TL;DR. Levent Alpöge, a mathematician at Anthropic, writes that he and Claude have proved the Riemann Hypothesis, the most famous open problem in mathematics. The evidence is not a paper but only code for the proof checker Lean: 17 million lines that no human will read any time soon. We downloaded the code and looked. The statement at stake fits on 28 lines and rests on a definition that was fixed beforehand. That is the lesson for every company that lets AI do work: once nobody can read the result any more, people have to fix the question in advance and hand the checking to a machine.
What is being claimed
On November 20, Alpöge put a four-page note online. It says: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.” The method, he writes, reaches even further, to the so-called Generalized Riemann Hypothesis. The conjecture from 1859 says where the zeros of the zeta function lie, and with that how evenly the prime numbers are distributed. A proof carries one of the seven Millennium Prizes of the Clay Institute.
The timeline in the note starts on October 6 and ends on November 16, six weeks later. Towards the end, Alpöge writes, Claude “far outpaced” him. There is no readable version. A generated long version came to more than 250,000 pages by his account, and he is holding it back as unreadable. All he has published is the Lean code, split into 14 archive parts.
Experts are cautiously impressed. James Maynard, a Fields medalist, told IBM he suspects the proof will turn out to be correct, but that even checking the Lean version will take time. Terence Tao points in his blog to the Lean community, which is taking on the checking right now. Anthropic itself has not commented so far. The note appeared one day after the company's stock market debut; Alpöge raises the timing himself and calls the obvious cynical reading false.
What we checked ourselves
On November 21 we downloaded all 14 parts, about 240 megabytes packed and just over one gigabyte unpacked. They contain 175,187 Lean files with a total of 17,374,215 lines and about 750,000 individual theorems and lemmas.
Two small files are what matters. The main file contains the theorem riemannHypothesis : RiemannHypothesis. The term RiemannHypothesis does not come from Alpöge's code but from Mathlib, the large community library for Lean. It was already there on August 21, in the version the project uses, so before the work began according to the timeline. The extension to the generalized conjecture is stated in a second file of 16 lines, also in Mathlib's terms.
Lines of Lean code: the whole proof vs. the statement a human has to check
Then we searched for the usual back doors that can fool Lean: extra assumptions (“axiom”), gaps left open (“sorry”), shortcuts around the checking kernel, a separate, deviating definition of the conjecture. We found none of them. Five small helper files modify the Lean environment; as we read them, they only put names on a block list of the search helper and do not touch the proofs.
What we did not do: compile the code. That takes hours and is the actual checking step. A user on Hacker News reports that all 175,187 modules ran through without errors. No independent party has confirmed that yet.
The catch
- Not checked yet. Whether the code really compiles and which basic assumptions the kernel counts in the end is being worked out in public by Lean experts right now. Until then it is a claim.
- Checkable does not mean understood. Even if Lean confirms everything, nobody knows why the proof works. Maynard calls that “somewhat frustrating”. Alpöge himself writes that he has not mastered parts of the mathematics used and may have built in many strange detours.
- Lean is software too. A bug in the checking kernel is unlikely, but not impossible. A proof of this size is unusual even for Lean.
- The prize comes in 2028 at the earliest. The Clay Institute requires publication in a qualifying outlet, then two years of waiting and general acceptance. A Lean file on a private website meets none of that yet.
Days from the first timeline entry to the earliest possible review by the Clay Institute
41 daysfinding, according to the timeline in the note
at least 731 daysClay Institute waiting period, counted from a publication that does not exist yet
Finding (Oct 6 to Nov 16, 2026)until the note (Nov 20)waiting period, at the earliest
What this means for mid-sized companies
No company will have to deal with the Riemann Hypothesis. The pattern behind it is another matter. AI more and more often produces more than a person can read: code for the ERP system, analyses across thousands of contracts, reports from machine data. Anyone who wants to check the result line by line can no longer keep up.
The case shows what still holds then. The question was written down precisely before the AI started, by other people, in a shared library. The checking is done by a program that believes nothing and recomputes every step. And in the end a person only has to read one thing: whether the 28 lines really ask what is meant. That is exactly the part the experts checked first.
Translated to your business: your acceptance criteria are worth more than the result. If an agent gets to write the criteria as well, in the end it checks against its own question.
What you should do now
- Fix the question in advance. Before an agent works, a person writes down what must hold at the end, short and checkable. The agent may read this file but not change it.
- Hand the checking to a tool. Tests, reconciliations and plausibility rules run automatically and independently of the agent that produced the result.
- Read the question, not the answer. When the result is too big, spend your review time on the criteria and the tests. That is where it is decided whether “green” means anything.
- Look for back doors. Skipped tests, disabled checks, exception lists: that is the “sorry” in your own code. Have someone search for them before you trust a result.
- Wait for confirmation. A breakthrough is only one once others have checked it. That applies to mathematics and to every vendor figure.
Whether the proof holds we will know in a few weeks; whether it is understood, perhaps only in years. If you want to work out how to make the work of your AI agents checkable, we are happy to talk.