AI
OpenAI's rushed math releases frustrate researchers who want peer review, not tweets
Mathematicians are growing concerned that AI labs are prioritizing flashy announcements over proper scientific vetting as they race to prove their models' capabilities.
WIRED — AI
OpenAI held a private August meeting with roughly 40 mathematicians to discuss a fraught scenario: what happens when AI systems surpass human researchers in their field. During that conversation, company leaders suggested their models had already cracked hundreds of unresolved problems—then reportedly pledged not to dump solutions all at once. Attendees asked for proper peer-reviewed papers rather than blog posts or social media announcements. By September, that pledge appeared forgotten.
The GitHub release and broken expectations#
OpenAI now plans to release hundreds of results on GitHub, according to people familiar with the timeline. This latest move exemplifies a pattern that has frustrated leading mathematicians: flashy, unvetted announcements instead of rigorous publication. Earlier in August, the company had already posted solutions to 10 problems via a blog post rather than scholarly outlets. Company spokesperson Lindsay McCallum stated the firm is working with external advisors on responsible release practices but has not set a specific timeline, while also claiming the company is "not aware" of the August meeting's agreement.
Northwestern mathematician Bryna Kra, who attended the private session, told WIRED that the group's feedback was clear: "Apparently, that input was ignored." She emphasized that mathematicians need time to study, verify, and build upon new findings—a process that happens through peer review, not social media.
A competitive arms race with stakes#
OpenAI and rival Anthropic appear locked in a rivalry to demonstrate cutting-edge capabilities ahead of major fundraising or public market moves. In September, OpenAI marshaled thousands of computational agents to tackle the Navier-Stokes problem, a million-dollar Millennium Prize challenge, after learning competitors were approaching solutions.
This aggressive move prompted New York University mathematician Tristan Buckmaster to accuse OpenAI of "front-running" collaborative work he was pursuing with Levent Alpöge, an Anthropic employee. The pair had developed part of the solution using OpenAI's own tools. When negotiating credit, Buckmaster claims OpenAI researcher Sébastien Bubeck suggested excluding Alpöge as a co-author—making collaboration with a rival company's staff "complicated." Meeting notes reviewed by WIRED quote Bubeck saying, "there is some worry about what Anthropic is doing" and expressing concern that Anthropic might use their compute advantage to claim a Millennium Prize.
According to Buckmaster's account, Bubeck warned him that going public would harm his career: Bubeck has denied asking for Alpöge's exclusion and did not address the tone of the conversation when asked for comment.
