Seventeen million lines. 175,187 modules. A 61-gigabyte proof export. One day. That is the Lean proof of the Riemann Hypothesis that Levent Alpöge, an Anthropic mathematician, released on Friday, and the time it took the Lean and Mathlib communities to certify it: "both the statement and the proof of the Riemann Hypothesis are correct."
The certification isn't the sentence to read. The last line of that statement is. The community says it looks forward to "all the activity in the mathematical community that will help explain the new ideas in this proof." A problem open since 1859 is settled, and the explanation is listed under future work.
For most of mathematics' history, checking a proof and understanding it were the same act. A referee confirmed a proof by following it, and following it was how the ideas spread. Alpöge's proof pulls the two apart. Certainty arrived first, from software. Understanding is now a separate project, started from a theorem already known to be true.
He is candid about it. By his own timeline he found the proof "working with Claude" in about six weeks, and near the end, he writes, "Claude far outpaced me." The argument runs through automorphic forms on higher-rank groups, where he says he has "little mastery." Generating a full written proof was easy; his ran past 250,000 pages. He called posting it "obviously inappropriate" and released only the Lean.
This puts us in the rather extraordinary situation this age has made possible, in which I am presenting the world with a complete argument before having mastered a massive neighbourhood around its mathematics.Levent Alpöge
The verification is just as clear about what it covers. Andrew Sutherland's check log at MIT, run with Claude Code, records that every module compiles, that the five headline theorems rest only on Lean's three standard axioms, and that independently written kernels accept the whole export. The fastest independent check finishes in under four minutes on 64 threads. Human experts vetted the statement, a few lines of Mathlib. The proof was read by programs.
That is by design. The Xena Project blog set out its tests for trusting AI-generated Lean on November 12, after OpenAI's dump of 722 papers. Its closing warning is that a passing check proves nothing about the paper beside it: the write-up of OpenAI's Navier–Stokes result and its Lean proof turned out to diverge. Alpöge's answer was to leave out the paper. That removes the thing that could be wrong, and with it the thing that could be read.
Engineers have a word for code that passes every check and that nobody on the team can explain. It's legacy. Mathematics has just shipped a legacy theorem on release day.
The obvious objection is that this has happened before. Appel and Haken's 1976 proof of the four-colour theorem leaned on a computer to grind through cases by the thousand; mathematicians grumbled, then moved on. Perelman posted his Poincaré proof in 2002 and 2003, and three teams spent about three years confirming it. Understanding has always lagged. That's true, and it misses the order. Appel and Haken understood their strategy, and the machine only checked cases they had already reduced the problem to. Perelman's checkers could verify only by understanding, so those three years left behind expositions the rest of the field could learn from. Checking used to produce understanding as a byproduct. This time the byproduct is a SHA-256 hash.
Checking used to produce understanding as a byproduct. This time the byproduct is a SHA-256 hash.
So the explaining has to be done on purpose, by people who already know the answer, against a proof no one has to read in order to trust. Alpöge says he will start: "I intend to do my best to understand and impart understanding of this argument." He'll need company. For 167 years, the hard part of the Riemann Hypothesis was finding out whether it was true. Now that takes four minutes on 64 threads. The hard part is the reading.