Terence Tao Had A Month Ago That AI Could Solve Riemann Hypothesis Only By 2050

A little over a month ago, on October 15, Terence Tao sat down with Brady Haran for a Numberphile interview titled “No More Heroes.” At one point, Haran asked the Fields Medalist the obvious question: was the Riemann Hypothesis, mathematics’ most famous unsolved problem, about to fall?

Tao’s answer was skeptical, bordering on dismissive of the AI labs’ chances in the near term. No, he said, even though he was fairly sure AI companies had “secretly devoted much more time to the Riemann hypothesis than what they’ve disclosed,” they still “couldn’t touch it.” In his view, “something beyond current AI” was needed, and the real progress might come only from “some combination of humans and AIs in 2050 or something.”

terrance tao

A quarter-century-out guess from one of the most respected living mathematicians. Yesterday, that guess was wrong by about 24 years.

On November 20, a note claiming a proof of the Riemann Hypothesis appeared on the website of Anthropic mathematician Levent Alpöge, crediting Claude as the main collaborator. “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis,” the note’s opening line reads. It claims to go further still, resolving the Generalized Riemann Hypothesis as well, with the author adding that he expects the Extended version to follow with more time.

The timeline Alpöge laid out suggests the breakthrough came together remarkably fast. According to the 20-step log he says Claude itself compiled, the push ran just 41 days, from early October to a final entry on November 16 marking the Generalized Riemann Hypothesis. That start date lines up almost exactly with the week Tao was telling Numberphile that AI “couldn’t touch” the problem.

It isn’t the first time Claude had been pointed at Riemann. Back in August, an unreleased research version of Claude was set loose on the hypothesis and, while it didn’t crack it, pushed a decades-old related bound from 41% up to 67% — a result Anthropic was careful to say wouldn’t scale into a full proof. Alpöge himself was one of the mathematicians who reviewed that earlier result, and he’s had a busy year, having also used Claude to disprove the Jacobian conjecture and the Exponential Time Hypothesis.

The new claim credits Claude with the parts of the argument Alpöge says he hasn’t personally mastered — automorphic forms on higher-rank groups — and with suggesting the chain length for a key theta-lift construction. The whole thing has reportedly been formalized in the Lean proof assistant.

None of this makes it official. The Clay Mathematics Institute requires a proof to be published, scrutinized, and accepted by the mathematical community for at least two years before a $1 million prize is even considered, and mathematicians online have been picking the claim apart since Friday, with some questioning whether the Lean formalization even states the theorem it claims to. Verification of a 250,000-page argument, by human standards, does not happen in a weekend.

Still, Tao’s “2050 or something” now reads very differently than it did a month ago. Whether what landed on Alpöge’s website this week survives the scrutiny ahead is still an open question — but it’s already forced the timeline for AI and the Riemann Hypothesis to be rewritten decades ahead of schedule.

Posted in AI