OpenAI publishes new math results from an internal frontier model
Lean proofs and ChatGPT Pro compute estimates follow advice from an outside math-AI group.
The facts
- OpenAI is releasing mathematical results from an internal frontier model via a GitHub repository, with protocols for paper revisions and citations
- The release includes formalizations of many proofs in Lean, a language that lets a computer check mathematical proofs, with more to be added over time
- OpenAI is publishing 10 summaries of the model's reasoning, compute estimates in terms of ChatGPT Pro usage, and statistics on attempted problems
- OpenAI says the average result used compute equivalent to roughly three hours of ChatGPT Pro thinking
- OpenAI consulted the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study and drew on its public recommendations
- OpenAI says it will fund workshops, conferences and special programs to help the community understand major AI-produced results, and is working to responsibly release the model itself
OpenAI is releasing a set of mathematical results from an internal frontier model. It is the same class of system behind the company's recently reported Navier-Stokes work. The results go out through a GitHub repository, not a journal. Protocols are attached for how the papers can be revised and cited. OpenAI says it is still exploring community-hosted alternatives that would satisfy guidelines set out by outside advisers.
Those advisers are the Advisory Group on Mathematics and Artificial Intelligence, an independent body at the Institute for Advanced Study. OpenAI says it consulted the group and drew on its public recommendations to shape the release. The repository includes Lean formalizations of many of the proofs. Lean is a programming language that lets a computer verify a proof step by step. OpenAI promises more formalizations as it obtains them.
Alongside the proofs, OpenAI is publishing ten summaries of the model's reasoning. It is also publishing statistics on how many problems were attempted. And it gives compute estimates in a unit any ChatGPT subscriber would recognise: Pro-tier usage. OpenAI says the average result consumed compute equal to roughly three hours of ChatGPT Pro thinking. It did not disclose how many results are in the release. Nor did it say how many problems were attempted in total, or how the ten summaries were chosen from the full set.
OpenAI says it will fund workshops, conferences and special programmes to help mathematicians understand major results produced by AI. Further details are promised later. The company also says it is working to responsibly release the model itself. That would let scientists use the capability directly, rather than wait for OpenAI to run the proofs on their behalf.
Why it matters
The compute disclosure matters more than the proofs. A ChatGPT Pro subscription now buys a researcher roughly the reasoning budget that produced these results. The bottleneck shifts from access to the model toward a mathematician's skill at posing and verifying the right problem. The Lean formalizations matter too: they let outsiders check OpenAI's claims directly, rather than trust a paper's prose. That check is the real test this release has to pass.
Primary source
OpenAI — openai.comWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.