OpenAI published hundreds of math manuscripts from an unreleased AI model. Outside experts must now assess the results.
OpenAI on Tuesday published a large set of new mathematical results produced by an internal AI model that the company has not released publicly. The company posted the work in a public GitHub repository, along with computer-checkable proofs for many of the results. Coverage of the repository describes hundreds of manuscripts, though outlets differ on the exact count.
The repository includes proofs written in Lean, a programming language that lets a computer verify mathematical arguments, and OpenAI said it will add more as they are completed. The company also published summaries of the model’s reasoning, estimates of the computing power used and statistics on how many problems were attempted. OpenAI said the average result took the equivalent of roughly three hours of ChatGPT Pro thinking time. The repository’s documentation cautions that some results that have not been formalized in Lean could contain issues, and OpenAI said it will fix any quickly.
OpenAI said it consulted the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study in Princeton, New Jersey, on how to release the results. The independent group of nine mathematicians, which includes Timothy Gowers and Edward Witten, was formed to advise on releasing a large volume of results. OpenAI said it is exploring community-hosted alternatives to GitHub and plans to fund workshops and conferences on understanding major AI-produced results. It added that it is working to release the model responsibly.
The release follows earlier OpenAI math announcements this year. In May, the company said an internal model had disproved a long-standing conjecture related to the Erdős unit distance problem, and outside mathematicians checked that work. In August, it reported results on 10 open problems in mathematics and theoretical computer science. Reaction among mathematicians to the latest release has been mixed, according to coverage of the announcement. The manuscripts are posted on GitHub rather than published in journals, and independent experts have yet to assess how significant they are.
Information for this report came from OpenAI, Gizmodo, Unite.AI and The Next Web.