Mathematicians Accuse OpenAI of Scooping Their Work, Demand Removal of Anthropic Employee's Name from Paper

动察 Beating AI Flash: New York University mathematicians Tristan Buckmaster and Levent Alpöge recently leveraged large models such as Claude and Codex to advance multiple fluid dynamics problems within a month, with some results already verified through Lean machine checking.

Before the papers were even fully released, the duo had already fallen out with OpenAI.

Buckmaster stated that in early September, their unpublished progress reached OpenAI. He proactively contacted OpenAI to avoid a collision between the two parties. Three days later, OpenAI researcher Sébastien Bubeck told him that OpenAI's internal model had already written approximately 100 pages of a Navier-Stokes proof. This is one of the seven "Millennium Prize Problems" in mathematics, with a $1 million reward offered by the Clay Mathematics Institute.

Buckmaster subsequently pressed OpenAI repeatedly on when they had started. According to his account, the other side eventually confirmed that the first prompt was only sent in recent days, at which point OpenAI already knew about their progress. OpenAI was also pursuing a highly overlapping route that had rarely been studied before, backed by a team and substantial computing power.

The duo's research drafts had previously been stored in Codex. Buckmaster therefore questioned whether this data had been used for training, but received no clear answer. He also acknowledged that there is currently no evidence proving OpenAI used this data.

The subsequent negotiations pushed the matter fully into the public eye. Buckmaster said OpenAI proposed that they release their own results first, followed by the Navier-Stokes paper the next day; another option was for Buckmaster to write OpenAI's paper, but without Alpöge as a co-author. Bubeck reportedly twice suggested removing Alpöge, citing his employment at Anthropic as the reason.

After Buckmaster refused, he stated that if OpenAI released under this plan, he would publicly disclose the entire process. He claimed the other side then asked him: "Why would you want to ruin your own career?" Later that day, Bubeck separately contacted Alpöge, saying he was unsure whether Buckmaster had been "fully rational" at the time. Alpöge declined the separate communication.

Buckmaster ultimately released his paper and statement early. He also stressed that he had not seen OpenAI's proof, did not know whether it held up, and was unaware if his Codex data had been used.

Bubeck subsequently publicly denied the allegations, calling them "false and inflammatory," and stated he would respond further.