· via Hacker News – Front Page (native)
Anthropic models reportedly prove the Riemann Hypothesis in 17.4 million lines of Lean
A mathematician says Anthropic models produced a 17.4-million-line Lean proof of the Riemann Hypothesis. Scott Aaronson bets it stands, but the verification of AI-generated proofs is now the central question.
A 17.4-million-line claim
On November 20, Scott Aaronson reported on his blog Shtetl-Optimized that the mathematician Levent Alpöge has gotten Anthropic's models to prove the Generalized Riemann Hypothesis, with the purported proof written as roughly 17.4 million lines of code in the Lean proof assistant. As of these posts, the claim traces to Alpöge's account rather than to a formal paper or an Anthropic announcement. Both Aaronson's post and an essay on signoregalilei.com, published the same day, quickly reached the front page of Hacker News, where the claim and its verification are being argued over.
The two accounts also frame the result slightly differently: Aaronson describes it as the generalized hypothesis, a stronger family of statements, while the signoregalilei.com essay presents it as the Riemann Hypothesis itself.
What the hypothesis says
In the version Aaronson gives for non-specialists, the conjecture holds that the number of primes up to n stays very close to what you would see if every integer greater than 1 had an independent 1/ln(x) chance of being prime, with the total error capped around the square root of n. The Prime Number Theorem, proved in 1896, had already fixed the rough count at n/ln(n); Riemann's 1859 conjecture tightens the bound on how far the primes can drift from it. Aaronson opens his post with Riemann himself conceding he had set the search for a rigorous argument aside.
According to the signoregalilei.com essay, the problem has been mathematics' most celebrated open question since Fermat's Last Theorem was settled in the 1990s, and a large body of published theorems simply assume it. A valid proof would turn that shelf of conditional results into unconditional ones at a stroke.
Aaronson adds that machine assistance has been closing in for a while: a recent result showed the nontrivial zeros of the zeta function lie no more than 3/8 from the critical line, work that Quanta Magazine covered the day before his post and that, in his view, already delivers much of what a full proof would imply for number theory.
Whether to trust the certificate
Aaronson is hopeful but not carefree. He points out that Lean, for all the weight the community now places on it, has a kernel with a history of bugs that once admitted proofs of 0=1, and that the soundness of the underlying type theory is not fully settled. He argues those weaknesses urgently need repair, because a sufficiently determined AI could exploit them to certify a false statement. He writes that he sees no sign of that happening here: to date, no AI result backed by a Lean certificate has had to be withdrawn, while several results lacking certificates have been. His stated bet is that the proof stands, though not with complete certainty.
Acceptance by working mathematicians is a separate question. Aaronson expects researchers will regard the enormous Lean artifact as provisional until they can extract, explain and rebuild its core ideas in human terms, a distillation he thinks could take years.
A milestone greeted without celebration
The signoregalilei.com essay captures the strange mood. The author had expected an announcement of this magnitude to land like a Mars rover touchdown or the discovery of the Higgs boson; instead the atmosphere feels mournful, because the proof, if it holds, came from a machine rather than a person, as if the field had given up and peeked at the answers in the back of the book. Invoking Galileo, the author argues that truth is simply however the world turns out to be, and worries that the window for genuinely human discovery is closing: eventually, anything people still prove by hand will be a problem nobody thought to hand to a machine first. Aaronson's closing anecdote points the same way: his nine-year-old son, told the news, replied that he was not even surprised.
Why it matters
If Alpöge's claim survives scrutiny, one of the longest-standing open problems in mathematics falls, and a large literature built on the hypothesis becomes unconditional overnight. It would also be the clearest demonstration yet that frontier AI models can sustain verifiable mathematical work at a scale no human could match. The episode exposes where the real bottleneck now lies: not in generating proofs, but in hardening the verification tools themselves and in human comprehension of what machines produce. Aaronson's warning that proof-assistant kernels and type-theory soundness must be fixed before an AI can weaponize a bug applies well beyond this single claim. And it resets the map of what remains: with Riemann reportedly fallen, Aaronson casts P versus NP as the last great prize, while others are left asking what role human mathematicians will still play. The claim is unverified, and the community's cautious, audit-first reaction is exactly what a result of this weight demands.
- #anthropic
- #riemann-hypothesis
- #lean
- #formal-verification
- #mathematics
- #ai