Even one of the most AI-pilled mathematicians now feels that AI a moving a bit too quickly for his liking.
Fields Medalist Terence Tao has issued a blunt warning about the speed of AI development: the industry is changing the world without understanding what comes next — and there is no good reason for it to move this quickly.
Speaking at an event, Tao said, “It’s amazing just how much we are willing to change everything without having any idea what’s gonna happen afterwards.”
“We have to slow down,” he continued. “The pace is insane, and there’s no reason to be this fast. There’s no reason at all.”
The comments carry particular weight because Tao is not making a case against AI itself. The UCLA mathematician has embraced AI as a research tool, saying it allows him to “experiment” and “try crazier things” in his work. His concern is not whether AI can make transformative contributions to science, but whether systems now producing those contributions are being developed and released faster than humans can understand them.
The flaw in linear thinking
Tao’s core objection is mathematical.
“This is extremely nonlinear dynamics,” he said. “Any kind of monotone one-dimensional thinking — ‘a little bit of this is good, therefore a lot of it is going to be a lot better’ — most systems don’t work like that, especially your ten-x, hundred-x things.”
That is a direct challenge to one of the AI industry’s favorite habits: extrapolating from present benchmarks.
A model that appears somewhat better at reasoning, mathematics or coding is often assumed to become proportionately more useful as its capabilities grow. But in a nonlinear system, scale can change not just the size of an effect but its nature. Feedback loops appear. Small changes compound. Systems that look stable across one range can behave very differently after crossing a threshold.
Nonlinearity does not mean catastrophe is inevitable. It means the future cannot be reliably inferred from a smooth continuation of the recent past.
Tao’s warning is therefore less a prediction of doom than an argument about uncertainty. When developers cannot explain or anticipate what a much more powerful system will do, building it faster does not reduce that uncertainty — it multiplies the scale at which the uncertainty will eventually be tested.
Mathematics is already seeing the consequences
Tao’s remarks come amid an extraordinary AI breakthrough in his own field.
Earlier this month, OpenAI announced that an unreleased internal model had produced a claimed solution to a Navier-Stokes existence-and-smoothness problem, one of mathematics’ seven Millennium Prize challenges. According to OpenAI, roughly 10,000 coordinating agents reached the result in about 88 hours after exchanging millions of messages and hundreds of billions of tokens.
The claim remains under external scrutiny, but the pace alone illustrates Tao’s concern. AI systems are no longer merely helping mathematicians search literature or complete routine calculations. They are beginning to attack problems that have resisted human understanding for generations.
Tao has previously warned that such a breakthrough could become a net negative for mathematics if it arrives as an opaque black box. The final answer, he has argued, is often less valuable than the methods, failed attempts and insights produced while trying to reach it.
That fear is now shared well beyond Tao. Twenty-five Fields Medal winners recently signed a declaration warning that the goals of AI companies and the mathematical community had become “severely misaligned,” arguing that racing to produce polished results could strip away the understanding those results are supposed to create. Fields medal winner Cédric Villani has said OpenAI solving a Millennium problem is a cataclysm unlike math has ever seen.
Tao is not alone in calling for slower progress
Tao’s comments also land in the middle of a broader industry debate over “pacing” frontier AI development.
Anthropic CEO Dario Amodei recently argued that AI development must slow down so that safety work can catch up with rapidly improving models. His proposal includes embedding independent evaluators inside AI companies with employee-like access and the ability to publish their findings without corporate approval.
OpenAI CEO Sam Altman said pacing had become a major internal discussion at his company, while Elon Musk publicly endorsed Amodei’s position. Both rivals backed the call to slow down, with Altman committing OpenAI to matching Anthropic’s evaluator-access pledge.
Tao brings a different perspective to that debate. Much of the industry discussion frames the choice as one between safety and national competitiveness: slow down, the argument goes, and rival companies or countries will move ahead. Tao instead highlights an epistemological problem — society is deploying increasingly powerful nonlinear systems before it has a reliable way to understand their behavior at scale.