Two dozen Fields winners warn AI math race threatens the field
The Journal reported that OpenAI published a solution this week to the Navier-Stokes problem, one of seven Millennium Prize Problems selected in 2000 with a $1 million prize offered for each solution. Within days of the release, OpenAI said it was close to solving another Millennium Prize problem, raising the possibility of two discoveries that had eluded mathematicians for decades arriving in the same week.
The Journal reported that OpenAI deployed the math effort after hearing rumors that Anthropic had already solved one of the prize problems. A swarm of as many as 10,000 AI agents worked for 88 hours, building on each other’s work and on prior human mathematics until they reached a solution. The effort required millions of dollars in computing resources.
Before the proof appeared online, social media lit up with suspicion that OpenAI’s model had appropriated unpublished work by human mathematicians, which the company denied, the Journal reported.
The Wall Street Journal characterized the Millennium Prize moment as “a friendlier version of the Hugging Face incident,” writing that both demonstrate what increasingly powerful AI systems can do when they are turned loose.
About two dozen Fields winners, including Terence Tao, warned that the AI industry’s obsession with solving math problems could “undermine the purpose of math itself.” Tao told the Journal that math is quickly approaching a “worst-case scenario” that threatens the long-term health of his field and society at large.
Tao wrote on his blog in 2023 that AI could become “radically transformative,” to the point where maintaining traditional mathematical practices and culture without adaptation “would become unsustainable.” He told the Journal he still values the models but is losing faith in the companies building them.
Timothy Gowers, another Fields medalist, said the result leaves open “the flickering dream that there’s still a human role in mathematics.” But, he added, “my guess is that if you’ve got a model that’s capable of doing this, it will be capable of doing lots of other things.”
“I don’t want to say it’s all over,” Gowers told the Journal, “but I certainly don’t want to say it’s not all over.”
Fields Medal-winning mathematician Alain Connes, who said in 2000 that the seven prize problems were “totally inaccessible to computers,” told the Journal he is “extremely positive” about AI’s progress in math.
For those who haven’t thought about math since high school, the recent events have introduced the public to Erdős problems, the Jacobian conjecture, and a phenomenon known as “finite-time blowup.”
Hours after the math news broke, Anthropic researcher Jacob Coxon resigned over safety concerns and posted publicly that people building AI “earnestly believe that it could kill us all.” Coxon had previously worked at both OpenAI and Anthropic.
Evan Hubinger, a colleague at Anthropic whose job is making sure AI doesn’t kill humans, wrote on X that he put the odds of annihilation above 10% — a probability he characterized as higher than Stephen Curry missing a free throw.
Last month, OpenAI hosted mathematicians at its San Francisco offices to discuss a future in which AI surpasses human mathematical ability. The summit began with that assumption and worked through implications for education, collaboration, the publication of ideas, and training the next generation of mathematicians, the Journal reported. The attendees didn’t come up with any solutions, but that wasn’t the point — their goal, the Journal reported, was selecting the right problems to work on, like the mathematicians who chose the Millennium Prize Problems.
A year ago, expert forecasters gave AI a roughly 20% chance of solving one of the Millennium Prize Problems by 2030. The breakthrough arrived sooner than those forecasts predicted.