Steven Strogatz Says AI Math Is Pushing Humans Toward Interpretation

Steven Strogatz sees recent AI advances in mathematics as more than a contest to solve famous problems. In a newly published WIRED interview, he argues that breakthrough mathematics may soon require AI, while human researchers could be pushed toward explaining results generated by systems they do not fully understand.

The immediate backdrop is a burst of prominent claims. OpenAI said it used tens of thousands of agents on a 90-year-old problem related to Navier–Stokes equations, carrying a $1 million prize. The proposed solution still needs independent verification and builds on a strategy developed by Spanish mathematicians Diego Córdoba and Luis Martínez-Zoroa.

Checking a proof is not the same as finding one

Another recent result illustrates an important distinction. Anthropic says Claude formalized an existing proof of Fermat’s Last Theorem in Lean, a programming language that can check the logical steps of a proof. The advance it describes is not a new proof of the theorem; it is the conversion of an established argument into a form a computer can verify.

That distinction matters because formalization and discovery solve different problems. A proof assistant can test whether every encoded link in a logical chain follows from its rules. But getting a proof into that form is difficult: human-written mathematics often leaves routine steps unstated, while a formal system requires those steps to be spelled out. Anthropic says Claude worked largely autonomously through a multi-agent effort, with occasional high-level human instructions.

11 daysTime Anthropic says Claude took

Anthropic says Claude worked largely autonomously for 11 days on the Fermat’s Last Theorem formalization.

29,500Intermediate theorems used

Anthropic says Claude proved 29,500 intermediate theorems used in the final formalization.

13 million linesLean code produced

Anthropic says the project generated 13 million lines of Lean code.

The figures are Anthropic’s own account of its project, but they show why Strogatz is focused on what happens after a machine-checkable result appears. He calls for “proof digestion”: explanations that let people understand and appreciate a result, rather than merely accept that software has verified it. He thinks humans may retain that translating role for a while, but doubts it will be permanent.

Speed also turns credit into a live issue

The Navier–Stokes episode shows that this transformation is not only technical. Mathematician Tristan Buckmaster alleged that OpenAI accelerated its effort after learning about work he had conducted with Anthropic researcher Levent Alpöge, and tried to influence who received credit. The allegation is unresolved, but it puts provenance at the center of an AI-driven research race.

The tradeoffs Strogatz identifies

• AI could broaden participation in mathematics, rather than reserve difficult work for a small group with years of specialist training.
• Researchers motivated by being first to solve hard problems may lose a central source of purpose if machines consistently arrive first.
• If AI performs work now associated with professional mathematicians, Strogatz questions why institutions would continue funding people to do it.

A result still needs a place in human work

Strogatz frames AI’s effect on pure mathematics as either devastation or revolution, depending on whether one values answers alone or the human challenge of finding them. His collaborator Alex Townsend offers a nearer-term version of that tension: he said ChatGPT helped him solve a decades-old numerical linear algebra problem, yet working with an AI agent felt different from personally standing at the frontier of knowledge.

For now, a checked formal proof and a claimed solution awaiting independent scrutiny leave different human jobs unfinished. One needs explanation people can learn from; the other still needs outside assessment. Strogatz’s warning is that even if AI makes both processes faster, mathematics may have to decide whether correct answers alone are enough to sustain a profession built around human understanding.