OpenAI has posted hundreds of mathematical manuscripts produced by an internal model on GitHub. Mathematicians are debating whether the format is sufficient for review and reuse, while the company says some proofs have not yet been formalized.
OpenAI has published a GitHub collection of 722 mathematical manuscripts grouped into 372 families of results. The company says they were produced by an internal model and are at different stages of verification: some do not have Lean formalizations, and some unformalized results could contain problems. OpenAI says it will address issues and update the repository.
The release followed an August meeting between the company and about 40 mathematicians. According to WIRED, participants urged OpenAI to publish full papers explaining methods and proofs instead of relying on blog or social-media posts, so other researchers could check, understand and build on the work. Northwestern University mathematician Bryna Kra said she felt that advice had been ignored.
WIRED reported that OpenAI was preparing to put hundreds of results on GitHub. OpenAI spokesperson Lindsay McCallum told the magazine she was not aware of assurances that the results would not be released all at once. She also said no release time had been set and the company was using recommendations from its advisory group on mathematics and artificial intelligence at the Institute for Advanced Study.
The disagreement follows earlier disputes. In September, OpenAI deployed thousands of AI agents to work on a Millennium Prize problem carrying a $1 million award. New York University mathematician Tristan Buckmaster accused the company of getting ahead of work on part of the problem that he had pursued with Anthropic employee Levent Alpöge. WIRED also described objections about negotiations over possible co-authorship with OpenAI researcher Sébastien Bubeck.
Some academics fear that technology companies are concentrating computing resources and changing established norms for attribution and publication. NYU professor Nestor Guillen told WIRED that such a perception exists among mathematicians. Kra argued that releasing proofs as short posts does little to help the research community develop the knowledge on which AI systems are trained.
After the collection went online, OpenAI said it would preserve version history and publish corrections as new versions. The company said most results came from one internal method applied to about 4,000 problems, with each result using an average of about three hours of ChatGPT Pro thinking compute. Some mathematicians continue to call for detailed academic papers, while Bubeck believes AI could expand researchers’ capabilities and help them take on more ambitious problems.