By now, most people have heard that OpenAI has solved a Millennium Prize Problem; specifically, a proof of finite-time blowup for the forced three-dimensional Navier-Stokes equations, announced yesterday afternoon at 1:20pm ET.@OpenAI, September 8, 2026, 1:20pm ET As someone who has followed the story somewhat closely, along with the surrounding discourse around AI in math, I wanted to write a short blog post that lays out a timeline of events, and (hopefully) contextualizes this result with the broader discourse over the past year.
Trust me bro
Towards the beginning of 2026, AI in mathematics was still treated as somewhat of a novelty. There were some interesting results, mostly on Erdős problems, mostly with commerically available AI models, and mostly from a few peoplePrimarily @AcerFur and @Liam06972452. on Twitter. Mathematicians were paying attention, though without much alarm. For example, after one such problem was solved in January, Terence Taothe Fields Medal-winning, MacArthur-recognized, field-defining mathematician, who kept a page tracking AI-assisted mathematical contributions,AI contributions to Erdős problems remarked that he was less interested in the solution itself than in "the emerging AI-powered capability to rapidly write and rewrite expositions of the solution."@tao, January 7, 2026.
This stopped being a novelty in May, when OpenAI announced that an internal model had disproved the Erdős unit distance conjecture,An OpenAI model has disproved a central conjecture in discrete geometry, May 20, 2026. an open problem in discrete geometry. This result was supplemented by a companion paper, written by nine mathematicians, that digested the argument and simplified it.Alon et al., Remarks on the disproof of the unit distance conjecture, May 20, 2026. This sort of focus on human understanding of these proofs, however, did not last very long.
By the summer, the character of the results being posted had changed quite a bit. On July 19, Levent Alpöge (a member of technical staff at Anthropic) posted a counterexample to the Jacobian conjecture:@__alpoge__, July 19, 2026.
hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final
((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)
Notably, this was a black-box result: no paper accompanied it, nor any explanation of how the counterexample had been found. Some understanding of it was recovered afterwards, but only after other people@davikrehalt, July 20, 2026. went and did that work (with the use of AI models). Subsequently, Alpöge posted a resolution of the Hadamard conjecture on August 12,@__alpoge__, August 12, 2026. and, on August 23, a proof that $S^6$ admits a complex structure.@__alpoge__, August 23, 2026..
The labs followed a similar trend. On August 1, OpenAI published ten results attributed to an internal version of Astra,Ten advances in mathematics and theoretical computer science, August 1, 2026. among them the first explicit construction of a non-sofic group. This construction (and the nine other results) did not come with a companion paper written by mathematicians with domain expertise, the way the unit distance result had been. Instead, the only proofs provided were entirely ones entirely AI-written. Again, understanding had to come from outside. Francesco Fournier-Facio, a mathematician at Cambridge, independently worked out the criterion underneath the OpenAI results and used that to produce a broader family of examples, including torsion-free onesA torsion-free non-sofic group, August 3, 2026.. Similarly, on August 10, Anthropic announced that an unreleased model had raised the lower bound on the proportion of zeros of the zeta function lying on the critical line from 41.6% to 67.2%, after a member of technical staff asked it to "take a real stab" at the Riemann hypothesis and then mostly sent it words of encouragement.Learning more about Claude's mathematical capabilities, August 10, 2026. Like others, this was not released with a companion paper, nor any focus on helping mathematicians understand this result, beyond the correctness of it.
Hype moments and aura
Around this time, some mathematicians started pushing back. Their issue wasn't that the proofs were wrong; in fact, most were accompanied by (compiling) Lean certificates. One complaint was the writing. In a talk at IPAM,Accelerating Math and Theoretical Physics with AI, IPAM, UCLA. Tao described how AI-generated proofs would spend pages on trivial details and then a single line on the key idea. The larger complaint was that mathematics was being used as a benchmark for AI companies, strip-mined@tao, September 2, 2026. for hype moments and aura by people with no particular interest in the field itself. One such critique, written by Hugo Duminil-Copin,A recipient of the 2022 Fields Medal. in a blog post sarcastically titled "Care for a little more AI?", lamented that:Care for a little more AI?, Proofs and Prompts, August 30, 2026.
...the current use of AI does not empower us, it petrifies us. These artificial discoveries risk decapitating entire fields before they have had time to develop to their full potential. Even worse, they nuke the mathematical landscape, making it increasingly difficult to inhabit after each blast.
Terence Tao (previously a vocal proponent of AI in math)How Terry Tao Became an Evangelist for AI in Math, Quanta, June 8, 2026. soon began making a similar case. In a sequence of posts on Mathstodon, written over four days from the evening of September 2 to the afternoon of September 5, he developed his argument.
In the first of these posts, Tao described the idea of contamination,A term from AI evaluation, where a problem whose solution is already public is no longer useful as a test. cited Duminil-Copin's article, and suggested that:@tao, September 2, 2026, 11:15pm ET.
It may become necessary to declare certain classes of mathematical problems off-limits to automated solvers, in order to preserve their broader value to the mathematical ecosystem (for instance, through the training of future mathematicians). Admittedly this can be hard to enforce when such tools are both powerful and widely available.
In a second post (whose importance will be obvious later) he gave a concrete example. Using the famous Navier-Stokes equations, Tao explained what could be lost if an important problem like this were solved without any insight into the process that produced the answer:@tao, September 3, 2026, 11:37am ET.
[T]here is now a scenario in which an autonomous AI harness, backed by an enormous amount of computational resources, performs this entire iteration internally, and ends up producing the final ansatz, and thence the solution to the Navier-Stokes regularity problem, while the AI company running the harness keeps the process to arrive at that ansatz almost completely out of public view. Technically, one of the most prominent open problems in mathematics would now be solved; but there would be almost no value added to mathematics as a consequence.
