The first theorem in Levent Alpöge’s experimental write-up began with something a schoolchild could check. A real number different from one half must be either bigger or smaller than one half. By Alpöge’s count, the document ran to more than 250,000 pages. On Friday, he chose to publish four pages and a link to the code instead.

The number theorist, who works at Anthropic, believes he and Claude have proved the Riemann hypothesis. It is a 167-year-old claim about the way primes scatter along the number line. Alpöge says the AI supplied a mathematical idea he needed near the end. An independent effort has begun checking the files. A verdict on the full proof is still to come. Alpöge’s announcement · Independent verification log

How Far Off Can a Prime Count Be?

Count the primes below a million and you get exactly 78,498. A classic formula that smooths out their irregular spacing gives about 78,628. It misses by roughly 130. Now move the finish line to a billion, a trillion, or a number with a thousand digits. How far from the true count can that smooth estimate wander? Prime counts and estimates

In November 1859, Bernhard Riemann found a way into that question in six handwritten pages. He connected the count to a function called zeta. The inputs at which it equals zero give corrections to the smooth estimate. Those inputs can be drawn as points on a plane. Riemann suspected that every one of the zeros that matter here sits on the same vertical line, at a horizontal coordinate of exactly one half. Riemann’s original manuscript

That sliver of geometry has a consequence you can express in ordinary numbers. If all the zeros stay on the line, the error in the count can grow no faster than the square root of the counting limit multiplied by a logarithmic factor. An off-line zero could force much larger swings far out among the primes. The first ten trillion zeros have been checked. The word that has resisted proof is all. Clay’s description of the problem · The prime-counting bound

The Move Alpöge Credits to Claude

Alpöge started from a setting where mathematicians already know how to prove a cousin of Riemann’s conjecture. There, an object can vanish several times over at the same point. That repetition gives a proof extra leverage. He wanted to reproduce the effect in the arithmetic of ordinary whole numbers, where the familiar technique no longer works.

His account singles out a decision about a chain of mathematical operations called theta lifts. Each link was meant to supply a little more of that missing repetition. How long should the chain be? For a sum extending to a number X, he says Claude suggested the square root of X. It is a specific choice another mathematician can investigate. The four-page note sketches where it enters the argument. Alpöge’s account of the collaboration

What Twelve Lines of Code Can Tell Us

Open the main file and the destination is visible almost at once. On line four, Alpöge’s code points to the Riemann hypothesis as it was already defined in Lean’s mathematics library. The file is just twelve lines long. Its short declarations draw on proofs stored elsewhere. Lean, a computer proof assistant, lets other people check whether each conclusion actually follows from the steps beneath it. Read the twelve lines

At 16:19 UTC on November 20, one independent verification project reported that its copy matched the released source and that the five target statements could be read by Lean on their own. Its scan also found no markers for unfinished proofs. The project still listed the full build as underway. Reading the destination and checking every step that gets there are two different jobs. Dated verification report

What Would Change if It Holds?

Consider the last digit of a prime. Above five, it must be 1, 3, 7 or 9. In the long run, each ending gets a quarter of the primes. Stop at a particular number, though, and one ending can lead another. Alpöge also claims a proof for a whole family of related functions, the Dirichlet L-functions. That result would give sharper guarantees on how far counts in these groups can stray from their expected shares, even when mathematicians sort primes into far more than four groups. A theorem whose only unproved assumption is this conjecture would then hold unconditionally. The generalized claim · What it implies for counting primes

The famous million-dollar prize runs on a slower clock. Before the Clay Mathematics Institute will consider a solution, it requires publication in a qualifying outlet, at least two years after that publication, and broad acceptance among mathematicians. Clay’s prize rules

There is already a glimpse of what might happen in the meantime. In August, Anthropic reported that Claude had raised the guaranteed share of zeros on the critical line from 41.6% to 67.2%. In September, mathematician Youness Lamzouri published a simpler argument for the two-thirds result. A second researcher had found a clearer route to a destination the model helped identify. The August result · Lamzouri’s proof

If Alpöge’s new argument survives, someone may eventually be able to open those six pages Riemann wrote in 1859 and explain, in a few pages more, why the zeros have to line up. That explanation is still worth looking for, even in a world where a computer might give the right answer first.