Altman says OpenAI will keep releasing AI math results as researchers struggle to check them

OpenAI CEO Sam Altman has said for the first time, in public, that the company's flood of AI-generated mathematics is hard on the people who have to absorb it. He also said it will not slow down. "I think, on net, the world deserves to understand how the universe works," Altman said on an episode of the Cheeky Pint podcast released Thursday. "But I do have empathy for the method."

This matters because a proof is only useful once someone has checked it and understood it. According to The Verge, OpenAI released hundreds of mathematical results this month alone, often with short writeups that researchers call poorly written and hard to follow. Mathematicians who spoke to the publication say they are the ones left doing that checking.

What Altman said

Altman argued that the public's right to new scientific knowledge "outweighs" the disruption OpenAI's releases cause. "I feel unequivocal on this point: Humanity deserves fundamental knowledge, and we should not be sitting on a major fundamental science discovery," he said. He conceded that the company needs to improve how it communicates its findings.

He was asked whether OpenAI has more than the hundreds of proofs it has already published. "We do. Of course," he replied.

Altman also said the company's pursuit of math problems did "get out of control" for a short period. He said this was deliberate and needed to probe the limits of OpenAI's models. He described an internal Slack channel called "Model Name-Hold My Beer," where employees "were just throwing that at the hardest problems they could."

He estimated that OpenAI has spent hundreds of millions of dollars on mathematical problems. That includes around $10 million, "maybe a little more," on its Navier-Stokes result.

The Navier-Stokes claim and how it is checked

OpenAI announced the Navier-Stokes result on September 8. The problem is one of the Clay Mathematics Institute's Millennium Prize problems. Quanta Magazine reported that about 10,000 autonomous agents, running an unreleased internal model, produced a proof of a singularity in the three-dimensional equations after 88 hours. Another model formalised the result in the proof assistant Lean about 17 hours later. At the time, OpenAI researcher Sébastien Bubeck put the compute cost at several million dollars. Altman's new figure is higher.

A Lean formalisation gives strong evidence that a proof is correct. It does not settle everything. As Quanta noted, humans still have to confirm that the formal statement matches the problem it is meant to solve. The result also arrived in a dispute over credit. It came roughly 12 hours after Tristan Buckmaster of NYU and Levent Alpöge said they had resolved closely related problems with AI help.

Criticism from both directions

Mathematicians who spoke to The Verge have repeatedly criticised how OpenAI announces results. They cite thin attribution, little context and insufficient verification. They say OpenAI keeps releasing new work while they are still checking the last batch.

OpenAI now also faces the opposite complaint: holding results back. The company recently announced three claimed breakthroughs, in mathematics, physics and computer science. Details of one, involving Yang-Mills theory, the mathematical basis of modern particle physics, have not been made public. OpenAI says it is following the advice of a hand-picked group of experts it consulted on how to release the result. The Verge reports that this goes against the advice of an earlier group of mathematicians OpenAI consulted. The company describes it as part of a new approach to communicating findings. Researchers told The Verge they see it as gatekeeping.

That puts the company in an awkward position. Altman says OpenAI should not sit on major discoveries. By his own account, it has unreleased proofs, and it is holding back at least one headline result.

What it means for work

AI models now produce candidate proofs faster than people can review them, and the cost lands on the people who do the reviewing. Formal checkers like Lean can confirm that a proof's logic is valid. Human experts still have to judge whether the formal statement means what it claims, place the result in the existing literature and sort out who deserves credit. Each of those jobs takes time, and a few-paragraph writeup makes them harder. Researchers also worry about money. Altman puts OpenAI's spending in the hundreds of millions of dollars, and researchers fear that scale could price them out of their own fields. Three questions are still open: who pays for verification, what standard of documentation AI labs will be held to, and whether journals, prize committees and institutions like the Clay Institute will build processes to handle work at this volume.