OpenAI Is Pissing Off a Bunch of Mathematicians—Again

According to WIRED, OpenAI has told people it plans to post hundreds of results on long-standing unsolved math problems to GitHub, after an August meeting where roughly 40 mathematicians asked it to publish proper papers instead of blog posts or tweets. Northwestern mathematician Bryna Kra, who attended, says that input was apparently ignored.

OpenAI's spokesperson Lindsay McCallum says a new internal model, which began training August 28, has resolved more than 100 long-standing open problems across most areas of mathematics, in addition to the Navier–Stokes Millennium Prize problem. She says the company is working to release the results responsibly with input from an advisory group at the Institute for Advanced Study, and that no release time has been set.

This is no longer a niche academic spat. It is a test of whether the companies with the most powerful reasoning tools will respect the norms of the fields they are now transforming.

Let's start with what's remarkable. If these results hold up, this is one of the most significant things AI has done for human knowledge. We want to say that clearly, because the frustration in this story is not about whether AI can do math. Kra herself says she's glad to have a powerful new tool. Her ask is modest: disclose results in a form mathematicians can verify, attribute and build on.

That is where OpenAI keeps falling short. A dump of results, however impressive, puts the cost of checking and digesting them on a community that never agreed to carry it. Papers exist for a reason. They explain the method, cite prior work and let others confirm the logic. Skipping that step treats a field's shared infrastructure as a free trellis for a marketing win. Kra put it well when she said math by press release is not how to nurture the ecosystem these models were trained on.

The harder part of WIRED's reporting is about power and conduct. NYU's Tristan Buckmaster accuses OpenAI of front-running work he was doing with Anthropic's Levent Alpöge, and meeting notes seen by WIRED describe OpenAI researcher Sébastien Bubeck saying he didn't have to be nice. Bubeck has publicly denied asking that Alpöge be dropped as an author. Another NYU mathematician, Nestor Guillen, says colleagues perceive "mobster behavior" from the labs, which OpenAI disputes. We can't adjudicate those private conversations, and we won't pretend to. But when respected researchers describe angst not about AI but about the companies, and about power accumulating in one place, that is a warning sign leaders should take personally.

We also don't want to make this an OpenAI-only story. WIRED notes mathematicians feel the field has become a showcase for both OpenAI and Anthropic as they head toward IPOs, and Alpöge himself announced a result by tweet. The incentive to be first is industry-wide.

Where does this lead if it scales? A future where the most important discoveries arrive as unexplained output, credited to whoever had the most compute, while the humans who trained the next generation are left to clean up. The better path is right there: tools like Hexagon and Palomar already exist for AI-generated and machine-verified math, and more than 4,000 mathematicians have signed the Leiden declaration spelling out standards. Using them would cost OpenAI little and earn it a great deal of trust.

OpenAI convened mathematicians in August and formed an advisory group in mid-September to guide how it assesses and communicates results. Earlier in August it announced ten solved problems by blog and social posts, and in September it deployed thousands of agents on a Millennium Prize problem, prompting the attribution dispute with Buckmaster.