▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=UA0UkPombNk
• Original title :: The Navier Stokes situation is insane

OpenAI claims to have solved the Navier-Stokes Millennium Prize problem with 10,000 agents running 88 hours, at a cost likely north of $10M for a $1M prize. Mathematician Tristan Buckmaster published a four-page statement implying OpenAI trained on his and Levent Alpagay's Codex session logs. Prime reads both sides.

Buckmaster's account

Buckmaster and Alpagay had been working the problem for a year using both Anthropic's Claude and OpenAI's Codex, putting all their drafts into Codex sessions. Progress was slow until August 15, when an LLM produced a proof he called the most horrendous he had ever read — verified in Lean on August 22.

On September 3 a rumour circulated that Anthropic had cracked one of the seven Millennium problems. OpenAI reacted. Buckmaster emailed to clarify it was just him and Alpagay, got a warm reply, then repeated pressure to join a call. He was told an internal OpenAI model had produced a proof of finite-time blowup for forced Navier-Stokes — stated in R3 and T3, the same unusual framing he had told almost no one about. That was his red flag.

He asked whether the model had been trained on or had access to their Codex sessions. He was told the model does not look up user data; on the training question he got no answer.

He was offered two options: post his result with OpenAI posting the next day, or write the paper himself acknowledging that an internal OpenAI model resolved it. The catch, per Buckmaster: OpenAI researcher Sebastian wanted Alpagay removed from authorship, apparently because Alpagay works at Anthropic. When Buckmaster said he would go public, the reply was "why would you ruin your career," and then "if you don't want me to be nice, then I don't have to be nice."

OpenAI's side

Sebastian Bubeck responded that he never asked for Alpagay's removal, and posted the texts: an offer to coordinate the release of concurrent discoveries, a statement that an internal model produced both formal and informal proofs with very little human input, and an insistence that academic credit go to Buckmaster and Alpagay, with willingness to share all human prompts.

Prime's read

He finds the generosity suspicious — if you genuinely solved one of the hardest open problems independently, you don't hand the credit away. That reads like a partial concession that the pair's data was in the training set. But he does not think OpenAI committed academic fraud. His theory: de-identified data may have leaked into training and gave the search a starting point, but the real unlock was believing it was solvable. The Anthropic rumour is what made OpenAI throw 10,000 agents and tens of millions of dollars at it, because they could not let Anthropic be seen winning.

Closing jab: if the compute works that well, run it again on P vs NP.