One Line, and Everything Hanging From It

Write out the whole numbers and start crossing out. Two survives; strike every even number after it. Three survives; strike every third. Five, seven, eleven, thirteen. What is left is the primes, and the first thing anyone notices is that they arrive without any pattern you can name. Between 1 and 100 there are twenty-five. Between 1,000,000 and 1,000,100 there are six. They thin out, but not steadily; they clump, then leave long gaps; and no formula anyone has ever found will tell you where the next one falls. Euclid proved there are infinitely many of them around 300 BC, and for the next two thousand years that was very nearly the whole of what could be said.

Now do something slightly different. Do not ask where the primes are. Ask only how many there are so far—count them as you walk up the number line and plot the running total. What you get is a staircase: flat for a while, then a step up, flat, flat, step. Locally it is as ragged as the primes themselves. But stand back far enough and the raggedness disappears into a smooth, gently bending curve, and the whole of modern number theory lives in the gap between the staircase and the curve. The Riemann hypothesis is a precise statement about how small that gap is allowed to get, and virtually everything we would like to know about the primes turns out to hang from it.

⤢ enlarge

The Curve the Staircase Follows

Gauss noticed the curve as a boy. Given a table of primes and too much time, he observed that the density of primes near a large number x falls off roughly like one in every log x numbers—so the count up to x should be about x divided by log x. He then improved it: instead of a single average density, add up the local density all the way along, giving the logarithmic integral, written Li(x). That refinement matters. Up to a trillion, x/log x undercounts the primes by about four percent; Li(x) is right to within a few parts in a million.

That the two really do converge—that the ratio of the true count to the estimate tends to 1—is the prime number theorem, conjectured by Gauss and Legendre and finally proved in 1896, independently, by Hadamard and de la Vallee Poussin. But a theorem about a ratio tending to 1 is a weak thing. It permits the actual error to grow enormous, so long as it grows more slowly than the count itself. Knowing the primes follow the curve on average tells you almost nothing about whether they might, somewhere out past the numbers anyone has ever checked, wander badly off it. Everything sharp in this subject is a statement about the error term, and the error term is where Riemann comes in.

A Sum Over Integers That Is Secretly a Product Over Primes

In 1859, elected to the Berlin Academy and obliged to submit something, Bernhard Riemann wrote eight pages titled On the Number of Primes Less Than a Given Magnitude. It is the only paper he ever wrote on number theory, it is famously compressed, and it changed the subject permanently. His instrument was a function Euler had studied a century earlier: take a complex number s, and add up one over every whole number raised to the power s. That is the zeta function, and for s greater than 1 it converges to something finite.

Euler’s discovery about it is the hinge on which this entire essay turns, and it is worth seeing why it is true. That infinite sum over every whole number can be rewritten as an infinite product over only the primes—one factor for each prime, and nothing else. The reason is the fundamental theorem of arithmetic. Every whole number factors into primes in exactly one way, so when you multiply out the product, each whole number appears exactly once, from exactly one combination of prime factors. Unique factorisation, which sounds like a triviality about arithmetic, is precisely the statement that makes the two sides equal. The zeta function is the primes, wearing the costume of the integers—and that is why interrogating it tells you about them.

Riemann’s move was to extend zeta beyond the region where the sum converges, out to the whole complex plane, by a process called analytic continuation. This is less mystical than it sounds: a well-behaved complex function has essentially only one possible continuation, so extending it is not an act of invention but of discovery—there was only ever one way to do it. The continued function blows up at s = 1 and is perfectly finite everywhere else, and it satisfies a beautiful symmetry, a functional equation relating its value at s to its value at 1 - s. That symmetry pins a mirror down the plane, and the mirror sits at real part one-half.

The Strip, the Line, and the Hypothesis

Now ask where the continued function equals zero. Some zeros are dull: they sit at every negative even integer, -2, -4, -6, and onward, and they fall straight out of the functional equation. Everyone calls them trivial and ignores them. All the others—and there are infinitely many—are confined to a vertical band of the complex plane where the real part lies between 0 and 1. This is the critical strip, and those zeros are where the primes are hiding.

Riemann computed a few of them, found they sat exactly on the line down the middle of the strip, and wrote that it seemed very likely all of them did. He then added, with an offhandedness that has irritated mathematicians for a century and a half, that a rigorous proof would be desirable but that he had set the attempt aside, since he did not need it for his immediate purpose.

