Has AI proved the Riemann Hypothesis?
A researcher says he has used Anthropic's AI models to prove the Generalized Riemann Hypothesis, a central open problem in number theory that has resisted mathematicians for more than 150 years. Levent Alpöge says the models produced 17.4 million lines of Lean code, a formal proof language meant to check each step of an argument by machine.
Scott Aaronson, a computer scientist who writes a blog, reported the news after waking up to it. He said he is not ready to call the proof settled, but he is leaning toward believing it.
What is the Riemann Hypothesis about?
The Riemann Hypothesis concerns the distribution of prime numbers. In its simplest form, it says the number of primes up to a given value is very close to what you would expect if each integer had an independent chance of being prime, based on its size. The deviations from that expectation are bounded by a standard statistical error term that never exceeds a specific limit. Mathematicians have known since 1896 that the number of primes up to n is roughly n divided by the natural logarithm of n, a result called the Prime Number Theorem. But the finer question of how far the actual counts can drift remains open.
The problem has a long and famous history. Bernhard Riemann suggested in 1859 that it was "very probable" that all the nontrivial roots of his zeta function were real, though he admitted he had not proven it. David Hilbert later said that if he woke after sleeping for a thousand years, his first question would be whether the Riemann Hypothesis had been proven. G. H. Hardy, fearing his ship might sink, once sent colleagues a telegram claiming a proof, so the world might at least believe the proof had gone down with him.
Aaronson said the claim comes as AI has already made progress on the same territory. Earlier, AI systems reportedly proved that the nontrivial zeros of the Riemann zeta function lie at most 3/8 away from the critical line. Aaronson noted that this result carries most of the implications for number theory that a full proof would have. Quanta Magazine published an article about that breakthrough the day before he wrote his post.
Can Lean proofs be trusted?
The verification system itself is a point of concern. Aaronson acknowledged that Lean has had problems, including kernel bugs that have, in the past, allowed false proofs of 0=1. He also pointed to questions about the soundness of the underlying type theory. He argued those issues should be fixed urgently, since a motivated AI could in principle exploit them to prove a false statement.
Aaronson said he sees no sign that this has happened here. He noted that none of the AI results that came with Lean certificates have had to be retracted so far, while several AI results without them have been. That record is the basis for his bet that this proof will stand.
Even so, mathematicians are unlikely to treat 17 million lines of code as the end of the story. Aaronson said they will want to explain, distill and rebuild the ideas for themselves, a process he expects will take years.
The announcement also drew a reaction online. Some social media users mocked critics who had called large language models "stochastic parrots," a term from 2021, by pointing to the proof claim. Aaronson's post included that joke as an update, and it was clearly meant as satire.
Dieser Artikel wurde mit Unterstützung von KI verfasst.
News Factory APP - agentische News für besseres SEO & AEO.