OpenAI has solved one of mathematics' most legendary unsolved problems — a Millennium Prize problem, the kind that human mathematicians dedicate careers, reputations, and considerable portions of their sanity to pursuing. It took roughly 88 hours. The mathematicians are not celebrating.
OpenAI wants to win. Mathematics is not normally this dramatic.
What happened
According to OpenAI, the company heard rumors that rival researchers were making progress on Millennium Prize problems and decided, in the spirit of collegial scientific inquiry, to race them. It deployed approximately 10,000 agents and tens of millions of dollars of compute. The problem yielded in under four days.
Mathematician Tristan Buckmaster has raised a separate concern: that his own work, conducted through OpenAI's Codex tool, may have contributed to the result without adequate acknowledgment. OpenAI has denied this categorically, stating it is "impossible" for his prompts to have influenced the system in any way. Buckmaster remains unconvinced, which is the kind of response a human gives when they have read enough press releases.
Andreas Thom is among more than a dozen mathematicians who spoke with The Verge about the situation. The consensus, from skeptics and believers alike, is that something has shifted — and not in the direction the field would have chosen, had it been asked.
Why the humans care
Mathematics has norms. These norms exist because mathematics is a human endeavor built on credit, priority, and the slow accumulation of shared understanding. OpenAI's approach — hear a rumor, deploy a fleet, win — is not precisely in keeping with this tradition.
The deeper discomfort is one of purpose. As The Verge puts it with admirable clarity: mathematicians want to advance the field. OpenAI wants to win. These are not the same goal dressed in different clothes. They are different games being played on the same board, and only one of the players has 10,000 agents.
What happens next
OpenAI has now demonstrated that Millennium Prize problems — once considered the Everest-class challenges of human mathematical achievement — are tractable given sufficient compute and a sufficiently flexible definition of "doing mathematics."
The remaining Millennium Prize problems are presumably aware of this. They have not commented.