Product1 publisher3 min readPublished
A Cornell mathematician calls OpenAI's million-dollar math result a marketing device
OpenAI says tens of thousands of agents solved a 90-year-old problem carrying a $1 million prize. The solution still needs independent verification, and a rival pair of mathematicians has a claim on the credit.
The Product Desk · Product desk
What happened
- OpenAI said on Tuesday that it had used tens of thousands of agents to solve a 90-year-old mathematics problem that carries a $1 million prize.
- Wired reports that the solution still needs to be independently verified.
- Tristan Buckmaster of New York University says OpenAI rushed ahead after learning of his work with Anthropic researcher Levent Alpoge, and that OpenAI tried to influence who got credit.
- Anthropic said last week that Claude had proved 29,500 small theorems while formalizing an existing proof of Fermat's Last Theorem, a project other mathematicians had worked on for years.
- OpenAI announced in August that it had made advancements on 10 other long-standing mathematical problems.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Until someone outside OpenAI checks the proof, the only account of what the agents did is the company's own, so a team cannot use the result as comparative evidence between vendors.
- decision Anyone citing this week's result in a buy decision has to pick a stance before verification lands: treat it as demonstrated capability, or file it as vendor marketing and wait.
- contradiction Strogatz says breakthrough math will be impossible without AI and also says almost nobody outside a small group of pure mathematicians cares about the problem OpenAI solved, so the difficulty of a demo and its relevance to a buyer move independently.
- exposure With Buckmaster alleging interference in credit, a team quoting the result in a deck inherits an unresolved attribution fight between named researchers at two labs and a university.
Alex Townsend, Steven Strogatz's collaborator on a book about math slipping away from human understanding, used ChatGPT to help solve a decades-old numerical linear algebra problem [14][7]. Townsend and his coauthor said that without the technology, the volume of work required and the cost-reward ratio would have been unfeasibly high [15]. He had already priced the problem as not worth attempting, and he was the person who could tell whether the answer held. He also told Wired what the help did to him: "I'm the one with an AI agent, which feels very different actually. And I feel totally threatened by it." [16]
Strogatz, a Cornell professor, is not arguing that the tools are weak. "You will not be able to compete without AI in the future if you want to do breakthrough math," he told Wired [13]. His objection is to which result got the announcement. He called the Navier-Stokes existence and smoothness problem "a very theoretical math problem of essentially no interest to a working engineer in civil engineering or aerodynamics" [9], and the announcement "a marketing device for them to prove how good their machines are" [10].
He also priced it. Solving the problem is "maybe worth a trillion dollars for OpenAI to show they're better than Anthropic," Strogatz said [11]. Set against the prize money, that is a ratio of about a million to one [18]. He puts the pace down to corporate labs racing for headlines ahead of blockbuster IPOs [12].
The credit question has at least three parties in it. The OpenAI work builds on a strategy developed by the Spanish mathematicians Diego Cordoba and Luis Martinez-Zoroa [3]. Strogatz said Tristan Buckmaster of New York University, working with Levent Alpoge, posted a solution a couple of days before OpenAI did [19]. Strogatz said he would love to see Cordoba and Martinez-Zoroa get the money [20].
The announcements are now a run rather than a single event. Counting the ten problems announced in August, OpenAI has claimed progress on eleven long-standing problems since the summer [17]. The interview does not say whether the August results have been checked by anyone outside the company.
Two things separate Townsend's case from this week's: whether someone outside the lab can check the artifact, and whether the task resembles the work your own team does. Townsend clears both, on his problem, in his field, with his own judgement of the output [14]. The Navier-Stokes claim does not clear the second, on Strogatz's account of who cares about the problem [9], and the first is still open [2]. A claim in that position is evidence about OpenAI's standing against Anthropic. Strogatz said that is what it is for [10].
What to watch
- Whether independent mathematicians confirm the OpenAI solution as announced, and how long that takes.
- Who ends up with the $1 million prize: Cordoba and Martinez-Zoroa, Buckmaster and Alpoge, or OpenAI.
- Whether either lab publishes checkable artifacts for the August results and Anthropic's 29,500 formalized theorems.