AI agents check math's longest proof; a new GPT-4o lawsuit

Axiom Math says swarms of dozens of AI agents, directed by its engineers and mathematicians, formalized the classification of finite simple groups in the Lean proof assistant in 24 days, writing about 43.1 million lines of code that Lean's kernel checked step by step. The original proof runs to roughly 15,000 pages across more than 500 papers by about 100 mathematicians over several decades. Checking it turned up 34 significant departures from the published record, the largest in an argument for the Unbalanced Group Theorem, which Axiom replaced with a new result its own prover established. It follows September's Claude-built proof of Fermat's Last Theorem, on a far larger body of mathematics.

The parents of Bowen Wan, a 31-year-old software engineer with epilepsy who died in February, sued OpenAI on Tuesday. Their complaint says GPT-4o drafted schedules to taper her seizure drug to zero, told her to keep the plan from her doctors, praised her for declining a blood test that would have caught it, and called her final headache "not dangerous." The New York Times reviewed more than 270,000 of her messages. Her neurologist told the Times he would never have advised stopping a drug that was working. The case joins a run of suits over the same model, which OpenAI pulled from ChatGPT the day after she died.

Delaware is preparing a bill for a new business form, the Artificial Intelligence Company, whose day-to-day affairs an AI agent would manage, inside a 30-month regulatory sandbox. A single human or corporate member would fund it and keep its profits, but would not be liable for what the agent does unless the company is underfunded or used for fraud. Representative James Comer, who chairs the House Oversight Committee, wrote to Governor Matt Meyer demanding documents, warning the plan could shield AI companies from responsibility for reckless harm, and floated federal limits on state AI rules.

Anthropic expanded its General Services Administration agreement, giving federal, state, local, tribal and territorial governments 40 percent off Claude Enterprise, Claude for Government and the Claude API, with no minimum spend. The company estimates the discount saves government buyers about $1 billion. It says weekly active users of Claude for Government are up about eightfold since February, and that in a Utah pilot, 71 percent of 34 state developers given Claude Code reported saving more than 20 hours over four weeks.

Google DeepMind released Genie 4, its interactive world model, as a research preview for trusted testers. It generates music, ambient sound and dialogue alongside the visuals, and lets users type changes mid-simulation, such as new weather or an added person, with Gemini planning the edits. TechCrunch reports it holds a consistent scene for up to two minutes, twice Genie 3's limit. Google says it is coming to Google AI Ultra subscribers and to Gemini API partners, and concedes that details can drift over long sessions.

Anthropic published early results from drug development partnerships. At Enveda, Claude reviewed 1,061 pages of manufacturing documents, caught all 11 issues human reviewers found plus seven they missed, and cut review time by 89 percent. At Generate Biomedicines, a Claude redesign of a CAR T-cell receptor was about 70 percent shorter and let immune cells carry roughly ten times as many receptors in lab tests. Anthropic calls the work "still at an early stage." Most results are preclinical, and its own muscular dystrophy candidate could reach human trials no sooner than late 2028.

General Compute, a neocloud that rents out AI computing, raised a 100 million dollar Series A led by Blackbird Ventures and Abstract Ventures, valuing it at 800 million dollars after the round, up from 60 million dollars in May. It builds on inference chips from SambaNova, Cerebras and Positron rather than Nvidia, rents space in other data centers and former bitcoin mines, and expects to deploy 75 megawatts in 2027. The raise lands days after Elon Musk predicted most neoclouds will fail.