Eight signals in three days: the AI capability funnel, 30,000 merges a day, and a Chrome full chain in six days

In the three days to October 14, 2026, eight items from the AI and tech accounts this channel follows were worth carrying. Andrej Karpathy reopened an argument he published in April and concluded that the gap in how people understand AI capability has turned into a funnel. Anthropic described defences that have cut the reasoning its models leak to distillation campaigns by an order of magnitude, and explained why it now runs its own version-control system. XBOW took a Chrome full-chain exploit from a standing start to a $250,000 award in six days. Google's Gemini 4 Argon taught a humanoid to play cornhole from a video, and François Chollet set out what he thinks the ideal outcome for science would be. The window runs three days because the editions of October 12 and October 13 were not published.

Source scope: the fixed set of public AI and tech accounts this channel monitors, while the personal X connection is not linked. One of those accounts, Benedict Evans, can no longer be read on X, so its posts are absent from this edition.

Karpathy: the gap in how people understand AI capability has become a funnel Andrej Karpathy's post on X · 42,902 likes and 6.38 million views on the post at capture · October 11

Andrej Karpathy, a founding member of OpenAI, revived an argument he first made in April because the gap he described then has widened 1. Where he once saw two groups talking past each other, he now describes a sharp funnel. His rough arithmetic puts about 6 billion people, some 75 per cent of the world, in the group that has barely met a large language model.

A second group of 1 to 2 billion uses free tiers occasionally and sees a chatbot a little better than a search box. About 20 million people, 0.2 per cent, work with frontier models professionally and watch agent runs finish projects that used to take weeks. Around 5,000 people, 0.00006 per cent, have internal access to frontier systems 1.

Karpathy gives three drivers for the funnel: impact grows with the ambition and time horizon of a problem, capability clusters in work that is digital, verifiable and economically valuable, and access splits three ways. The figures are his own napkin maths, an argument about distribution. Hacker News carried the post as a story with 717 points 2.

Ethan Mollick, a Wharton professor who writes about AI adoption, added a qualification under the post. Of the roughly 5,000 people inside the labs with access to the most advanced models, many have yet to push them far, and their use concentrates on cybersecurity, mathematics and AI research itself 3.

A Chrome full chain in six days: $40,000 of model time and a $250,000 award XBOW's exploit writeup, carried by OpenAI co-founder Greg Brockman on X · 4,997 likes and 420,182 views on Brockman's post at capture · October 13

XBOW, a security company that runs AI agents against software, built a full-chain exploit of Chrome between Sunday August 2 and Friday August 7. Its agents used OpenAI's gpt-5.6-sol model for vulnerability search, proofs of concept and the individual exploit stages, while XBOW researchers chose targets and decided which pieces to join 4. Reconstructing usage from archived Codex logs put the six-day campaign at roughly $40,000 in API charges, and Google paid the chain the first $250,000 full-chain reward under the bug-bounty rules it introduced in April 2026 4.

For scale, XBOW says a four-person team needed six to nine months for a Chrome full chain before language models were in the loop, and its own first submission took two months. A full chain crosses both of Chrome's boundaries, giving an attacker code execution on the host. This full chain used three bugs: a V8 Maglev compiler flaw that handed the exploit read and write access across the V8 pointer-compression cage, an animated-WebP decoding bug that produced one chosen native write, and a Crashpad bug in the unsandboxed crash handler that turned that write into execution 4.

The middle stage turns on an assumption Blink makes about V8's sandbox. XBOW exploited a property the sandbox deliberately leaves unenforced: a transferred ArrayBuffer can still be changed by an attacker holding an in-cage view, which Blink assumed it could trust. Chromium fixed each stage, and XBOW points to root-cause reports in Chromium issues 547819997 and 548509153. OpenAI co-founder Greg Brockman framed the result as a reason for defenders to use the same tools, and put the population exposed by such a vulnerability at an estimated 1 billion users 5.

Anthropic says its distillation defences cut the reasoning distillers can extract by 10 to 100 times Anthropic's policy post, and its post on X · 6,132 likes and 2.24 million views on Anthropic's post at capture · October 12

Distillation attacks are covert, industrial-scale campaigns that harvest one model's outputs to train a rival, usually through networks of fraudulent accounts. The most valuable material is the model's written reasoning, because it teaches the copy how the original reached an answer 6. Anthropic reports that its current defences reduced the data distillers harvested by around 90 to 99 per cent, and that the reasoning tokens extracted on its own platform now run 10 to 100 times lower than before the new safeguards.

The defences work in three places. New API accounts can no longer edit a model's prior context, a change Anthropic calls preserved thinking, and one model's reasoning is no longer readable by earlier models with weaker safeguards. A two-stage classifier blocks extraction attempts with a false positive rate under 0.1 per cent, the same class of safeguard Anthropic uses against bioweapon and cyber misuse. Accounts that keep trying lose quota and then get banned, which Anthropic says took under two hours in tens of thousands of cases 6.

Anthropic's stated example is Zhipu, which distilled cyber capabilities from United States models while building GLM-5.3. Zhipu tried Claude Fable 5, found that its cyber safeguards degraded the attacks, and switched to an older Claude model and another lab's model 6. The 90 to 99 per cent figure is Anthropic's estimate of traffic on its own platform, and the company says it has yet to see a sophisticated, persistent attempt to get around the system.

