Four threads run through today’s issue, and the first is a pile of big claims waiting on checkers. Anthropic says a Claude model beat a famous computing barrier, Axiom says its agents formalized a 15,000-page proof, and Cognition’s tougher test cut the best coding score to 45.5 percent.
The second is who owns the chips. Oracle will sell Google’s TPUs, with OpenAI reportedly a buyer, while Nvidia and AMD are putting $4 billion into Meta stock.
The third is Claude moving into medicine. Anthropic shared results from six drug developers, and a founder says it bought his one-employee biotech.
The fourth is AI getting eyes, hands and memory, from Google’s Genie 4 to a robot planner now on Google Cloud.
Proof on Trial: Big Claims Waiting for Outside Eyes
Two labs say AI cracked problems that stumped people for decades, and a new coding test shows how far the scores can fall when the grading gets stricter.
- Anthropic says its Claude model also broke a famous computing barrier. Anthropic posted a paper and a Lean proof saying an internal Claude model found a fast algorithm for 3-SAT, a day after OpenAI made a similar claim. Outside experts have not confirmed either, and Anthropic’s own page says the result tells us nothing about solving real formulas.
- AI agents checked math’s longest proof and found 34 problems. Axiom Math says its agents turned a roughly 15,000-page proof about symmetry into computer-checked code in 24 days, and that 34 departures from the published text needed real fixes. The company mentions no outside audit yet.
- Cognition’s tougher coding test drops the best AI score to 45 percent. On FrontierCode 2.0, all eight models Cognition lists scored below their old marks, and the top one, Claude Opus 5.5, fell from 54.6 to 45.5 percent. Cognition sells the Devin coding agent and has not published a Devin score on the new test.
Follow the Chips: Who Sells Them, Who Funds Them, Who Gets Them
The money around AI hardware is moving in odd directions: a rival cloud now sells Google’s chips, and suppliers are buying shares in their own customer.
- Google lets Oracle sell its AI chips, and OpenAI is reportedly a buyer. Alphabet says Oracle will offer Google’s TPU chips to its customers, and CNBC confirmed after The Information that OpenAI plans to buy them. TD Cowen models $62.5 billion in TPU sales for 2027, while Wells Fargo models $99 billion.
- Nvidia and AMD are buying $4 billion of Meta stock. Nvidia is putting in $3 billion and AMD $1 billion, CNBC reported, and both already sell Meta the hardware behind its AI. Neither has said whether the investments are tied to how many chips Meta buys.
- Nearly three-quarters of the world’s new AI computing power went to the US. Epoch AI’s customs-based tracker estimates the US took 72 percent of AI compute delivered between 2022 and June 2026, with China at 10 percent once its home-built chips are counted. Epoch says the method misses places like Taiwan, where chips never cross a border.
Claude Goes to the Lab: Anthropic Moves Into Medicine
Anthropic spent this week showing off its drug-development work, and a prominent critic has a theory about when the company acts on principle and when it acts like everyone else.
- Anthropic shows what Claude did inside six drug developers’ labs. Divergence Labs got about 35,000 antibody candidates in a month, and Enveda says it cut document review time by 89 percent. Anthropic says nearly all of it is still preclinical, and every figure comes from Anthropic and its partners.
- Anthropic buys a biotech that had one employee and a lot of Claude. Cody Tranbarger, the biotech poster known as Professor Oak, says Anthropic acquired Mithril Medicine and hired him to find out how many drug programs one person can run at once. His thread gives no price or deal terms.
- A critic says Anthropic acts like two companies at once. Zvi Mowshowitz argues Anthropic’s costly stands follow CEO Dario Amodei’s personal priorities, and that where Amodei is less worried the company behaves like any other business. It is one commentator’s reading, and some of it rests on people he does not name.
Eyes, Hands and Memory: AI Steps Out of the Chat Window
Models are learning to simulate worlds, direct robots, remember more and shrink onto phones, though most of it is still a demo, a preview or a lab result.
- Google’s Genie 4 builds playable worlds with sound and talking characters. Google DeepMind’s new world model renders explorable scenes as you type, with its own sound and characters who talk back. Only trusted testers have access, and the post offers no benchmark scores.
- Google puts its robot-brain AI on Google Cloud for businesses. Gemini Robotics ER 2 watches a robot’s video feed, judges whether each step worked, and directs the machine. It is in public preview, with no pricing or general availability date, and the performance figures are Google’s own.
- Google researchers give AI a sleep cycle so it stops forgetting. Google Research describes a method where a deployed model pauses to turn what it just learned into lasting memory, and reports gains on long-context, fact and math tests. The post does not name the size of the model tested.
- Liquid AI ships six small models that see, hear and decide on-device. Liquid AI’s LFM3 family has six open-weight models of about 2 billion parameters each. Its own tests put the text model ahead of Alibaba’s larger Qwen3.5-9B on an average across agent, coding, knowledge and long-context tests, though Qwen wins on coding and science.
Today’s Quick Hits
- A video AI that guesses what you want before you ask. A 16-author arXiv paper pairs a new test set with a model that watches a live video stream and updates its guess about the user’s goal. The reported 10.7 percent gain comes from the authors’ own benchmark.
- Meta releases a virtual home to train AI that sees through your eyes. Meta’s research group open-sourced a simulator where one AI plays the wearer and another the assistant, and a task passes only if the 3D world ends up right. The license is non-commercial, so it is mostly for labs.
- Visual Studio now compares old commits against your uncommitted code. Microsoft’s October update adds bulk merge-conflict fixes, a rebuilt Copilot chat and a Hot Exit mode that keeps unsaved files across restarts. The new chat and Hot Exit both stay off until someone switches them on.
- AI agents can now buy SEO and sales data one lookup at a time. Glasser gives coding agents one key to more than 2,000 paid data endpoints and bills per call. Every price comes from Glasser’s own page, and it is unclear whether the data providers signed off.