“A rigorous proof of this would certainly be desirable.”—Bernhard Riemann, 1859

⤢ enlarge

That is the Riemann hypothesis, in full: every non-trivial zero of the zeta function has real part exactly one-half. It has resisted since 1859. It is one of the seven Millennium Prize Problems, carrying a million dollars, though nobody working on it is working on it for that. Hilbert placed it in his famous 1900 list of problems and is supposed to have said that if he awoke after a thousand years, his first question would be whether it had been settled.

Why a Line Controls the Primes

Here is the part that is genuinely startling, and the reason this is not just a curiosity about a particular function. Riemann derived what is now called the explicit formula, and it does something that ought to be impossible: it writes the prime counting staircase exactly, with no error term at all, as the smooth curve minus a correction term for every single non-trivial zero.

Each zero contributes an oscillation—a wave along the number line, with its own frequency and its own amplitude. Add the first few and the smooth curve begins to develop bumps in roughly the right places. Add hundreds and the bumps sharpen. Add all of them, infinitely many, and the waves interfere in exactly the right way to reconstruct the staircase, every step landing precisely on a prime. The primes are not random and they are not patterned; they are a chord. The zeros are the frequencies it is built from, which is why people call them the music of the primes, and the metaphor is not decoration—it is what the formula literally says.

⤢ enlarge

Now the punchline. The amplitude of a zero’s wave is governed by its real part. A zero sitting at real part one-half produces a wave whose size grows like the square root of x. A zero further to the right—at 0.6, say, or 0.7—produces a wave that grows faster, and its oscillation would push the staircase further from the curve. So the Riemann hypothesis is not an abstruse statement about a function. Translated, it says: the primes are as evenly distributed as they could possibly be. Every deviation from the smooth curve is no larger than the square root of the count, which is the smallest error the structure permits. One zero off the line, anywhere, and the primes are lumpier than we think—not fatally, but measurably, and infinitely often.

The hypothesis is the claim that the primes are not hiding anything. Every deviation they make from the curve is the smallest one available to them.

What It Would Unlock, and What It Would Not

The practical answer to why anyone cares is unusual: a proof would not so much add a theorem as discharge a debt. Thousands of published results in number theory are stated conditionally—true if the Riemann hypothesis holds, or if its generalisation to the wider family of L-functions holds. An enormous amount of modern mathematics has been built on a floor nobody has yet shown is solid. Proving it would not reveal those results; it would ratify them, all at once, and every ‘assuming RH’ in the literature could quietly be deleted.

Concretely, it would give the sharpest possible error bound on prime counting; tighten what we can say about the gaps between consecutive primes; sharpen the effective bounds in Dirichlet’s theorem about primes in arithmetic progressions; and firm up estimates for class numbers and character sums that currently rest on the generalised hypothesis. There is also a substantial industry of results that run the other way, proving statements equivalent to RH—among the prettiest being a criterion of Lagarias phrased entirely in terms of the divisors of ordinary whole numbers, with no complex analysis in sight at all.

Now the correction, because this is the claim that gets mangled everywhere. A proof of the Riemann hypothesis would not break encryption. RSA rests on the difficulty of factoring large numbers, and RH says nothing whatever about how to factor; it constrains how primes are distributed, not how composites come apart. The related myth—that RH is what makes fast primality testing possible—was already obsolete in 2002, when Agrawal, Kayal and Saxena produced a deterministic polynomial-time primality test that assumes nothing. The honest statement is that RH is load-bearing for the theory and almost entirely irrelevant to the cryptography. It matters enormously; it does not matter in that way.

The Evidence, and Why Evidence Is Not Enough

The case for believing it is strong. More than ten trillion zeros have been computed and every one sits on the line. Hardy proved in 1914 that infinitely many zeros lie on it—which sounds decisive until you notice that infinitely many is entirely compatible with almost none, since there are infinitely many to go around. So the real work has been on proportions: what fraction of all the zeros can be proved to lie on the critical line? Selberg showed in the 1940s that it was a positive fraction. Levinson got it above a third in 1974. Conrey pushed it past two-fifths in 1989, and that number crept upward in small increments over the following decades, reaching roughly 41.6 percent—where it sat, essentially, for a generation.

