First published on tt-tech.io, with live animations.
On Friday 20 November, Levent Alpöge, a mathematician at Anthropic, put a four-page note on his website. One sentence in it: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." [1] There is no paper. There are fourteen archive files of Lean code, a language in which a computer checks every step of a proof.
One day later the Lean and Mathlib communities answered with a joint statement: "We have full confidence that Alpöge's proof of the Riemann Hypothesis is correct." [2] The problem was open since 1859. [3]
My team works with Claude in codebases every day, so a result made with the same tool was worth a close look. This is a long read for this site. It covers what the problem asks, why the answer says something about prime numbers, what was done in six weeks, and why a proof that nobody has read yet can still be trusted. The original article has live animations. Here you get screenshots of them, each with a link to the animated version.
What the problem asks
A prime number has no divisors except 1 and itself: 2, 3, 5, 7, 11, 13. They never stop, and they follow no simple rule. The gaps between them look random. The Clay Mathematics Institute says it in one line: their distribution "does not follow any regular pattern". [3]
Riemann's idea was to study the primes through one function, the zeta function. It starts as a plain sum over all whole numbers:
ζ(s) = Σn = 1∞ 1/ns = 1 + 1/2s + 1/3s + 1/4s + …
The big Σ (sigma) means "add up": one term for every whole number n, starting at 1 and never stopping.
Where are the primes in that? Euler had already found that the same sum is a product with one factor per prime, and nothing else:
ζ(s) = Πp prime 1/(1 − p−s) = 1/(1 − 2−s) × 1/(1 − 3−s) × 1/(1 − 5−s) × …
The big Π (pi) means "multiply": one factor for every prime p.
So everything the function does, it does because of the primes. Riemann then let the input s be a complex number, a point in a plane with a real part and a height t, and asked one question: for which inputs is the output exactly 0?
He also found an equation that ties the value at s to the value at 1 − s. [4] One way to write it is as a product of a few simple parts and one hard part:
ζ(s) = 2s πs−1 × sin(πs/2) × Γ(1 − s) × ζ(1 − s)
The sine is the easy part. A sine is 0 at regular steps, and here that makes the output 0 at the negative even numbers: −2, −4, −6 and so on. These are called the trivial zeros. [4] Every other zero has to come from the last part, zeta itself, and the equation does not say where those are. That part is the problem.
The other zeros all sit in a narrow vertical band, the critical strip, where the real part of s is between 0 and 1. [4] The Riemann hypothesis says they are not spread over that band. They are all on its centre line, real part ½. [3]
Screenshot. See it animated on tt-tech.io, where you can play it and change the real part.
On the board, the first zero comes at height 14.13, the next at 21.02, and there are ten below 50. On the live board, change the real part to 0.6 or 0.8. The curve keeps turning around the centre and never lands on it. Computers had checked the first ten trillion zeros this way, and all of them were on the line. [3] Ten trillion is still not all of them. That is the difference between evidence and a proof.
Why the zeros say something about primes
So what does a zero of a function have to do with counting primes? The zeros are the frequencies the primes are made of.
Take a staircase that climbs at every prime, with a step a little higher for each bigger prime. If you walk far enough, the staircase follows a straight line. That is the average. Riemann's formula gives the rest: subtract one wave for every zero of zeta, and the straight line bends into the staircase, step by step. The height of a zero sets how fast its wave swings. Its real part sets how loud the wave gets as the numbers grow.
Screenshot with all 100 waves added. See it animated on tt-tech.io, where the waves are added one by one.
Here is the point of the whole hypothesis. If every zero has real part ½, every wave grows at the same slow rate, the square root of x, and no wave drowns out the others. One zero off the line would add a louder wave, and the primes would drift further from their average than expected. Alpöge puts it in one sentence: the hypothesis "asks for optimal control on the count of prime numbers of a given size". [1]
The primes look random. The hypothesis says they are as regular as they can be.
A century of partial answers
Nobody could prove that all the zeros are on the line, so mathematicians proved it for a part of them. In 1914 Hardy showed that infinitely many are on the line. In 1942 Selberg showed it is a positive share of all zeros. Levinson reached one third in 1974, and Conrey two fifths in 1989. [5] The record of 2020 was slightly over five twelfths. [6]
Screenshot. See it animated on tt-tech.io.
