OpenAI, the company behind ChatGPT, claims that a new artificial intelligence model, supported by a fleet of 10,000 AI agents, managed to solve part of one of the world’s most difficult mathematical problems in just 88 hours.
The problem involves the so-called Navier-Stokes equations, which describe the motion of fluids. For around 90 years, key aspects of the problem have lacked an adequate mathematical proof.
OpenAI has described the result as a milestone and one of the clearest demonstrations yet of the extraordinary progress being made in artificial intelligence.
The new system consisted of roughly 10,000 AI agents working almost autonomously. Together, they reportedly found a solution to a Navier-Stokes-related case that is considered important for understanding turbulence — a phenomenon that remains only partially understood in physics.
The broader mathematical problem is connected to the prestigious Millennium Prize Problems established by the Clay Mathematics Institute. Solving it requires proving four separate statements, and the AI system reportedly managed to resolve two of them.
Although a correct solution to the full problem carries a $1 million prize, OpenAI says it has no intention of claiming the money.
“Our goal in publishing this result is to report significant progress in the capabilities of our AI models, not to pursue the Millennium Prize,” the company said.
To reach the result in only 88 hours, the AI agents exchanged nearly three million messages, processed and generated an astonishing 130 billion tokens of text and code, and consumed computing resources that would reportedly cost around $10 million at commercial rates.
However, the achievement has not come without controversy.
New York University mathematics professor Tristan Buckmaster and his colleague Levent Alpöge, who works for rival AI company Anthropic, claim they had been working on the same problem in parallel using Codex.
Buckmaster says that on September 3, he learned that information about their progress had allegedly “leaked” to OpenAI.
Interestingly, OpenAI reportedly deployed its 10,000-agent system on September 1, after “hearing rumors that the problem had been solved,” before quickly publishing its own result.
“I feel compelled to publicly state what I was told and when, because the alternative is to remain silent while something I know to be false is being published,” Buckmaster said.
OpenAI quickly responded to the allegations, rejecting them while congratulating the two mathematicians on their work. The company categorically denies having access to their research or using their private customer data.
However, one detail in OpenAI’s statement has left some room for speculation.
“While unlikely, we cannot rule out the possibility that anonymized data resulting from their use of our products may have contributed to improving our models. Nevertheless, our proofs and precise results differ significantly,” the company said.
The proposed solution is still awaiting independent verification.
Discussion(0)