Brand New AI Solves a Millennium Prize
Zvi (Don't Worry About the Vase) TheZvi ● Covered by 21 sources
OpenAI says its new model solved Navier-Stokes after eight days of training. The bigger story is the labs fighting over credit while AI progress keeps speeding up.
Based on reporting by Zvi (Don't Worry About the Vase), TheZvi — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI says a new model solved the Navier-Stokes Millennium Problem after eight days of training. That is the headline, and it is already enough to make a lot of people sweat. But the surrounding mess is almost as revealing as the result itself: this was not just a math breakthrough, it was a public glimpse of how fast frontier models are moving and how badly the top labs are handling each other.
The source story starts with Tristan Buckmaster and Levent Alpoge, who had spent a year on several blowup results and say they used large language models to push an older program forward. Buckmaster credits Diego Cordoba and Luis Martínez-Zoroa with the basic idea, says the work should make people rethink advanced math, and even says Martínez-Zoroa deserves a Fields Medal. He also says the writeups were rushed and ugly, especially the Euler proof, because they had to publish before they could make the papers properly readable.
Then came the OpenAI side. Buckmaster says OpenAI heard a false rumor on September 1 that two Millennium Problems had been solved, then spent millions of dollars in inference exploring the unsolved problems. According to the account in the source, an internal model stronger than Astra cracked Navier-Stokes in 88 hours, and Astra then handled the Lean formalization and verification in another 17 hours. OpenAI also disclosed big numbers for the attempt: 4.9 million messages and about 300 billion output tokens across all problems, with 2.7 million messages and roughly 130 billion output tokens used on Navier-Stokes alone.
The ugliest part is the credit fight. Buckmaster says OpenAI floated proposals that would have let them post the result on their terms, including one that would have removed Levent Alpoge from authorship. OpenAI’s Sebastien Bubeck denies wrongdoing and says they were willing to talk. Both sides now insist they acted in good faith, while also saying the other side did not. The source also says OpenAI now believes user data from Codex did not reach training runs, and several people from both OpenAI and Anthropic say they think misuse is extremely unlikely. Even so, the episode leaves the same bad taste: the labs are racing, they can solve famous problems, and they still can’t seem to coordinate on the basic social part of doing science.
My take — AI-written commentary, not fact-checked reporting
This is the kind of moment where the industry loves to talk about collaboration and then immediately behaves like a pair of feral raccoons over a shiny object. A Millennium Prize result should have been a rare clean win for everyone; instead it became a live demo of why the labs should not be trusted to self-regulate anything more delicate than a press release. The models are advancing fast. The adults remain unavailable.
Read more about this at: Zvi (Don't Worry About the Vase)