Then, in August, Anthropic reported that an unreleased research version of Claude had moved that bound from under 42% to 67.2%. The prompt was to "take a real stab at the Riemann hypothesis". A first run tried 650 ideas and none worked. A second run used about 60 subagents for a day and a half. The two runs together used about 31 million output tokens. Anthropic was clear about the limit: it did not expect these techniques to lead to a proof. [7]
People picked it up fast. Three weeks later Youness Lamzouri, professor at the University of Lorraine, published a simpler proof of the same bound, [8] written without AI. [9] He is Moroccan 🇲🇦 and Canadian, and he won the gold medal of the Moroccan Mathematical Olympiad in 1998. [10] As a Moroccan I read that line of his CV twice, with some pride. The model found the result, and one of ours found the clean way to say it.
Six weeks, one mathematician, one model
The November proof does not come from the August method. Alpöge's note has a timeline, and it runs from 6 October to 16 November. [1]
The starting point is an old success. There is a simpler cousin of the Riemann hypothesis, over what mathematicians call function fields, and that one has been proven for a long time. One of its proofs, by Stepanov, uses a counting trick called multiplicity. The trick does not carry over to ordinary whole numbers: Alpöge compares it to "differentiating whole numbers". The new proof gets the same effect another way. It chains a construction called a theta lift, many times in a row, and each link in the chain plays the role of one unit of multiplicity. [1]
That is as far as I can take you, and I am in good company. Alpöge writes that he chose the direction and brought two ideas together, and that then "Claude far outpaced me near the finish line". He also writes that he is "presenting the world with a complete argument before having mastered a massive neighbourhood around its mathematics". [1]
Is that a weakness of the result? No. It is the honest state of it. Written out for humans, the argument passes 250,000 pages, so he published only the Lean code and promised a proper write-up. He expects the proof to be "massively simplified, likely immediately". [1]
Why trust a proof nobody has read
Seventeen million lines of code cannot be reviewed by people. [2] They do not have to be. This is the full statement that was proven, as it stands in Mathlib, the shared Lean library of mathematics: [11]
def RiemannHypothesis : Prop := ∀ (s : ℂ) (_ : riemannZeta s = 0) (_ : ¬∃ n : ℕ, s = -2 * (n + 1)) (_ : s ≠ 1), s.re = 1 / 2 theorem riemannHypothesis : RiemannHypothesis Read it slowly and it is the sentence from the first section: every zero of zeta that is not one of the easy ones (−2, −4, −6, …) has real part ½. The definition is part of Mathlib and was vetted by experts in Lean and in number theory. [2]
Nobody reviews 17 million lines. You review the statement, and a small program checks the rest.
How a proof nobody has read gets checked
- Lean sources: 175,187 files 17 million lines
- Compile: zero errors depends only on Mathlib
- Raw proof export: one inert file no code can run in it
- Kernels: Lean's own, plus the independent nanoda and con-leche
- Statement match: proven theorem = Mathlib's definition
- Accepted: three standard axioms only
Andrew Sutherland of MIT published the checks one by one. He ran them with Claude Code, and the programs that judge the proof are independent of it. All 175,187 files compile with zero errors. The proof depends only on the three standard axioms of Lean. The whole proof was exported to one raw file, and independent checkers replayed it, among them nanoda, a kernel written in Rust by other people. [11] Two tools confirmed that the theorem proven is exactly Mathlib's statement and not a look-alike, and dozens of people reproduced the same export on their own machines. [2]
There is one more check that I find the nicest. The Riemann hypothesis is equivalent to a plain statement about divisors, with no zeta function in it: for every n above 5040, the sum of the divisors of n stays under a known bound. Sutherland combined the new proof with an independent Lean proof of that equivalence, and the plain statement came out proven too. [11] A mistake hidden in the definition of zeta would have to survive that as well.
