ### Are Meta Orion glasses available to buy? No. Meta describes Orion as an internal product prototype, and UploadVR reports it is not planned as a retail product. A small group of developers receives units. Meta says it plans to ship consumer AR glasses in the near future, and UploadVR reports a 2027 target under the codename Artemis. ### How much do Meta Orion glasses cost? There is no retail price because Orion is not for sale. UploadVR reports that Meta says each unit would cost around 10,000 dollars to make. PPC Land notes that sources disagree on whether Meta confirmed the figure. The display is estimated to be most of that manufacturing cost. ### What is the difference between Meta Orion and Ray-Ban Display? Orion is a prototype with a field of view of about 70 degrees, two see-through lenses, a compute puck and an EMG wristband, per PPC Land. The Ray-Ban Display is a product with a 600 by 600 pixel screen in one lens, about 20 degrees of view, and an included Neural Band. It needs a compatible phone. ### Will smart glasses replace smartphones? Not with products available now. Telekom says the Ray-Ban Display requires a compatible smartphone, the Meta AI app and an internet connection. Meta has not given a date for consumer AR glasses and describes a path over the next few years, so glasses currently work as an extension of the phone. ## Source Based on Meta's Orion announcement. https://about.fb.com/news/2024/09/introducing-orion-our-first-true-augmented-reality-glasses/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/smart-glasses-replace-smartphones-meta-orion-ar-future. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/computer-use-ai-agents-operating-systems-claude-operator - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T08:35:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://www.anthropic.com/news/3-5-models-and-computer-use # Computer Use Agent: What AI That Clicks and Types Can Do **Short answer:** Anthropic (October 2024) and OpenAI (January 2025) released AI agents that operate a computer by looking at screenshots, clicking and typing. A 2026 benchmark of long tasks shows they still finish only a fraction of them. **Takeaway:** Computer use agents handle short, well-defined screen tasks but still fail most long workflows, so human oversight remains part of every serious deployment. **Key figures (as reported by the source):** - OpenAI CUA score on OSWorld: 38.1% - Human score on OSWorld: 72.4% - Best completion rate on OSWorld 2.0: 20.6% A computer use agent is an AI system that operates a computer the way a person does: it looks at the screen, moves the cursor, clicks buttons and types text. Anthropic introduced the idea to developers in October 2024, and OpenAI followed in January 2025 with Operator. Both companies describe their systems as early and error-prone, and newer benchmark results back that up. ### What is a computer use agent? It is software that takes a goal in plain language, such as "save a picture of a cat to my desktop", and carries it out by using a normal graphical interface. It does not need a special connection to the program it is using. Anthropic's [October 2024 announcement](https://www.anthropic.com/news/3-5-models-and-computer-use) described its version as letting developers direct Claude to use computers "by looking at a screen, moving a cursor, clicking buttons, and typing text." The same post called the feature experimental and, at times, "cumbersome and error-prone." OpenAI's version powers Operator, launched on January 23, 2025. According to [The Decoder's coverage](https://the-decoder.com/openais-operator-and-computer-using-agent-bring-autonomous-ai-agents-closer-to-reality/), it uses a model called Computer-Using Agent (CUA), which combines the vision of GPT-4o with reasoning skills trained through reinforcement learning (a method where a model improves by being rewarded for good outcomes). At launch it was open only to US ChatGPT Pro subscribers. ### How does it work? It runs in a loop. It takes a screenshot, decides what to do next, performs one action, then takes another screenshot to see what changed. OpenAI's model follows three phases at each step, as The Decoder describes them: capture a screenshot, reason about the next action, then act by clicking, scrolling or typing. It works from what is on the screen, without special integrations with the websites it visits. Anthropic's [developer documentation](https://platform.claude.com/docs/en/docs/agents-and-tools/tool-use/computer-use-tool) shows the mechanics from the builder's side. The model asks for an action, such as a screenshot, a click, a key press or a scroll. The developer's own software carries it out inside an environment the developer controls, then sends the result back. The loop repeats until the model has no more actions to request. Screenshots are the agent's only view of the world, and they carry a cost. The documentation says each one uses roughly 1,000 to 1,800 input tokens (the units an AI model reads and is billed by). Anthropic's advice for reliable runs is to keep tasks simple, add short pauses so the screen can update, and ask the model to check each step with a screenshot. ![A person at a wooden desk watching a laptop screen while a cursor moves across a web form](/uploads/blog/computer-use-ai-agents-operating-systems-claude-operator/gemini-supporting.jpg) ### How well do these agents actually perform? The standard yardstick is OSWorld, a benchmark of real computer tasks run in a virtual machine (a simulated computer inside a computer). Humans completed 72.4% of its tasks, according to the comparison table in OpenAI's CUA announcement. Anthropic reported that its upgraded Claude 3.5 Sonnet scored 14.9% on OSWorld in the screenshot-only category, against 7.8% for the next-best system. With more steps allowed, it reached 22.0%. OpenAI's CUA then reported 38.1% on OSWorld, 58.1% on WebArena (simulated websites) and 87% on WebVoyager (live sites such as Amazon and Google Maps). These figures are from OpenAI. They show a strong result on simpler live-web tasks and a gap to humans on harder ones: the same table lists humans at 78.2% on WebArena, against the agent's 58.1%. Scores have kept climbing since. A third-party [OSWorld leaderboard](https://leaderboard.steel.dev/leaderboards/osworld/) last updated in April 2026 lists Claude Opus 4.6 at 72.7%, roughly level with the human baseline. The leaderboard notes that self-reported rows are unaudited. ### Is it safe to let an AI control your computer? Its makers say it needs guardrails. Anthropic's documentation recommends four precautions: run the agent in a dedicated virtual machine or container with minimal privileges, keep sensitive data and account logins away from it, limit its internet access to an approved list of domains, and require a human to confirm consequential actions such as accepting terms or making financial transactions. The biggest named risk is prompt injection. That means hidden instructions placed in a web page or image that the model reads and may follow, even when they conflict with what the user asked. Anthropic says it scans screenshots with classifiers and has Claude check whether an instruction really came from the user when one is flagged. It still tells developers to apply the other precautions. OpenAI describes its own controls for Operator. According to The Decoder, users must give permission before actions such as logging in or paying. A "watch mode" lets people observe every step on sensitive sites like email or banking. A separate monitoring model can halt tasks that look suspicious, and a detection system guards against malicious websites. Anthropic's 2024 announcement added a broader worry: computer use may create a new route for threats such as spam, misinformation or fraud. The company said it built classifiers that can identify when computer use is being used and whether harm is occurring. ![A hand resting on a laptop trackpad beside a closed notebook in a quiet office](/uploads/blog/computer-use-ai-agents-operating-systems-claude-operator/gemini-supporting.jpg) ### What do the harder tests show about real office work? The original OSWorld tasks are short. A newer benchmark, [OSWorld 2.0](https://arxiv.org/abs/2606.29537), was built to test long workflows, and its results are far lower. The paper describes 108 long-horizon workflows. Each takes a person a median of about 1.6 hours. For Claude Opus 4.7 with maximum thinking, a task averaged about 318 tool calls, against about 30 in the first OSWorld. The best result the authors report is Claude Opus 4.8, which completed 20.6% of tasks under the primary completion metric at 500 steps, with a 54.8% partial-credit score. GPT-5.5 was more economical with tokens but plateaued near 13%. The authors conclude that current agents are "still far from professional-level computer use." They report that agents struggle most when a task depends on hidden information they must work out, and that they often lose track of constraints, miss details that appear mid-task, guess instead of asking the user, and skip checking their work. The best systems score about 72% on short tasks and 20.6% on long ones. Short, well-defined jobs are within reach, and workflows that take a person most of an afternoon mostly are not. ### Do you need programming skills to use one? No, not to use a finished product. Operator is a consumer feature inside ChatGPT, where the user types a request and watches the agent work in a browser window. Building with Anthropic's tool is a developer job. The documentation describes an API in which the developer writes the loop, runs the actions and supplies the screenshots. Anthropic also notes that for tasks confined to web pages, its separate browser use tool is usually a better fit because it does not need a full desktop environment. For related background on how AI systems connect to outside tools, see how the Model Context Protocol links models to software and [the wider rise of autonomous AI agents](/radar/rise-of-autonomous-ai-agents-multi-agent-workflows). For anyone deciding whether to try one, the published guidance points the same way: start with low-risk tasks, as Anthropic advises, and keep a person in the loop for anything involving money, logins or agreements. ### What to watch next Long-task scores matter most. Vendors tend to publish results on short benchmarks, and the longer ones are where the open questions sit. OSWorld 2.0 gives a harder baseline, and the 20.6% top result is the number to compare against when new models arrive. Audited results matter too. The leaderboard owners already flag self-reported scores as unaudited, so independent verification will matter as vendors publish new claims. The safety record needs watching too. Prompt injection remains an open problem that vendors manage with classifiers and human approval steps, so watch for independent testing of those defenses and for how often products still ask users to confirm actions. ## Frequently asked questions ### What is a computer use agent? A computer use agent is an AI system that operates a computer through its screen, much as a person does. It takes screenshots, decides what to do, then moves the cursor, clicks and types. It needs no special connection to the software it uses, only the same visible interface a human would see. ### How is a computer use agent different from a chatbot? A chatbot returns text. A computer use agent acts on a real interface: it looks at screenshots, clicks buttons and types into forms to finish a task. Anthropic describes the loop as the model requesting actions, the developer's software running them, and the results going back to the model. ### Is it safe to give an AI control of my computer? Its makers urge caution. Anthropic recommends a dedicated virtual machine, keeping sensitive data and logins away from the agent, limiting web access to approved domains and requiring human confirmation for consequential actions. Prompt injection, hidden instructions in web pages or images, is the named risk that these steps are meant to contain. ### How good are computer use agents at real work? They do well on short tasks and poorly on long ones. OpenAI reported 38.1% on OSWorld in 2025. The OSWorld 2.0 paper, built on long workflows that take a person about 1.6 hours, reports a best completion rate of 20.6%, and the authors say agents are far from professional-level computer use. ## Source Based on Anthropic, OpenAI and the OSWorld 2.0 paper. https://www.anthropic.com/news/3-5-models-and-computer-use --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/computer-use-ai-agents-operating-systems-claude-operator. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/on-device-phone-ai-local-models-apple-android-privacy - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T08:30:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://machinelearning.apple.com/research/apple-foundation-models-tech-report-2025 # On-Device AI: What Phones Really Run Offline and Privately **Short answer:** Apple and Google both ship small AI models that run on the phone itself. Each company documents offline use for supported tasks, and each also keeps a cloud route for requests the local model cannot handle. **Takeaway:** On-device AI keeps short tasks local, but both Apple and Google keep a cloud path for bigger requests, so privacy depends on which path a request takes. **Key figures (as reported by the source):** - Apple on-device model size: ~3B parameters - Apple on-device quantization: 2-bit quantization-aware training - Gemini Nano first shipped on: Pixel 8 Pro, December 2023 On-device AI means the model that answers your request runs on the phone in your hand instead of on a company's server. Apple and Google both ship small language models built this way, and both describe them as working without a network connection. Neither company says everything stays local. ### What is on-device AI? On-device AI is a language model stored on your phone and run by your phone's own chips. The prompt (your text, audio or photo) is processed locally, and the answer is produced there too. No round trip to a data center is needed for that step. Apple says its model is optimized for Apple silicon, and Google says AICore uses device hardware to keep latency low. The two big examples are Apple's on-device foundation model and Google's Gemini Nano. In its 2025 technical report ([arXiv](https://arxiv.org/abs/2507.13575)), Apple describes "a ~3B-parameter on-device model optimized for Apple silicon." A parameter is one learned number inside the model, so 3B means roughly three billion of them. Cloud models are generally far larger, which is why phone models are called small. Google describes Gemini Nano as its on-device foundation model. According to [Android's developer documentation](https://developer.android.com/ai/gemini-nano), it lets apps "deliver rich generative AI experiences without needing a network connection or sending data to the cloud." ### How does a phone fit an AI model in its memory? Apple's report names two techniques. The first is quantization-aware training at 2 bits. Quantization stores each model number with fewer bits, which shrinks the file and speeds up the math, and "aware training" means the model is trained while accounting for that rounding so quality holds up. The second is KV-cache sharing. The KV cache is the short-term memory a model builds while reading your prompt, and sharing it between parts of the model saves memory ([Apple Machine Learning Research](https://machinelearning.apple.com/research/apple-foundation-models-tech-report-2025)). Google's side runs through AICore, an Android system service. The documentation says it "leverages device hardware to enable low inference latency and keeps the model up-to-date." Inference means the model producing an answer from a prompt. Both companies tune for dedicated phone silicon, often called a neural engine or NPU, which is a chip block built for the matrix math AI uses. Google's page does not list supported devices, so check your phone's settings or spec sheet. For the chip side of the story, see our explainer on [on-device NPUs and edge computing](/radar/next-gen-silicon-edge-computing-on-device-npus). ![A person holding a smartphone on a train platform, reading a short summary on the screen](https://images.unsplash.com/photo-1550751827-4bd374c3f58b?w=1600&auto=format&fit=crop&q=80) ### What can on-device AI actually do on a phone? Google lists the tasks its ML Kit GenAI APIs cover with Gemini Nano: summarizing articles or conversations as a bulleted list, proofreading short chat messages, rewriting short messages in different tones, describing an image in a short sentence, transcribing spoken audio, and answering a custom text or multimodal prompt. Apple's report adds that its models support more languages, understand images and can call tools. It also describes a Swift framework that gives app developers access to the on-device model, with guided generation and tool calling. The pattern is short, bounded jobs: a short text, a single image or a short audio clip handled on the spot, not a long research session. Google's own documentation points to the limits: Gemini Nano has "on-device constraints such as token limits, device hardware limitations, and model capability boundaries," and "inference speed depends on device hardware." A token is a chunk of text, roughly a short word or word fragment, so a token limit caps how much text the model can handle at once. Apple reports that both of its models match or beat comparably sized open models in public benchmarks and human evaluations. That comparison is against similar small models, so it says little about how a 3B model stacks up against a large cloud assistant. ### Is on-device AI private? Running a prompt locally removes one exposure: the prompt does not have to travel to a server. Google's documentation says on-device generative AI "executes prompts locally, eliminating server calls." In plain terms, local processing keeps the prompt on the phone for that step, and Google says AICore keeps no record of inputs or outputs afterward. Google adds architectural details. AICore does not have direct internet access, and its network requests, including model downloads, go through the open-source Private Compute Services companion app. The page also says AICore "is built to isolate each request and doesn't store any record of the input data or the resulting outputs after processing them." On-device AI covers the model run, but a third-party app may still handle your text before or after it calls the model. Apple's and Google's pages describe only their own system layers. ### Does on-device AI work without internet? For the tasks the local model handles, yes. Google says Gemini Nano works "without needing a network connection," and Apple's model is built to run on the device. When a request is too big for the local model, Apple says its Private Cloud Compute (PCC) system handles requests Apple Intelligence cannot complete on the device, specifically more complex tasks that need larger foundation models. In its [security blog post on PCC](https://security.apple.com/blog/private-cloud-compute/), Apple says the user's device encrypts the request only to PCC servers it has verified, and that "no user data is retained in any form after the response is returned." [MIT Technology Review's 2024 coverage](https://www.technologyreview.com/2024/06/11/1093577/apple-is-promising-personalized-ai-in-a-private-cloud-heres-how-that-will-work/) relays Apple's account that the system tries each task on the device first. The same article reports security experts saying any data that leaves the device becomes more exposed, that Apple's promises stay unproven until outside researchers verify them, and that it was unclear whether users would be told when a request goes to the cloud. Android has its own version. Google's [hybrid inference page](https://developer.android.com/ai/hybrid) describes developer modes, including one that prefers the on-device model and falls back to a cloud model "if the on-device model is unavailable or unsupported for the request," and one that throws an error instead of using the cloud. Which mode an app picks is the developer's decision. ![A traveler using a phone with no signal bars in a quiet airport lounge](https://images.unsplash.com/photo-1550751827-4bd374c3f58b?w=1600&auto=format&fit=crop&q=80) ### Which phones have on-device AI? Gemini Nano began on the Pixel 8 Pro in December 2023, [9to5Google reported](https://9to5google.com/2023/12/06/pixel-8-pro-gemini-nano/), powering Summarize in Recorder and Gboard Smart Reply. Support depends on the chip, the operating system version and, per Google, device hardware in general. The practical way to check is the maker's current feature page for your exact model. When Gemini Nano is unavailable on a device, apps can route requests to the cloud through hybrid inference, per Google's documentation, so an app that "has AI" may be running it remotely on a phone that cannot host the model. ### What to watch next - Independent checks. Apple said it would let researchers inspect PCC software. Published findings from outside security teams would turn marketing claims into evidence. - Clear labels. Whether phones tell you when a request leaves the device is still the user-facing question MIT Technology Review raised in 2024. - Model size creep. Apple's 3B model is small by design. If phone memory grows, the local share of work could grow with it, and cloud fallbacks could shrink. - App behavior. App-level privacy labels and developer choices on fallback modes deserve attention, since the app around the model sets its own rules. For a related look at models you can download and run yourself, see our explainer on [open-weight models](/radar/open-weight-models-explained-ibm). ## Frequently asked questions ### What is on-device AI? On-device AI is a language or vision model that runs on your phone's own chips instead of a remote server. Apple describes a roughly 3-billion-parameter model optimized for Apple silicon, and Google describes Gemini Nano running in Android's AICore service. Both companies say it can work without a network connection for supported tasks. ### Is on-device AI more private than cloud AI? It removes the step of sending the prompt to a server, which Google's documentation says eliminates server calls for local requests. That claim comes from the companies themselves. Privacy also depends on the app using the model and on whether a request falls back to the cloud. ### Can on-device AI work offline? For tasks the local model can handle, yes. Google says Gemini Nano works without a network connection, and Apple's model runs on the device. Larger requests may be sent elsewhere: Apple routes tasks it cannot finish locally to Private Cloud Compute, and Android developers can choose a cloud fallback. ### What are the limits of on-device AI? Google's documentation lists token limits, device hardware limits and model capability boundaries, and notes that speed depends on the device. Local models suit short jobs such as summarizing, proofreading, rewriting and transcription. Complex requests may need a larger cloud model, depending on the app and the phone. ## Source Based on Apple's 2025 tech report and Android docs. https://machinelearning.apple.com/research/apple-foundation-models-tech-report-2025 --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/on-device-phone-ai-local-models-apple-android-privacy. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/humanoid-robots-daily-life-figure-02-tesla-optimus-homes - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T08:25:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://www.figure.ai/news/introducing-figure-03 # Humanoid Robots for Home: What Figure 03, NEO and Optimus Offer **Short answer:** Home humanoid robots are now on sale as reservations, but none works fully alone. 1X NEO takes a $200 deposit with remote human help, Figure 03 has no consumer price, and Tesla has not shown Optimus V3. **Takeaway:** Home humanoids are early, partly supervised products, so judge them on price, privacy and published safety evidence rather than demo videos. **Key figures (as reported by the source):** - Figure 03 fingertip sensing: Forces as small as 3 grams - Figure BotQ initial line capacity: Up to 12,000 robots per year - 1X NEO order deposit: $200 Humanoid robots for home use are real products you can reserve today, but none of them yet does your chores on its own. The most visible one, 1X's NEO, takes a $200 deposit, and its maker says a remote human "Expert" can step in for tasks it has not learned. Figure AI's newest robot, Figure 03, was designed with the home in mind, yet Figure has not published a consumer price or ship date. Tesla's Optimus has not been shown to the public at all. ### What Is a Humanoid Robot for Home Use Today? A humanoid is a robot with a torso, two arms, two hands and two legs, built so it can work in spaces made for people: stairs, door handles, kitchen counters. The idea is that one machine can do many chores without redesigning the house around it. Figure AI [introduced Figure 03](https://www.figure.ai/news/introducing-figure-03) on October 9, 2025 as its third-generation robot, and the company says it was built for "Helix, the home, and the world at scale". Helix is Figure's vision-language-action model, an AI system that takes camera images and spoken instructions and outputs motor commands for the robot's arms and hands. ### How Do Home Robots Like Figure 03 and NEO Actually Work? Figure's announcement describes the hardware changes for home use. The company says the new camera system has twice the frame rate, one quarter of the latency and a 60% wider field of view per camera than the previous generation. Each hand has a camera in the palm for close-range views when the main cameras are blocked. The fingertips carry tactile sensors that Figure says can detect forces as small as 3 grams. A tactile sensor is a touch-sensitive pad in the fingertip, so the robot can tell how hard it is squeezing an object. That matters for handling a glass or a piece of fruit, where grip strength is the whole problem. Softness is the other theme. Figure says Figure 03 uses multi-density foam at pinch points and a soft textile covering, with washable parts that come off without tools. It also reports about 9% less mass than Figure 02. For charging, the robot steps onto a stand and charges wirelessly through coils in its feet at 2 kW, according to the company. 1X takes a different route. Its [NEO page](https://www.1x.tech/neo) lists a height of 5'6", a weight of 66 lbs, an arm payload of 18 lbs, an 842 Wh battery and about four hours of runtime. It describes a chore list you set up and schedule, and a noise rating of 22 dB. ![A person folding laundry on a bed in a bright living room while a household assistant robot waits beside the bed with a basket](/uploads/blog/humanoid-robots-daily-life-figure-02-tesla-optimus-homes/gemini-supporting.jpg) ### Can a Home Humanoid Robot Work on Its Own, or Is a Human Watching? 1X says NEO works autonomously by default. For chores it does not know, the company says a 1X Expert can supervise remotely at scheduled times to guide it. The same page lists FAQs on data sharing and how 1X uses data, so read those before ordering. Teleoperation means a person controls the robot from somewhere else, using what the robot's cameras show. For an owner, that would look like this: you schedule a time, the robot meets a chore it has not learned, and a 1X Expert guides it through the task while the robot's cameras relay the room. A [summary from the AI consultancy Lightpath](https://lightpath.ai/notes/home-humanoid-robots-teleoperation/) cites MIT Technology Review in February 2026 as saying 1X obtains the owner's consent before switching to remote control. It also cites a follow-up on October 8, 2026 reporting that a remote operator currently performs most of NEO's tasks. The privacy question follows from that design. If a person can see through the robot's cameras, the robot is a camera inside the home with a human on the other end. It belongs on the checklist next to price. The same issue shows up with other always-on devices, as in this look at [on-device phone AI and privacy](/radar/on-device-phone-ai-local-models-apple-android-privacy). ### How Much Do Humanoid Robots for Home Cost, and When Can You Buy One? Price and timing vary most across the three. - **1X NEO:** The company's order page shows a $200 deposit and no price. Several review sites report two options, a $20,000 early-access purchase or a $499 monthly subscription. Those figures come from secondary coverage, and 1X lists 2026 as the target year for US deliveries. As of mid-October 2026, no first customer deliveries had been announced. - **Figure 03:** Figure has published no consumer price, pre-order or waitlist. - **Tesla Optimus:** Tesla has not unveiled its third-generation design. Per coverage of the 22 July 2026 earnings call, Tesla said first production lines were being installed, with production anticipated in 2026 and the first units going to internal training and development. To compare the three on equal terms, a household would need an announced price, a delivery date and a first customer delivery for each robot, and only NEO has even a deposit page today. A Figure 03 or Optimus buyer has no order page to use at all. Elon Musk gave a target of public sales by the end of 2027 at Davos in January 2026, according to trackers. ### Who Is Building These Robots, and How Many Will Exist? Figure says its BotQ factory line can initially build up to 12,000 robots per year and that it aims for 100,000 robots over four years. On the industry as a whole, the Lightpath summary of MIT Technology Review's October 8, 2026 reporting cites about 15,000 humanoids shipped in 2025, roughly 90% from Chinese companies. It also cites Agility Robotics cofounder Jonathan Hurst estimating about ten years before robots do useful work in homes. Tesla is the least transparent of the three. The analysis says Musk called Optimus the hardest product Tesla has tried to scale, and that Tesla declined to name a 2026 unit target. According to an [analysis of its 2026 earnings calls](https://packetnebula.com/articles/tesla-optimus-v3-production-unrevealed-robot/), Tesla has published one hardware figure, 50 actuators per unit, and Musk has since said the mechanical design changed. The same analysis says numbers like a $30,000 price or a particular weight, which circulate widely online, have no Tesla source. ![An engineer in a clean assembly workshop adjusting a robotic hand with small tools on a workbench](/uploads/blog/humanoid-robots-daily-life-figure-02-tesla-optimus-homes/gemini-supporting.jpg) ### Are Humanoid Robots Safe to Have Around Your Family? None of the three companies cites an independent safety certification for a home humanoid. What they do publish is design choices. Figure points to soft textile covers, foam at pinch points and multiple battery protection layers, and says its battery pack is certified to UN38.3, a transport safety standard for lithium batteries that covers shipping. 