Navier-Stokes – Tristan Buckmaster [pdf]

Original Article ↗ Hacker News Discussion ↗

An NYU mathematician has publicly alleged that an AI lab used his private Codex sessions and research direction to “scoop” him on a Navier–Stokes breakthrough, then pressured him to drop his coauthor who works for a rival lab in exchange for shared credit on a potential Millennium Prize solution. Commenters dissect timelines, terms-of-service and training-data practices, raising broader concerns about plagiarism, academic norms, power imbalances, and whether frontier AI labs can be trusted with sensitive research. Many see this as a turning point for how mathematicians and other researchers should use commercial LLMs, and for how credit and ethics will work when AI is deeply embedded in scientific discovery.

Alleged timeline and dispute

  • Two researchers reportedly used multiple LLMs over ~a year on fluid‐dynamics problems, claiming progress on Euler, porous media, Boussinesq, and a related “hypo‑dissipative” Navier–Stokes problem (not itself the Millennium version).
  • Rumors spread that a rival AI lab had solved a Millennium Prize problem; one of the researchers contacted another lab to clarify this was independent work involving an employee of the first lab.
  • Shortly after, the contacted lab said its new internal model had found a Navier–Stokes blowup result (apparently prize‑qualifying) via a similar “forced” approach, after only a few days of focused effort and heavy compute.

Use of chat logs and training data

  • Central worry: the second lab’s model may have benefited from the first team’s private Codex/ChatGPT sessions (drafts, ideas, prompts).
  • The researchers report being told “the model does not look up user data,” but received no clear answer about training on those sessions.
  • The lab’s public statement later said: no specific user data was accessed to solve the problem, but “cannot rule out” that de‑identified data from product usage improved the models.
  • Some commenters find deliberate snooping implausible (access controls, business risk); others consider it entirely consistent with current LLM‑training practices and the company’s incentives.

Authorship, academic norms, and alleged threats

  • According to the statement, the second lab proposed coordinated publication and offered to let the academic lead a paper on the full result, but wanted to exclude the coauthor employed at a rival lab.
  • Commenters note that removing a substantial contributor due to affiliation violates academic norms.
  • The statement claims a senior person warned that going public would “ruin” the academic’s career and that they “didn’t have to be nice”; many readers interpret this as intimidation.
  • A researcher at the lab later acknowledged the “ruin your career” phrasing as a serious mistake, while disputing other allegations.

Status and nature of the mathematical results

  • Thread consensus: the academics’ current papers are major but do not solve the official Navier–Stokes Millennium problem; the internal lab result allegedly does (with forcing).
  • Details of the full proof, correctness, and whether it meets the Clay Prize specification remain unclear and will need community vetting.
  • Separate announcements on Euler blowup (forced vs. unforced) add to confusion about what exactly has been proved.

AI’s role in mathematics and future incentives

  • Many see this as a “Deep Blue” moment for math: LLM‑driven automation plus huge compute can rapidly close long‑standing gaps once a promising direction is known.
  • Several mathematicians in the thread argue that if AI “solves” big problems via opaque, brute‐force, Lean‑formalized proofs, this may advance formal results but harm human understanding and the culture of open collaboration.
  • There is strong concern that merely hearing rumors of a promising approach now triggers an arms race of compute from frontier labs, incentivizing secrecy about research directions.

Privacy, ToS, and practical takeaways

  • Commenters repeatedly emphasize:
    • Consumer accounts often default to allowing training on chats; many academics don’t realize this.
    • Opt‑out semantics and “de‑identified data” clauses are seen as dark‑patterned and ambiguous.
    • For work where IP and priority matter, many recommend self‑hosted or contractually guaranteed zero‑retention setups rather than relying on general‑purpose cloud LLMs.

Community sentiment

  • Opinions split between:
    • Viewing the labs’ behavior as predictable cut‑throat competition plus poor communication; and
    • Interpreting it as systemic plagiarism, bullying, and proof that frontier labs cannot be trusted with sensitive research.
  • Enthusiasm about a potential Navier–Stokes breakthrough coexists with deep skepticism about the narrative of “AI just solved it by itself” and alarm about the longer‑term impact on mathematical research.