OpenAI made a big move this week: it announced the solve of a million-dollar Millennium Prize Navier-Stokes problem, while being publicly accused by a mathematician of replicating his year-long research path in under a week. The latter is what deserves our attention.

What This Is

Simply put, this is a dispute over whether AI used your unpublished thinking to train a model that replaces you.

The protagonist is NYU mathematician Buckmaster. He and his collaborators spent a year working on the Navier-Stokes problem, using several large models throughout—including OpenAI's Codex (OpenAI's code LLM)—with drafts and conversations stored in the cloud. By his account, OpenAI traversed the same path in under a week.

He also revealed that OpenAI had offered a condition: he and his colleagues could publish first and claim the million dollars, provided they removed a colleague who works at Anthropic (OpenAI's chief competitor) from the discoverer credit. After he refused, he went public with the whole thing, and said he received replies that strongly implied he was being "bought off." OpenAI called the claims "false and inflammatory" on X.

The unavoidable real question: did OpenAI train its models on Buckmaster's cloud data? OpenAI has been evasive on this point.

Industry View

The math community's reaction is split. Supporters see this as the first time AI for Math has demonstrated real reasoning at the Millennium Prize level; critics counter that replicating a year's path in one week looks more like back-tracing from a known answer than an independent discovery.

The developer community's silence is more telling. The original author observed that most engineers didn't discuss the incident, for a blunt reason: "I've spent months maintaining a system, and AI can generate the same thing in hours"—a position developers know all too well. But what concerns us more is the other side: developers are highly tolerant of "feeding API keys (the credentials used to call AI models) and conversations to unvetted relay services." The earlier anti-distillation (preventing others from using your model's outputs to train new models) discourse has long faded.

That's exactly where the risk lies. If OpenAI really did train its models on unpublished research, today it's a mathematician; tomorrow it could be lawyers, doctors—any knowledge worker who uploads their intellectual labor to the cloud. A high-frequency commenter who reposted the piece on X put it bluntly: this is training a tool that replaces you on your own labor.

Impact on Regular People

  • For enterprise IT: Compliance boundaries for API keys, internal drafts, and customer conversations uploaded to the cloud need re-examination. "We used AI" is no longer a liability shield.
  • For individual careers: Pause a second before feeding unpublished solutions or client data to AI assistants—you may be helping a competitor train a tool that replaces you.
  • For the consumer market: No short-term impact, but the longer "training replacement models on your data" is tolerated, the more the trust foundations of consulting, legal, and medical knowledge services will erode.