Peer review, meet the firehose. OpenAI on Tuesday evening posted 722 math manuscripts, organized into 372 “families” of related results, produced by an internal frontier model the public can’t use yet, according to OpenAI’s announcement and its openai/math GitHub repository.
The numbers: over the evaluation, the model was posed about 4,000 problems. On average, each result used the equivalent of roughly three hours of ChatGPT Pro thinking compute, OpenAI says. The Decoder reports nearly every result came from a single prompt to a single agent — a far cry from last month’s Navier-Stokes claim, which took a swarm of up to 10,000 agents and millions of dollars in compute.
What’s in the pile: work across fields including number theory and algebraic geometry, and results the Wall Street Journal says touch three of the remaining Millennium Prize Problems, among them the Riemann hypothesis, per Quartz. Many — but not all — of the papers come with Lean formalizations, computer-checkable proofs, and the repository itself warns the collection “includes results at different stages of verification.”
The bold move is where it lives: GitHub, not journals. OpenAI says it consulted the Institute for Advanced Study’s Advisory Group on Mathematics and AI — but Quartz reports it broke with one request, that labs stop testing frontier math problems on proprietary models. OpenAI says it will keep doing that.
Mathematicians aren’t all cheering. In an open letter titled “A Severe Misalignment of AI in Mathematics,” 25 Fields Medal winners warned that mass-producing true statements could destroy fertile ground rather than create new ideas, The Decoder reports. OpenAI says it will fund workshops on AI-produced results and is working to “responsibly release” the model. The humans now have homework — 722 papers of it.
Sources: OpenAI · GitHub: openai/math · The Decoder · Quartz
Leave a Reply