In a year of unthinkable math progress, as increasingly powerful AI systems have cracked conjectures that had long stumped humans, OpenAI says it reached a breakthrough on Tuesday that represents the most significant advance yet.
Back in 2000, seven of the deepest and most difficult unresolved questions in mathematics were enshrined as the Millennium Prize Problems. Since then, some of the world's brightest minds have pursued the most elusive answers in math-and the $1 million promised for any solution.
Despite the intellectual firepower aimed at these notorious problems, only one had ever been solved.
But now OpenAI has published what it says is the solution to a second Millennium Prize Problem. And this one was produced by AI.
The lab said on Tuesday that one of its internal models has solved the Navier-Stokes existence and smoothness problem, a fundamental question about the rules that describe how liquids and gases move.
For nearly a century, mathematicians have tried to determine whether smooth solutions to the equations describing three-dimensional fluid motion always remain smooth or if they ever break down. To put it less mathematically: On paper, can water blow up? OpenAI says its model found a scenario in which "an initially smooth fluid at rest can develop a singularity in a finite time." In the proof, a spaghetti-like vortex gets smaller while spinning faster and faster, its speed growing without limit in what mathematicians call "finite-time blowup."
OpenAI said it would publish its findings along with the 165-page proof and a formalization that verifies the argument step by step. The company says it took as many as 10,000 AI agents working together on the problem about 88 hours to find the solution. It also took "millions of dollars" in computing resources.
The race to solve a Millennium Prize Problem had become so ferociously competitive that OpenAI's researchers did not circulate the results with outside mathematicians before publication.
Over the weekend, the math community on social media erupted with chatter that a Millennium Prize Problem had been solved. In the hours before Tuesday's announcement, the mathematical drama took another turn when competing researchers challenged OpenAI's version of events.
By the time mathematicians prepared to read the proof, they could agree on one thing: It appeared to be the most striking example of artificial intelligence doing something that lots of brilliant humans have been trying and failing to do for a very long time.
"It's undeniable that symbolically, it's a big moment-and the next in a natural chain of big moments," said Timothy Gowers, a Fields Medal-winning mathematician who had not yet reviewed the paper.
Not long ago, the smartest models struggled with basic math. In recent years, they have become smart enough to earn a silver medal and then gold at the International Mathematical Olympiad. By this year, they were solving major open problems. Their rapid improvement has left researchers inside frontier labs both surprised and alarmed by the pace of development.
"Even a month ago, I don't think we would have predicted that we would be talking about a Millennium Prize Problem," said Jakub Pachocki, OpenAI's chief scientist. "We see this as a demonstration of just how far AI has gotten-and how quickly."
In recent months, the AI companies have flaunted the capabilities of their models by trumpeting advances in combinatorial geometry and algebraic number theory, turning math into an unlikely form of advertising. Despite those strides, the possibility of a frontier lab taking the giant leap it would require to solve a Millennium Prize Problem seemed improbable.
This latest discovery was made by an unreleased model that OpenAI says is "significantly more capable" than GPT-6 Astra, which the lab released just last week.
When the company tested both models on its own benchmark of math problems, Astra was able to solve about 10%. The unreleased model scored closer to 50%.
"Basically, you throw at it almost any open problem," said OpenAI researcher Sebastien Bubeck, "and it's a coin flip whether the model can solve it."
This model has been in training since late August, the company said. On Sept. 1, with the math community buzzing about rumors that Anthropic had solved two Millennium Prize problems, OpenAI unleashed the model on all six open questions. Even then, researchers expected it would produce the same results as the other times they have pointed their models at the infamous Riemann hypothesis, the P vs. NP problem and the Navier-Stokes equation.
"We didn't expect it to solve any," said OpenAI researcher Noam Brown.
After 50 hours of work, the AI had made enough progress on a problem related to Navier-Stokes that the humans intervened. They made a calculation of their own to shift the agents away from the other five problems-and put all of the effort's computing power into this one.
With that investment, the AI increased the size of its army of agents from 100 to as many as 10,000. To resolve Navier-Stokes, they sent 2.7 million messages and used 130 billion output tokens-the rough equivalent of a million books. The company says their work was completed on Saturday and computer-verified on Sunday.
The news of a result surprised even the wonkiest researchers inside OpenAI, who scrambled to check it by using their own AI agents.
Dan Roberts was at a park with his kids when he got word of the breakthrough on Slack.
Cheuk Hei Chu found out when he looked at his phone after emerging from an escape room.
OpenAI's chief research officer Mark Chen was asleep in Tokyo and woke up to several frantic calls.
"I couldn't sleep for the rest of the night," he said.
But it wouldn't be the last sleepless night for the OpenAI team.
In recent months, solving the kind of math problems that most people find unintelligible has become a subject of vicious competition between the world's most valuable AI companies.
And in the hours leading up to publication, that rivalry blew up in a very human way.
OpenAI says it began the furious sprint to a Navier-Stokes resolution last week, after hearing rumors that a Millennium Prize problem had been solved-by Anthropic.
As it turned out, the rumors concerned work by Anthropic researcher Levent Alpöge and New York University math professor Tristan Buckmaster on a related question about fluid dynamics. OpenAI says it contacted them over the weekend to discuss publishing the results at the same time. From there, the accounts diverge. Before OpenAI released its proof, Buckmaster published his work with Alpöge.
Buckmaster said in a statement that he and Alpöge used models from both Anthropic and OpenAI in their work, and he asked OpenAI whether its new model had been trained on or had access to those sessions. OpenAI said that neither its agents nor researchers saw the competing unpublished work until it was released publicly and congratulated them on the accomplishment.
The proof that was slated for release on Tuesday dates back to the 1800s, when Claude-Louis Navier and George Gabriel Stokes developed a set of partial differential equations that would later help us design airplanes, forecast the weather and understand how blood flows through our bodies.
In 2000, the Clay Mathematics Institute made the Navier-Stokes equation one of the seven problems worthy of a million-dollar bounty. The prizes were meant to draw attention to how little we knew about math-no matter how many of us have suffered through high-school algebra.
In 2010, a reclusive Russian mathematician Grigori Perelman was honored for proving the Poincaré conjecture, the first Millennium Prize Problem to be solved. Then he proved just how unpredictable humans can be when he declined the money.
OpenAI won't be collecting the $1 million, either.
"Our goal in releasing this result is to report on the substantial progress of our AI models," the company said. "We do not intend to claim the Millennium Prize."
Comments