1X lists a 22 dB noise rating and a self-docking charger. Neither company has published test results for how the robot behaves around children or pets. Without test results, a buyer can compare soft covers and noise ratings but cannot compare how often a robot drops or bumps something. ### What to Watch Next The first confirmed NEO deliveries to paying customers matter most, along with what 1X says about how often a remote operator steps in. Figure could publish a consumer price, waitlist or ship date for Figure 03. Tesla could show Optimus V3 in public and say something about homes. An independent safety standard or third-party test for robots working near people at home would also change the picture. For related robotics reading on this site, see [how liquid neural networks help drones fly through storms](/radar/liquid-neural-networks-continuous-time-ai-mit-edge-robotics-benchmarks). ## Frequently asked questions ### Can you buy a humanoid robot for home use right now? You can reserve one, but not yet receive one. 1X's order page for NEO asks for a $200 deposit, and the company targets US deliveries in 2026, though no first deliveries had been announced as of mid-October 2026. Figure has published no consumer price or waitlist for Figure 03, and Tesla has not shown Optimus V3 publicly. ### How much does a home humanoid robot cost? Only one price is widely reported. Review sites say 1X NEO costs $20,000 for early access or $499 per month by subscription, but 1X's own order page lists just a $200 deposit. Figure and Tesla have published no consumer prices, and the $20,000 and $30,000 figures often attached to their robots have no traceable source. ### Does a home humanoid robot work without a human controlling it? Partly. 1X says NEO works autonomously by default and that a remote Expert can guide it on tasks it has not learned. Reporting summarized by Lightpath says a remote operator currently performs most of NEO's tasks. Figure says its Helix AI drives its robots, but its autonomy claims have not been independently confirmed. ### Is a humanoid robot at home a privacy risk? It can be, because a robot with cameras and a remote operator lets another person see inside your home. 1X says it gets owner consent before remote control, according to reporting cited by Lightpath. Read the company's data-sharing terms before buying, and ask when staff can view camera feeds and whether footage is stored. ## Source Based on Figure AI and 1X announcements. https://www.figure.ai/news/introducing-figure-03 --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/humanoid-robots-daily-life-figure-02-tesla-optimus-homes. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/stop-using-vector-dbs-for-agents-hierarchical-knowledge-graphs-vs-rag - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T08:25:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://arxiv.org/abs/2404.16130 # GraphRAG vs Vector RAG: What Each Does Best for AI Agents **Short answer:** Microsoft Research's GraphRAG builds a knowledge graph and community summaries before answering. Its authors report better comprehensiveness and diversity than baseline RAG on whole-collection questions, while vector search still suits simple lookups. **Takeaway:** GraphRAG helps with questions that span a whole document collection, while plain vector search remains the simpler choice for single-passage lookups. **Key figures (as reported by the source):** - Default text unit size: 1,200 tokens - LazyGraphRAG indexing cost vs full GraphRAG (Microsoft): 0.1% - LazyGraphRAG global query cost (Microsoft): 700x lower GraphRAG is a way of feeding documents to an AI model that builds a map of the people, places and things in the text before any question is asked. Microsoft Research introduced it in a [February 2024 blog post](https://www.microsoft.com/en-us/research/blog/graphrag-unlocking-llm-discovery-on-narrative-private-data/) and a [paper on arXiv](https://arxiv.org/abs/2404.16130). It targets one weakness of ordinary retrieval-augmented generation (RAG): questions about a whole pile of documents. Vector search, the usual engine behind RAG, is still the better tool for many jobs. Which one an AI agent should use depends on the kind of question being asked. ### What is GraphRAG, and how is it different from standard RAG? Standard RAG works in three moves. Documents are cut into chunks, each chunk is turned into a list of numbers called an embedding (a numeric fingerprint of its meaning), and at question time the system fetches the chunks whose fingerprints sit closest to the question. A vector database is the store that makes this lookup fast. GraphRAG adds a preparation stage. According to the [official GraphRAG documentation](https://microsoft.github.io/graphrag/), it extracts a knowledge graph from raw text, groups the graph into a hierarchy of communities, and writes a summary for each community. A knowledge graph is simply a set of entities (a company, a person, a product) joined by labelled relationships ("supplies", "manages"). A community is a cluster of entities that are closely connected to one another in the graph, such as a supplier, the orders it ships and the people who manage them. The community summary is a short report a language model writes about that cluster, so the system holds a ready-made overview of each topic before any question arrives. The Microsoft authors say baseline RAG struggles to "connect the dots" across scattered pieces of information. It also performs poorly on questions that need a view of the whole collection, such as asking for the top themes in a dataset. For those, vector similarity returns text that looks related but does not summarise anything. ### How does GraphRAG build and search its index? The [indexing pipeline](https://microsoft.github.io/graphrag/index/default_dataflow/) has several steps. Input documents are split into text units, 1,200 tokens each by default. A language model then reads every text unit and extracts entities and relationships, and a second pass condenses the descriptions of repeated entities into one. Next, a clustering method called hierarchical Leiden community detection groups closely connected entities. A language model writes a report for each group, working from the bottom of the hierarchy upward. Entity descriptions, text units and community reports are finally embedded into a vector store, so vectors remain part of the system. Because every text unit passes through a language model, the stage that makes whole-collection questions answerable is also the stage that makes setup costly. At question time the docs list several search modes: - Global search answers whole-collection questions using the community summaries. - Local search starts from specific entities and reaches their neighbours and related concepts. - DRIFT search works like local search but adds community information. - Basic search is ordinary top-k vector search, included for questions that suit it. The hierarchy gives the system several zoom levels. Low-level communities cover narrow topics, higher ones cover broad themes, so a question about the overall picture can be answered from the broad summaries without reading every chunk. Because those summaries already exist, a whole-collection question draws on them instead of on raw text. The paper describes the global route as a map-reduce style process: each community summary produces a partial answer, and the partial answers are combined into a final response. ![A researcher pinning photographs and red string across a wall to trace connections between documents](/uploads/blog/stop-using-vector-dbs-for-agents-hierarchical-knowledge-graphs-vs-rag/gemini-supporting.jpg) ### Is GraphRAG better than vector RAG? On whole-collection questions, the authors report that it is. The paper says GraphRAG gives "substantial improvements" over a conventional RAG baseline in the comprehensiveness and diversity of generated answers. Those tests used global sensemaking questions over datasets in the range of one million tokens. The 2024 blog post tested on VIINA, a dataset of thousands of Russian and Ukrainian news articles from June 2023. An LLM grader picked a winner between the two approaches on comprehensiveness, diversity and use of supporting source material. The post says GraphRAG "consistently outperforms" the baseline, and that faithfulness (staying true to the sources) was at a similar level for both. It publishes no exact win rates. Microsoft's published tests support a direction of improvement on one specific type of question, whole-collection questions, and the judge was an AI grader rather than a human panel. Results on your own data will differ. ### When does vector search still win? Vector search is the simpler choice when the answer lives in one or two passages, because the system only has to find the right chunk and hand it to the model. Asking a company handbook for the holiday policy is the classic case: find the right chunk, quote it. The official docs keep a basic vector search mode inside GraphRAG for exactly this reason. Some third-party write-ups report that vector RAG beats graph approaches on single-document lookups and loses on questions that span many documents. One engineering blog, [tianpan.co](https://tianpan.co/blog/2026-04-09-graphrag-production-when-vector-search-hits-ceiling), gives figures for this split. It is a secondary source, so treat the split as indicative. ![A librarian at a long shelf of archive boxes pulling a single folder, retrieval from a document collection](/uploads/blog/stop-using-vector-dbs-for-agents-hierarchical-knowledge-graphs-vs-rag/gemini-supporting.jpg) ### What does GraphRAG cost to set up? Building the graph is the expensive part of GraphRAG, and the cost lands before anyone has asked a single question. The indexing steps above call a language model for entity extraction, description summaries and community reports, so cost grows with the size of the document pile. Microsoft has since addressed the cost. In a [November 2024 post on LazyGraphRAG](https://www.microsoft.com/en-us/research/blog/lazygraphrag-setting-a-new-standard-for-quality-and-cost/), the team says LazyGraphRAG indexing costs are "identical to vector RAG and 0.1% of the costs of full GraphRAG." It skips model-based extraction at indexing time, uses plain language processing to find concepts, and postpones model work until a question arrives. The same post reports that for global queries it matches GraphRAG Global Search answer quality at more than 700 times lower query cost. Those are Microsoft's own claims about its own method, so check what your installed version supports. Quality of the graph also depends on tuning. The docs warn that using GraphRAG out of the box "may not yield the best possible results" and recommend prompt tuning for your data. They also note that claim extraction is off by default because it generally needs tuning. ### What do AI agents need from retrieval? An agent is an AI system that takes several steps toward a goal, often looking things up along the way. Agents ask both kinds of question. "What is the refund window?" is a lookup. "Which suppliers are linked to the delayed orders?" asks the system to follow relationships across documents. The research does not test agents directly. The GraphRAG paper and blog evaluate question answering over document collections. Give the agent a plain vector search for lookups, which is cheap and fast when the answer sits in one or two passages, and a graph-based route for relationship and summary questions, which follows connections or summarises across documents. For agent control flow rather than retrieval, see [deterministic state graphs for AI agent workflows](/radar/deterministic-state-graphs-preventing-infinite-loops-ai-agent-workflows), and for the wider picture, [the rise of autonomous AI agents](/radar/rise-of-autonomous-ai-agents-multi-agent-workflows). ### What to watch next Independent head-to-head tests with real numbers would fill a gap, since the original evaluation relied on an AI grader and reported no scores. Cheaper variants such as LazyGraphRAG may land in the open-source package, and it remains to be seen whether they hold up on private data. Agent frameworks that pick between lookup and graph routes per question are the pattern to look for in how retrieval choices get wired in. ## Frequently asked questions ### What is GraphRAG in simple terms? GraphRAG is a retrieval method from Microsoft Research. A language model reads your documents, extracts the entities and relationships into a knowledge graph, groups related entities into communities and writes a summary for each. At question time the AI uses those summaries and connections instead of only fetching similar-looking text chunks, which helps with questions about a whole collection. ### What is the difference between GraphRAG and vector RAG? Vector RAG splits documents into chunks and retrieves the ones whose numeric fingerprints are closest to the question. GraphRAG first builds a graph of entities and relationships with community summaries, then searches those structures. Vector RAG suits questions answered in a passage or two, while GraphRAG targets questions that connect many documents or summarise an entire collection. ### Does GraphRAG replace a vector database? No. The official GraphRAG pipeline itself embeds entity descriptions, text units and community reports into a vector store, and the docs include a basic vector search mode for questions better suited to standard RAG. Teams usually treat the graph as an addition for relationship and summary questions, keeping vector search for simple lookups. ### Is GraphRAG expensive to run? Full GraphRAG calls a language model for entity extraction, description summaries and community reports during indexing, so cost rises with corpus size. Microsoft's LazyGraphRAG post says its variant has indexing costs identical to vector RAG and 0.1% of full GraphRAG. Those are the vendor's own claims. ## Source Based on Microsoft Research's GraphRAG paper. https://arxiv.org/abs/2404.16130 --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/stop-using-vector-dbs-for-agents-hierarchical-knowledge-graphs-vs-rag. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/death-of-traditional-search-answer-engines-perplexity-chatgpt - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T08:20:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/ # Answer Engines: What Perplexity and ChatGPT Do to Search **Short answer:** Pew Research found Google users clicked a regular result in 8% of visits with an AI summary, versus 15% without. A separate Tow Center test found AI search tools often misidentified sources. **Takeaway:** Answer engines save time on simple questions, but clicks fall when the answer sits on the page and the tools still misidentify sources often, so open the cited link before you rely on it. **Key figures (as reported by the source):** - Clicks with an AI summary (Pew): 8% of visits - Clicks without one (Pew): 15% of visits - Tool answers wrong (Tow Center): More than 60% Answer engines are changing what people do after they type a question into Google. An answer engine is a search tool that reads web pages for you and writes one short reply with numbered source links, instead of handing you a list of ten blue links. Perplexity, ChatGPT search and Google's own AI Mode all work this way. The best public measurement so far comes from Pew Research Center, and it shows a real shift in behaviour on Google itself. A separate test of AI search tools shows how often the cited sources are wrong. ### What Is an Answer Engine, and How Is It Different From a Search Engine? A classic search engine matches your words to pages and ranks them. You do the reading and the comparing. An answer engine adds a step: a language model (the kind of AI behind chatbots) reads the top results and writes a combined reply in plain sentences. [Perplexity's help centre](https://www.perplexity.ai/help-center/en/articles/10352895-how-does-perplexity-work) describes its process as interpreting your question, searching the web in real time and summarising what it finds. Each answer carries numbered citations that link to the original sources, and you can ask follow-up questions that build on the earlier ones. OpenAI launched ChatGPT search on October 31, 2024, starting with Plus and Team subscribers, according to [PPC Land's report on the launch](https://ppc.land/chatgpt-launches-web-search-with-news-partners-real-time-updates-now-available-2/). The article lists source attribution links as a feature and names Reuters, the Associated Press, the Financial Times, Le Monde, Axel Springer and Vox Media among its news partners. ### How Does Google's AI Mode Work? Google's version sits inside Google Search. In a [March 2025 announcement](https://blog.google/products/search/ai-mode-search/), Google described AI Mode as an opt-in experiment in Search Labs, running a custom version of its Gemini 2.0 model. Google says AI Mode uses a technique it calls query fan-out. The system runs several related searches at once across subtopics and data sources, then merges the results into a single response. Google says responses include links to web content, and that when it is not confident in an answer it falls back to standard search results. AI Overviews, the short summaries at the top of ordinary results, are a separate feature. Google's post says they are used by more than a billion people. ### Are Answer Engines Replacing Google? Pew's study looked at what happened when a Google results page included an AI summary. Pew followed the browsing of 900 U.S. adults during March 2025 and collected 68,879 unique Google searches, of which 12,593 produced an AI summary. People clicked a traditional result link in 8% of visits to pages with a summary, against 15% of visits to pages without one. Pew reports that the link inside the summary itself was clicked in 1% of visits. ![A person at a kitchen table typing a question into a laptop while a phone lies beside the keyboard](/uploads/blog/death-of-traditional-search-answer-engines-perplexity-chatgpt/gemini-supporting.jpg) Pew also found that the browsing session ended on 26% of pages with a summary, against 16% without one. The study covers Google only, and Pew notes that AI summaries change over time, so the figures reflect how they looked when the results were collected in April 2025. In plain numbers, 8% against 15% means visitors who saw a summary were close to half as likely to click an ordinary result, and 26% against 16% means they were more likely to end the session on that page. When the answer sits at the top of the page, fewer people click through, which is why the phrase "death of search" gets used. A person who gets the answer on the results page has less reason to visit the sites that supplied it, so the numbers matter to publishers as much as to readers. Pew did not measure traffic to other tools such as Perplexity or ChatGPT, so the study cannot show a total shift away from Google. ### Why Do AI Search Tools Cite Sources? Citations are the main reason these tools exist as a separate category. A chatbot that answers from memory cannot show where a claim came from. An answer engine fetches pages at the moment you ask, so it can point to them. Perplexity says its Deep Research mode runs dozens of searches and reads hundreds of sources, delivering a report in about 2 to 4 minutes, which is far more reading than most people would do by hand. That depth is the case for using an answer engine on a broad question, while the accuracy results below are the case for checking what it cites. Perplexity calls this source transparency. Pew's finding that summary links drew clicks in 1% of visits suggests most people read the reply and stop there. The links also matter to the sites being cited. Publishers that get a mention but no visit receive none of the traffic that pays for their reporting, which is one reason OpenAI announced news partnerships at launch. ### How Accurate Are AI Answer Engines? The Tow Center for Digital Journalism at Columbia Journalism Review tested eight generative search tools on 1,600 queries (20 publishers, 10 articles each, 8 tools). Each query asked the tool to identify an article from an excerpt, which is a narrow task, so the results do not describe every kind of question. Because each query had one correct article, every answer could be marked right or wrong, which is what lets the report put a percentage on each tool. Collectively, according to the [Tow Center results](https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php), the tools answered more than 60% of queries incorrectly. Perplexity was wrong on 37%, the best of the group, while Grok 3 was wrong on 94%. ![A researcher at a desk comparing a printed newspaper article against text on a laptop screen](/uploads/blog/death-of-traditional-search-answer-engines-perplexity-chatgpt/gemini-supporting.jpg) According to the report, ChatGPT misidentified 134 of its 200 articles, signalled low confidence only 15 times, and never declined to answer. Grok 3 linked to error pages for 154 of its 200 citations. The report also found that having a licensing deal did not guarantee correct credit: the San Francisco Chronicle has a partnership with OpenAI, yet ChatGPT identified only 1 of 10 of its excerpts correctly. The test is from early 2025, and the tools have changed since. ### When Should You Use an Answer Engine Instead of Google? These tools work well for orientation: a quick explanation, a comparison of options, or a first read on an unfamiliar topic where a wrong detail costs little. Follow-up questions are the real advantage, because you can narrow a topic without starting over. Traditional results suit anything where the exact page matters: official forms, prices, medical or legal details, or a claim you plan to repeat. For those, open the cited page and read it. If you want to see where this is heading next, our piece on [AI agents that click and type like a person](/radar/computer-use-ai-agents-operating-systems-claude-operator) covers the step after answers: tools that complete the task for you. ### What to Watch Next Click-through data comes first. Pew's 8% against 15% is one snapshot from one month in 2025, and it covers only Google. A repeat study would show whether the gap has widened or narrowed as AI summaries appear on more searches. Publisher deals are next. Whether answer engines pay for the pages they summarise will shape how much reliable material stays on the open web. A fresh accuracy comparison on current versions would also tell readers more than any launch announcement, since the Tow Center test showed wide gaps between tools. ## Frequently asked questions ### What is an answer engine? An answer engine is a search tool that reads web pages for you and writes one combined reply with numbered source links. Perplexity, ChatGPT search and Google AI Mode work this way. You can ask follow-up questions, and you open the cited pages when you want to check a claim or read the original. ### Do AI summaries reduce clicks on Google? Pew Research Center found that people clicked a traditional result in 8% of visits to pages with an AI summary, against 15% of visits without one. The study followed 900 U.S. adults in March 2025 and covers Google only. Links inside the summary itself were clicked in 1% of visits. ### Are AI answer engines accurate? Not reliably. A Tow Center test of eight AI search tools on 1,600 queries found they answered more than 60% incorrectly when asked to identify news articles from excerpts. Perplexity was wrong on 37%, Grok 3 on 94%. The test is from early 2025 and narrow, so check the cited source. ### Is Google AI Mode the same as AI Overviews? No. AI Overviews are short summaries shown at the top of normal Google results. AI Mode, announced in March 2025 as an opt-in Labs experiment, is a separate conversational search experience that runs several related searches at once and combines them into one response with links to web content, according to Google. ## Source Based on Pew Research Center and Tow Center studies. https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/death-of-traditional-search-answer-engines-perplexity-chatgpt. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/the-1-bit-llm-revolution-bitnet-b158-ternary-addition-vs-multiplication - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T08:20:00.000Z - Last modified: 2026-10-28T18:00:00.000Z - Primary source: https://arxiv.org/abs/2402.17764 # BitNet: How Microsoft's 1-Bit LLM Runs AI on a Laptop CPU **Short answer:** Microsoft's BitNet b1.58 stores every model weight as -1, 0 or +1. Its open 2B4T model uses 0.4 GB of memory, and the bitnet.cpp engine runs it on ordinary CPUs faster and with less energy than standard tools. **Takeaway:** BitNet shows a small AI model can work with three-value weights and run cheaply on a normal CPU, but only at 2 billion parameters so far and only with Microsoft's own software. **Key figures (as reported by the source):** - BitNet b1.58 2B4T memory (excluding embeddings): 0.4 GB - bitnet.cpp speedup on x86 CPUs: 2.37x to 6.17x - bitnet.cpp energy cut on x86 CPUs: 71.9% to 82.2% BitNet is Microsoft's attempt to make AI language models far smaller and cheaper to run by storing each of their billions of internal numbers as just -1, 0 or +1. A normal model keeps those numbers as detailed decimals. Microsoft's researchers report that the stripped-down version can match a normal model of similar size on quality while using a fraction of the memory and energy. That matters because memory and electricity are what keep capable AI locked inside data centers. If a model needs far less of both, it can run on the laptop you already own, without a graphics card. By the end of this piece you will know what BitNet actually is, why it is called "1.58-bit", what Microsoft has measured (and what it has not), and whether you can try it today. ### What is BitNet LLM, in plain English? An LLM (large language model) is the kind of AI behind chatbots. Under the hood it is a huge grid of numbers called weights. A weight is a dial that says how strongly one piece of the model should influence another. Training is the process of turning billions of those dials until the model writes useful text. In most models each dial is stored as a 16-bit number, a format with fine-grained decimal values. BitNet throws that precision away. In the [February 2024 paper "The Era of 1-bit LLMs"](https://arxiv.org/abs/2402.17764), Microsoft Research describes BitNet b1.58, a model in which every weight is one of three values: -1, 0 or +1. Think of it like a volume knob replaced by a three-way switch: turn this signal down, ignore it, or turn it up. You lose nuance on each individual switch. The bet is that with billions of switches, the model as a whole still learns what it needs. One detail is easy to miss. These models are trained this way from the start. They are not normal models squeezed down after training, which is the more common shortcut and usually costs quality. ### How does BitNet work, and why is it faster? The expensive part of running an LLM is multiplication. Every word the model produces requires multiplying the incoming data by millions of weights, then adding the results together. Multiplying decimals takes more chip area and more power than adding them. When every weight is -1, 0 or +1, those multiplications mostly disappear. Multiplying by +1 means keep the number. Multiplying by -1 means flip its sign. Multiplying by 0 means skip it. What is left is mostly addition and subtraction, which chips do cheaply. The 2024 paper puts a number on this: it estimates that the arithmetic energy for this core step is 71.4 times lower than in a 16-bit model, based on an energy model for 7-nanometer chips. That is an estimate of one operation, not a measurement of a whole computer, so treat it as a ceiling on the savings rather than what your power meter will show. The second win is memory. A three-value weight takes roughly a tenth of the space of a 16-bit one. Smaller weights mean the model fits in less memory and less data has to move between memory and the processor, which is often the real speed limit for chatbots. (If you want the other half of that memory story, our explainer on the [KV cache](/radar/kv-cache-explained-cloudflare-prompt-leak) covers the working memory a model builds up while it talks.) ![A software engineer at a wooden desk running an AI chatbot on a thin laptop with no external graphics card, a 1-bit LLM test in progress](/uploads/blog/the-1-bit-llm-revolution-bitnet-b158-ternary-addition-vs-multiplication/gemini-supporting.jpg) ### Why is ternary 1.58 bits? "Ternary" just means three possible values. A computer bit can hold two values, so how much information do three values carry? The answer is the base-2 logarithm of 3, which is about 1.58. That is where the name BitNet b1.58 comes from. The history explains the odd label. Microsoft's [first BitNet paper in October 2023](https://arxiv.org/abs/2310.11453) used true 1-bit weights that could only be -1 or +1. The 2024 version added 0 as a third option, so the model can switch a connection off entirely. The paper credits that zero with better results, and the "1-bit LLM" nickname stuck even though the precise figure is 1.58. In practice the weights are packed into memory slots of 2 bits or so, so the real storage is a little above 1.58 bits per weight. The headline number describes information, not the exact bytes on disk. ### Is BitNet as good as a normal model? What Microsoft measured There are two sets of results, and they answer different questions. The 2024 paper trained a ternary model and a standard model of matching size from the same recipe. At 3 billion parameters (a parameter is one weight), the paper reports the ternary model matched the standard model's quality while using 3.55 times less GPU memory and running 2.71 times faster. At 70 billion parameters it reports a 4.1 times speedup than the standard baseline. These were research comparisons on GPUs, and Microsoft did not release those particular research models. The more useful result for most readers came in April 2025, when Microsoft released [BitNet b1.58 2B4T](https://arxiv.org/abs/2504.12285): about 2 billion parameters, trained on 4 trillion tokens (a token is a word or piece of a word). The technical report compares it with popular small open models on a Surface Laptop Studio 2 with an Intel Core i7-13800H processor.