Anthropic: small models may be months from a self-replicating worm on cheap IoT devices Anthropic's Frontier Red Team report, carried by Ethan Mollick on X · 527 likes and 34,737 views on Mollick's post at capture · October 12–13

William Zhou and Daniel Freeman, on Anthropic's Frontier Red Team, tested the least-studied step of an AI worm: getting a model running on hardware it has never seen 7. A 9B model set up inference on 8 of the 9 hardware types tested. To check whether cheap devices carry enough compute, Anthropic spread a 26B model across 15 off-brand Android TV boxes it bought for about $55 each, and served 0.39 tokens per second.

Anthropic's assessment is that models of 30B parameters and under already complete many of the individual replication steps, and that GLM-5.2 and stronger models manage the full cycle in simplified evaluations. If small models keep improving at their historical rate, Anthropic expects them to reach that level by the end of 2026 7.

A worm like this has no central point to switch off, and it could spread with a doubling time measured in days. Ordinary malware spreads where it already knows the device; this kind can work out unfamiliar hardware on its own. These are Anthropic's own simplified evaluations of a threat model, and the report sets aside the exploit stages of replication on the grounds that earlier work covers them 7. Ethan Mollick drew the wider point: capability thresholds matter more than the steady climb, and this one is the level needed for a self-replicating worm on devices like off-brand televisions 8.

Anthropic replaced git because its agents now merge 30,000 changes a day Anthropic's engineering blog, and its post on X · 20,552 likes and 5.06 million views on Anthropic's post at capture · October 13

Anthropic merged around 2,500 code changes a day in May 2026. By September it was merging more than 30,000 a day, and the bottleneck had moved from writing code to the developer infrastructure around it 9. Git stopped fitting that load: the main branch's history alone is nearly 100 GB, a compressed working copy is over 25 GB, and Anthropic runs on the order of 25 million code checkouts a day for CI alone.

The replacement is Scion, an in-house version control system built on the assumption that clients are always online. The server answers questions about file history, creates commits and performs merges, while the client keeps a bounded cache and checks out only the directories a job asks for. Anthropic migrated its monorepo off git in five months, and reports that its source-control infrastructure now needs 200 times less CPU and memory than the several hundred git mirrors it once ran 9.

Checkouts went from minutes to seconds, which freed 25 per cent of CI CPU capacity on its own. A team of seven built Scion and Antique, the code review tool beside it, and Anthropic says Claude rewrote a hand-written ingestion pipeline into one using 20 times less memory and running four times faster. Anthropic's own conclusion is that the design suits enormous monorepos, that it has no version of Scion to ship, and that it is offering the ideas to git and the platforms built on it 9.

Paul Graham: nearly all the standard startup advice still applies in the age of AI Paul Graham's thread on X · 1,000 likes and 115,361 views on the opening post at capture · October 13

Paul Graham, co-founder of Y Combinator, answered the question founders keep asking at YC's office hours: how much of the standard startup advice still applies now that AI can do much of the work. At this point nearly all of it, he wrote, with the caveat that he cannot promise the same in ten years 10.

Two things changed in his account. A founder no longer always needs a cofounder to do the work, because AI can take a lot of it, although doing the work was never the main reason for a cofounder; sharing the stress was, and AI cannot do that 11. Telling founders to move fast because they have nothing to lose carries more weight now than it used to 12.

His concrete application is moats: do not pass on an opportunity to grow only because it has no moat, he wrote, since moats often emerge later 13. These are the judgments of an investor who talks to founders constantly; no data sits behind them.

Gemini 4 Argon taught an Apptronik humanoid to play cornhole from a video Google's developer blog and Google DeepMind's post on X · 2,358 likes and 162,941 views on the cornhole post at capture · October 13

Google made Gemini 4 Argon globally available on October 13, and put the same model into its robotics work the same day 14. On the Embodied Reasoning benchmark, which tests physical and spatial problems in 3D environments, Google says Argon scores 87 per cent from reasoning alone and beats every other frontier model 15. In the demonstration, Argon mapped the physics from a video, wrote its own simulation and taught Apptronik's Apollo humanoid to sink seven throws in a row 16.

The rest of Google's post shows the same capability across a robotics stack. A Boston Dynamics Spot practises stepping on a foot-pedal bin inside a MuJoCo simulation before trying it on the real bin. A camera feed, an egocentric human video, or a hand-drawn sketch becomes a calibrated simulation. Google reports its own figure for the payoff on a tabletop manipulation task: after Argon generated more than 4,500 domain-randomized training episodes, a vision-language-action model's real-world success rate rose from 56 to 75 per cent and completion time fell from 35 seconds to 24 15.

Both numbers are Google's, one from a benchmark it built and one from a task it designed, and the cornhole run is Google's own demonstration. The new material in this post is the physical reasoning and the robotics workflows built on it.

Chollet: the ideal outcome for science is millions of AI-assisted scientists François Chollet's post on X · 1,185 likes and 33,001 views on the post at capture · October 14

François Chollet, who created the Keras deep-learning library and co-founded the ARC-AGI benchmark, wrote that the ideal outcome for science would be millions of human scientists doing more ambitious work with AI 17. He set that against a smaller version of the same technology: a couple of labs spending millions of dollars to crack trophy problems.

The question underneath is who ends up holding the capability and what they do with it. Chollet has been making the neighbouring argument for months: science is a recursively self-improving system whose inputs grow exponentially while its measured output has been roughly flat since the industrial revolution 18.

Chollet states a preference here and reports no result, and the post carries no supporting document. His earlier posts on the same subject make the case with numbers; this one makes it with a position.