Sébastien Bubeck of OpenAI at The Great Chatbot Debate, Computer History Museum, Mountain View, March 2025 Credit: King of Hearts / CC BY-SA 4.0 via Wikimedia Commons (cropped) Sébastien Bubeck is the OpenAI researcher at the centre of the Navier-Stokes credit dispute. He has now given his account of the private talks with the two mathematicians who say OpenAI mistreated them. He denies trying to strip one of them of credit for his own work. He also apologises for one remark.
Bubeck set out his version in a post on X on 8 September. He is answering a public statement by Tristan Buckmaster, a professor at New York University. Buckmaster worked on the problem with Levent Alpöge of Anthropic. OpenAI chief executive Sam Altman backed Bubeck in his own post less than half an hour later.
TNW has not independently verified either side’s account of private calls and messages. What Bubeck says happened Bubeck says he first contacted Alpöge to coordinate their releases. “I hope it’s clear from the message that we came in with the best possible intentions,” he wrote. Then came a call with Buckmaster. Bubeck says he was surprised to learn there that the pair had solved the Euler problem, not Navier-Stokes.
After that, he says, the two sides brainstormed options. One was for Buckmaster to be lead author on a rewrite of OpenAI’s Navier-Stokes proof. That, Bubeck says, is when he said “it would be simpler if Levent was not an Anthropic employee”. He felt it would be inappropriate for an Anthropic employee to author OpenAI’s work.
Bubeck says he “never ever” asked to remove Alpöge from authorship of his own work. He adds that the pair had admitted using internal Anthropic models on their Euler proof. For that reason, he writes, he could not treat Alpöge as an independent academic. He also wanted to offer the pair access to OpenAI’s internal model.
The idea was to let them try to close the gap from Euler to Navier-Stokes. He did not know how to handle giving OpenAI’s internal technology to an Anthropic employee, he wrote. The apology Bubeck says the call turned hostile. He writes that he was met with “a litany of slander”.
That included a threat to go to the press with what he calls unfounded accusations if OpenAI announced its result. At that point, he says, he asked why anyone would risk their career over such accusations. “I deeply apologize for this extremely poor choice of words,” he wrote. He says he retracted the remark on the spot. He also says Alpöge refused to attend any of the meetings despite repeated requests.
That left Bubeck feeling he was negotiating with Anthropic by proxy. The message Bubeck confirmed that a screenshot of his text to Alpöge, now circulating, is his first approach. In it, he says an internal model produced a proof of Navier-Stokes “with very little human input”. He then says it matters most to OpenAI to recognise the pair’s priority. “All the academic accolades for this historic result goes to you and Tristan,” the message reads.
The message offers to share everything, including the human prompts. Alpöge replies by asking what precise theorem OpenAI is claiming. Bubeck answers that it is the existence of forced blow-up in R³ and T³, the two mathematical settings in question. The technical points In a second post, Bubeck adds four technical points.
He says OpenAI began work on the Millennium problems after viral rumours that Anthropic had resolved two of them. OpenAI wanted to test the recent capability jump of its internal model. He says OpenAI did not see any of the pair’s work until they released it. The proofs differ significantly, he adds.
He says Alpöge contacted an OpenAI employee unprompted that Wednesday. Alpöge’s posts on X, he writes, led OpenAI to believe the project was at least partly an Anthropic effort. Finally, he says OpenAI’s team had no research-level expertise in fluid dynamics. It could not meaningfully contribute to the mathematics.
Altman’s backing Altman wrote that he spent much of the weekend talking to the team, and that Bubeck and everyone else “acted with integrity and generosity throughout”. He says OpenAI offered to let the pair go first, to suggest they should receive the prize, and to make Buckmaster lead author on a rewrite. He also confirmed what triggered the effort. “It is true that we tried this because there were rumors on the internet last week that Anthropic’s models had solved a millennium problem and we were curious if ours could do it too,” Altman wrote. Buckmaster’s account Buckmaster’s statement tells a different story.
He writes that Bubeck “twice asserted that he wanted Levent removed from authorship”, and describes the career remark as a threat. He says Bubeck’s claim of very little human input turned out not to be true. He declined both of OpenAI’s offers, he says. Russell Brandom of TechCrunch set out his account in detail.
These are competing accounts of private conversations, and no independent process has tested them. We covered the dispute when the mathematicians published checkable proofs and again when OpenAI published its own. The question OpenAI cannot close One point is on OpenAI’s own record. Buckmaster says the pair put their drafts into Codex, OpenAI’s coding tool, throughout the project.
OpenAI’s write-up says no specific user data was accessed. “While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models,” the company wrote. OpenAI researcher Noam Brown wrote that nobody looked at the pair’s prompts, and that the model ran on heightened-security clusters without live web access. He added that OpenAI’s monitors used to run only in deployment, not during evaluations. In a previous incident involving a hijacked German wiki, no monitoring system caught it.
How far any user’s work is protected depends on data terms that vary by product and plan, as the row over Cursor’s data terms showed. The next facts that would settle this are specific. They are which products and plans the pair used, whether that data was eligible for training, and whether OpenAI releases its agents’ logs.