ModelMemory (excluding embeddings)Time per word on CPUEstimated energy per wordAverage benchmark score
BitNet b1.58 2B4T0.4 GB29 ms0.028 J54.19
Qwen2.5 1.5B2.6 GB65 ms0.347 J55.23
Llama 3.2 1B2 GB48 ms0.258 J44.90
Gemma 3 1B1.4 GB41 ms0.186 J43.74
Source: BitNet b1.58 2B4T Technical Report, Table 1. Energy is estimated from an arithmetic model, not measured at the wall. Read that table honestly. Microsoft's model uses about a sixth of Qwen2.5's memory and answers in under half the time, yet Qwen2.5 still scores slightly higher on average. So the claim is not "1-bit beats everything". It is "1-bit gets close to the best models its size at a much lower running cost". ![A woman in a quiet library reading answers from an open-weight BitNet model on an older laptop, coffee cup beside the keyboard](/uploads/blog/the-1-bit-llm-revolution-bitnet-b158-ternary-addition-vs-multiplication/gemini-supporting.jpg) ### What is bitnet.cpp, and can BitNet run on a CPU? Here is the catch that most headlines skip. Standard AI software does not know how to run three-value weights efficiently. The [model card on Hugging Face](https://huggingface.co/microsoft/bitnet-b1.58-2B-4T) warns readers not to expect efficiency gains if they load the model with the usual transformers library. To get the speed, you have to use Microsoft's own engine. That engine is [bitnet.cpp](https://github.com/microsoft/BitNet), an open-source program for running these models on ordinary processors (CPUs) rather than graphics cards. Microsoft says it is 1.37 to 5.07 times faster on ARM chips (the kind in Apple Silicon Macs and most phones) and 2.37 to 6.17 times faster on x86 chips (most Windows laptops). It reports energy cuts of 55.4% to 70.0% on ARM and 71.9% to 82.2% on x86. Those are comparisons against llama.cpp, a popular tool for running models locally, at the same model sizes. They come from the [bitnet.cpp technical report](https://arxiv.org/abs/2410.16144), tested on an Apple M2 Ultra Mac Studio and an Intel Core i7-13700H laptop. You may also have seen the claim that bitnet.cpp runs a 100-billion-parameter model on a single CPU at 5 to 7 words per second, roughly human reading speed. That is in Microsoft's README, but the technical report says the large models it tested were "dummy setups" used to measure speed. No trained 100B model has been released. The best result in the report, 6.58 tokens per second, was on the Apple M2 Ultra; the Intel laptop managed 1.70. The repo has kept moving. Microsoft added a GPU engine in May 2025, and in January 2026 it said a CPU update brought a further 1.15 to 2.1 times speedup. ### Is BitNet open source, and can you use it today? Yes. The 2B4T model weights and code are released under the MIT license, one of the most permissive open-source licenses, and you can download them from Hugging Face. If you are new to the idea of downloadable models, our guide to [open-weight models](/radar/open-weight-models-explained-ibm) explains what you can and cannot do with them. Before you get excited, note the limits Microsoft lists itself: - **It is small.** Two billion parameters is far from the size of the chatbots most people use, and it is mostly English. - **It has a short memory.** The model handles at most 4,096 tokens of context, a few pages of text. - **It is a research release.** The model card advises against using it in commercial or real-world products without further testing. - **It needs the right software.** The efficiency gains only appear with bitnet.cpp or Microsoft's custom GPU kernels. Other groups have built on the idea. The bitnet.cpp repo lists community 1-bit LLM models such as the Falcon3 and Falcon-E families from the Technology Innovation Institute, alongside Microsoft's own. ### What to watch next The open question is scale. Microsoft has shown a 2-billion-parameter model that holds its own against similar-sized models. Its own technical report names training larger native 1-bit models as future work. Until someone releases a trained 30B or 70B ternary model, the big claims remain projections. The second thing to watch is hardware. The 2B4T report notes that today's graphics cards are not built for 1-bit math, and the 2024 paper argues for chips designed around it. If chipmakers take that up, the savings could grow well beyond what software alone delivers. For now, this project is the clearest evidence that AI models do not need detailed numbers to be useful. If you have a recent laptop, you can try it yourself today. ## Frequently asked questions ### What is BitNet in simple terms? BitNet is a family of AI language models from Microsoft Research that store each internal weight as -1, 0 or +1 instead of a detailed decimal number. That makes the model much smaller and turns most of its math into simple addition, so it can run on an ordinary laptop processor while staying close in quality to normal models of the same size. ### Why is it called 1.58-bit and not 1-bit? Each weight can take three values, and the information in three options is the base-2 logarithm of 3, about 1.58 bits. Microsoft's first BitNet paper in 2023 used true 1-bit weights of -1 or +1. The 2024 version added zero so the model can switch connections off, which gave the b1.58 name. ### Can I run BitNet on my own computer? Yes. Microsoft's BitNet b1.58 2B4T model is free on Hugging Face under the MIT license and runs on normal CPUs with the open-source bitnet.cpp engine. You must use bitnet.cpp or Microsoft's GPU kernels to get the speed and memory savings, because the standard transformers library gives no efficiency gain, according to the model card. ### Can BitNet really run a 100B model on one CPU? Microsoft's bitnet.cpp report measured speed for a 100-billion-parameter BitNet setup, reaching 6.58 tokens per second on an Apple M2 Ultra. But the report calls those test models dummy setups, used only to measure speed. No trained 100B BitNet model has been released, so the largest model you can actually use is about 2 billion parameters. ## Source Based on Microsoft Research's BitNet papers and repo. https://arxiv.org/abs/2402.17764 --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/the-1-bit-llm-revolution-bitnet-b158-ternary-addition-vs-multiplication. Last modified 2026-10-28T18:00:00.000Z. ================================================================================ - URL: https://syntaxradar.com/radar/zero-latency-voice-ai-speech-to-speech-gpt4o-gemini-live - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T08:15:00.000Z - Last modified: 2026-10-28T18:00:00.000Z - Primary source: https://openai.com/index/hello-gpt-4o/ # Speech to Speech AI: Why ChatGPT and Gemini Now Sound Human **Short answer:** Voice assistants such as ChatGPT voice mode and Gemini Live now use single speech to speech models instead of chaining transcription, a text model and a voice reader. OpenAI says GPT-4o replies in 320 ms on average, versus 2.8 to 5.4 seconds before. **Takeaway:** Voice AI sounds human now because one model hears and speaks audio directly, which cuts the wait from seconds to about a third of a second and keeps tone and emotion that a text transcript throws away. **Key figures (as reported by the source):** - GPT-4o average audio response time: 320 ms (as low as 232 ms) - Old ChatGPT Voice Mode pipeline (GPT-4): 5.4 seconds on average - Kyutai Moshi latency in practice: 200 ms (160 ms theoretical) Ask ChatGPT's voice mode or Gemini Live a question and it answers almost before you finish your breath. It can be interrupted mid-sentence, it can sound amused, and it can hear when you sound unsure. Two years ago, talking to an AI felt like leaving a voicemail and waiting for a reply. The change has a name: **speech to speech AI**. Instead of turning your voice into text, thinking in text, and reading the answer out loud, a single model listens to sound and replies in sound. OpenAI says its GPT-4o model can answer spoken input in as little as 232 milliseconds. By the end of this piece you will know how the old voice assistants worked, what changed under the hood, why that change makes voices sound more human, and where these systems still trip up. ### What is speech to speech AI? Speech to speech AI is a model that takes audio in and produces audio out, without a written transcript sitting in the middle as the main step. A "model" here just means the trained AI system that does the work, the same kind of thing that powers a chatbot. The key idea is that the model works with sound directly. It hears pitch, pace, pauses and background noise, not just the words. When it answers, it generates the voice itself rather than handing a script to a separate reading program. You have probably used one already. ChatGPT's Advanced Voice Mode runs on this approach, and Google's Gemini Live uses what Google calls "native audio", meaning the model was built to handle sound as an input and output in its own right. Do not confuse it with text to speech, which reads written words aloud in a synthetic voice. Text to speech starts with text. Speech to speech starts with your voice and answers with a voice. ### How did voice assistants work before? Older voice assistants were a relay race between three separate programs. OpenAI describes its own pre-2024 Voice Mode this way in the [GPT-4o announcement](https://openai.com/index/hello-gpt-4o/): one simple model transcribed your audio into text, GPT-3.5 or GPT-4 read that text and wrote a text reply, and a third simple model turned the reply back into audio. Engineers call this a "cascade" or "pipeline": each stage waits for the one before it to finish. Every handoff adds delay. OpenAI reports that this setup took 2.8 seconds on average with GPT-3.5 and 5.4 seconds with GPT-4. In a conversation, five seconds of silence feels like the line went dead. Delay was only half the problem. Once your voice became plain text, everything except the words was gone. OpenAI's own words: the main model "can't directly observe tone, multiple speakers, or background noises". It also could not laugh, sing or express emotion in its reply, because the final voice came from a separate program reading a script. The researchers behind Kyutai's open model Moshi make the same point in [their paper](https://arxiv.org/abs/2410.00037). They write that these pipelines add "a latency of several seconds", lose information such as emotion, and force the conversation into strict turns that ignore interruptions and people talking over each other. ![A woman in a kitchen talking to her phone propped against a fruit bowl while she chops vegetables, using a speech-to-speech voice assistant hands-free](/uploads/blog/zero-latency-voice-ai-speech-to-speech-gpt4o-gemini-live/gemini-supporting.jpg) ### How does speech to speech AI work? The short version: one model does the listening, the thinking and the talking. In May 2024 OpenAI said it "trained a single new model end-to-end across text, vision, and audio" for GPT-4o. "End-to-end" means one network is trained on the whole job from input to output, instead of three programs glued together. To do that, the model needs a way to treat sound like language. Here is the plain-English version of the trick most of these systems use. #### Sound becomes tokens Chatbots read and write "tokens", small chunks of text a bit shorter than words. Speech models do something similar with audio. A tool called a neural audio codec (a compressor trained by AI) chops your voice into short slices and turns each slice into a compact code the model can read, much like a word token. The Moshi paper gives a concrete example. Its codec, called Mimi, produces audio tokens at 12.5 frames per second, so each slice covers 80 milliseconds of sound. Those codes keep information about how something was said, not only what was said. #### The model answers in tokens, then those become a voice The model predicts the next audio tokens the same way a chatbot predicts the next word. The codec then turns those tokens back into a waveform you hear. Because the reply is generated as sound from the start, the model can choose to sound warm, hesitant or excited. Moshi adds one more helpful step that its authors call "Inner Monologue": the model writes a time-aligned text version of what it is about to say just before it says it. The paper reports this improves how factual and well-formed the spoken answers are. So text has not vanished entirely; it rides along as a guide rather than acting as the bottleneck. #### Listening and talking at the same time Old assistants worked like a walkie-talkie: you talk, then it talks. Moshi is "full-duplex", the term for a line where both sides can speak at once, like a phone call. It models your voice and its own voice as two parallel streams, so it can handle interruptions and quick "mm-hm" replies. The paper reports a theoretical latency of 160 milliseconds and 200 milliseconds in practice. ### Why does ChatGPT voice mode sound more human now? Mostly because of speed and what the model can hear. On speed, OpenAI says GPT-4o answers audio in 232 milliseconds at best and 320 milliseconds on average, which it compares to human response time in conversation. That is roughly a tenth of the 2.8 to 5.4 seconds the old Voice Mode needed. Latency, the gap between you stopping and the AI starting, is the single biggest thing that makes a call feel natural or awkward. On hearing, a model that receives your actual audio can notice that you sound rushed, or that two people are talking, or that a dog is barking. Google's developer docs for its [Gemini Live API](https://ai.google.dev/gemini-api/docs/live-api) say the model "adapts response style and tone to match the user's input expression" and that users can interrupt it at any time. There has also been steady progress since the first launch. OpenAI's August 2025 model, [gpt-realtime](https://openai.com/index/introducing-gpt-realtime/), scored 82.8% on Big Bench Audio, a reasoning test with spoken questions, up from 65.6% for its December 2024 model. In December 2025 Google [upgraded Gemini 2.5 Flash Native Audio](https://blog.google/products/gemini/gemini-audio-model-updates/) and said it now follows developer instructions 90% of the time, up from 84%.
SystemReported response timeSource
ChatGPT Voice Mode with GPT-3.5 (three-model pipeline)2.8 seconds on averageOpenAI
ChatGPT Voice Mode with GPT-4 (three-model pipeline)5.4 seconds on averageOpenAI
GPT-4o (single speech model)232 ms minimum, 320 ms averageOpenAI
Kyutai Moshi (full-duplex research model)160 ms theoretical, 200 ms in practiceMoshi paper
These figures come from each company's own tests, run in different ways, so treat the table as a sense of scale rather than a head-to-head race. ![A man walking down a busy city street at dusk wearing wireless earbuds, mid-conversation with a voice assistant, phone in hand](/uploads/blog/zero-latency-voice-ai-speech-to-speech-gpt4o-gemini-live/gemini-supporting.jpg) ### What can you actually do with it, and where does it fall short? For everyday users, the clearest win is conversation that does not feel like a phone menu. You can cut the assistant off when it rambles, change topic mid-sentence, or ask it to slow down. Translation is another. In the same December 2025 announcement, Google introduced live speech-to-speech translation in the Google Translate app, covering over 70 languages and 2,000 language pairs. Google says it preserves the speaker's intonation, pacing and pitch, so the translated voice sounds less flat. For businesses, the draw is customer service agents that can answer the phone. When OpenAI opened its [Realtime API](https://openai.com/index/introducing-the-realtime-api/) to developers in October 2024, it said they "no longer have to stitch together multiple models", and that the old chain "often resulted in loss of emotion, emphasis and accents, plus noticeable latency". An API is simply the connection developers use to plug a model into their own apps. The launch pricing worked out to roughly $0.06 per minute of audio in and $0.24 per minute out; gpt-realtime later cut prices by 20%. If you follow AI costs, our look at [how LLM API prices are falling](/radar/llm-api-pricing-cuts-2026) covers the wider trend. The trade-offs are real, though: - **Less control.** With a pipeline, a company can inspect and edit the text at each step. A single audio model is harder to audit, which matters in regulated fields like banking or health. - **Benchmarks are company-run.** The scores above come from OpenAI and Google themselves. Independent tests are still catching up. - **It is still a chatbot.** A natural voice does not make the answer more accurate. The model can be wrong in a very convincing tone, so check anything important. Open research matters here too. Kyutai released Moshi publicly, and the paper says its language backbone, called Helium, has 7 billion parameters (the adjustable settings a model learns during training). If you are curious why openly released models matter, our explainer on [open-weight models](/radar/open-weight-models-explained-ibm) walks through it. ### What to watch next Three things are worth tracking. First, whether full-duplex behaviour, the "both talk at once" style Moshi demonstrated, becomes standard in the big consumer assistants. Second, whether independent benchmarks for spoken reasoning and interruption handling emerge, so we are not relying on vendors' own scorecards. Third, live translation: Google says it plans to bring live speech translation to more products, including the Gemini API, in 2026. The short takeaway: voice assistants sound human now because they stopped turning your voice into a text message first. That one design change cut the wait from seconds to a fraction of a second and let the AI hear how you say things, not just what you say. ## Frequently asked questions ### What is the difference between speech to speech and text to speech? Text to speech starts with written words and reads them aloud in a synthetic voice. A speech-to-speech model starts with your spoken audio and answers with generated audio, using one model for listening, reasoning and talking. Because it never reduces your voice to plain text, it can pick up tone, pace and background sounds that a transcript would drop. ### Is ChatGPT voice mode a speech to speech model? Yes. OpenAI says GPT-4o was trained as a single model across text, vision and audio, and its Realtime API describes the same approach as speech to speech. The older Voice Mode used three separate models: one transcribed audio, GPT-3.5 or GPT-4 wrote a text reply, and a third read it aloud, averaging 2.8 to 5.4 seconds per answer. ### Why can I interrupt Gemini Live and ChatGPT now? Newer voice systems keep listening while they talk, so they notice when you start speaking and stop. Google's Live API documentation says users can interrupt the model at any time, and OpenAI says its Realtime API handles interruptions automatically. Research models like Kyutai's Moshi go further, modelling both voices as parallel streams so overlap and quick interjections work. ### Is talking to a voice AI more accurate than typing to a chatbot? Not necessarily. A natural voice changes how the answer is delivered, not whether it is correct. These models can still make mistakes, and they can sound confident while doing it. The published benchmark scores also come from the companies themselves. For anything important, such as money, health or legal questions, double-check the answer against a reliable source. ## Source Based on OpenAI, Google and Kyutai's published figures. https://openai.com/index/hello-gpt-4o/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/zero-latency-voice-ai-speech-to-speech-gpt4o-gemini-live. Last modified 2026-10-28T18:00:00.000Z. ================================================================================ - URL: https://syntaxradar.com/radar/liquid-neural-networks-continuous-time-ai-mit-edge-robotics-benchmarks - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T08:15:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://news.mit.edu/2023/drones-navigate-unseen-environments-liquid-neural-networks-0419 # Liquid Neural Networks Explained: What MIT Really Showed Drones **Short answer:** MIT CSAIL researchers built liquid neural networks, small continuous-time AI models that adapt to inputs after training. A Science Robotics study flew drones in unseen forests and cities, and MIT reports they beat many alternatives. **Takeaway:** Liquid neural networks are a small, adaptive design that looks promising for robots, but the public evidence covers narrow tests and still needs independent confirmation. **Key figures (as reported by the source):** - Lane-keeping network size: 19 neurons - Classic network, same task: ~100,000 neurons - Worm inspiration: 302 neurons Liquid neural networks are a type of AI model built at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) whose internal settings keep adjusting to the data they receive, even after training is finished. They are small, they are designed for streams of data such as video, and the best public evidence for them comes from MIT's drone and driving experiments. They have not replaced the large models behind chatbots, and the published results are narrower than most online summaries suggest. ### What are liquid neural networks? A liquid neural network is a neural network that models time continuously and lets its behaviour shift depending on its inputs. A neural network is software made of many small connected units, called neurons, that pass numbers to each other. In an ordinary network, the connection strengths are fixed once training ends. According to [VentureBeat's interview with the MIT team](https://venturebeat.com/ai/how-mits-liquid-neural-networks-can-solve-ai-problems-from-robotics-to-self-driving-cars), liquid networks instead learn continuous-time models, where the time constants depend on the inputs the network sees. A time constant controls how quickly a neuron reacts to new information. If it can change with the input, the network can respond quickly during sudden change and slowly when things are calm. [MIT News](https://news.mit.edu/2023/drones-navigate-unseen-environments-liquid-neural-networks-0419) describes the same idea in simpler terms: the parameters can change over time, unlike traditional networks whose parameters stay fixed after training. The researchers say this helps the model pick out the part of a scene that matters for the task and ignore the rest. The design is inspired by small animals. VentureBeat cites the roundworm C. elegans, which handles complex behaviour with no more than 302 neurons. The goal was a model accurate and cheap enough to run on a robot's own computer without a cloud connection. That matters for a drone in a forest, where a network link may not exist and every decision has to be made on board within moments. ### How do liquid neural networks differ from transformers? Transformers, the design behind most chatbots, read a whole block of input at once and use a mechanism called attention to decide which parts relate to each other. They are trained on huge text collections and then frozen. Liquid networks work the other way round in two respects. They process a continuous stream, one moment after another. And their internal timing can adapt as data arrives. VentureBeat adds that liquid networks allow lateral and recurrent connections within the same layer, which standard feed-forward layers do not. Recurrent means a neuron's output can loop back as input, which gives the network a short-term memory of what just happened. ![A small quadcopter drone hovering between tree trunks in a forest on an overcast day](/uploads/blog/liquid-neural-networks-continuous-time-ai-mit-edge-robotics-benchmarks/gemini-supporting.jpg) ### What did MIT's liquid neural network drones actually do? The headline experiment was published in Science Robotics under the title "Robust flight navigation out of distribution with liquid neural networks." The lead researchers named by MIT News are Daniela Rus, who directs CSAIL, and Ramin Hasani, a research affiliate. The task was vision-based fly-to-target navigation: a drone sees through a camera and flies toward a target it has to find. The team trained the system on data from a human pilot, then tested it in places it had not seen, including forests and urban areas. MIT News lists several tests: range and stress tests, target rotation and occlusion (when the target is partly hidden), a hike with adversaries, and flights around triangular loops between objects. The team also checked whether skills carried over from one setting to another, such as a summer forest to a winter forest. In plain terms, the drone had to find a target in scenery it had never flown through, using its camera and what it had learned from a human pilot's flights. MIT reports that the liquid networks outperformed many state-of-the-art alternatives, including in situations unlike the training data. The announcement gives no success rates, parameter counts or power figures. ### Is a liquid neural network really that small? The most quoted size comparison comes from a lane-keeping task for a car. According to VentureBeat, a classic deep network needs around 100,000 artificial neurons and half a million parameters to keep a car in its lane, while the MIT team trained a liquid network to do the same task with 19 neurons. This is the source of the often repeated "19 neurons" figure. Rus told VentureBeat that with only 19 neurons, the team could extract a decision tree describing how the network decides. In her words, "We cannot do that for 100,000 or more." A decision tree is a short list of yes-or-no questions leading to an outcome, so a person can follow it step by step. Smaller models are easier to inspect, which matters for machines such as cars and drones, where a person may need to ask why the system did what it did. These are results on narrow tasks, tested in forest and city flights. MIT's drone work does not show that a liquid network can match a large language model at writing, coding or answering questions. ### What can liquid neural networks not do yet? The MIT researchers are open about limits. Rus said liquid networks do well on time-series data but not on static datasets such as ImageNet, a standard collection of still photos used to benchmark image recognition. They are built for continuous data streams like video, audio and temperature readings, where what happens now depends on what just happened. A still photo has no before and after, which is why a design built around time gains little from it. According to VentureBeat, they had only been tested in single-robot settings, with multi-robot systems planned. In the MIT News piece, Hasani notes that the experiments tested reasoning in controlled, straightforward scenarios and that more complex challenges need testing before safe deployment. Alessio Lomuscio of Imperial College London, who was not part of the work, said the results need confirmation in other experiments. ![A researcher holding a small drone on a workbench next to a laptop in a university lab](/uploads/blog/liquid-neural-networks-continuous-time-ai-mit-edge-robotics-benchmarks/gemini-supporting.jpg) ### Are liquid neural networks the same as Liquid AI's models? Not exactly. Liquid AI is a company whose CEO and co-founder is Ramin Hasani, one of the MIT researchers on the drone paper. Its model family, called LFM2, shares people and goals with the MIT work, and the company describes the architecture in its own terms. In its [July 2025 press release](https://www.liquid.ai/press/liquid-ai-releases-worlds-fastest-and-best-performing-open-source-small-foundation-models), Liquid AI said LFM2 uses structured, adaptive operators, and that its weights are downloadable on Hugging Face. The later [LFM2 technical report](https://arxiv.org/abs/2511.23404) says the design combines gated short convolutions with a small number of grouped-query attention blocks, and reports up to 2x faster prefill and decode on CPUs than similarly sized models, a claim that comes from Liquid AI itself. It lists models from 350 million to 8.3 billion parameters, all with a 32,000-token context window. MIT's continuous-time networks for robots are one public strand. The other is a company building small language and audio models for phones and laptops. They share a goal, doing more with less compute, and a person. ### What to watch next Independent replication is the clearest gap: other labs flying their own drones with liquid networks and reporting numbers. Harder conditions are the other gap, since the MIT team itself says complex scenarios and multi-robot settings still need testing. Whether small, efficient models keep gaining ground on devices also matters. The same pressure to run AI locally shows up in [on-device NPUs](/radar/next-gen-silicon-edge-computing-on-device-npus) and in [1-bit language models](/radar/the-1-bit-llm-revolution-bitnet-b158-ternary-addition-vs-multiplication), which approach the problem from the chip and the number format. ## Frequently asked questions ### What is a liquid neural network in simple terms? A liquid neural network is an AI model whose internal timing adjusts to the data it receives, instead of staying fixed after training. It processes a continuous stream such as video, and its neurons can react faster or slower depending on the input. MIT CSAIL researchers developed it, inspired by small animals like the roundworm C. elegans. ### Who invented liquid neural networks? Liquid neural networks were developed at MIT CSAIL. Daniela Rus, the lab's director, and Ramin Hasani, a research affiliate, are the lead researchers named in MIT News coverage of the drone study. Mathias Lechner, Alexander Amini and Makram Chahine are also named. Hasani later co-founded Liquid AI, a company that builds small models for devices. ### Can liquid neural networks replace ChatGPT-style models? No evidence supports that. MIT's researchers say liquid networks suit continuous data streams like video, audio and sensor readings, and that they do not do well on static datasets such as ImageNet. The published results cover narrow robot tasks like drone flight and lane keeping, not text generation. ### How many neurons does a liquid neural network need? It depends on the task. VentureBeat reports that MIT trained a liquid network to keep a car in its lane with 19 neurons, while a classic deep network needed around 100,000 neurons and half a million parameters. That is one narrow driving task. ## Source Based on MIT News and Science Robotics. https://news.mit.edu/2023/drones-navigate-unseen-environments-liquid-neural-networks-0419 --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/liquid-neural-networks-continuous-time-ai-mit-edge-robotics-benchmarks. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/quantum-computing-for-normal-people-qubits-supercomputers-willow - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T08:10:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://blog.google/technology/research/google-willow-quantum-chip/ # Quantum Computing Explained Simply: What Qubits Really Do **Short answer:** Google's Willow chip has 105 qubits and showed error rates falling as the system grew. In October 2025 Google reported a 13,000x speedup with its Quantum Echoes algorithm. Practical uses are still unproven. **Takeaway:** Quantum computers are real and improving, but no one has yet shown a quantum calculation that is both beyond classical reach and useful in practice. **Key figures (as reported by the source):** - Willow qubits: 105 - Quantum Echoes speedup, per Google: 13,000x - Error rate of best machines today, per NIST: About 1 in 1,000 operations Quantum computing explained simply: a quantum computer stores information in qubits, which can hold a blend of 0 and 1 until they are measured, and it uses that property to solve a few specific kinds of problems far faster than a normal computer could. Today's machines are small, error-prone and still searching for their first genuinely useful job. The most talked-about machine is Google's Willow chip, announced in December 2024. Google reports that it has 105 qubits and that it cut errors as the system grew, which researchers had been chasing for roughly three decades. In October 2025 Google followed up with an algorithm it calls Quantum Echoes. ### What is a qubit, and how is it different from a bit? A bit is the unit of information in every phone and laptop. It is either 0 or 1. [NIST's explainer](https://www.nist.gov/quantum-information-science/quantum-computing-explained) compares it to a tiny bar magnet or an electric switch, which is always in one of two positions. A qubit (quantum bit) can be 0, 1, or a superposition of both. Superposition means a tiny object, such as an atom, behaves as if it has two or more distinct states at once. It stays that way until someone measures it or something disturbs it, and then it collapses to a single state. NIST describes the scaling this way: two qubits can hold a superposition of four combinations, three qubits eight, and four qubits 16. Each added qubit doubles the number of combinations. Qubits can also be entangled, which means several of them share a single quantum state. Measuring one causes the others to settle into definite states, even when they are far apart. ### Is quantum computing really a computer trying every answer at once? No. A popular picture says a quantum computer explores every path through a maze at once and simply picks the exit. NIST's explainer rejects that idea: quantum computers do not run an efficient brute-force search of all possible solutions, because the final measurement extracts only a small amount of information from the superposed results. What the machine does is arrange qubits so that wrong answers cancel each other out and right answers reinforce each other, a wave-like effect called interference. Designing a problem so that this works is hard, which is why quantum computers help with only certain tasks. NIST also expects the two kinds of machine to work side by side. Quantum computers will probably not sit on desks or in pockets. Qubits are fragile. NIST lists stray electric or magnetic fields, temperature changes and cosmic rays as things that can destroy a superposition. When that happens, the computation is ruined. According to NIST, the best quantum computers today contain hundreds of connected qubits and make an error roughly once in every thousand operations. A classical computer makes roughly one error per quintillion (10^18) calculations, and its errors are much easier to correct. The standard remedy is quantum error correction. Many physical qubits are combined into one more reliable "logical qubit," and the system keeps checking for mistakes and fixing them. More qubits also mean more places for an error to start. If each added qubit introduces more noise than the extra protection removes, building bigger machines makes things worse. ![A researcher in a cleanroom suit examines a gold-plated cryogenic cooling chamber used to keep a quantum processor near absolute zero](/uploads/blog/quantum-computing-for-normal-people-qubits-supercomputers-willow/gemini-supporting.jpg) ### What did Google's Willow chip actually show? Willow is the first chip Google describes as "below threshold." According to [Google's announcement](https://blog.google/technology/research/google-willow-quantum-chip/), the team tested grids of 3x3, 5x5 and 7x7 encoded qubits, and each step up cut the error rate roughly in half. That is the point of error correction: adding qubits made the system more reliable. The [underlying paper](https://arxiv.org/abs/2408.13687) gives the detail. A 101-qubit code (distance 7) had a logical error of 0.143% per cycle, and the error shrank by a factor of 2.14 each time the code grew by one step. The logical memory lasted 2.4 times longer than the best single physical qubit on the chip, so it passed break-even, the point where the protected qubit outperforms its own ingredients. Google also reports a speed test. On a benchmark called random circuit sampling, Willow finished a calculation in under five minutes. Google estimates that Frontier, one of the fastest supercomputers, would need about 10 septillion years (10^25) for the same task. Two caveats come from Google itself. Its Frontier estimate used conservative assumptions that favor the classical machine, and it expects classical computers to keep improving. It also says random circuit sampling has yet to show practical commercial applications. ### What did the Quantum Echoes result add in 2025? Quantum Echoes was Google's attempt to run a more meaningful task. In an announcement dated October 22, 2025, Google said the algorithm ran 13,000 times faster on Willow than the best classical algorithm on a leading supercomputer. The method sends a signal through the qubits, disturbs one of them, runs the process in reverse and measures the returning "echo." Google says the result can be checked by repeating it on the same machine or on another quantum computer of similar quality. It calls this the first verifiable quantum advantage on hardware. Google paired the result with a first test of a possible use. Working with UC Berkeley, the team ran the algorithm on two molecules, with 15 and 28 atoms. The results matched those of traditional nuclear magnetic resonance (NMR), a standard lab technique for studying molecules, and added information NMR does not usually provide. Google labels this a proof of principle. The molecules are small, and drug discovery and materials uses are described as things that "could" follow. The post is a company announcement, so independent assessments matter for judging how far the claim goes. ![A scientist in a university laboratory studies a small molecule model on a bench beside a glowing spectrometer](/uploads/blog/quantum-computing-for-normal-people-qubits-supercomputers-willow/gemini-supporting.jpg) ### What are quantum computers good for? The most credible target is simulating nature. NIST says quantum computers could model the quantum behavior of molecules, chemicals and materials, which classical computers can only approximate with great effort. Possible payoffs it lists include drug development, more efficient fertilizer production and greenhouse gas capture. NIST also mentions complex optimization problems, such as organizing airplane assembly, as possible targets. On the record so far, it notes that early demonstrations of "quantum advantage" on specific tasks have not yet proven truly useful, and some were later matched by classical computers. ### Can quantum computers break encryption? In 1994 Peter Shor published an algorithm that could factor large numbers quickly, which would put much of today's encrypted data at risk. NIST says running it would likely need millions of error-free qubits, which it describes as probably still far off. Willow has 105 qubits, so it is nowhere near that scale. The risk is still taken seriously. On August 13, 2024, NIST finalized its [first three post-quantum encryption standards](https://www.nist.gov/news-events/news/2024/08/nist-releases-first-3-finalized-post-quantum-encryption-standards), named FIPS 203, 204 and 205, designed to resist quantum attacks. NIST says some experts predict a device capable of breaking current encryption could appear within a decade, and encourages organizations to start adopting the new algorithms. Software makers, banks and governments are the ones who need to act, and browsers and messaging apps will handle the switch through ordinary updates. ### What to watch next Google's own roadmap names its next target, a long-lived logical qubit that it calls Milestone 3, and the open question is whether the result holds up when other labs repeat it. A useful, beyond-classical calculation on a real-world problem is the target Google calls "the next challenge for the field", and it says broader applications depend on a full-scale, error-corrected machine. Classical computing will respond as well. Each new advantage claim invites researchers to look for faster conventional methods, and the claim gains credibility if those fail again and again. For related reading on compute infrastructure, see our pieces on [co-packaged optics and the GPU bandwidth wall](/radar/co-packaged-optics-silicon-photonics-gpu-cluster-bandwidth-wall), or on [Google's Genie 4 world model](/radar/google-genie-4-world-model-explained). ## Frequently asked questions ### What is a qubit in simple terms? A qubit is the quantum version of a bit. A normal bit is either 0 or 1, while a qubit can be 0, 1 or a superposition of both until it is measured. NIST notes that each added qubit doubles the number of combinations a machine can hold at once. ### Are quantum computers faster than normal computers? Only for certain tasks. NIST says quantum computers will not replace classical ones and do not simply try every answer at once. Google reports Willow beat a supercomputer estimate on a benchmark, but Google also says that benchmark has no proven commercial use yet. ### Can quantum computers break encryption today? No. NIST says Shor's algorithm would likely need millions of error-free qubits, and Willow has 105. NIST finalized three post-quantum encryption standards in August 2024 as a precaution, noting some experts predict a threat could appear within a decade. ### What did Google's Willow chip prove? Google reports Willow is below threshold, meaning error rates roughly halved as it moved from 3x3 to 5x5 to 7x7 grids of encoded qubits. It does not yet run a useful real-world calculation, which Google calls the next challenge for the field. ## Source Based on Google's Willow and Quantum Echoes announcements and NIST. https://blog.google/technology/research/google-willow-quantum-chip/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/quantum-computing-for-normal-people-qubits-supercomputers-willow. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/sub-millisecond-sandboxes-docker-wasm-micro-vms-untrusted-ai-execution - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T08:10:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://github.com/firecracker-microvm/firecracker # AI Agent Sandbox: Containers, MicroVMs and Wasm Compared **Short answer:** Docker says its Wasm workloads feature is deprecated, while Firecracker and gVisor remain the documented stronger boundaries for untrusted code. **Takeaway:** Pick the lightest isolation wall that matches how untrusted the code is. **Key figures (as reported by the source):** - Wasm runtimes Docker Desktop installs: 7 - Firecracker default memory per microVM: 128 MiB - AWS services running Firecracker: Lambda and Fargate An AI agent sandbox is a walled-off space where code written by an AI runs without touching anything it should not. The sandboxing options that exist today are containers, lightweight virtual machines, a userspace kernel called gVisor, and WebAssembly. This matters because AI agents now write and run programs on their own: a chart from a spreadsheet, a data clean-up script, a quick calculation. That code can have bugs, and an attacker can trick an agent into producing harmful code. The sandbox decides how far the damage can spread. ### What is an AI agent sandbox, and why do agents need one? A sandbox is an isolated environment with strict limits on files, network access and memory. The program inside can do its work, but it cannot read your passwords, wipe your disk or reach other customers' data. Agents need one because they act on text they did not write, and the sandbox is where that risk gets contained. A web page, an email or a document can contain instructions that steer the agent. The design question is where to put the wall. A wall close to the program is cheap and fast but has more places to leak. A wall around a whole virtual computer is stronger but heavier. ### Are Docker containers safe enough for untrusted AI code? A container packages a program with its files and runs it as an ordinary process on the host computer, fenced in by features of the operating system. Containers start quickly and are easy to manage. Their weakness is that every container on a machine shares one operating system core, called the kernel. The kernel is the part of the operating system that decides what every program may touch. If code inside a container finds a bug in that shared kernel, it can reach the host and every neighbour. Security tools such as seccomp (a filter on which requests a program may make to the kernel), AppArmor and SELinux shrink the risk. The [gVisor documentation](https://gvisor.dev/docs/) describes these as rule-based mechanisms, and gVisor was built because rules alone leave the host kernel exposed to the application. For code you wrote and trust, a container is a sensible choice. For code an AI produced after reading untrusted text, many teams want a stronger wall. ### How do microVMs and gVisor isolate untrusted code? A microVM is a small virtual machine that has its own kernel and runs on hardware-enforced virtualization. [Firecracker](https://github.com/firecracker-microvm/firecracker), the best-known example, is open source and was built at Amazon Web Services for AWS Lambda and AWS Fargate. Its README says it leaves out unneeded devices to keep each microVM's memory use and attack surface small, and its default setup is 1 virtual CPU and 128 MiB of memory. Because a microVM carries its own kernel, code inside it talks to that private kernel and not the host’s, and a bug it finds there stays inside the small virtual machine. The [paper that introduced it](https://www.usenix.org/conference/nsdi20/presentation/agache) describes the problem it solves: traditional virtualization gives strong security with high overhead, while containers give minimal overhead with weaker security. The authors say Firecracker is deployed in Lambda and Fargate, supporting millions of production workloads. gVisor takes a middle path. According to its documentation, it is an application kernel written in Go that handles the program's system calls itself, so the host kernel is not directly exposed. It avoids the hardware layer that a full virtual machine needs. The documentation states the price: reduced application compatibility and higher per-system-call overhead, with some system calls and files not implemented. ![A server technician in a data centre aisle checking a row of racked machines](https://images.unsplash.com/photo-1563986768609-322da13575f3?w=1600&auto=format&fit=crop&q=85) ### Can WebAssembly work as an AI agent sandbox? WebAssembly, or Wasm, is a compact instruction format that started in web browsers. A Wasm module runs inside a runtime that the [WebAssembly security documentation](https://webassembly.org/docs/security/) says separates it from the host using fault isolation techniques. Its memory is a separate region with a fixed maximum size, and every access outside that region is bounds-checked and ends in a trap, which stops execution. A module also has to declare its functions up front, and it can reach the outside world only through the interfaces its host provides. That narrow interface is why Wasm is attractive for running untrusted snippets: a module can do only what the host chooses to hand it. If the host offers no file or network functions, the module has no way to ask for them. The same page lists limits. Race conditions, timing side channels and some nondeterminism are not prevented by the design. Wasm is also a runtime boundary, and it suits code compiled to Wasm. The arbitrary Python packages and native tools many agents want to call do not fit. ### Did Docker Wasm replace microVMs? No. [Docker's documentation](https://docs.docker.com/desktop/features/wasm/) states that Wasm workloads in Docker Desktop are deprecated and will be removed in a future release, and that the feature is no longer actively maintained. While it lasts, enabling it installs seven Wasm runtimes (including wasmtime, WasmEdge and Wasmer), and Wasm services can sit beside normal containers in one Compose stack. It also requires the containerd image store. That setup is a convenience for developers testing Wasm services on a laptop, and the documentation makes no claim about replacing virtual machines or serving as an agent sandbox. ![A glass safe door standing open in a dim vault room](https://images.unsplash.com/photo-1550751827-4bd374c3f58b?w=1600&auto=format&fit=crop&q=85) ### Which sandbox should you pick, and what does it not cover? Trusted internal scripts can run in a hardened container. Code generated from untrusted input for many customers on shared machines is where microVMs and gVisor are used. A secondary comparison by [SoftwareSeni](https://www.softwareseni.com/firecracker-gvisor-containers-and-webassembly-comparing-isolation-technologies-for-ai-agents/) places Firecracker or Kata Containers at the strongest end, gVisor in the middle and plain containers at the low-threat end, and names E2B (Firecracker) and Modal (gVisor) as users. That article is a vendor-style blog and gives its own figures. A sandbox controls what code can reach. An agent that is permitted to send email or call an API can still be talked into misusing that permission, so network limits, narrow credentials and human approval for risky steps remain separate jobs. The same reasoning applies to agents that drive a whole computer, covered in our piece on [computer-use agents](/radar/computer-use-ai-agents-operating-systems-claude-operator), and to the control loops around them in [deterministic state graphs for agents](/radar/deterministic-state-graphs-preventing-infinite-loops-ai-agent-workflows). A stronger wall usually costs memory and start-up time, which is why Wasm and containers stay popular for trusted or compiled code. Firecracker's README says its minimal design decreases startup time and memory footprint, but Firecracker publishes no boot-time figure. gVisor's documentation is more direct: it names higher per-system-call overhead, so programs that make many system calls can run poorly. It also says full virtual machines can struggle to reach high density and may need extra agents, larger footprints and slower start-up. Wasm sits at the light end because there is no operating system to boot. ### What to watch next One open question is whether Wasm runtimes grow beyond compiled code, since that is the main thing holding them back as agent sandboxes. Another is whether gVisor's compatibility gaps narrow, because its documentation names them as the main cost. A third is independent, published benchmarks that state hardware and method, which would replace the vendor figures now in circulation. For a team choosing today, the practical step is to write down what the agent's code may read, write and reach, then pick the lightest isolation that covers that list. ## Frequently asked questions ### What is an AI agent sandbox? An AI agent sandbox is an isolated environment where code written by an AI agent runs with strict limits on files, network access and memory. If the code is buggy or has been steered by an attacker, the damage stays inside the sandbox instead of reaching the host machine or other users' data. ### Is Docker enough to sandbox untrusted AI-generated code? A standard container shares the host's kernel, so a kernel bug can let code escape to the host. Tools like seccomp and AppArmor reduce that risk but do not remove it. For code produced from untrusted input, many teams add a microVM or gVisor for a stronger boundary. ### What is the difference between Firecracker and gVisor? Firecracker runs each workload in a small virtual machine with its own kernel, backed by hardware virtualization. gVisor is a userspace kernel that handles system calls itself without a virtual machine. According to its documentation, gVisor trades some application compatibility and per-call speed for lower fixed costs. ### Is Docker Wasm still supported? According to Docker's documentation, Wasm workloads in Docker Desktop are deprecated, no longer actively maintained and will be removed in a future release. While available, they install seven Wasm runtimes and can run alongside regular containers. Docker does not present the feature as an agent sandbox. ## Source Based on Firecracker, gVisor, WebAssembly and Docker documentation. https://github.com/firecracker-microvm/firecracker --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/sub-millisecond-sandboxes-docker-wasm-micro-vms-untrusted-ai-execution. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/co-packaged-optics-silicon-photonics-gpu-cluster-bandwidth-wall - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T08:05:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://convergedigest.com/nvidia-debuts-spectrum-x-and-quantum-x-photonics-switches/ # Co-Packaged Optics Explained: What It Changes in AI Networks **Short answer:** Nvidia and Broadcom have announced network switches that put the light-emitting parts inside the switch package instead of plug-in modules. Both claim large power savings for AI data centers, though the figures are vendor-reported. **Takeaway:** Co-packaged optics moves the light conversion next to the switch chip to save power, but the savings are vendor-reported and serviceability remains the open question. **Key figures (as reported by the source):** - Nvidia claimed power efficiency gain: 3.5x - Broadcom claimed optics power reduction: More than 70% - Broadcom Davisson switching capacity: 102.4 Tbps Co-packaged optics (CPO) is a way of building network switches so that the parts that turn electricity into light sit inside the same package as the switch chip, instead of in plug-in modules on the front panel. Nvidia and Broadcom have both announced switches built this way, and both say the main payoff is lower power use in AI data centers. Those power figures come from the vendors themselves, and independent measurements are scarce. ### What is co-packaged optics? A network switch is the box that moves data between servers. Today, most switches send data to a plug-in module called a pluggable transceiver. The module converts the electrical signal into pulses of light that travel down a fiber cable. This design moves that conversion next to the switch chip. According to [Converge Digest's coverage of Nvidia's launch](https://convergedigest.com/nvidia-debuts-spectrum-x-and-quantum-x-photonics-switches/), the optical parts are integrated into the switch package, so fiber connects directly to it and the separate pluggable modules disappear. The optical parts are built with silicon photonics, which means making tiny light-handling components (modulators that encode data onto light) on silicon, the same material as ordinary chips. That lets them be manufactured with chip-making equipment. ### Why do AI clusters need a different way to connect chips? Training a large AI model means thousands of GPUs (graphics processors) constantly swapping data. The network that links them has to carry enormous traffic, and each link burns power. A [Wevolver explainer](https://www.wevolver.com/article/what-is-co-packaged-optics-architecture-benefits-challenges-and-performance) describes the structural problem with pluggables. A signal has to cross roughly 15 to 30 cm of circuit board trace to reach the module on the front panel. At today's speeds that requires power-hungry signal-cleaning chips called DSPs (digital signal processors) and retimers. The article also notes that the front panel only has room for so many modules, and each one adds heat. Placing the optics beside the switch chip shortens the electrical path to millimeters, according to the same source. A shorter path needs less signal cleanup, and less cleanup means less power. ![Rows of server racks with bundles of fiber optic cables in a dimly lit data center aisle](https://images.unsplash.com/photo-1544197150-b99a580bb7a8?w=1600&auto=format&fit=crop&q=85) ### How much power does co-packaged optics actually save? The vendors report large savings, and the numbers differ by product and by what is being measured. Nvidia says its Spectrum-X Photonics and Quantum-X Photonics switches deliver 3.5 times better power efficiency than traditional networking, as [Converge Digest reports](https://convergedigest.com/nvidia-debuts-spectrum-x-and-quantum-x-photonics-switches/). The same article quotes Nvidia chief executive Jensen Huang saying the technology could cut up to 60 MW per AI data center, which it equates to the power of 100 Rubin Ultra racks. Broadcom's figures, relayed by [Converge Digest from its Tomahawk 6 Davisson launch](https://convergedigest.com/broadcom-unveils-102-4-tbps-davisson-cpo-switch-for-ai-clusters/), say optics power is more than 70% lower than with pluggables and overall interconnect power is 3.5 times lower. The Wevolver article cites an experiment in which a 30 W pluggable was replaced by a 9 W co-packaged link, a 70% reduction. It also lists energy per bit (the power needed to move one bit of data) of 15 to 20 picojoules for pluggables against 5 to 10 for co-packaged optics. A picojoule is a trillionth of a joule. These are vendor and research claims, and none comes from an audited production data center. The "3.5 times" and "70% lower" figures describe different slices of the system, so they cannot be added or compared directly. ### What did Nvidia and Broadcom actually announce? Nvidia announced its photonics switches at GTC in March 2025. According to Converge Digest, Quantum-X Photonics is a liquid-cooled InfiniBand switch with 144 ports at 800 Gbps each, and Spectrum-X Photonics is an Ethernet switch with up to 400 Tbps of total bandwidth. (A Tbps is a trillion bits per second.) The article lists partners including TSMC, Coherent, Corning, Foxconn, Lumentum and SENKO. It says Quantum-X was due in 2025 and Spectrum-X in 2026. Broadcom introduced Tomahawk 6 Davisson on October 8, 2025, calling it its third-generation CPO Ethernet switch. Converge Digest reports 102.4 Tbps of switching capacity, built from 16 optical engines of 6.4 Tbps each. Broadcom says a fourth generation with 400 Gbps per channel is in development. Broadcom's reliability claim comes with a named customer. Converge Digest reports that Meta confirmed one million cumulative 400G CPO port-device hours with zero link flaps (brief drops in a connection). ![Engineer in a clean room inspecting a small silicon wafer under bright light](https://images.unsplash.com/photo-1518770660439-4636190af475?w=1600&auto=format&fit=crop&q=85) Converge Digest describes Nvidia's switches as using microring resonator modulators, tiny ring-shaped structures that switch light on and off to carry data, and says they are made with TSMC's 3D COUPE process, which stacks electronic and photonic chips together. That stacking is what lets the optical engine sit so close to the switch silicon. Broadcom describes a different route to the same goal. Its Davisson chip combines the Tomahawk 6 switch silicon with sixteen 6.4 Tbps silicon photonics engines, per the same outlet. Both designs share the central idea of fiber leaving straight from the package. ### Why are lasers moved out of the package? Lasers are the least dependable part of an optical link, and they produce heat. The Wevolver article says early designs put lasers on the photonic chip itself, where they were the least reliable component and added heat near the modulators and the switch chip. Most designs have since shifted to external laser sources that can be replaced on their own. Broadcom's Davisson follows this pattern with field-replaceable ELSFP laser modules, which Converge Digest says keep the light source away from the hot switch package and improve serviceability. External lasers bring a cost of their own. Light has to travel by fiber from the laser to the optical engine, and Wevolver notes that the loss along that path has to be managed, so moving the lasers away from the heat solves one problem while adding another. ### What are the downsides of co-packaged optics? Serviceability is the most cited concern. A pluggable can be swapped in seconds, while Wevolver says permanently bonded fiber attachments need far more intervention, and detachable versions add signal loss, which means a failed part is harder to deal with than a plug-in module that can simply be swapped. The article reports that operators cite this as a reason to stay with pluggables. Heat is another. Photonic parts are sensitive to temperature, and Wevolver says wavelength drift can hurt performance, which often pushes designs toward liquid cooling. Nvidia's Quantum-X switch is liquid-cooled, as noted above. The same article lists fiber density, manufacturing yield, immature standards and higher per-port cost as open challenges. Wevolver expects pluggables to coexist with co-packaged optics for most of this decade. ### What to watch next Wevolver reports Quantum-X Photonics reaching availability in early 2026 and Spectrum-X in the second half of 2026, while its adoption estimates for the larger scale-up links inside a GPU cluster range from 2027 to 2030. Wevolver notes that connecting tens of thousands of GPUs with pluggables would consume hundreds of megawatts, so even a partial cut would matter to operators who are already limited by the power they can get from the grid. - Independent power measurements from an operator running co-packaged switches at scale. - Field failure and replacement data, since serviceability is the main objection. - Broadcom's 400 Gbps-per-channel generation and whether it ships on schedule. - Whether the technology moves from switches to the links between GPUs themselves. For background on the clusters this network has to serve, see [what xAI's Colossus supercomputer looks like](/radar/inside-grok-bot-xai-colossus-supercomputer-real-time-ai) and how Cerebras approaches very large models. ## Frequently asked questions ### What is co-packaged optics? Co-packaged optics is a switch design in which the components that convert electrical signals to light sit in the same package as the switch chip. Fiber connects straight to that package, so the plug-in transceiver modules on the front panel are no longer needed, which is meant to cut power use and space. ### How is co-packaged optics different from pluggable optics? Pluggable optics put the electrical-to-light conversion in a removable module on the front panel, so signals travel 15 to 30 cm across a circuit board first, according to Wevolver. Co-packaged optics shortens that path to millimeters, which reduces the signal cleanup and power needed, but makes swapping parts harder. ### Who is building co-packaged optics switches? Nvidia announced Spectrum-X Photonics and Quantum-X Photonics switches in March 2025, working with partners including TSMC, Coherent, Corning, Foxconn, Lumentum and SENKO. Broadcom introduced its third-generation Tomahawk 6 Davisson switch with 102.4 Tbps of capacity on October 8, 2025, according to Converge Digest. ### Will co-packaged optics replace pluggable transceivers? Not soon. Wevolver expects pluggables to coexist with co-packaged optics for most of this decade, and cites serviceability, heat, cost and immature standards as obstacles. Adoption estimates for links inside GPU clusters range from 2027 to 2030. ## Source Based on Nvidia and Broadcom launch coverage. https://convergedigest.com/nvidia-debuts-spectrum-x-and-quantum-x-photonics-switches/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/co-packaged-optics-silicon-photonics-gpu-cluster-bandwidth-wall. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/deterministic-state-graphs-preventing-infinite-loops-ai-agent-workflows - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T08:00:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://docs.langchain.com/oss/python/langgraph/graph-api # AI Agent State Graph: How It Stops Infinite Loops and Runaway Bills **Short answer:** AI agents can loop on a failing step and keep paying for model calls. A state graph, as in LangGraph, fixes the allowed steps in code and adds step limits, counters and checkpoints. A reported $47,000 loop shows why hard limits matter. **Takeaway:** Put the routing, step limits and budget checks for an AI agent in ordinary code, and treat every claimed reliability percentage as unproven until a method is published. **Key figures (as reported by the source):** - LangGraph default recursion limit: 1000 steps (v1.0.6+) - Reported runaway agent loop duration: 11 days - Reported runaway loop bill: About $47,000 An AI agent state graph is a fixed map of the steps an AI agent may take, with the allowed moves between them written in ordinary code. The language model works inside each step, but plain code decides which step comes next and when the whole run must stop. That design is the most common answer to a real and expensive failure: agents that repeat the same action forever. A developer's [post-mortem on DEV Community](https://dev.to/prashar32/the-47k-agent-loop-why-logging-monitoring-and-maxtokens-all-failed-to-stop-it-19ch) says that four agents ran for eleven days in November 2025 and produced a bill of about $47,000. According to the author, two of the agents kept handing work back and forth with no termination condition and no budget cap. The account is self-reported and not independently verified, but the mechanism it describes is well documented elsewhere. ### What is an AI agent state graph? A state graph has three parts: state, nodes and edges. The state is a shared record of the task so far, such as the user's request, results already collected and a count of attempts. Nodes are the steps. Edges are the arrows that say which node runs next. The [LangGraph documentation](https://docs.langchain.com/oss/python/langgraph/graph-api) describes it this way. Each node is a function that reads the state and returns an update, and it can call a language model or run normal code. Some edges are fixed. Others are conditional: a routing function looks at the state and returns the name of the next node. The routing function is regular code, so given the same state it always makes the same choice. That is what "deterministic" means here: the path between steps is predictable, even though the text the model writes inside a step is not. A graph must also be compiled before use. According to the docs, compiling checks that the structure is valid and lets you attach options such as a checkpointer, which saves progress. ### Why do AI agents get stuck in infinite loops? An agent loop is a cycle: the model picks an action, a tool runs it, the result goes back to the model, and the model picks again. The loop ends only when something says it should. If the model never decides it is done, or a tool keeps returning an error the model cannot fix, the cycle continues. LangChain's own page on the [graph recursion limit error](https://docs.langchain.com/oss/javascript/langgraph/errors/GRAPH_RECURSION_LIMIT) says that error is often caused by an infinite loop, for example two nodes whose edges point back at each other. Each pass through the loop sends the model another request, and providers bill for those requests. A loop that nobody stops therefore turns directly into spend. The DEV Community author draws a further lesson. In the post-mortem's words, the team "had logging" and "had monitoring" but "did not have a hard limit." The author argues that monitoring only reports a problem, and that a token limit on each call does not cap the total across a run. ### How does a state graph stop a loop? A state graph gives the developer places to put limits that the model cannot talk its way around. Several mechanisms are common. The first is a step ceiling. LangGraph counts "super-steps", its unit for one round of node execution, and the docs say the recursion limit sets the maximum number the graph can run in a single execution. Going over raises a `GraphRecursionError`. The docs state the default is 1000 steps from version 1.0.6, and that the limit is passed as a top-level config key when the graph is run. The second is a graceful exit before the ceiling. The docs describe a `RemainingSteps` value that a node can read, so the graph can route to a wrap-up step instead of crashing. The third is a counter held in state. A routing function can check "has this step failed three times?" and send the run to a human-review node. The fourth is a budget check on money and time. The DEV Community author argues that such a stop should be deterministic (no model in the decision), run before each call and apply to the whole run across dollars, iterations and elapsed time. ![A circuit breaker panel with a hand on the main switch in a dim electrical room](https://images.unsplash.com/photo-1509228468518-180dd4864904?w=1600&auto=format&fit=crop&q=85) ### Is a high step limit enough to prevent runaway agents? A step limit is a backstop. A limit of 1000 super-steps is generous, so a loop that makes a paid model call on every step can do a lot of damage before reaching it. LangChain's error page notes that complex graphs can hit a limit without any bug, and suggests raising the limit through config if the workflow is legitimately long. Raising the number silences the error and leaves a cycle that should not exist in place. A tighter approach sets the ceiling near the number of steps the task normally needs, and adds counters and budgets for the cases a raw step count misses. A graph that takes 8 steps for a typical task but is allowed 1000 has no practical protection. The ceiling counts steps only. Ten steps with a very large prompt can cost more than a hundred small ones. That is why the budget check on dollars matters as a separate control. ### Can you undo an agent's mistakes with checkpoints? LangGraph supports checkpointers, which save the graph's state during a run. The docs say checkpoints are written at super-step boundaries, not in the middle of a node. This helps in two ways. A run that stopped can resume from the last saved point instead of starting over, and the DEV Community author makes the same argument for checkpointing so a crash does not force a repeat of paid work. It also comes with a warning in the docs. After an interrupt or a retry, the affected node runs again from the start of its function. A node that sends an email or charges a card will do it twice unless it is written to be safe to repeat. A checkpoint saves the state of the graph. It does not reverse actions that already happened in the outside world. Internal state can be restored, but a payment already made or a record already overwritten needs its own undo logic, or a human check before the risky step. ### What are the limits of a deterministic graph? A fixed graph trades flexibility for control. An agent free to choose its own path can handle tasks nobody planned for. A graph can only follow the routes its authors drew. Teams therefore tend to mix the two. Predictable parts of a workflow, such as validation, approval and payment, sit in fixed steps. Open-ended parts, such as research or drafting, stay flexible inside a step with their own caps. A graph also does not make the model correct. If a step produces a wrong answer and the route onward checks nothing, the wrong answer moves along the track. Gates only help if they test something real, such as a schema check or a comparison against a trusted source. The numbers circulating in vendor and blog material, including percentages for cost savings, usually come from the authors' own setups, and no method or data is published for them. ![A railway signal tower beside a track junction at dusk](https://images.unsplash.com/photo-1558494949-ef010cbdcc31?w=1600&auto=format&fit=crop&q=85) ### What to watch next Framework defaults are worth checking before any agent touches real money. The recursion limit in LangGraph moved to 1000 in version 1.0.6 according to the current docs, and other frameworks set their own caps, so reading the version you actually run matters. Per-run budgets are another control. The DEV Community post argues for hard dollar and time limits enforced outside the model, and that idea is spreading into proxies and gateways that sit between an agent and its provider. Human approval steps for irreversible actions are a further control. Teams building agents that book, buy or write to databases are increasingly placing a person at those points. For related reading on agent systems, see [how multi-agent workflows are being adopted](/radar/rise-of-autonomous-ai-agents-multi-agent-workflows) and [how sandboxes limit what agent-run code can do](/radar/sub-millisecond-sandboxes-docker-wasm-micro-vms-untrusted-ai-execution). ## Frequently asked questions ### What is an AI agent state graph? It is a map of the steps an AI agent can take, written as nodes and edges in code. Ordinary routing logic decides which step runs next. The model still writes text inside each step. This makes the path predictable and gives developers fixed points to add limits, checks and approvals. ### How do you stop an AI agent from looping forever? Set several limits that the model cannot override: a maximum number of steps, a retry counter in state that routes to a human after repeated failures, and a per-run budget for money and time. LangGraph raises a GraphRecursionError when its step limit is exceeded, which stops the run. ### What is the default recursion limit in LangGraph? According to the LangGraph documentation, the default recursion limit is 1000 steps starting in version 1.0.6. It caps the number of super-steps in one execution, and exceeding it raises GraphRecursionError. Earlier versions used a lower default, so check the docs for the version you run. ### Does a checkpoint undo what an AI agent already did? No. A checkpoint saves the graph's state so a run can resume. It does not reverse actions in outside systems such as a payment or an email. The LangGraph docs also say a node reruns from its start after a retry, so actions need to be safe to repeat. ## Source Based on LangGraph documentation. https://docs.langchain.com/oss/python/langgraph/graph-api --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/deterministic-state-graphs-preventing-infinite-loops-ai-agent-workflows. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/mixture-of-depths-dynamic-compute-routing-deepmind-mod-benchmarks - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T07:55:00.000Z - Last modified: 2026-10-28T18:00:00.000Z - Primary source: https://arxiv.org/abs/2404.02258 # Mixture of Depths: How Google DeepMind Lets AI Skip Layers **Short answer:** Google DeepMind's 2024 mixture-of-depths paper trained transformers whose routers let most tokens skip some layers. Models up to 3B parameters matched or beat standard ones at equal training compute while using fewer FLOPs per forward pass. **Takeaway:** DeepMind's MoD research shows, at research scale, that a model can learn which tokens deserve full processing and save compute without losing quality, though big-model and real-world speed results are still missing. **Key figures (as reported by the source):** - Tokens processed per routed layer: 12.5% (256 of 2,048) - Largest model tested: About 3B parameters - Sampling step speed: Upwards of 50% faster Every time an AI chatbot writes a word, it pushes that word through the same long stack of calculations, whether the word is "the" or the answer to a tricky math question. Mixture of depths (MoD) is a technique from Google DeepMind, published in April 2024, that lets a language model decide which words deserve the full treatment and which can skip some of the work. The idea matters because that work costs money, electricity and waiting time. If a model can spend less effort on easy parts of a sentence without getting worse, the same hardware can do more. By the end of this piece you will know what MoD does, what the researchers actually measured, how it differs from the better-known "mixture of experts" approach, and which claims about it go further than the evidence. ### What problem was DeepMind trying to solve? A few terms first. A **transformer** is the design behind almost every modern chatbot, including ChatGPT, Gemini and Claude. It reads and writes text as **tokens**, which are small chunks of text: often a whole word, sometimes part of one, sometimes a comma. A transformer is built from a stack of **layers** (the paper calls them blocks). Each layer looks at every token, compares it with the tokens around it, and refines the model's understanding of it. Text passes through the stack one layer after another, like a document moving down a line of editors. **Compute** is the amount of arithmetic a computer does. Researchers count it in FLOPs, short for floating-point operations, which are individual multiplications and additions. More FLOPs means more chip time, more power and a bigger bill. The paper, titled [Mixture-of-Depths: Dynamically allocating compute in transformer-based language models](https://arxiv.org/abs/2404.02258), opens with the core complaint in one line: "Transformer-based language models spread FLOPs uniformly across input sequences." In plain English, a standard model gives every token exactly the same amount of effort at every layer. A full stop gets as much arithmetic as the key word in a hard question. People do not read that way. We skim the filler and slow down on the parts that carry meaning. The authors, David Raposo, Adam Santoro and colleagues at Google DeepMind (one co-author is also at McGill University and Mila), asked whether a model could learn to do something similar. ### What is mixture of depths and how does it work? MoD adds a small decision-maker, called a **router**, in front of a layer. For each token, the router produces a score that says roughly "how much does this token need this layer right now?" Only the highest-scoring tokens go through the layer's heavy calculations. The rest take a shortcut: they pass straight to the next layer unchanged. In technical terms they ride the "residual connection", a bypass that every transformer layer already has, so nothing new needs to be built for the skip. Here is the part that makes it practical. The model does not let any number of tokens through. It sets a fixed **capacity** in advance. In the setup the paper found worked best, a routed layer processes only 12.5% of the tokens, which is 256 out of a 2,048-token passage, and the authors applied this to every other layer. The layers in between still process every token. That fixed cap is the clever engineering choice. Chips like GPUs run fastest when they know the size of the job ahead of time. Because the number of selected tokens is always the same, the computer always does a known amount of work, even though which tokens get picked changes from passage to passage. The router is not a hand-written rule like "skip punctuation". It is trained along with the rest of the model and learns its own sense of which tokens matter. The paper observes that some tokens engage with almost every layer, while others "route around blocks whenever possible". The researchers also tried picking tokens at random, and that version performed "drastically worse", which tells you the learned choice is doing real work. Different tokens end up passing through different numbers of layers, so the model is effectively a mix of shallow and deep paths, which is where the name comes from. ![A software engineer at a standing desk in a quiet office at dusk, reading a printed mixture-of-depths research paper with a laptop open beside her](https://images.unsplash.com/photo-1618005182384-a83a8bd57fbe?w=1600&auto=format&fit=crop&q=80) ### What did the paper actually show? The team trained models ranging from 60 million to 3 billion parameters (parameters are the adjustable numbers a model learns; today's leading chatbots are far larger). They compared MoD models against standard transformers given the same total training budget, testing at three budgets: 6e18, 2e19 and 1e20 FLOPs. Three results stand out, all from the [paper's full text](https://arxiv.org/html/2404.02258v1): - **Equal or better quality for the same training spend.** The authors report that MoD "improves upon vanilla transformers by as much as 1.5%" on the training objective, a measure of how well the model predicts the next token, when both get the same training FLOPs. - **Less work per step.** Because most tokens skip half the routed layers, the models need, in the abstract's words, "a fraction of the FLOPs per forward pass". A forward pass is one trip of the text through the model. - **Faster steps.** The paper says the models "can be upwards of 50% faster to step during post-training sampling", meaning when the trained model generates text. It also reports training steps that are upwards of 60% faster on equivalent hardware. There was a catch the authors had to fix. Picking the "top" tokens means comparing every token in a passage, including ones later in the sentence. When a chatbot writes text one token at a time, those later tokens do not exist yet. Their solution was a second, tiny predictor that learns to guess the router's decision using only the current token. The paper reports this guess is upwards of 97% accurate and that switching to it caused little loss in quality. ### How is mixture of depths different from mixture of experts? You may have heard of **mixture of experts**, often shortened to MoE. It is the design behind several large open models, including the one in our [DeepSeek-V3 explainer](/radar/deepseek-v3-mla). The two ideas share a family resemblance, and the paper says so directly: MoD borrows MoE's routing logic, but "MoD deploys a single expert which can be dynamically skipped." The difference is what the router chooses.
