The Prime Suspects

Levent Alpöge and Claude may have proved the Riemann hypothesis. Lean lets others check their work. Explaining it to the rest of us could take a little longer.

I usually write about artificial intelligence over at gpt.gekko.de. What brought the subject over here was a programming language. On November 20, 2026, the mathematician Levent Alpöge announced that he and Claude had, he believed, proved the Riemann hypothesis. He published a four-page note and a formal argument in a language called Lean. Within hours, Samuel Albanie had taken stock of the situation on X.

has ai progress stalled? i'm personally not aware of any major breakthroughs since riemann a few hours ago

Mathematics had waited 167 years. The timeline had given it until lunch.

A question of where to draw the line

Count the primes up to 100 and you get 25. Count them up to a million and you get 78,498. Now suppose I give you a number with fifty digits and ask roughly how many primes lie below it. You can give me an excellent estimate without trying every candidate. The prime number theorem tells us how the count grows. Riemann’s question concerns the error in that estimate. How far can the primes wander from the expected count?

In 1859, Bernhard Riemann approached this counting problem by investigating a function now named after him. For a real number ss greater than one, the zeta function starts innocently enough. Add up 1+1/2s+1/3s+1/4s1 + 1/2^s + 1/3^s + 1/4^s, and keep going. As Enrico Bombieri explains in his account of the problem, the same quantity can be written as a product with one factor for each prime. Every positive integer has a unique prime factorization, which is how the primes got into a sum that appeared to invite everybody.

Extend this function into the complex plane and ask where its value is zero. Some zeros sit at the negative even integers. The others are the interesting ones. The Riemann hypothesis says that every single one of those has real part exactly one-half. On a graph, that puts them all on one vertical line. Their positions enter a formula for the fluctuations in the prime count. A zero off the line would allow larger swings than Riemann predicted. Trillions of zeros have submitted to inspection. Unfortunately, every is an expensive word in mathematics.

Lean has one small objection

Lean is an open-source programming language and a proof assistant. You can write an ordinary program in it. You can also give it a mathematical statement and construct an object whose type is that statement. If the object checks out, it constitutes a proof. This sounds like an eccentric use of a type system until you catch yourself dividing by zero.

import Mathlib

example (x : ℝ) (hx : x ≠ 0) : x / x = 1 :=
  div_self hx

That says that a real number divided by itself equals one, provided it is nonzero. The interesting character is hx. It gives a name to the evidence that xx is nonzero, and the last line hands that evidence to a theorem in Mathlib, Lean’s community mathematics library. Delete the condition and try to finish the argument. You now owe Lean a reason why an arbitrary real number cannot be zero. Take your time.

At research scale, authors use commands called tactics to build proofs from smaller steps. A relatively small part of Lean, its kernel, checks the resulting proof. This gives Claude something unusually valuable for a language model inclined to finish its sentences. It can suggest a step, see precisely what remains to be proved, and try again. A beautifully worded paragraph about why xx is probably nonzero leaves the obligation exactly where it was.

What happened on November 20?

Alpöge gives a rather good account of being one of Claude’s “meat proxies,” if we insist on calling a number theorist who had been preparing for this problem since his student days a proxy. In his account of the collaboration, he describes steering the work from a simpler problem toward a setting where two ideas might work together. He credits Claude with a startling suggestion about how many times to repeat a particular construction. He also admits that parts of the resulting argument took him outside his expertise.

An attempt at writing the whole thing out ran past 250,000 pages. Its first theorem observed that a real number unequal to one-half must be either greater or smaller. One begins to see where the pages went.

The claim extends to a family of related functions called Dirichlet L-functions. And the evidence goes considerably further than a green tick on somebody’s laptop. By 22:05 UTC on November 20, MIT mathematician Andrew Sutherland had published independent checks showing that all 175,187 modules compiled. Two independently written checkers had accepted the full exported proof. The five headline theorems depended only on Lean’s three standard axioms, and the Riemann theorem used Mathlib’s existing definition of the hypothesis. Several further checks were still running when Sutherland posted that update.

What do we get?

If the result stands, mathematicians can revisit papers containing the words assuming the Riemann hypothesis and discharge that assumption. The generalized result would also sharpen estimates for primes in particular arithmetic progressions. Take the last digit. Every prime larger than five ends in 1, 3, 7, or 9. Over the long run, each ending gets an equal share. At any particular cutoff, though, one ending can be ahead. The generalized hypothesis would give us much tighter control over the size of that lead.

The dividend I’m most curious about is still being earned. Somewhere in this enormous argument are choices another mathematician might use on a different problem. Which construction does the heavy lifting? Why repeat it that many times? As Terence Tao observed on the day of the announcement, mathematicians were already trying to extract the strategy. Lean can check every use of an idea without circling the moment a reader ought to put down the paper and go for a walk.

That is what brought me over from the AI blog. A mathematician and a machine may have settled a question from 1859. Thanks to Lean, a stranger could begin checking their work that very afternoon. I’m happy to give the meat proxies a little longer to explain the good bit.

Do yourself a favour and watch Petr Lebedev’s excellent video, Did AI Just Prove The Riemann Hypothesis?. He puts into words exactly how I feel about all this. I’d happily give him the last word, which, after a thousand words of my own, is remarkably generous of me.

No comments yet