Product1 publisher3 min readPublished
OpenAI's 88-hour Navier-Stokes run burned roughly 880,000 agent-hours
OpenAI says the run took 88 hours and cost millions, which tells a small research group more about whether it can compete than the proof itself does, and its wording on user data shows the limit of what any vendor can promise.
The Product Desk · Product desk
What happened
- OpenAI said in a Tuesday blog post that one of its unreleased models took 88 hours to find a solution to the Navier-Stokes problem, which has stumped human researchers for close to 90 years.
- The company said it got there by focusing a swarm of roughly 10,000 AI agents powered by its internal model on the task, and called the result a milestone.
- A day earlier, NYU mathematics professor Tristan Buckmaster had published findings on a related problem with Levent Alpoge, an Anthropic researcher who was not working on behalf of his employer.
- Buckmaster says OpenAI urged him to publish the work with credit to its internal model and to drop Alpoge as a coauthor.
- Buckmaster had used Codex while working on the problem, and says that when he asked OpenAI whether it had accessed those sessions, the company's responses became evasive and then hostile.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- exposure Anyone running unpublished work through a vendor's assistant is left relying on that vendor's own account of what the logs did, since OpenAI's wording covers direct access but not what de-identified usage data may have done to a model.
- decision Choosing an assistant for pre-publication work turns into a contracting question about retention and training exclusion, rather than a preference about which tool writes better code.
- constraint There is no adjudicator in this dispute: the Clay Mathematics Institute awards a prize rather than ruling on bylines, and OpenAI has taken itself out of even that process by declining the money.
- precedent The fear researchers have named is a chilling effect, and its practical form is a small group choosing not to mention progress at a seminar until the paper is out.
OpenAI's post says that no specific user data was accessed in order to solve the problem [10]. The same post says the company cannot rule out that de-identified data derived from the pair's use of its products helped improve its models, while stressing that the two proofs differ significantly [11]. Someone on the other side of that has no way to test either sentence, and nor does a reviewer, because the provenance of AI-generated work is hard to establish even in calm conditions, as The Verge puts it [17].
The figure that will travel is 88 hours [1]. The figure that matters to a group deciding whether to keep competing is the product of the two numbers OpenAI published: about 10,000 agents held on the task for 88 hours is on the order of 880,000 agent-hours [1]. OpenAI says the effort was hurried and cost millions of dollars [13]. The bounty attached to the problem is $1 million [3], the Clay Mathematics Institute has not awarded it, and OpenAI says it does not intend to claim it [14]. On the plainest reading of "millions", the run cost at least twice what the prize pays [2], which is consistent with the company's stated goal of reporting on the substantial progress of its models [15].
The thing being pitched is a milestone in mathematics [2]. The thing actually done is a throughput demonstration aimed at a problem the company shows no public record of working on before September [16], after word of another group's progress reached it, according to The Verge's account of how the effort began [19]. The conduct around credit is the part with no paperwork. By Buckmaster's account an OpenAI researcher asked why he would ruin his career by going public, and added that if Buckmaster did not want him to be nice, he did not have to be nice [7]. Abhishek Saha, a mathematics professor at Queen Mary University of London, told The Verge this is the kind of thing mathematicians generally will not do [12]. That is a statement about manners, and manners are the enforcement mechanism.
For a team that has to decide something on Monday, two axes do the work. One: whether your claim to an output depends on being first to publish or file. Two: whether the vendor you are typing into is a plausible competitor for that same output. Where both are true, the work belongs behind a written retention and training exclusion, or off the assistant entirely, and the cost of that is real, because the cheap tier is usually the good tier. Where neither is true, the consumer product is fine and the provenance argument is theatre. The awkward quadrant is priority-sensitive work at a vendor who is not a competitor today, where the honest answer is dated records of your own, kept outside the tool, so that you can show your timeline without needing the vendor to confirm it.
No evidence shows that anyone at OpenAI read a Codex session. The company can say no human looked, but it cannot say nothing was retained, and Buckmaster has no independent record of either.</body_markdown> </invoke>
What to watch
- Whether the Clay Mathematics Institute comments on the result or the prize that OpenAI says it will not claim.
- Whether Buckmaster and Alpoge publish with the coauthorship intact and how OpenAI is credited, if at all.
- Whether a second lab points comparable compute at another Millennium Prize problem, which would make this a pattern rather than one hurried run.