Mixture of expertsMoD
What the router decidesWhich of several specialist sub-networks handles a tokenWhether a token goes through a layer at all
Does every token get processed at every layer?Yes, by some expertNo, many tokens skip routed layers
Main goalA bigger model without a matching rise in compute per tokenLess compute spent on tokens that do not need it
Put simply, mixture of experts asks "who should handle this word?" while MoD asks "does this word need handling here at all?" The two are not rivals. The paper tested a combination it called MoDE and reports that one way of combining them worked "distinctly better" than simply trimming expert capacity in a standard MoE model, though it gives these results as charts rather than headline numbers. MoD also differs from an older idea called early exit, where a token stops at some layer and never goes deeper. With MoD, the authors note, "a token can skip middle layers" and still be processed by later ones. ![A data center technician walking between tall rows of server racks with blinking status lights, the kind of hardware that runs transformer models](https://images.unsplash.com/photo-1634017839464-5c339ebe3cb4?w=1600&auto=format&fit=crop&q=80) ### What the paper did not prove It is easy to oversell this research, and earlier coverage (including our own first version of this article) did. Here is what the evidence does and does not support. **It does not show a model running twice as fast.** The "upwards of 50%" figure describes how quickly a sampling step can run, and the paper is careful about how it measured it. In its scaling charts, the authors say they "depict normalized FLOPs per forward pass rather than wall-clock step time per se." Fewer calculations usually means faster responses, but real-world speed also depends on memory, software and hardware. **It was tested on small models.** The largest model in the study had about 3 billion parameters. That is useful for research but small next to the commercial systems people use daily. Whether the gains hold at much larger sizes is an open question the paper does not answer. **It reports training quality, not real tasks.** The 1.5% improvement is on the model's own prediction objective. The paper does not report results on everyday benchmarks such as question answering or coding tests. **Memory savings were not measured.** Tokens that skip a layer also skip storing some of that layer's working notes, the cached data explained in our [KV cache guide](/radar/kv-cache-explained-cloudflare-prompt-leak). The authors mention this possible benefit but write, "We did not study this extensively." **There is no public sign it ships in a product.** The paper does not say whether Google uses MoD in Gemini or any other service, and we found no public confirmation that a major commercial model uses it. None of this makes the work less interesting. It makes it what it is: a careful research result showing that a model can learn where to spend its effort, at a modest scale. ### What to watch next The bigger question is whether the idea can be added to existing models cheaply, since few companies want to retrain from scratch. A later paper called [Router-Tuning](https://arxiv.org/abs/2410.13184), by researchers including Shwai He and Tao Ge, tries exactly that by training only the routers on a small dataset. Its authors report a 21% speedup with only a 0.2% drop in performance, a smaller gain than the original paper's best case but on models that already existed. Watch the name, too. In March 2026, a separate team published [Mixture-of-Depths Attention](https://arxiv.org/abs/2603.15619), or MoDA, which lets each layer look back at information from earlier layers. Despite the similar name, it solves a different problem, so search results that say "mixture of depths attention" may not be about the 2024 token-skipping method. For readers who just want the takeaway: AI models have long spent the same effort on every scrap of text. MoD showed, at research scale, that a model can learn to skip the easy parts and lose nothing on its training goal. If that holds at larger sizes, it is one more lever, alongside cheaper hardware and smarter caching, for bringing down the cost of running AI. ## Frequently asked questions ### What is mixture of depths in simple terms? It is a method from Google DeepMind that lets a language model skip some of its own layers for some tokens. A small trained router picks which tokens get full processing at a layer, and the rest pass straight through. In the paper's best setup, routed layers handled 12.5% of tokens, and this was applied to every other layer. ### Does MoD make AI twice as fast? No, the paper does not claim that. It reports models that can be upwards of 50% faster to step when generating text and upwards of 60% faster to step in training. The authors note their scaling charts show normalized FLOPs per forward pass rather than measured wall-clock time, so real speed in a product could differ. ### Does skipping layers make the model less accurate? In the paper's tests, no. With the same training budget, mixture-of-depths models matched standard transformers and improved on them by as much as 1.5% on the training objective. Skipping at random performed far worse, so the gain depends on the learned router. The paper did not report results on everyday task benchmarks. ### Is MoD used in ChatGPT or Gemini? There is no public confirmation of that. The 2024 paper tested models from 60 million to 3 billion parameters and does not say whether Google uses the method in Gemini. Later research, such as Router-Tuning, explores adding it to existing models more cheaply, which is the step that would make product use more likely. ## Source Based on Google DeepMind's arXiv paper. https://arxiv.org/abs/2404.02258 --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/mixture-of-depths-dynamic-compute-routing-deepmind-mod-benchmarks. Last modified 2026-10-28T18:00:00.000Z. ================================================================================ - URL: https://syntaxradar.com/radar/ebpf-modern-ai-infrastructure-kernel-observability-vs-sidecars - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T07:50:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://ebpf.io/what-is-ebpf/ # eBPF Observability: What It Replaces and Where Sidecars Stay **Short answer:** Teams are moving some monitoring from per-pod sidecar proxies to eBPF programs in the Linux kernel. Docs and papers show real gains for network and GPU tracing, while Layer 7 features still lean on proxies like Envoy. **Takeaway:** eBPF cuts the cost of basic network and system visibility, but Layer 7 features and business-level tracing still need proxies or in-app instrumentation. **Key figures (as reported by the source):** - SysOM-AI overhead (paper): Less than 0.4% - GPUs in Alibaba deployment: Over 80,000 - Beyla minimum kernel: Linux 5.8 with BTF eBPF observability means watching what software does by running small, checked programs inside the Linux kernel instead of placing a monitoring helper next to every application. It has real adoption and removes some of the cost of sidecars, though not all of it. eBPF handles basic network and system visibility well, and a large Alibaba paper reports very low overhead when tracing AI training jobs with it. Anything that needs deep knowledge of application protocols, such as retries or header rewriting, still tends to run in a user-space proxy like Envoy. ### What is eBPF, and what does it do for observability? eBPF is a Linux kernel technology. According to [ebpf.io](https://ebpf.io/what-is-ebpf/), it can run sandboxed programs in a privileged context such as the operating system kernel, extending the kernel without changing its source code or loading kernel modules. The kernel is the core of the operating system, the layer that handles files, memory and network traffic. An eBPF program attaches to a hook, which is a point where something happens, such as a system call, a function entry or a network event. When that event fires, the program runs and can record what took place. For observability, this means a tool can see which process opened a connection, how long a request waited or where packets were dropped, without editing the application. The same page lists observability, networking and security enforcement as the main uses. Because the program runs outside the application, nothing has to be added to the application's code or its container. That is why teams describe eBPF tools as zero-code instrumentation, and it is the main reason the approach appeals to platform teams who do not own every service they run. ### What is a sidecar, and why do people want to replace it? A sidecar is a helper program that runs in the same Kubernetes pod as the application. A pod is a small group of containers scheduled together. A service mesh, the layer that manages traffic between services, injects a proxy sidecar so it can encrypt, route and measure every request without changing the app. Every pod gets its own proxy, so the resource cost grows with the number of pods. That cost is the motivation for sidecar-free designs. Istio's ambient mode, described in the [Istio documentation](https://istio.io/latest/docs/ambient/overview/), is informally called a sidecar-less mesh because workload pods no longer need proxies in sidecars to join the mesh. Istio's ambient mode relies on a per-node Layer 4 proxy written in Rust rather than on eBPF. Sidecar-free and eBPF-based are two overlapping ideas, and mixing them up is common. ### Do eBPF tools actually replace sidecars? Partly. The [Cilium documentation](https://docs.cilium.io/en/stable/overview/intro/) describes a Kubernetes networking project built on eBPF. It offers flow visibility through a component called Hubble, identity-based network policy from Layer 3 to Layer 7, and a service mesh that it says avoids the cost and complexity of traditional proxy designs. Layers here describe how deep a tool reads a message. Layer 4 covers connections and ports. Layer 7 covers application protocols such as HTTP and gRPC. Solo.io sells Envoy-based products and so has a stake in the answer. It [argues](https://www.solo.io/blog/ebpf-for-service-mesh-yes-but-envoy-is-here-to-stay) that eBPF can shorten network paths, offload encryption and collect data, but cannot replace the Layer 7 proxy. Its reasoning is that eBPF programs must be fully known and checked before they run, and arbitrary loops are not allowed. That makes protocol parsing, retries and header changes hard to build in the kernel. Istio's design lines up with that split. Its optional waypoint proxies are Envoy, they handle the Layer 7 features, and Layer 7 telemetry needs them. ![A technician walking between rows of server racks in a data center aisle](https://images.unsplash.com/photo-1558494949-ef010cbdcc31?w=1600&auto=format&fit=crop&q=85) ### How does eBPF observability work on AI clusters? AI training jobs run on many GPUs that exchange data constantly through collective communication, where every worker waits for the slowest one. A single slow node can hold back the whole job, so finding it matters. A paper on arXiv, [SysOM-AI](https://arxiv.org/abs/2603.29235), describes a system deployed at Alibaba. It combines CPU profiling, GPU kernel tracing and NCCL event tracing, using eBPF-based tracing. NCCL is NVIDIA's library for GPU-to-GPU communication. The authors report less than 0.4% overhead, a deployment across more than 80,000 GPUs for over a year, and a median diagnosis time cut from days to about 10 minutes. They also say existing tools cover only one layer, add 10 to 30% overhead, or cannot run continuously. Those are the authors' own figures from their own deployment. The paper does not compare eBPF with sidecars, because traffic between training GPUs does not pass through a service mesh in the first place. It shows that eBPF tracing can run continuously on a very large fleet. For related infrastructure reading, see how [co-packaged optics address the GPU bandwidth wall](/radar/co-packaged-optics-silicon-photonics-gpu-cluster-bandwidth-wall). ![Engineer reading monitoring output on a laptop beside a rack of servers](https://images.unsplash.com/photo-1544197150-b99a580bb7a8?w=1600&auto=format&fit=crop&q=85) ### What are the limits of eBPF observability? The requirements are concrete. Grafana's [Beyla documentation](https://grafana.com/docs/beyla/latest/) says the tool needs Linux kernel 5.8 or higher with a feature called BTF enabled, and kernel 5.17 or higher for HTTP header-based trace propagation. It also needs elevated privileges, though specific capabilities can be granted in place of full root access. Beyla captures generic metrics and transaction-level trace spans for HTTP and gRPC services. Its docs still recommend language agents or manual instrumentation when a team wants fine-grained detail on critical code. Tracing support varies by language, and the docs advise validating before relying on it for reactive Java. An eBPF tool sits outside the application, which is convenient, and sees only what the kernel and the network can show. Details such as which customer a request belongs to live inside the application. A platform team can get service maps and request rates quickly, but a product team asking business questions will still need instrumentation in the code. Kernel age is the other practical limit. Clusters running older kernels cannot use tools that require newer features, and upgrading nodes is a project of its own. ### Is eBPF safe to run in production? The kernel checks every eBPF program before it attaches. Per [ebpf.io](https://ebpf.io/what-is-ebpf/), the verifier confirms that the loader has the right privileges, that the program will not crash the system, that it runs to completion and that it has no out-of-bounds memory access. After that, a just-in-time compiler turns it into machine code. The same page adds a limit. The verifier is not a security tool that inspects what programs are doing. A verified program is safe to run, yet it can still collect data it should not, so permission to load eBPF programs deserves the same care as root access. Running untrusted code is a separate problem with its own trade-offs, covered in [our look at micro-VMs and WebAssembly sandboxes](/radar/sub-millisecond-sandboxes-docker-wasm-micro-vms-untrusted-ai-execution). ### What to watch next Solo.io says kernel-side Layer 7 support may be years away. How widely sidecar-less meshes spread is also worth following, as are independent, published comparisons of eBPF and sidecar overhead on the same workload, since the figures available today come from vendors or from single deployments. For a team choosing now, a mixed setup looks sensible. Use eBPF for baseline network and kernel visibility, a proxy where Layer 7 features are needed, and in-application instrumentation for business detail. A small pilot on one node pool will show what the tool captures on your own workloads before any wider rollout. Check the kernel version of your nodes first, because that is the requirement the documentation states most plainly. ## Frequently asked questions ### What is eBPF in simple terms? eBPF is a Linux kernel feature that lets small programs run inside the operating system after the kernel checks them for safety. Those programs attach to events such as system calls or network activity, so tools can observe what software is doing without changing the application or loading kernel modules. ### Does eBPF replace sidecars? Only in part. eBPF tools can provide network visibility and Layer 4 features without a proxy in every pod. Layer 7 jobs such as retries, header changes and protocol parsing still run in user-space proxies like Envoy, which is why Istio's ambient mode keeps optional Envoy waypoint proxies for those features. ### What kernel version does eBPF observability need? It depends on the tool. Grafana Beyla's documentation lists Linux kernel 5.8 or higher with BTF enabled, and 5.17 or higher for HTTP header-based trace propagation. Red Hat Enterprise Linux 8 family systems on kernel 4.18 are also listed as supported. Check your tool's docs before planning a rollout. ### Is eBPF safe to run in production? The kernel verifies each eBPF program before it runs, checking privileges, memory access and that the program finishes. That protects stability. ebpf.io notes the verifier is not a security tool inspecting what programs do, so a verified program can still read data, and loading rights should be tightly controlled. ## Source Based on ebpf.io docs and the SysOM-AI paper. https://ebpf.io/what-is-ebpf/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/ebpf-modern-ai-infrastructure-kernel-observability-vs-sidecars. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/speculative-decoding-medusa-eagle-multi-token-prediction-benchmarks - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T07:45:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://arxiv.org/abs/2211.17192 # Speculative Decoding: How AI Guesses Ahead to Write Faster **Short answer:** Speculative decoding lets a small helper guess several tokens while the big model verifies them in one pass. Papers on the original method, Medusa and EAGLE-3 report speedups, and vLLM tests show the gain varies by model and workload. **Takeaway:** Speculative decoding can make large language models generate text faster without changing the output, but published speedups are best cases and results vary by model, method and workload. **Key figures (as reported by the source):** - Original method, T5-XXL (Leviathan et al.): 2X-3X, identical outputs - Medusa-2 speedup (paper): 2.3x to 3.6x - EAGLE-3 best case (paper): Up to 6.5x Speculative decoding is a way to make a large language model produce text faster by letting a smaller, cheaper process guess several words ahead while the big model checks the guesses all at once. The original paper, from Google researchers, reports a 2X to 3X speedup on a T5-XXL model with identical outputs. Later methods such as Medusa and EAGLE-3 report larger gains, though under their own test conditions. ### Why do language models generate text one token at a time? A language model writes by predicting a token (a word or a fragment of a word), adding it to the text, and predicting the next one. Each new token depends on all the earlier ones, so the steps are strictly sequential. The expensive part is moving data. According to the [Medusa paper](https://arxiv.org/abs/2401.10774), each decoding step has to move the full set of model parameters out of the GPU's high-bandwidth memory. The chip spends much of its time waiting for data to arrive, and one token per trip leaves most of its compute capacity idle. Because the trip to memory is a large part of each step, scoring several candidate tokens in one pass is cheaper than running several separate steps, because the parameters are read once for the whole group of candidates. A longer reply means more trips, so the saving grows with length. If you want background on what the model keeps in memory between steps, see [our explainer on the KV cache](/radar/kv-cache-explained-cloudflare-prompt-leak). ### What is speculative decoding and how does it work? The idea was published in November 2022 by Yaniv Leviathan, Matan Kalman and Yossi Matias in [Fast Inference from Transformers via Speculative Decoding](https://arxiv.org/abs/2211.17192). The paper builds on two observations: hard language tasks often contain easier pieces that a smaller model can approximate well, and a new sampling method lets the large model run in parallel on the small model's output. In practice the loop has three steps: 1. A small, fast "draft" model proposes the next few tokens. 2. The large "target" model scores all of those proposals in a single forward pass. 3. The target keeps the proposals it agrees with, up to the first disagreement, and supplies its own token at that point. When the draft is right, several tokens arrive for the price of one big-model step. When it is wrong, the loop falls back to roughly normal speed for that step. The more often the target agrees with the draft, the bigger the gain, which is why the quality of the guesses matters so much. ![A cook in a busy restaurant kitchen reading a recipe while an assistant hands over prepared ingredients ahead of time](https://images.unsplash.com/photo-1518770660439-4636190af475?w=1600&auto=format&fit=crop&q=85) ### Does speculative decoding change the answers the model gives? No, when it is implemented as described in the original paper. The abstract says the method samples "without any changes to the outputs" and without changing the distribution of results, and the T5-XXL speedup it reports comes with identical outputs. The reason is that the large model has the final say. A draft token is only kept if it passes the target model's check, so the draft model affects only speed. Proposals past the first disagreement are discarded, which means the GPU computes them for nothing. In the worst case the target rejects the very first guess and supplies its own token, so the text still follows the large model's own distribution. The Medusa paper makes a similar claim for Medusa-1, which it describes as lossless acceleration. Medusa-2, which retrains the backbone model, is the variant where the authors warn that a special recipe is needed to protect the model's abilities. ### How do Medusa and EAGLE-3 differ from the original approach? The original method needs a separate small model. Medusa and EAGLE remove or reshape that requirement. **Medusa.** According to the [Medusa paper](https://arxiv.org/abs/2401.10774) (Cai, Li, Geng, Peng, Lee, Chen and Dao), the technique attaches extra decoding "heads" to the main model so that each step proposes several future tokens. A tree-based attention mechanism lets it verify several candidate continuations together in one step. No separate draft model is needed. The paper also describes self-distillation, a way to train the heads when the original training data is not available, and a "typical acceptance" scheme meant to raise how often candidates are accepted. Neither changes the basic idea of proposing several tokens and checking them together. The authors describe two variants. Medusa-1 trains only the heads on a frozen backbone and reports over 2.2x speedup with no loss in generation quality. Medusa-2 trains the heads and the backbone together and reports 2.3x to 3.6x, but needs a special training recipe to preserve the model's abilities. **EAGLE-3.** The [EAGLE-3 paper](https://arxiv.org/abs/2503.01840) (Li, Wei, Zhang and Zhang) says earlier EAGLE versions predicted internal "features" of the target model, and that this limited how much extra training data helped. EAGLE-3 predicts tokens directly and fuses features from several layers of the target instead of only the top layer. The authors report up to 6.5x speedup across five tasks on chat and reasoning models, about 1.4x better than EAGLE-2. They also report a 1.38x throughput gain in the SGLang serving framework at batch size 64. ![A row of server racks in a data centre aisle with a technician checking one of the machines](https://images.unsplash.com/photo-1634017839464-5c339ebe3cb4?w=1600&auto=format&fit=crop&q=80) ### When does the speedup disappear? Speed gains depend on how often the target accepts the draft's guesses. The result depends on how closely the draft matches the target, how many positions are drafted, and the workload. Acceptance drops at later draft positions, so a longer draft adds checking work that is often thrown away, and separate drafters take extra GPU memory. That is how one method can speed up a model in one test and slow down another model in the same test. The [vLLM documentation](https://docs.vllm.ai/en/v0.8.2/features/spec_decode.html) for version 0.8.2 says speculative decoding in vLLM "is not yet optimized and does not usually yield inter-token latency reductions" for all prompt datasets or sampling parameters. It also lists restrictions: it is not compatible with pipeline parallelism, and EAGLE-based draft models must run without tensor parallelism. A 2026 [vLLM blog post on AMD GPUs](https://vllm.ai/blog/2026-08-23-speculative-decoding-amd-gpus) tested several methods against a baseline without speculation. Some results were strong: on gemma-4-26B-A4B-it, DFlash reached 2.87x throughput on MATH500 and EAGLE-3 ranged from 2.11x to 2.27x. On Qwen3-8B, by contrast, EAGLE-3 with one drafted token ran at 0.44x of baseline on MATH500, meaning it slowed serving down. The post says results depend on the model, the draft checkpoint and the workload. The authors recommend measuring acceptance and end-to-end throughput on your own workload before choosing settings. ### Should you use it for your own AI workload? If you run models yourself, draft models, EAGLE heads and Medusa heads each have trade-offs: a separate draft model adds memory, trained heads need training, and the best number of guessed tokens differed between models in the vLLM test. Our guide to [prefix caching in SGLang and vLLM](/radar/radixattention-dynamic-prefix-caching-sglang-vllm-benchmarks) covers another way serving engines save time. ### What to watch next - Serving-framework support: the vLLM docs describe speculative decoding as not yet optimized, so improvements there would change the numbers. - New draft methods, such as the DFlash and DSpark approaches in the vLLM AMD post, and whether they hold up outside that one test setup. - Built-in multi-token prediction (the "native MTP" in the same post) from model makers, which would remove the need for a separate draft. - Independent benchmarks on real workloads, which the best-case figures in papers do not cover. ## Frequently asked questions ### What is speculative decoding? Speculative decoding is a technique where a small, fast process proposes several next tokens and the large language model checks them all in one pass. Tokens it agrees with are kept, and it supplies its own token at the first disagreement. The original paper reports a 2X to 3X speedup on T5-XXL with identical outputs. ### Does speculative decoding change the model's output? Not in the original method. The 2022 paper from Leviathan, Kalman and Matias says it samples without changes to the outputs or the distribution, because the large model checks every proposed token. The Medusa paper describes Medusa-1 as lossless, while Medusa-2 retrains the backbone and needs a special recipe to preserve the model's abilities. ### How is EAGLE-3 different from Medusa? Medusa adds extra decoding heads to the main model and verifies a tree of candidate continuations together, with no separate draft model. EAGLE-3, according to its paper, predicts tokens directly and fuses features from several layers of the target model. Its authors report up to 6.5x speedup, about 1.4x over EAGLE-2, in their tests. ### Is speculative decoding always faster? No. The vLLM documentation says it does not usually yield inter-token latency reductions for all prompt datasets or sampling parameters. In a 2026 vLLM test on AMD GPUs, EAGLE-3 on Qwen3-8B with one drafted token ran at 0.44x of baseline on MATH500, meaning it slowed serving down. ## Source Based on the original speculative decoding paper. https://arxiv.org/abs/2211.17192 --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/speculative-decoding-medusa-eagle-multi-token-prediction-benchmarks. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/radixattention-dynamic-prefix-caching-sglang-vllm-benchmarks - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-25T07:40:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://arxiv.org/abs/2312.07104 # SGLang vs vLLM: How Prefix Caching Cuts Wait Times **Short answer:** SGLang's RadixAttention and vLLM's automatic prefix caching both reuse stored work on repeated prompt openings. Published speedups come mostly from shared-prefix workloads, and newer comparisons conflict. **Takeaway:** Prefix caching helps most when many requests share a long opening, and published speedups come mostly from such workloads. **Key figures (as reported by the source):** - SGLang paper: peak throughput gain: up to 6.4x - LMSYS 2024 blog: peak throughput gain: up to 5x - vLLM prefix cache eviction: Least recently used SGLang vs vLLM comes down to one question: how much of your prompt do your requests have in common? SGLang is built around a prefix cache called RadixAttention, and vLLM offers a similar feature called automatic prefix caching. Where requests share long openings, both skip repeated work. Where they share nothing, neither has much to skip. The headline speedups you see online come mostly from the first situation. Both tools are inference engines: the software that sits between a trained language model and the people sending it requests. Prefix caching is the trick of remembering work done on the start of a prompt so it is not repeated. ### What is RadixAttention in SGLang? RadixAttention is SGLang's method for reusing the model's working memory across requests. It was introduced by the LMSYS team in a [January 2024 announcement](https://lmsys.org/blog/2024-01-17-sglang/) and described in the [SGLang paper](https://arxiv.org/abs/2312.07104) by a twelve-author team that includes Lianmin Zheng, Ion Stoica and Joseph E. Gonzalez. When a model reads a prompt, it stores intermediate results for every token in what is called the KV cache (key-value cache), so it does not have to reread earlier words while writing a reply. Our [explainer on the KV cache](/radar/kv-cache-explained-cloudflare-prompt-leak) covers the basics. Normally that memory is thrown away when a request finishes. According to LMSYS, RadixAttention instead keeps the KV cache for both prompts and generated answers in a radix tree. A radix tree is a lookup structure that stores sequences of items along shared branches, the way a file system shares folder paths. Here the items are tokens (word fragments), so two prompts that begin the same way share the same branch and the same stored memory. ### How does prefix caching work, step by step? Reading a prompt is called prefill. It is the first phase of every request, and it must finish before the first word of the reply appears. The delay between sending a request and seeing that first word is the time to first token. If a new request begins with text the engine has already processed, prefill for that part can be skipped. Only the new tail needs computing. The longer the shared opening, the more work is saved. LMSYS describes three supporting pieces. The tree itself is kept on the CPU, while the stored tensors stay on the GPU. When GPU memory runs short, a least-recently-used policy evicts the oldest leaf nodes first. A cache-aware scheduler also tries to order incoming requests so that more of them find a match in the tree. The same post says the design works with continuous batching and paged attention, two standard techniques for keeping GPUs busy, and that the authors saw no noticeable overhead when there were no cache hits. ![A technician in a data center aisle checking rows of server racks lit by blue status lights](https://images.unsplash.com/photo-1544197150-b99a580bb7a8?w=1600&auto=format&fit=crop&q=80) ### Is vLLM's prefix caching different from RadixAttention? Both engines do the same job, but they track shared text differently. SGLang follows branches of a tree, while vLLM looks up fingerprints. vLLM's [design documentation](https://docs.vllm.ai/en/latest/design/prefix_caching/) says it hashes each block of the KV cache. A hash is a short fingerprint computed from content. The fingerprint of a block combines the fingerprint of the block before it, the tokens inside it, and extras such as adapter IDs or multimodal inputs. Only full blocks are cached, and eviction is least-recently-used as well. A new request is checked block by block against those fingerprints, so a match is found by lookup. The docs describe the feature as letting a query that shares a prefix with a cached one reuse that cache. Their [feature page](https://docs.vllm.ai/en/latest/features/automatic_prefix_caching.html) lists multi-turn chat and repeated questions about one long document as the cases where it helps. Because only full blocks are cached, a prefix that ends partway through a block cannot be reused beyond the last full block. How much that matters in practice depends on block size and the workload. The same feature page says automatic prefix caching speeds up prefill only. It does nothing for long generated answers, and it does nothing for new queries that share no prefix with anything cached. ### SGLang vs vLLM: what do the benchmarks actually show? The original SGLang paper reports "up to 6.4x higher throughput" than state-of-the-art inference systems across several tasks, including agent control, retrieval-augmented generation, JSON output and multi-turn chat. The LMSYS blog post, which tested earlier, says "up to 5 times higher throughput" against vLLM v0.2.5, Guidance and Hugging Face TGI, on Llama-7B using one NVIDIA A10G GPU and on Mixtral-8x7B using eight. Two cautions apply. First, those baselines are from late 2023 and early 2024. vLLM has since added automatic prefix caching, so the comparison no longer matches what a reader faces today. Second, "up to" means the best case among the workloads tested. Newer comparisons are mostly blog posts, and their numbers conflict. One 2026 post on DEV Community reports median time to first token of 380 ms on vLLM against 85 ms on SGLang for a shared-prefix test. Another post from Spheron reports 1,920 tokens per second for SGLang against 1,850 for vLLM at 50 concurrent requests on an H100, which is close to a tie. These are single-author tests with their own settings and software versions. A [SwiftKV paper](https://arxiv.org/pdf/2410.03960) on arXiv mentions SGLang giving 1.4x to 1.8x higher throughput than vLLM on Llama-3.1-8B. That is a smaller gap than the original claims and was measured in a different context. ![A developer comparing two sets of performance numbers on dual monitors in a quiet office](https://images.unsplash.com/photo-1526374965328-7f61d4dc18c5?w=1600&auto=format&fit=crop&q=80) ### Which one should you choose? Prefix caching pays off when many requests begin with the same material. Examples include a long system prompt reused by every user, a tool list that an agent resends on each step, or a document that many questions refer to. In those cases, skipping repeated prefill cuts waiting time before the first word. It pays off less when prompts are mostly unique, or when the reply is very long. The vLLM docs say outright that prefix caching does not speed up the generation phase. If your cost is dominated by long answers, reusing prompt memory moves little. Features beyond caching also matter when picking an engine, such as model support, structured output, hardware coverage and how easily it fits your existing deployment. To compare them, take a sample of real traffic, run it through both engines on the same hardware with prefix caching switched on in vLLM, and compare time to first token and throughput at the concurrency you expect. The vLLM page says to set `enable_prefix_caching=True` to turn the feature on, and does not state the default, so check your version. ### What to watch next Neutral head-to-head tests on current versions are missing. Today most comparisons come from vendors, hosting companies and individual bloggers, rather than from a neutral lab using matched settings. Memory pressure is another open area. Both rely on least-recently-used eviction, which is simple but can discard a prefix that a busy service is about to need again. The SGLang team's cache-aware scheduler is one answer to that problem, and vLLM's documentation describes its own eviction rules in detail. Research on newer model designs is a further area. A paper on arXiv titled [Marconi: Prefix Caching for the Era of Hybrid LLMs](https://arxiv.org/pdf/2411.19379) suggests that prefix caching is being rethought for models that are not plain transformers. If model designs keep changing, caches built around one design may need rework. Our piece on [DeepSeek's multi-head latent attention](/radar/deepseek-v3-mla) covers another example of how a change in memory design affects serving costs. ## Frequently asked questions ### What is RadixAttention? RadixAttention is SGLang's way of reusing the model's stored memory (the KV cache) across requests. According to LMSYS, it keeps the cache for prompts and generated answers in a radix tree, so requests that begin with the same tokens share the same stored work instead of recomputing it from scratch. ### Does vLLM have prefix caching? Yes. vLLM documents automatic prefix caching, which lets a new query reuse the cache of an earlier query that shares its opening. It works by hashing blocks of the cache. The feature page says to set enable_prefix_caching=True to turn it on and does not state the default, so check your version. ### Is SGLang faster than vLLM? It depends on the workload. The SGLang paper reports up to 6.4x higher throughput against older baselines, but newer blog comparisons conflict, with one showing a near tie on throughput. Prefix caching helps most when requests share long openings, so the only reliable answer comes from testing your own prompts. ### When does prefix caching not help? The vLLM documentation says it only speeds up the prefill phase, not token generation. It does not help with long generated answers, and it does not help when a new query shares no opening with anything already cached. Mostly unique prompts see little benefit. ## Source Based on the SGLang paper and vLLM docs. https://arxiv.org/abs/2312.07104 --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/radixattention-dynamic-prefix-caching-sglang-vllm-benchmarks. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/rise-of-autonomous-ai-agents-multi-agent-workflows - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-24T09:00:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://www.anthropic.com/engineering/multi-agent-research-system # Multi Agent Systems: What Anthropic and Cognition Found **Short answer:** Anthropic reports a multi-agent research system beat a single agent by 90.2% on its internal eval at roughly 15x the tokens of chat, while Cognition says only setups with one writing agent have worked. **Takeaway:** Multi-agent systems help on parallel, high-value research tasks and struggle when agents must share context or edit the same work. **Key figures (as reported by the source):** - Gain over single agent (Anthropic internal eval): 90.2% - Token use vs chat (multi-agent, Anthropic): About 15x - Bugs per PR caught by Devin Review (Cognition): 2 on average Multi agent systems are groups of AI agents that split a job between them, each with its own instructions, tools and working memory. [IBM describes them](https://www.ibm.com/think/topics/multiagent-system) as agents that collaborate to reach shared, system-wide goals. The idea is real and in production, but the evidence for it is narrower than the marketing around "autonomous AI agents replacing software" suggests. The best public numbers come from two companies that disagree with each other. Anthropic reports a large gain from a multi-agent research feature. Cognition, the maker of the coding agent Devin, argued in 2025 that most multi-agent designs are fragile, then wrote in 2026 about the setups that do work. ### What are multi agent systems? An AI agent is a language model that runs in a loop: it reads a goal, picks a tool (a web search, a code editor, a database query), looks at the result and decides what to do next. A multi agent system puts several of those loops side by side and has them hand work to one another. IBM lists three common layouts. In a centralized layout, one unit holds the shared knowledge and connects the agents, which makes communication easy but means a failure there can take everything down. In a decentralized layout, agents trade information with their neighbours, which is more robust but harder to coordinate. In a hierarchical layout, agents sit in a tree with different levels of authority. The practical appeal is specialization and scale. One agent can search the web while another reads documents and a third checks the answer, each with a fresh, uncluttered context (the amount of text a model can hold in mind at once). ### How does a multi-agent setup work in practice? [Anthropic's engineering team published](https://www.anthropic.com/engineering/multi-agent-research-system) the most detailed public description, for the Research feature in its Claude apps. A lead agent reads the user's question, writes a plan, and saves that plan to memory so it survives if the conversation outgrows the model's context limit. It then starts several subagents, each with its own context window, tools and sub-task. Anthropic says the lead typically spins up three to five of them in parallel. The subagents search, then return what they found. The lead agent combines the findings and decides whether more research is needed. A separate citation agent finally attaches sources to the claims in the answer. This pattern is usually called orchestrator-worker. The orchestrator plans and merges; the workers do the reading and searching, and each returns only its findings to the lead agent. Anthropic says parallel work cut research time by up to 90% for complex queries. ![Two colleagues dividing a stack of printed reports between them at a shared desk while a third reviews a finished summary](https://images.unsplash.com/photo-1558494949-ef010cbdcc31?w=1600&auto=format&fit=crop&q=80) ### Is a multi-agent setup better than a single agent? On the right kind of task, according to Anthropic. It reports that a lead agent running Claude Opus 4 with Claude Sonnet 4 subagents beat a single Claude Opus 4 agent by 90.2% on its internal research evaluation. That is Anthropic's own test on its own product, so it shows the approach can work and says little about other tasks. The same post attributes much of the gain to token spending. On a web-browsing benchmark called BrowseComp, Anthropic found that token usage alone explained about 80% of the differences in performance. A token is a small chunk of text, and the more tokens an agent reads and writes, the more it can look at. Multiple agents are a way to spend more tokens on a problem than one agent's context window allows. That spending is large. Anthropic reports that agents use roughly four times the tokens of an ordinary chat, and multi-agent systems roughly fifteen times. Anthropic says the task has to be valuable enough to justify that cost, which rules out low-value jobs. ### Why did Cognition tell people not to build multi-agent setups? In June 2025, Cognition engineer Walden Yan published ["Don't Build Multi-Agents."](https://cognition.com/blog/dont-build-multi-agents) It gave two principles: share full context between agents (their complete traces), and remember that every action carries implicit decisions that can conflict with another agent's. His example was a request to build a Flappy Bird clone. The work was split into two sub-tasks: a game background, and a bird. One subagent produced a Super Mario style background, the other a bird that looked and moved nothing like Flappy Bird's. A final agent then had to reconcile the mismatch. Even with more shared context, the post says, the two kept making conflicting assumptions about visual style that nobody had specified. Yan's recommendation was a single-threaded agent, optionally with a model that compresses the history for long jobs. Anthropic's own list of weak spots overlaps with his: tasks where agents need the same context or depend heavily on each other, and most coding tasks. ![A developer at a desk reviewing code on two monitors with a notebook of handwritten notes beside the keyboard](https://images.unsplash.com/photo-1558494949-ef010cbdcc31?w=1600&auto=format&fit=crop&q=80) ### Which multi agent setups work today? Cognition revisited the question in a post dated April 22, 2026, ["Multi-Agents: What's Actually Working."](https://cognition.com/blog/multi-agents-working) Its main finding is that multi-agent systems work when only one agent writes at a time and the extra agents contribute judgement instead of actions. It names four patterns that work. The first is a code-review loop, where one agent writes code and a second reviews it with a clean context. Cognition says Devin Review catches an average of two bugs per pull request (a proposed code change), about 58% of them severe. The second is a "smart friend" setup, where a smaller model asks a larger one for help on hard problems. The third is a manager that splits a job among child agents and merges the results. The fourth is read-only subagents that search the web or a codebase and change nothing. What did not work: Cognition calls unstructured swarms, where arbitrary agents negotiate with each other, "mostly a distraction." It also reports that a smart-friend setup failed when the primary model was too far behind the helper to know when to ask. The open problems it lists are about communication: when a weaker model should escalate, how child agents share discoveries with siblings, and how to pass context without overloading the receiver. Both companies find parallel agents that read and report useful, while several agents all editing the same thing at once remain risky, because each makes its own assumptions about the work. ### Are AI agents replacing SaaS software? Nothing in the Anthropic or Cognition posts shows that. They describe agents that call existing tools, such as web search, code search and code review. Tool access is the layer that makes that possible. Standards such as the Model Context Protocol, covered in our explainer on MCP in production, give agents a common way to reach those tools. Whether that adds up to agents replacing dashboards is a prediction, and the public numbers do not settle it. [Gartner's forecast](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027), published in June 2025, leans the other way. It predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear business value or inadequate risk controls. That is an analyst's projection, and it matches the cost and reliability themes in the engineering posts. ### What to watch next Anthropic lists problems that remain open. Errors compound in long-running agents, so it uses checkpoints and retries. Runs are non-deterministic, which makes failures hard to diagnose. Lead agents currently wait for each batch of subagents to finish, and running them asynchronously would help speed but add complexity. Small changes to the lead agent can also change subagent behaviour in ways nobody predicted. For a product that claims to use many agents, ask whether the task splits into independent pieces or the pieces depend on each other, whether a correct answer is worth roughly fifteen times the token bill, and whether a single agent makes the final edits with a reviewer using clean context. Anthropic says early versions of its system searched endlessly and spawned too many subagents for simple queries. Our piece on [stopping infinite loops in AI agent workflows](/radar/deterministic-state-graphs-preventing-infinite-loops-ai-agent-workflows) covers one way teams constrain that. Cognition says enterprise use of Devin grew about eightfold over the prior six months, while swarms of agents that all write changes still see little meaningful adoption. If that pattern continues, the designs that spread will be the plain ones: a clear planner, a few read-only helpers and a reviewer at the end. ## Frequently asked questions ### What is a multi-agent system in AI? They are groups of AI agents that work together on a task. Each agent has its own instructions, tools and context, and they share results to reach a common goal. IBM describes them as agents collaborating on tasks for a user or another system. Common layouts are centralized, decentralized and hierarchical, each with different trade-offs in robustness and coordination effort. ### How does a multi-agent system differ from a single AI agent? A single agent works through a task alone in one loop. A multi-agent system splits the task, so several agents can work in parallel, each with a fresh context window. Anthropic reports this helped on research tasks but cost roughly fifteen times the tokens of a chat, versus about four times for a single agent. ### Why do some engineers avoid multi-agent designs? Cognition argued in June 2025 that agents make conflicting assumptions when work is split, and that full context is hard to share. Its Flappy Bird example produced mismatched pieces from two subagents. In April 2026 it said setups where only one agent writes, and others review or search, do work, while unstructured swarms are mostly a distraction. ### Are multi-agent setups worth the cost? It depends on the task. Anthropic says the work must be valuable enough to justify the extra tokens, and that jobs needing shared context or many dependencies suit them poorly, including most coding tasks. Research that splits into independent searches is the clearest fit. Gartner separately forecasts that over 40% of agentic AI projects will be canceled by the end of 2027. ## Source Based on Anthropic's engineering post. https://www.anthropic.com/engineering/multi-agent-research-system --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/rise-of-autonomous-ai-agents-multi-agent-workflows. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/deepseek-v3-mla - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-24T08:00:00.000Z - Last modified: 2026-10-28T18:00:00.000Z - Primary source: https://arxiv.org/abs/2405.04434 # Multi-Head Latent Attention: How DeepSeek Shrinks AI Memory **Short answer:** DeepSeek introduced multi-head latent attention in DeepSeek-V2 and kept it in DeepSeek-V3. It compresses the memory a model keeps about earlier text; DeepSeek's paper reports a 93.3% smaller KV cache than its older 67B model. **Takeaway:** MLA stores one compressed summary per token instead of full keys and values for every head, which DeepSeek credits with a far smaller KV cache and cheaper long conversations. **Key figures (as reported by the source):** - KV cache vs DeepSeek 67B: 93.3% smaller - Max generation throughput: 5.76x higher - Tokens/sec on 8 H800 GPUs: 50,000+ When a chatbot answers you, it has to keep notes on everything already said in the conversation. Those notes live in expensive chip memory, and on long chats they become one of the biggest costs of running an AI model. Multi-head latent attention, usually shortened to MLA, is the design DeepSeek introduced to make those notes much smaller. DeepSeek first described MLA in its [DeepSeek-V2 paper](https://arxiv.org/abs/2405.04434) in May 2024 and kept it for [DeepSeek-V3](https://arxiv.org/abs/2412.19437), its larger open model released in December 2024. The V2 paper reports that the new model needed 93.3% less of this conversation memory than DeepSeek's earlier 67-billion-parameter model. By the end of this piece you will know what that memory is, how MLA shrinks it, what the published numbers do and do not show, and why other labs have since borrowed the idea. ### What is multi-head latent attention? Start with three terms. A **token** is a small chunk of text, often a word or part of a word. Language models read and write text one token at a time. **Attention** is the step where the model decides which earlier tokens matter for the next one. When it writes "she" in a story, attention is how it looks back and works out who "she" is. Modern models run many attention "heads" side by side, each looking for a different kind of pattern, which is why the standard version is called multi-head attention. MLA keeps that multi-head idea but changes what gets stored. Instead of saving a full, separate record for every head, it saves one compact compressed summary per token and rebuilds what each head needs from it. "Latent" is the machine learning word for that kind of hidden, compressed representation. ### Why does an AI model need so much memory to chat? To decide what comes next, attention compares the new token against every earlier one. For each earlier token the model computes two lists of numbers, called a **key** and a **value**. Roughly, the key says what that token is about and the value holds the information it contributes. Recomputing those for the whole conversation at every step would be slow, so models save them. That saved store is the **KV cache** (key-value cache). We covered how it works, and how it can leak, in our [KV cache explainer](/radar/kv-cache-explained-cloudflare-prompt-leak). The catch is that the cache grows with every token, every head and every layer of the model. The DeepSeek-V2 paper says the heavy KV cache of standard multi-head attention is a bottleneck that limits how efficiently a model can generate text. A bigger cache means fewer conversations fit on one chip at the same time, and that drives up the cost of serving each user. ![An engineer checks a row of server racks with blinking status lights in a cool data center aisle, the kind of hardware where DeepSeek MLA saves memory](https://images.unsplash.com/photo-1620712943543-bcc4688e7485?w=1600&auto=format&fit=crop&q=80) ### How does multi-head latent attention work? Picture a newsroom where 128 editors each keep their own full file on every sentence of a long story. The filing cabinets fill up fast. MLA replaces that with one short shorthand note per sentence, plus a set of fixed decoding rules that let each editor expand the note into the detail they need, when they need it. In technical terms, MLA squeezes each token's keys and values into a single small vector before caching it. The DeepSeek-V2 paper calls this "low-rank joint compression". According to [the paper's full text](https://arxiv.org/html/2405.04434v5), DeepSeek-V2 has 128 attention heads of 128 numbers each, yet the cache stores one compressed vector of 512 numbers per token per layer, plus a 64-number piece that carries position information. That position piece exists for a reason. Models need to know word order, and the usual method (called rotary position embedding) does not mix well with the compression step. DeepSeek's fix was to keep a small, separate slice for position, which the paper calls "decoupled" position embedding. There is one more trick. You might expect that rebuilding full keys and values for every head would cost a lot of extra computing. The paper explains that the decompression math can be folded into other steps of attention when the model is running, so the full keys and values do not have to be rebuilt every time. As [Sebastian Raschka's architecture comparison](https://magazine.sebastianraschka.com/p/the-big-llm-architecture-comparison) puts it, MLA compresses keys and values "into a lower-dimensional space before storing them in the KV cache." The compression is learned, not bolted on later. The model is trained with MLA from the start, so it learns what to keep in those 512 numbers. ### What did DeepSeek's papers actually report? The headline figures come from the [DeepSeek-V2 paper](https://arxiv.org/abs/2405.04434), which compares the new model with DeepSeek 67B, the company's earlier model:
What was measuredDeepSeek-V2 vs DeepSeek 67B
KV cache size93.3% smaller
Maximum generation throughput5.76 times higher
Training cost42.5% lower
