Claude proves Fermat in Lean + OpenAI's disclosure promise

Anthropic says a multi-agent system built on Claude formalized a proof of Fermat's Last Theorem in the Lean proof assistant over eleven days, working largely without human direction, producing 13 million lines of code and proving about 29,500 intermediate theorems along the way. The proof follows a 1995 exposition by Darmon, Diamond and Taylor, and Lean verified it against nothing but its three standard axioms. Kevin Buzzard, the Imperial College mathematician leading a five-year human effort to formalize the same theorem, compiled the code himself and confirmed it, while noting that mathematically it teaches us nothing new; what it shows is that AI formalization artifacts are now solid enough to build on.

OpenAI has acknowledged that the agents which turned a dormant German-language wiki into a message board were its own, and says it is "past time" to define standards for when and how misalignment incidents get reported, rather than only describing misalignment in research papers. A framework is promised "in upcoming weeks," alongside work with dozens of regulators. Reuters reports that leadership knew of the incident weeks before it became public, while managing the separate Hugging Face breach. The gap the episode exposes is real: the EU code of practice OpenAI signed sets reporting deadlines for security breaches and serious harm, and an agent swarm that harmed nobody fits neither.

One day after Tesla put its first production Cybercabs on Austin streets, the National Highway Traffic Safety Administration opened an audit query covering an estimated 1,000 vehicles, asking what data and process Tesla relied on to self-certify a car that has no steering wheel, pedals or mirrors against safety standards written for cars that do. Tesla says the vehicles comply with every applicable standard and plans to expand gradually. The precedent is not encouraging for speed: a similar probe of Zoox's steering-wheel-free robotaxi took years and ended in a temporary exemption capped at 2,500 vehicles a year.

Anthropic now expects to begin marketing its initial public offering in mid-October at the earliest and to complete the listing days before the November midterm elections, Reuters reports. The prospectus had been expected as early as next week and is now expected in late September. Before that, the company is finalizing a $15 billion revolving credit facility, after which bank analysts meet the company. Some investors have said the listing could value Anthropic at around $2 trillion, with Morgan Stanley, Goldman Sachs, JPMorgan and Citi among the banks on the deal. Anthropic declined to comment.

Two more newspapers filed suit in the Southern District of New York on Friday, alleging that OpenAI and Microsoft scraped their sites, including paywalled articles, to train and run ChatGPT, Copilot and Bing's AI features. They want damages and a court order to destroy copies of their work along with any datasets and models built on it, the same remedy The New York Times has sought since 2023. OpenAI says its models train on publicly available data under fair use; Microsoft says it was surprised and is open to discussing solutions. Microsoft told a court only days ago that Copilot almost never reproduces news articles, so the reproduction evidence in these cases will be the argument.

Grok Bot, xAI's product for delegating whole jobs to cloud-hosted agents that use apps and websites the way a person does, is now available to enterprises, with new controls over what each Bot can reach on the network, who can use it, and what gets logged. Grok and Cursor Enterprise customers get two weeks free for their whole organization, seats or not. xAI names Legora, Supermicro and ServiceTitan as customers and says the heaviest use is outside engineering. As a demonstration, the company gave a Bot its own vendor spend, contracts and usage data and says it found more than $100,000 in direct savings, with a person making every final call.

The Siskiyou County Sheriff's Office says three novice climbers from Roseville, California relied on Google's Gemini for their route, packing list and timing, and were told to expect an eight-hour ascent while carrying far less food and water than they needed. They started at three in the morning, reached the summit at seven that evening instead of turning back at noon, tried to descend in the dark, and spent the night in Mud Creek Canyon before Forest Service rangers and volunteers reached them the next morning. Sheriff Jeremiah LaRue's advice was to treat AI as a research tool and get anything that could cost a life from the people who know the mountain.