OpenAI Says It Solved a Millennium Prize Problem in 88 Hours — and Mathematicians Are Alarmed

Tuesday’s announcement from OpenAI that one of its unreleased models had solved the Navier-Stokes problem in just 88 hours was a real accomplishment. Yet the claim has quickly been eclipsed by charges that OpenAI pushed the work forward to beat rival researchers, possibly crossing ethical boundaries in the process.

Claims of scooping, spying, and breaches of academic standards have emerged from the dispute. The episode may dampen the field, according to mathematicians. A mathematics professor at Queen Mary University of London, Abhishek Saha, “OpenAI has engaged in the kind of things that mathematicians will generally not do,”.

An 88-Hour Solution

Tuesday brought news from OpenAI that its model had solved Navier-Stokes, a longstanding fluid dynamics problem that has left human researchers searching for answers for close to 90 years. The company described deploying a swarm of roughly 10,000 AI agents, each running on its internal model, to work on the problem. OpenAI referred to the accomplishment as a “milestone.”.

A reward of $1 million dollars is attached to this issue, managed by the Clay Mathematics Institute. It remains unclaimed. OpenAI has stated that it has no desire to lay claim to it. The firm said its sole objective “is to report on the substantial progress of our AI models.”

Millions of dollars were spent on the project, according to OpenAI’s own statement. The company describes the work as rushed. It appears the company did not spend much time on Navier-Stokes prior to September, or at least has not discussed any such work publicly.

Buckmaster’s Account

The timing drew attention. It came one day before the rival company that competes with OpenAI’s announcement, New York University mathematics professor Tristan Buckmaster published findings on a related problem with Levent Alpöge, a researcher at Anthropic, OpenAI’. Alpöge was not acting on behalf of his employer in this instance.

After Buckmaster learned that OpenAI had become aware of their work, he reached out to the company and asked when it started working on the issue and what training data its model had been fed. According to him, the conversation then took a negative turn.

A researcher from OpenAI questioned him about “Why would you ruin your career?”, after he declared he would make his work public. Buckmaster then asked what harm making his research known would cause to his standing, and the researcher responded with “If you don’t want me to be nice, then I don’t have to be nice.”.

OpenAI pushed Buckmaster to put the paper out and give credit to its internal model while cutting Alpöge from the list of authors. Sébastien Bubeck, the OpenAI researcher mentioned by Buckmaster, has challenged parts of that account. Bubeck denied ever asking Buckmaster to take Alpöge off the author list.

Codex And The Data Question

Buckmaster claimed he questioned OpenAI about whether it had looked at his sessions on Codex, the coding tool he used during his work on the problem. According to him, OpenAI became more evasive and even hostile with each reply.

In a blog post, OpenAI stated it has not used any particular user data. “We (the researchers and the agents) did not see any of their work through any means until they released it publicly — in particular, no specific user data was accessed in order to solve this problem,”

But OpenAI could not conclusively rule out an indirect influence. “While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models,” it said, while stressing that the two proofs differ significantly.

Researchers from OpenAI have also rejected those allegations. When asked about Buckmaster’s other claims, the company pointed The Verge toward its blog.

Why The Rush?

The reason OpenAI gave for the rush is straightforward: the company said it started working on the issue after learning that other researchers were reportedly making headway on Millennium Prize problems. It found those rumors “on Twitter.”, which prompted the work.

“We have such a strong model. Why don’t we try to solve also a Millennium Prize problem?” Bubeck said at a press briefing reported on by Science. OpenAI said it only later realized the rumors concerned Alpöge and Buckmaster.

Saha noted that scooping is not how mathematics is typically conducted, and it does occur, though it is not simple. The reason is that cutting-edge work demands such deep and specialized knowledge that few researchers are positioned to swoop in, even if they wished to do so.

Openness is a deeply embedded virtue in the discipline. “Mathematics depends heavily on an informal norm of trust,” said Matthew Ball.

Sorting out what happened is hard, given tangled timelines and overlapping research. Provenance for AI-generated material is already difficult, if not impossible, to trace under normal circumstances. So it should come as no surprise that different players might race to solve one of the world’s most famous mathematical problems, especially when there is a large prize attached.

Millions were spent on an issue that OpenAI has no intention of claiming a prize for. Researchers were beaten in a race, even though the company says it did not copy their work. And OpenAI admits it cannot rule out that its models were indirectly improved by that research. These points together raise questions about the company’s account.

The result stands. The questions around it do not.

Key Facts Box

• 88 hours: Time OpenAI’s unreleased model took to solve the Navier-Stokes problem.
• $1 million: Bounty attached to the Millennium Prize problem, not yet awarded.
• ~10,000: AI agents OpenAI said it deployed to tackle the problem.
• Close to 90 years: How long the Navier-Stokes problem has stumped human researchers.
• Millions of dollars: What OpenAI said the hurried effort cost.

Timeline

Source: theverge.com

AI ethics · mathematics · Millennium Prize · Navier-Stokes · OpenAI · research controversy