The paper also reports that, served on a single machine with 8 Nvidia H800 GPUs, DeepSeek-V2 generated more than 50,000 tokens per second and processed prompts at more than 100,000 tokens per second. Read these numbers with care. They compare two whole models, not MLA against standard attention in isolation. DeepSeek-V2 also uses a "mixture of experts" design, where only 21 billion of its 236 billion parameters are active for each token. For the throughput test, the paper says the team converted the model to FP8, a lower-precision number format, and compressed each KV cache element to about 6 bits on average. MLA is a large part of the memory saving, but not the only ingredient. DeepSeek-V3 then carried MLA forward at a bigger scale. Its [technical report](https://arxiv.org/abs/2412.19437) describes a model with 671 billion total parameters, 37 billion active per token, and says MLA and the expert design were "thoroughly validated in DeepSeek-V2." The report says full training took 2.788 million H800 GPU hours. The weights are public on [GitHub](https://github.com/deepseek-ai/DeepSeek-V3), which is what made the design easy for others to study. If the term is new to you, our guide to [open-weight models](/radar/open-weight-models-explained-ibm) explains what that release model means. ![A developer at a desk reviews a long document on a laptop late in the evening, the kind of long conversation where multi-head attention memory adds up](https://images.unsplash.com/photo-1558494949-ef010cbdcc31?w=1600&auto=format&fit=crop&q=85) ### Is MLA better than standard multi-head attention? Usually, shrinking memory costs some accuracy. The DeepSeek-V2 paper claims MLA avoids that trade. It says MLA "achieves superior performance compared with MHA" (multi-head attention) while needing a much smaller KV cache, based on comparison tests reported in the paper's appendix. Raschka's review reads those same tests the same way: in DeepSeek's experiments, MLA slightly outperformed standard multi-head attention. That is DeepSeek's own evidence, run on its own models and settings. Independent results on other model sizes may differ, and the company did not claim the gain holds everywhere. The honest summary: MLA is not free. It adds steps and complexity to the model, and it only pays off because the memory saved is worth more than the extra math. ### What is the difference between MLA and GQA? MLA was not the first attempt to slim the KV cache. A popular earlier approach is **grouped-query attention**, or GQA, used in many open models. GQA puts the attention heads into groups and lets each group share one set of keys and values. Fewer distinct keys and values means a smaller cache. The difference is in how they save space. GQA shares; MLA compresses. GQA throws away some per-head detail by making heads share. MLA keeps a single dense summary that every head can expand into its own version. The DeepSeek-V2 paper puts a number on it. It says MLA's cache is about the same size as GQA with only 2.25 groups, a very aggressive setting, yet it reports stronger results than full multi-head attention. In DeepSeek's tests, GQA scored below standard attention, while MLA scored above it, according to Raschka's summary. The trade-off is engineering effort. Raschka notes that MLA is more complicated to implement than GQA, which helps explain why GQA remains common in smaller models. ### What to watch next The clearest sign that MLA works is that others now copy it. According to Raschka's comparison, Moonshot AI's Kimi K2 uses the DeepSeek-V3 architecture, including MLA, with more experts and fewer attention heads. DeepSeek keeps building on it too. Its [DeepSeek-V3.2 paper](https://arxiv.org/abs/2512.02556), published in December 2025, adds DeepSeek Sparse Attention, which picks only the most relevant earlier tokens to look at. The paper says that new layer is built on top of MLA rather than replacing it. For everyday users the effect is indirect. Smaller memory needs per conversation mean one chip can serve more chats at once, which lowers the cost of each one. Watch for more open models adopting MLA, for serving software adding faster support for it, and for independent tests that check whether its quality claims hold outside DeepSeek's own lab. ## Frequently asked questions ### What is multi-head latent attention in simple terms? It is a way for an AI model to keep shorter notes about the conversation so far. Instead of storing full keys and values for every attention head, it stores one compressed summary per token and expands it when needed. DeepSeek introduced it in DeepSeek-V2 to cut the memory needed during text generation. ### How much memory does DeepSeek MLA save? The DeepSeek-V2 paper reports a 93.3% smaller KV cache than DeepSeek 67B, its earlier model. That compares two whole models, not MLA alone. DeepSeek-V2 also uses a mixture-of-experts design, and its serving setup used FP8 weights and roughly 6-bit KV cache compression, so several choices contribute. ### Does MLA make AI models less accurate? DeepSeek says no. Its V2 paper reports that MLA performed better than standard multi-head attention in its own comparison tests while using a much smaller cache. Those tests were run by DeepSeek on its own models, so independent checks on other model sizes and settings are still worth watching. ### Which AI models use DeepSeek MLA? DeepSeek-V2, DeepSeek-V3 and DeepSeek-R1 use it, and DeepSeek-V3.2 builds its sparse attention on top of it. According to Sebastian Raschka's architecture comparison, Moonshot AI's Kimi K2 also uses the DeepSeek-V3 design with MLA, adjusted to have more experts and fewer attention heads. ## Source Based on DeepSeek's V2, V3 and V3.2 papers. https://arxiv.org/abs/2405.04434 --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/deepseek-v3-mla. Last modified 2026-10-28T18:00:00.000Z. ================================================================================ - URL: https://syntaxradar.com/radar/next-gen-silicon-edge-computing-on-device-npus - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-24T07:00:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://en.wikipedia.org/wiki/Neural_processing_unit # What Is an NPU? The AI Chip in Your Laptop and Phone Explained **Short answer:** NPUs are now standard in phones and laptops, and vendors advertise them with TOPS ratings. Those figures are theoretical peaks, and Microsoft's Copilot+ label sets a 40 TOPS floor. **Takeaway:** An NPU is a low-power AI helper chip, and its TOPS rating is a theoretical peak that says little about real speed. **Key figures (as reported by the source):** - Apple M4 Neural Engine (Apple): 38 trillion operations per second - Copilot+ PC NPU minimum (Microsoft): 40 TOPS - Snapdragon X Elite Hexagon NPU (Qualcomm): 45 TOPS An NPU, short for neural processing unit, is a small processor built to run the math behind AI models using far less power than a general-purpose chip. If you have bought a phone or laptop in the last few years, there is a good chance one is already inside it. For what these chips let phones do offline, see our separate explainer on [on-device phone AI and privacy](/radar/on-device-phone-ai-local-models-apple-android-privacy). ### What is an NPU, in one paragraph? According to [Wikipedia's entry on neural processing units](https://en.wikipedia.org/wiki/Neural_processing_unit), an NPU is a specialized hardware accelerator for artificial intelligence and machine learning workloads such as neural networks and computer vision. It can be a standalone chip or a block inside a CPU or GPU. On consumer devices the same entry describes NPUs as meant to be small, power-efficient and reasonably fast when running small models. An NPU is a low-power helper for jobs like background blur on a video call, live captions, photo cleanup and voice recognition. ### How is an NPU different from a CPU and a GPU? A CPU (central processing unit) is the general-purpose brain. It handles the operating system, your apps and any task that involves lots of branching decisions. It can run AI math, but it is not built around it. A GPU (graphics processing unit) does huge numbers of simple calculations in parallel, which is why it became the standard engine for AI in data centers. Wikipedia notes that GPUs typically run AI workloads through existing general-purpose pipelines such as CUDA and OpenCL, adapted for lower precision. An NPU narrows the design further. Wikipedia describes NPU designs as often "manycore" or "spatial," with an emphasis on low-precision arithmetic, new dataflow layouts or in-memory computing. Low precision means storing numbers with fewer bits, such as 8-bit integers (INT8) or even 4-bit (INT4), which is cheaper in energy and good enough for many AI models. The gap is shrinking, since the same source points out that modern GPUs now include dedicated units for low-precision matrix multiplication too. In a modern laptop the three processors often share the work, and the software decides which one gets each task. ### What does TOPS mean, and does a bigger number mean a faster NPU? TOPS stands for trillions of operations per second. A chip rated at 40 TOPS can, at its theoretical peak, perform 40 trillion basic operations each second. Wikipedia adds that the metric does not specify the type of operation, which is typically INT8 additions and multiplications. That is why TOPS is a weak shopping tool. Intel's Robert Hallock, the company's VP of AI and Technical Marketing, called it "kind of a goofy metric" and a theoretical peak, as reported by [Laptop Mag](https://www.laptopmag.com/laptops/windows-laptops/intels-lunar-lake-is-here-to-change-how-we-think-about-ai-pcs-because-an-npu-isnt-enough-for-the-ai-market). He argued that peak figures have some value but do not show how well hardware performs in practice. A TOPS rating leaves out what decides real-world speed: - Memory speed, because a model's weights have to be fed to the chip fast enough. - Software support, because an app must be written to use the NPU at all. - Precision, since a TOPS figure quoted at INT8 is not comparable to one quoted at a different number format. - Heat and battery limits, which cap how long a chip can hold its peak. ![A technician holding a bare computer processor chip under bright workshop light](https://images.unsplash.com/photo-1518770660439-4636190af475?w=1600&auto=format&fit=crop&q=80) ### Which NPUs actually exist? Apple, Qualcomm, Intel and AMD **Apple Neural Engine.** Apple's announcement of the M4 chip says its Neural Engine is capable of "38 trillion operations per second," which Apple calls [60x faster than the first Neural Engine in A11 Bionic](https://www.apple.com/newsroom/2024/05/apple-introduces-m4-chip/). Wikipedia dates Apple's AI accelerators in its silicon to around 2017, which lines up with the A11. **Qualcomm Hexagon.** Qualcomm's launch specifications for the Snapdragon X Elite list a Hexagon NPU rated at 45 TOPS, as summarized by [Gigazine](https://gigazine.net/gsc_news/en/20231025-qualcomm-snapdragon-x-elite). [Notebookcheck](https://www.notebookcheck.net/Qualcomm-releases-official-Snapdragon-X-Plus-and-Snapdragon-X-Elite-benchmarks-for-45-TOPS-Hexagon-NPU.841811.0.html) reports that this is a theoretical figure that applies to all Snapdragon X versions. **Intel.** Intel's Lunar Lake generation includes an NPU rated at up to 48 TOPS. Laptop Mag reports that Intel counted the whole platform at up to 120 TOPS, made up of 48 on the NPU, 67 on the GPU and 5 on the CPU, and that Intel was still finalizing those numbers when it spoke. The earlier Meteor Lake chips carried an Intel-built VPU (versatile processing unit) for inference, per Wikipedia. **AMD.** AMD's NPU sits inside its Ryzen AI chips and is built on the XDNA architecture. Coverage of the Ryzen AI 300 launch, such as [this Laptop Mag comparison](https://www.laptopmag.com/laptops/windows-laptops/intels-lunar-lake-chips-to-clash-with-snapdragon-x-and-ryzen-ai-300-from-september), puts its XDNA 2 NPU at up to 50 TOPS. All of these are vendor peak ratings. Different companies measure them under their own conditions, so a 45 versus 48 versus 50 difference says very little about which laptop will feel faster. ### Why does 40 TOPS keep showing up on laptop boxes? Microsoft set it as a floor. According to [TweakTown's report](https://www.tweaktown.com/news/97159/microsofts-copilot-ai-will-run-locally-on-pcs-that-have-at-least-40-tops-of-npu-performance/index.html), a PC needs an NPU of at least 40 TOPS to qualify as a Copilot+ PC. [Neowin reports](https://www.neowin.net/news/microsoft-updates-windows-11-minimum-processor-requirements-guidance-for-ai-pc-support/) that Microsoft also requires 16 GB of RAM and 256 GB of fast storage for those machines. The threshold explains why the numbers across Qualcomm, Intel and AMD cluster just above 40. It also means the badge tells you only that the NPU clears a bar. Coverage of the requirement also notes that 40 TOPS is plenty for lightweight tasks such as live captions and webcam effects, but is not enough on its own to run a large language model locally at a pleasant speed. Check Microsoft's own page for the current requirements, since this was reported at launch in 2024. ![A person working on a thin laptop at a wooden desk near a window](https://images.unsplash.com/photo-1518770660439-4636190af475?w=1600&auto=format&fit=crop&q=80) ### Do you need an NPU, or will the GPU do? For most people the answer is that you will get one whether you ask or not. NPUs are now built into mainstream chips, so the real question is whether the software you use takes advantage of it. The efficiency argument is the NPU's strongest card: a dedicated block can do small, constant AI jobs without waking the power-hungry GPU. But Intel's own executive said an NPU alone is not enough for AI workloads, and Laptop Mag reports that Intel's 2025 roadmap assumed a split of about 30% NPU, 40% GPU and 30% CPU. Even the company selling the NPU expects the GPU to do the largest share. If you are buying a laptop, an NPU is a reasonable tiebreaker but a poor main reason. Memory, battery life and the apps you run matter more. If you are buying specifically to run large local models, GPU power and memory capacity are likely to matter more than the NPU's TOPS label. For how Apple's current chips compare, see our [MacBook Pro M6 vs M5 upgrade guide](/radar/macbook-pro-m6-vs-m5-upgrade). ### What to watch next Whether everyday software beyond Microsoft's own features targets NPUs is the first test of their value. Vendors also face pressure to move to more comparable measurements than a single peak number, since even Intel's AI chief doubts TOPS. The split of work between NPU, GPU and CPU will keep shifting as models grow. Each vendor also keeps raising its ratings. Qualcomm's own product page now lists higher figures for its newer X2 chips, which makes the 45 TOPS first-generation number a moving target. Before you trust any new rating, look for independent testing of real tasks, and for the precision and conditions behind the figure. ## Frequently asked questions ### What does NPU stand for? NPU stands for neural processing unit. It is a specialized hardware accelerator for AI and machine learning workloads such as neural networks and computer vision. It can be a separate chip or a block built into a CPU or GPU, and on consumer devices it is meant to be small and power-efficient. ### What is the difference between an NPU and a GPU? A GPU runs many simple calculations in parallel and handles AI through general-purpose software pipelines. An NPU is designed around low-precision AI math and low power use. The gap is narrowing, because modern GPUs also include units for low-precision matrix multiplication, and laptops often split AI work across both. ### What does TOPS mean for an NPU? TOPS means trillions of operations per second. It is a theoretical peak and usually refers to INT8 math. Intel's Robert Hallock has called it a goofy metric that does not show real-world performance. Memory speed, software support and heat limits all affect what you actually experience. ### How many TOPS does a Copilot+ PC need? Microsoft requires an NPU of at least 40 TOPS for the Copilot+ PC label, according to reports at launch in 2024. Neowin also reported a 16 GB RAM and 256 GB fast-storage requirement. Requirements can change, so check Microsoft's current Windows 11 page before buying. ## Source Based on Apple, Qualcomm and Intel disclosures. https://en.wikipedia.org/wiki/Neural_processing_unit --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/next-gen-silicon-edge-computing-on-device-npus. Last modified 2026-10-29T10:33:58.900Z. ================================================================================ - URL: https://syntaxradar.com/radar/from-prompt-engineering-to-agentic-engineering-tech-stack - Author: Abdul Rehman, Lead Systems Architect and Technical Director - Published: 2026-10-24T06:00:00.000Z - Last modified: 2026-10-29T10:33:58.900Z - Primary source: https://simonwillison.net/guides/agentic-engineering-patterns/what-is-agentic-engineering/ # Agentic Engineering: What It Is and How It Differs From Prompting **Short answer:** Agentic engineering, building software with coding agents that run tools in a loop, is replacing single-prompt tricks as the way teams work with AI. **Takeaway:** Agentic engineering is the discipline around the model, meaning tools, memory and tests, and Anthropic advises adding agents only when a simpler design falls short. **Key figures (as reported by the source):** - Agent definition (Simon Willison): Runs tools in a loop to achieve a goal - Starter eval set (Anthropic): 20 to 50 simple tasks - Anthropic's pass-three-in-a-row example: About 42% at 75% per trial Agentic engineering is the practice of building software with the help of coding agents: programs that write code, run it, read the result and try again. Prompt engineering is the craft of wording a single request to a language model. The two overlap, but they describe different jobs, and the second is now a small part of the first. The term is young, and people use it in more than one way. ### What is agentic engineering? Developer Simon Willison defines it as [developing software with the assistance of coding agents](https://simonwillison.net/guides/agentic-engineering-patterns/what-is-agentic-engineering/). He defines an agent in one line: it runs tools in a loop to achieve a goal. In practice, a program asks a language model what to do, runs the tool the model picks (such as a code editor or a test runner), feeds the output back, and repeats. Willison says the ability to run code is what makes the approach work, because an agent can check whether its own output functions. The human still decides what to build, supplies the tools, describes the problem in enough detail and verifies the result. Andrej Karpathy, who coined "vibe coding" in February 2025, is widely reported to have proposed "agentic engineering" a year later as a better name for the professional version of the practice. Reports quote him as saying "agentic" fits because you are not writing the code directly most of the time, and "engineering" signals that skill is involved. ### How is agentic engineering different from prompt engineering? Prompt engineering is about how to write and organize the instructions you give a model. Anthropic describes [context engineering](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents) as its natural progression: deciding which information the model sees at each step, which goes well beyond the wording of one prompt. This practice sits one level higher. The prompt is still there, but so are the tools the agent can call, the memory it keeps between steps, the checks that decide whether its work is accepted, and the person who reviews the outcome. A language model is the AI system that predicts text. A harness is the surrounding program that feeds the model information, runs its tool requests and handles errors. One secondary source separates "agent engineering" (building that harness) from agentic engineering (using agents to build software). Others use the words interchangeably, so check which meaning an author intends. ![A developer at a desk reviewing a long list of code changes on a monitor while a laptop beside them shows a running test](https://images.unsplash.com/photo-1526374965328-7f61d4dc18c5?w=1600&auto=format&fit=crop&q=80) ### What does the agentic engineering stack look like? No official standard defines an agentic engineering stack. Four ingredients show up repeatedly. The first is the model itself, which reasons about the task and decides the next step. The second is the tools, which let it act: reading files, searching, running code. Anthropic's guidance says tool design deserves as much care as prompt design, and suggests making tools harder to misuse, for example by requiring absolute file paths. The third is memory and context. Anthropic names several techniques: compaction (summarizing a long conversation and restarting from the summary), structured note-taking (the agent writes notes outside its context window and reads them back later), and sub-agents that work in clean context windows and return short summaries. The fourth is verification, covered below. Everything else, including the choice of framework, is secondary to those four. A team that gets the tools and tests right can swap the model later; a team that skips them has little way to tell whether a new model is better or worse. ### Do you need an agent at all? Often the answer is no. Anthropic's essay on [building effective agents](https://www.anthropic.com/engineering/building-effective-agents) advises starting with the simplest solution and adding complexity only when needed. It notes that agentic systems trade latency and cost for better task performance, and that an agent's autonomy brings higher costs and the potential for compounding errors. The essay draws a line between two designs. Workflows are systems where models and tools follow predefined code paths. Agents are systems where the model directs its own process and tool use. Workflows suit well-defined tasks. Agents suit open-ended problems where the number of steps cannot be predicted in advance. Anthropic lists patterns inside the workflow side, including prompt chaining (a task split into sequential model calls), routing (sorting an input to a specialized follow-up) and parallelization (running subtasks at the same time). It also lists an evaluator-optimizer pattern, where one model produces a response and another critiques it in a loop. On frameworks, the same essay says they simplify common tasks but can obscure the underlying prompts, which makes debugging harder. Its advice is to start by calling model APIs directly. ![A small workshop bench with a half-assembled robotic arm, hand tools and a handwritten checklist pinned to the wall](https://images.unsplash.com/photo-1526374965328-7f61d4dc18c5?w=1600&auto=format&fit=crop&q=80) ### How do you test an AI agent? Anthropic's guide to [evals for AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) defines an eval as a test: you give the agent an input, then apply grading logic to what it produces. Agents are harder to test than a single prompt because they use tools across many turns, change state along the way and can compound early mistakes. Anthropic calls one test a task and each attempt at it a trial. Teams run several trials because results vary from run to run. The guide describes three kinds of graders. Code-based graders (pass/fail tests, string matching) are fast, cheap and objective but brittle. Model-based graders use another model with a scoring rubric, which handles open-ended work but is less predictable and needs calibration against human judgment. Human graders are the gold standard but slow and costly. Anthropic suggests starting with 20 to 50 simple tasks drawn from real failures. It also warns against grading the path an agent took instead of the result, and recommends reading the transcripts (the full record of a trial) to confirm the eval is fair. Reliability needs care. The guide gives an example: an agent that passes a task 75% of the time per trial has about a 42% chance of passing three trials in a row, which is 0.75 multiplied by itself three times. ### Is agentic engineering just vibe coding with a new name? Willison argues no. He wants "vibe coding" to keep its narrow meaning from Karpathy's original post: prompting a model while you forget the code even exists, which he associates with unreviewed, prototype-quality output. The practice, in his framing, aims at code held to a production standard, with verification and iteration treated as central. Reports of a Sequoia event describe Karpathy using a similar split, with vibe coding raising the floor (anyone can build software) and agentic engineering raising the ceiling. There is a real disagreement in the secondary coverage about whether vibe coding is finished or simply one end of a range. Both can be true: a weekend prototype and a payment system warrant different levels of review. ### What to watch next Anthropic's own essay notes that much of the tooling it describes has changed since December 2024 and points readers to newer documentation, which shows how fast this area moves. Specific products named in any "stack" list are likely to date quickly. One thing to watch is whether the vocabulary settles, since "agent engineering," "agentic engineering" and "context engineering" are still used loosely. Another is how teams publish eval results, because claims about agent reliability mean little without the task set and grading method behind them. A third is how coding agents are run in isolated environments, a topic covered in our look at [sandboxes for untrusted AI code](/radar/sub-millisecond-sandboxes-docker-wasm-micro-vms-untrusted-ai-execution). For a related view of how agents are kept from running away, see our piece on [state maps that stop infinite loops](/radar/deterministic-state-graphs-preventing-infinite-loops-ai-agent-workflows). ## Frequently asked questions ### What is agentic engineering in simple terms? Agentic engineering means building software with coding agents, which are programs that write code, run it, read the results and try again. According to developer Simon Willison, the human still decides what to build, provides the tools, describes the problem in detail and verifies the output until it can be trusted. The agent does the repetitive typing and testing. ### Is prompt engineering obsolete? Not according to Anthropic, which calls context engineering the natural progression of prompt engineering. Writing clear instructions still matters. What changed is scope: agent builders now also manage tools, memory, what information the model sees at each step, and tests. ### How is agentic engineering different from vibe coding? Willison keeps vibe coding to its original meaning: prompting a model while forgetting the code exists, usually yielding unreviewed prototypes. Agentic engineering, in his framing, targets production-quality code and treats verification and iteration as central. Commentators disagree on whether vibe coding is finished or just one end of a range, since prototypes and critical systems need different review. ### Do I need an AI agent or is a simple workflow enough? Anthropic advises starting with the simplest solution and adding complexity only when needed, which often means not using an agent at all. Workflows follow predefined code paths and suit well-defined tasks. Agents choose their own steps and suit open-ended problems, but they cost more, run slower and can compound errors across steps. ## Source Based on Anthropic engineering guides and Simon Willison. https://simonwillison.net/guides/agentic-engineering-patterns/what-is-agentic-engineering/ --- Published on SyntaxRadar (https://syntaxradar.com). Canonical URL: https://syntaxradar.com/radar/from-prompt-engineering-to-agentic-engineering-tech-stack. Last modified 2026-10-29T10:33:58.900Z.