Buckmaster says LLM-assisted forced blowup work triggered dispute with OpenAI
- Tristan Buckmaster and Levent Alpöge publicly released finite-time blowup results with smooth forcing for incompressible porous media, Boussinesq, and 3D incompressible Euler, while saying their hypo-dissipative Navier-Stokes result is not ready for release or Lean verification
- Buckmaster says Claude, OpenAI Codex, and especially GPT-5.6 Sol helped extend Diego Córdoba and Luis Martínez-Zoroa's rough-forcing program to smooth forcing and Euler; Astra was used only for writeups and auditing
- Buckmaster says the first LLM-generated proof Alpöge sent was extremely poor in presentation, though they verified it in Lean on August 22 and then worked to understand and rewrite it
- He says results for smooth-forced Boussinesq and Euler arrived on August 15 after nearly a year of slower progress, and argues that a mathematician working with an LLM can now complete this scale of work in about a month
- The accompanying discussion focuses on Buckmaster's allegations that OpenAI had a team pursuing a similar forced approach and misrepresented its human input, while commenters stress that the statement is one side of an unresolved dispute
Hacker News opinions
The PDFs sound more dramatic than a scheduling dispute. Buckmaster seems to suspect OpenAI reached the same forced-blowup approach from their private chats and planned to claim the result.
I think we should distinguish this from the Millennium problem. The announced results concern forced equations, while Buckmaster only says they believe they have hypo-dissipative Navier-Stokes blowup and have not released a paper.
Why was OpenAI discussing its internal result with him at all? If their model used the same approach, the unanswered question is whether it came from independent work, researchers' chats, or other human input.
If OpenAI said it merely gave a model the problem, but had a whole team pursuing the work and drew on outside researchers, that would be a serious misrepresentation. The alleged remarks about ruining Buckmaster's career make it worse, if accurately reported.
I would not jump straight to a model breaking privacy controls. OpenAI's terms may permit training on private chats, and researchers could also have found clues by following public posts or mathematicians' work.
Even contractual isolation would look much less credible if private research can enter model behavior or internal research workflows. For sensitive work, self-hosted open-weight models may be the safer option.
The statement says Astra had little role in the mathematics. Buckmaster says Claude, Codex, and especially GPT-5.6 Sol did most of the work, while Astra helped with writeups and argument auditing.
The story is really about allegations that OpenAI was dishonest about an AI-assisted Navier-Stokes result, rather than a confirmed solution of Navier-Stokes.
I am angry at the possibility that a lab used researchers' work, then sought the public credit and allegedly threatened them. But this is still Buckmaster's account alone, so OpenAI's response matters.