When AI Solves Your Life’s Work

OpenAI said an unreleased frontier model produced 372 novel mathematical results supported by 722 papers, using average compute equivalent to about three hours of ChatGPT Pro thinking per result. It fully solved 90 of 500 problems on ProofAtlas’s AI-generated list, described as a reasonable sample of well-known open problems, and produced dozens of partial proofs, including work on the Riemann Hypothesis and the Hodge conjecture. The results and partial proofs still need scrutiny, and mathematicians were separately verifying a Navier–Stokes result.

OpenAI coordinated the release with its Mathematics Advisory Committee and published reasoning traces for 10 problems. The committee had recommended disclosing the resources used for each result, but OpenAI gave only an average. The release raised the prospect that mathematicians may shift from pursuing some problems for decades to checking AI-generated proofs. Alex Konturovich said the quasi-Riemann result would merit an immediate Fields Medal if a human had produced it, while Levent Alpogee called the release the most significant moment in mathematical history.

The central uncertainty is whether mathematically correct outputs will become usable knowledge. Francesco Maggi warned that proofs could remain “lettera morta” unless mathematicians can understand, connect, explain, challenge, and reuse them. He and Steve Sue raised the possibility that AI could produce mathematics faster than people can absorb it, leaving researchers dependent on selected explanations from systems they do not fully understand. It is not yet clear whether the tools will erode the value of mathematicians’ hard-won learning or let them pursue work they could not previously imagine.

Many of the theoretical problems, including most Erdős problems and the Navier–Stokes proof, were described as having no obvious practical application, though engineers already use simplified Navier–Stokes models. Stephen Strogetz pointed to a more direct connection: a matrix-multiplication exponent of no more than 2.25, improving on a prior record of about 2.37, matters because matrix multiplication underpins AI inference. Cryptography was notably underrepresented among the 722 published results. Justin Drake flagged that gap, and Scott Aronson said sources told him AI companies were discreetly testing whether internal models could break important protocols and primitives. A successful attack on major cryptographic schemes could have implications far beyond Bitcoin, affecting cryptography used across the digital world.

The transcript cautioned that mathematics may not be a template for automation elsewhere. Payman Millenfar argued that progress slows as domains become less structured: hard sciences depend on noisy measurements and experiments, markets on adaptive human behavior, and medicine on fragmented signals and idiosyncratic biology. Law, culture, art, and strategic leadership also involve ambiguity, tacit knowledge, judgment, or moral trade-offs. The speaker expects artists, lawyers, doctors, and CEOs to remain for a long time, even as AI changes parts of their work.

Coding was described as one of the fields most changed by AI, but demand for coders appears to be rising. In The Information’s small survey of a highly engaged AI-user audience, two-thirds of respondents worked at organizations developing AI applications; 35% said returns on AI spending were multiples of their spend, while 30% said it was too early to tell. Almost half said AI had not changed hiring practices, compared with 22% who said they were hiring less and 10% who said they were hiring more.

The Financial Times reported that OpenAI told investors it had roughly $50 billion in annualized revenue at the end of September, below the widely reported $68 billion figure. That larger figure reportedly came from an investor using gross revenue to compare OpenAI with Anthropic, whereas OpenAI reports net revenue after revenue sharing; how the figure was annualized remained unclear. OpenAI also reported third-quarter run-rate growth of 77% and enterprise-business growth of 107%. Nvidia shares fell 3%, Oracle fell 5.5%, and some neocloud stocks fell more.