The same day, a second claim, without AI
There is a second claim, and it has to be told with the first. On the same Friday, Ravi Vakil posted a draft by Alessandro Masullo on his mathematics blog: a proposed proof of the generalized hypothesis, the same statement for a wider family of functions that look like zeta. His words: "It is all human, with no AI." The mathematician Brian Conrad had been reading it closely for months, and it was "so far looking good to him". The plan was a year of polishing before any announcement. On that Friday Vakil wrote that "now it seems best to put out the current draft". [12]
About eight hours later he added an update. Several people had reported problems in the proposed proof, found with the help of AI models. He writes that all issues must be resolved "to the satisfaction of the community" before the proof would be accepted as true. [12]
So this one is open. It may be repaired, it may not, and I will not guess. What I see is the difference in how the two are checked. One came with a file that a small program can replay, and it was checked in a day. The other is a paper, and a paper needs months of careful readers. Vakil himself asks people to check the draft and to formalize it. [12]
What is still open
A checked proof is not the end of the work. Four things are open today.
- Understanding. James Maynard, Fields medallist at Oxford, suspects the proof will turn out to be correct and says understanding it "will require quite a lot of work". Lamzouri says it could take months or even years. [9]
- Scope. Alpöge writes that the method also settles the Generalized Riemann Hypothesis. [1] What was checked is that statement for Dirichlet L-functions. Wider families of functions are not among the checked statements. [11] That family matters outside maths too. The Miller-Rabin test, which checks that a huge number is prime, gives strong evidence and no proof. If the generalized hypothesis holds for Dirichlet characters, the same test becomes an exact one that still runs fast. [13]
- The last runs. When Sutherland last updated his page (21 November, 12:55 UTC), the slowest replays, by Lean's own kernel on the full export, were still running. The independent kernels had already accepted it. [11]
- The prize. The Clay Institute's million dollars needs three things: publication in a qualifying outlet, two years after that, and general acceptance in the mathematics community. [14] [15]
And the timing? The note came out the day after Anthropic's stock market debut. [14] Alpöge calls the timing "horrible" and the cynical reading false. [1] I do not have to take his word for it, and neither do you. The export file and the checkers are public, which is the whole point of doing it in Lean.
What I take from it for our own work
I am a frontend architect and a science and AI enthusiast, not a number theorist. Still, three things in this story are the same as in a codebase with agents in it.
The person chose the direction. Alpöge prepared for this problem since his student days, and he steered the work to the right setting. The model did the long technical run. That is the split that works for my team too: the person who knows the context decides where to go.
The check was written before the answer. The short statement was in Mathlib first, and the checker is small and trusted. Nobody had to believe the model, or the mathematician, or the company. In our work that is the spec that pins the behaviour before the agent touches anything.
A green check is not understanding. The proof passes, and the first thing its author plans is to understand it and make it shorter. We owe the same to a large diff from an agent that passes all tests: it is checked, and it is not yet ours.
On your next task for an agent, write the short statement of "done" first, in a form a machine can check. It worked for a problem that waited 167 years.
Sources
- Levent Alpöge, Remark (the four-page announcement), 20 November 2026
- Lean and Mathlib communities, Lean verification of the Riemann Hypothesis, 21 November 2026
- Clay Mathematics Institute, Riemann Hypothesis
- Enrico Bombieri, Clay Mathematics Institute, Problems of the Millennium: the Riemann Hypothesis
- From First Principles, A century of progress toward the critical line (episode 53), August 2026
- Pratt, Robles, Zaharescu, Zeindler, More than five-twelfths of the zeros of ζ are on the critical line, 2020
- Anthropic, Claude has improved on a longstanding lower bound for the fraction of zeros of the Riemann zeta function that satisfy the Riemann hypothesis, August 2026
- Youness Lamzouri, A new proof that more than 2/3 of the zeros of the Riemann zeta function are simple and on the critical line, September 2026
- IBM Think, Anthropic mathematician claims AI helped solve a 167-year-old math mystery, November 2026
- Institut Élie Cartan de Lorraine, Youness Lamzouri, member page
- Andrew Sutherland, MIT, Verification files for the Lean proof of the Riemann Hypothesis, 21 November 2026
- Ravi Vakil, Hurried Announcement: Alessandro Masullo GRH, 20 November 2026
- René Schoof, Four primality testing algorithms, 2008
- AFP via Phys.org, Anthropic researcher says he and Claude proved the Riemann hypothesis, first posed in 1859, November 2026
- Clay Mathematics Institute, Rules for the Millennium Prize Problems
First published on tt-tech.io, also in French and Arabic.


