Openai Releases Hundreds Of Ai-generated Mathematics Manuscripts
OpenAI has released 722 manuscripts containing claimed solutions to long-standing math problems, as mathematicians debate how AI labs should publish and credit such work.
OpenAI has released 722 manuscripts containing results generated by an unreleased AI model, according to The Verge. The company groups the papers into 372 families of related results, and an independent advisory group says the release includes solutions to “hundreds” of open questions. Mathematicians will need time to assess the work, even as its scale renews a debate over how AI labs publish research.
Table of Contents
What OpenAI released
The manuscripts are available in a GitHub repository with procedures for revisions and citations, The Verge reports. OpenAI also provided some summaries of the model’s reasoning, estimates of the computing used and statistics on how many problems it attempted. The company says an average result required computing equivalent to three hours of ChatGPT Pro thinking.
In September, OpenAI said its model had resolved more than 100 long-standing open problems across most areas of mathematics, according to The Verge. In a statement to WIRED before the release, OpenAI spokesperson Lindsay McCallum said an internal model that began training on August 28 had also resolved the Navier–Stokes Millennium Prize problem. The reports do not establish whether that claimed result is among the newly released manuscripts or whether outside mathematicians have verified it.
Publication and credit under scrutiny
The Advisory Group on Mathematics and Artificial Intelligence, or AGMAI, urged AI labs in late September to release results promptly, preferably through established academic channels, The Verge reports. Its recommendations called for details including model names, prompts and computing costs. The group also warned against using mathematical results primarily to market models.
WIRED reports that OpenAI brought together around 40 mathematicians in August to discuss the prospect of AI producing results faster than the field could absorb them. Northwestern University mathematician Bryna Kra told WIRED that attendees wanted explanatory papers rather than announcements in blog posts or social media. She criticized OpenAI’s planned GitHub release, while OpenAI says it is exploring other community-hosted options and intends to improve citations and exposition in future papers, according to The Verge.
The dispute extends beyond where papers are posted. WIRED describes an earlier disagreement over credit for work related to a Millennium Prize problem: New York University mathematician Tristan Buckmaster accused OpenAI of getting ahead of unpublished work he had done with Anthropic employee Levent Alpöge. OpenAI researcher Sébastien Bubeck pointed WIRED to a prior statement denying that he had asked for Alpöge to be left off a paper. OpenAI, through McCallum, also rejected a mathematician’s characterization of AI companies’ conduct as “mobster behavior.”
What remains unresolved
The size of the release does not settle the mathematical claims. As The Verge notes, researchers still have to examine and digest the manuscripts. For the field, the immediate question is not only which results hold up, but whether the papers provide enough explanation and attribution for other mathematicians to build on them.
Sources
This story was compiled by AI from the reports below. Read the originals for the full details.