Artificial Intelligence • • 5-8 minutes

Claude and the Riemann Hypothesis: The Proof That Compiles and No One Has Read

Diego Cortés
Diego Cortés
Full Stack Developer & SEO Specialist
Share:
Claude and the Riemann Hypothesis: The Proof That Compiles and No One Has Read
Image generated with AI

A mathematician at Anthropic says he and Claude proved the Riemann hypothesis. He shipped it as 17 million lines of machine-checkable code, with no paper. It compiles. Understanding it is another matter.

What Was Announced, and What Was Not

On Friday, November 20, Levent Alpöge — a mathematician at Anthropic, on leave from Harvard — posted a four-page note on his personal site. The line that travelled: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis". He adds that the method extends to the Generalized Riemann Hypothesis and that, near the finish line, "Claude far outpaced me".

Lean Instead of a Paper: 175,187 Files

What he published is Lean code, the language of a proof assistant that checks every step: about 17 million lines across 175,187 files. By his own note, the human-readable version would run past 250,000 pages, and he is not publishing it — that is the length of the text the model produced, not a proof sitting in a drawer. Alpöge admits he does not yet master all the mathematics involved.

What Anthropic Did Not Say

The company made no statement and Nature got no immediate reply. An employee's post on his own site is not a company announcement, and that distinction matters when the news arrives glued to the listing: the IPO priced at $83 on November 18. Alpöge wrote that there is "an obvious cynical view, to do with the timing" and that it is false.

What It Means for Lean to Say Yes

Lean is not an assistant with opinions. It is a language where every claim follows from definitions, earlier results and axioms, and a small program — the kernel — verifies the derivation. If the kernel accepts, there is no logical gap.

Kernel, Axioms and the Word "sorry"

Lean lets you write sorry where a proof is missing: it marks "this is not done" and compiles anyway. That is why a successful build is not a headline by itself. What was checked here: all 175,187 modules compile without errors, the five main theorems depend only on Lean's three standard axioms (propext, Classical.choice, Quot.sound) and there is no sorry anywhere. That is the difference between compiling and being proved.

The Lean FRO and Mathlib Statement, and What Was Still Running

On November 21, the Lean and Mathlib communities said they have "full confidence" that the proof is correct, after matching the statement against Mathlib's and compiling the proof. The public record is kept by Andrew Sutherland (MIT): independent checkers, nanoda among them, accept the exported proof, and an isolated recompilation produced an identical result. The next day, two long runs were still going — the lake comparator and the kernel replay. Sutherland: the chance that anything is fishy is "approximately zero", with the caveat that a pass depends on Mathlib's definitions meaning what they should.

Checked, Understood and Accepted Are Not the Same

This is the core of the story, and it is a familiar engineering problem: the machine saying yes does not mean anyone knows why.

Understood: "We Don't Even Know Why It Is True"

Kevin Buzzard of Imperial College put it better than any headline: "At the minute, we don't even know why the theorem is true". People are translating chunks of the code into English to reconstruct the argument. Terence Tao calls it "very initial analysis"; Andrew Granville described the four-page note as "garbled". James Maynard, by contrast, believes that if it fully checks out it is a milestone.

Accepted: the Clay Prize and Its Two Years

The million dollars stays out of reach, and not out of scepticism: the rules require publication in a refereed journal of worldwide reputation, at least two years later, and general acceptance. Clay does not take direct submissions. The problem is still listed as open, 167 years after Riemann posed it, and a passing build triggers none of that paperwork.

The Real Scope of What Was Verified

The verified generalisation covers Dirichlet L-functions, not every generalisation mathematicians use. And there is an old lesson attached: roughly 12.36 trillion zeros have been checked by computer on the critical line, and that says nothing about infinity. Finite does not imply infinite, which is exactly why formal verification carries weight here.

The Other Proof From the Same Day: 940 Pages, No AI

Alessandro Maria Masullo, who did his doctorate at Stanford under Brian Conrad, released a 940-page preprint, "The Music of the Primes", after about twelve years of work. Ravi Vakil — a Stanford mathematician described in coverage as president of the AMS — presented it as a "hurried announcement": the plan was another year of polishing before going public, but releasing the current draft seemed better. Hours later, several mathematicians reported problems, found with help from different AI models. The text is in major revision and Vakil makes no claim about whether the strategy is correct. This is not a refutation: it is a portrait of a process where reviewing 940 pages by hand is simply slow.

Why It Matters Even If the Proof Is Right

Because the bottleneck moved. Generating 17 million lines is cheap for a machine; reviewing them is brutally expensive for a human. Lean works as continuous integration for mathematics: if the kernel passes, the result is as reliable as its axioms and definitions. It is the same pattern as agents that hand you a finished patch to review, and it explains why a faster model with more context such as Claude Fable 5.2 barely moves the needle: the limit is not writing, it is checking.

Conclusion: Compiling Is Not Understanding

There is a proof of the Riemann hypothesis that a machine accepts and nobody has read. That does not make it false, and it does not make it useless: it makes it the first large-scale case where mathematical knowledge moves faster than its explanation. That is the question we keep chasing on this blog — the same one as always in new clothes: do not trust the result, verify the process.

Categories