Later that day, he put the concern more generally:@tao, September 3, 2026, 2:53pm ET.
But the currently fashionable practice of pointing a powerful AI tool at the task of answering a problem $X$, unguided by any human expert in the field $X$ resides in, has created an unprecedented divergence between the production of answers, and the production of insight, to the point where the two questions have become negatively correlated.
He followed this up with a case study that had unfolded a few days earlier. On August 31, Julia StadlmannJ.L. Doob Research Assistant Professor at the University of Illinois. released a preprint lowering the best known bound on gaps between primes.arXiv:2608.31126, August 31, 2026. Tao described how, immediately after this:@tao, September 3, 2026, 2:53pm ET.
...we were treated to the unedifying spectacle of no fewer than three separate AI companies racing to announce their own improvement on the bounded gaps between prime result that I mentioned yesterday. These results are numerically stronger than Stadlmann's improvement, but I am very glad that Stadlmann was able to complete her analysis just in time before the problem became contaminated [...]
It's fun to make predictions
Unbeknownst to Tao, prior to his post on Navier-Stokes, rumors had already been circulating about an AI lab solving two Millennium Prize Problems. Motivated by these rumors, and Tao's post, Andrew Curran tweeted the following:@AndrewCurran_, September 4, 2026, 10:26pm ET.
It's fun to make predictions. Here's a new one: Anthropic has solved a Millennium Prize Problem. And I'll be even more specific. Claude has solved Navier-Stokes. It is out for expert review. And to give myself a hard deadline, they will announce it before the IPO.
The next morning, Tao added a clarification to his own thread:@tao, September 5, 2026, 11:19am ET.
[I]n response to recent rumors about a possible solution to the Navier-Stokes problem: I am not aware of any significant developments in this regard; the above discussion is hypothetical, but not completely implausible at the current level of development of AI technology.
The rumors continued regardless.
Let ten thousand agents bloom!
At 11:58pm ET on Monday, September 7, Tristan BuckmasterProfessor of mathematics at NYU's Courant Institute. posted three results, in a joint collaboration with Levent Alpöge: finite-time blowup with smooth forcing for incompressible porous media, for Boussinesq, and for 3d incompressible Euler, with Lean formalizations for each.@tristanbuckmaster, September 7, 2026, 11:58pm ET. Alongside them, he posted a four page statement,statement.pdf. explaining the following.
On September 3, with the rumors circulating, Tristan had emailed a mathematician at OpenAI to correct the record, and over the following days was asked to meet with increasing urgency. On a call on Sunday, September 6, he was told that an internal OpenAI model had proved finite-time blowup for forced Navier-Stokes:
I asked when the first prompt was sent by them. This question was not answered directly by OpenAI for some time. Eventually it was agreed that it had been sent in the past few days, after information about our work had reached OpenAI. I asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project. I was told the model did not look up user data. I asked again, about training, and I did not get an answer.OpenAI says its researchers "did not see any of their work through any means until they released it publicly," but adds that "while unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models." On the Navier-Stokes Millennium Prize Problem.
He was then given an ultimatum:
Two proposals were offered to me. The first was that we post our Euler result, and that OpenAI post its Navier-Stokes result the next day. The second was that, after posting Euler, I alone write a paper presenting the Navier-Stokes result, acknowledging that an internal OpenAI model had resolved it. Sebastien twice asserted that he wanted Levent removed from authorship, and said it would all be simple if only it were not the case that, and it was so annoying that, Levent works at Anthropic. [...] I declined both offers.
OpenAI's announcement came thirteen hours later. Using ten thousand agents on an internal model, running for 88 hours, they had produced a proof and a Lean formalization that a fluid can develop a singularity in finite time. In other words, a "solution" to Navier-Stokes.@OpenAI, September 8, 2026. Shortly after, Sébastien Bubeck posted their account of how the effort had started:@SebastienBubeck, September 8, 2026, 1:39pm ET.
We began working on the Millennium problems due to viral twitter rumors that Anthropic had resolved 2 Millenium problems.
He said they saw none of Buckmaster and Alpöge's work until it was released, and that they had assumed Anthropic was behind the rumors because Alpöge's posts use Anthropic models. He made sure to emphasize that:
although a group of people was involved in our efforts, we collectively had no research-level expertise in fluid dynamics and the Navier-Stokes problem, and therefore were unable to meaningfully contribute to the mathematical content.
The difference from May is shocking. The unit distance result had been posted alongside a companion paper by nine mathematicians, invited to review the proof and develop a more comprehensive understanding of it. The Navier-Stokes result, rather than the result of collaboration, came out of direct competition with two people already working on the problem. At no point did understanding seem to be even a minor consideration.
⁂
Tao's post, ironically enough, turned out to be a self-fulfilling prophecy. By choosing Navier-Stokes as his example of how an AI-generated solution could inhibit the development of a field, he inadvertently fueled rumors about the problem being solved, that in short fashion led to the problem being solved, more or less in the way he described.
After the announcement from OpenAI, he made a new post, reflecting on the state of mathematics:@tao, September 8, 2026, 4:32pm ET.
We have now seen that even the rumor of someone working on a problem can trigger a massive amount of AI-powered effort to flatten it before the original research project has time to reach its full potential. The incentives may now be pointing in the direction of no longer sharing any promising research directions with the broader community, which would reverse centuries of traditions of open science and do serious long-term damage to the future of the field.
If this is all a bit too depressing, I recommend reading The 92-Year-Old Mathematician and the Teenage Apprentice, published a few days ago, about Joan Birman and the fifteen-year-old she took on as a mentee at the age of 92.