And yet number theory has a specific, humbling history of numerical evidence being a liar. Li(x) exceeds the true prime count for every value anyone has ever computed, which for a long time looked like a theorem waiting to be proved. In 1914 Littlewood proved that the difference changes sign infinitely often—the count overtakes the curve and falls back, endlessly—and the first crossing is somewhere out past 10^316, a number with no physical meaning whatsoever. The Mertens conjecture had even better numerical support and was disproved in 1985. In this subject, ten trillion confirmations establish that you have not yet reached the interesting part of the number line.

August 2026: Two-Thirds

On the tenth of August this year, Anthropic published a result obtained by an unreleased research version of its Claude model: the proportion of zeta zeros provably on the critical line raised from about 41.6 percent to 67.2 percent. It is, by a wide margin, the largest single jump that bound has ever taken, and it moved a number that had inched forward by fractions of a percent since the 1980s.

The mathematics is not a new idea so much as a new connection. The result works by assembling a suitable space of functions carrying a quadratic form induced by Weil, and then examining the positive- and negative-definite subspaces that arise from zeros on the line and off it. What made it possible was joining three bodies of work that had not previously been put together: a 2019 result of Aryan, a recent series of papers by Baluyot, Goldston, Suriajaya and Turnage-Butterbaugh, and a paper of Bombieri from 2000. The pieces were all in the literature. Nobody had read them in the same week.

⤢ enlarge

The process is worth stating plainly, because it is unlike how mathematics has been done before. Two sessions of work, roughly a day and a half. Around 650 initial ideas generated and triaged, then some sixty parallel subagents coordinating, running about 2,400 shell commands and hundreds of Python scripts to test candidate approaches numerically before committing to them. Thirty-one million output tokens. The proof was then reviewed internally, checked by two of Anthropic’s own mathematicians, Levent Alpoge and Ralph Furman, examined by two outside experts in exactly this area, Brian Conrey and Dan Goldston—Conrey being the author of the record it broke—and formalised in Lean, where it passes the standard validation tool.

Anthropic’s own framing of the significance is notably restrained, and the restraint is the most credible thing in the announcement: the company states plainly that it does not expect the techniques Claude used to lead to a proof of the Riemann hypothesis. They are right, and the reason is structural rather than modest.

Every proof in this family works by constructing a mollifier—a cleverly chosen auxiliary function that damps the zeta function’s wild behaviour enough to count zeros on the line. Better mollifiers give better percentages. But the method has an internal ceiling: it counts zeros it can see on the line and can never account for the ones it cannot, so it cannot reach 100 percent by refinement. Sixty-seven percent is a very long way from all, and ‘all’ is the entire content of the hypothesis. A proof of RH will require an idea that does not yet exist, not a sharper version of an idea that does.

What Actually Changed

So what should one take from it? Not that the machines are about to finish mathematics. The honest lesson is narrower and, I think, more interesting: the binding constraint here was reading, not reasoning. The three papers that had to be combined were public. Any mathematician could have read them. But a working number theorist reads within a specialty, and these sat in adjacent specialties, and the field now produces far more than any person can follow. What the model supplied was not a flash of genius but an implausible breadth of attention—the ability to hold three separate literatures in view at once and notice that the quadratic form in one was the object the other needed.

That is a genuine and slightly uncomfortable diagnosis of where mathematics currently is. The specialisation that lets the subject go deep also guarantees that connectable results sit unconnected, sometimes for decades. There is no reason to think this was the only such pair. And it says something about verification too: a proof produced this way arrives faster than the ordinary machinery of peer review can absorb, which is precisely why the Lean formalisation matters more than the expert endorsements. A machine-checkable proof does not care who or what wrote it—and in a world where proofs can be generated faster than they can be read, that property stops being a curiosity and starts being the foundation.

The zeros were not found by thinking harder. They were found by reading more widely than a human career permits.

And the hypothesis itself stands exactly where Riemann left it. Somewhere out along that line, ten trillion zeros deep and infinitely further, every one so far has landed precisely at one-half—as if the primes were being careful, as if the whole ragged, unpredictable sequence were obeying a constraint so tight that no counting anyone can do has ever caught it slipping. Two-thirds of them are now proven to be there. The remaining third is not a gap in a percentage. It is the possibility that somewhere past every number we will ever write down, a single zero steps off the line, and the primes turn out to have been keeping a secret after all.