I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.
That’s from the announcement Levent Alpöge posted yesterday – just days after I published my article on AI doing mathematics.
Claimed proofs are posted all the time.1 Even I have received a few messages over the years from people who thought they had made some great mathematical discovery, and who somehow thought that I would understand it better than the “academic establishment” – perhaps because I write a blog instead of publishing in journals.
Why is this one different?
For one thing, Alpöge has quite the track record by now. On the very same day, he and Tristan Buckmaster also announced a proof of finite-time blow-up for the unforced Navier–Stokes equations – another Millennium Prize-sized result.
For another: Lean! Several outside sources have already compiled and verified the formal proof, using different implementations, which greatly reduces the likelihood of a bug.
The fact that we can be confident this is indeed a correct proof after a couple of hours is itself a stark contrast to, say, Andrew Wiles’ proof of Fermat’s Last Theorem. There, we had to wait a couple of years until the proof was verified, including the dramatic turn of an actual gap in his first approach. Wiles announced his proof in Cambridge in June 1993. The gap turned up during refereeing, and it took him about a year, and the help of his former student Richard Taylor, to fix it.
So, has it fallen? The general rule of thumb is that the answer to any headline ending in a question mark is no.2 But in this case, it looks like we really have a yes on our hands.
A whole bag of mixed emotions
I’m feeling a whole bag of mixed emotions right now: euphoria, confusion, loss, curiosity, excitement about the future. (Euphoria? This blog had been silent for nine years, and this is my second post in a week.)
I didn’t expect this announcement in my lifetime. And now AI has got there within a couple of years. Posing the challenge of proving RH was almost like asking someone in 1950 to run a mile in under four minutes: it was assumed to be so far out of reach as to be de facto impossible.
Just a month ago, in a Numberphile interview called No More Heroes, Terence Tao was asked whether the Riemann Hypothesis was going to fall soon. It “hasn’t fallen, but it’s been dented quite a bit”, he said (more on that dent below). He was sure the AI companies had secretly devoted much more time to it than they’d disclosed, “and they couldn’t touch it. I mean, something beyond current AI is needed.” Maybe “some combination of humans and AIs in 2050 or something” would make further progress.
When I was working on my previous article, I also discussed my undergrad thesis on prime-generating sequences with Claude. It pointed out that Cloitre’s conjecture, one of the open questions in the thesis, implies that for every prime \(p\) there’s a prime \(q\equiv-1\pmod p\) below \(p^2\) – and even GRH only gets you to about \(p^2\log^2p\). Given the resources OpenAI threw at Navier–Stokes, it put the odds of a full proof in the low single digits. Even AI wasn’t AI-pilled enough to believe any of this was ripe! And yesterday’s proof, which covers GRH too, doesn’t settle it either: Cloitre’s conjecture needs more than that.
I am excited about what this means for maths. But I do feel like we were deprived of that Wiles moment, when he could announce his proof from a blackboard to the mathematical community – and finish with “I think I’ll stop here.”
Instead, we now have to trust gigabytes of code. (Alpöge also produced 250,000 pages of AI slop, but mercifully didn’t post them. Theorem 1, he says, began by “noting that because the reals are totally ordered if \(\sigma\neq1/2\) then \(\sigma>1/2\) or \(\sigma<1/2\)”.)
So the loss is palpable. The loss of the guiding light3 that was always considered the holy grail of mathematics, maybe of all human knowledge. And the loss of “the old ways”.
A confession: whilst I was happy to heavily rely on AI to help me write up the details of my previous AI experiment, I really had to do this one by hand. Maybe it’s the last bastion of human understanding. Or rather, the last desperate clinging to it.
Human understanding attached
The announcement just dropped, it’s massive and there isn’t even an attempt at a human-digestible write-up. So we couldn’t possibly hope to go through any of this material on this very amateurish blog. Quite the admission for a blog called Understanding the Riemann Hypothesis.
But we can take a look at the previous milestones from this autumn – the \(7/8\), the \(11/12\) and the no-Siegel-zeros proofs – which do have human understanding attached.
A quick recap first. The Riemann Hypothesis says that all non-trivial zeros of the zeta function have real part \(1/2\). Since 1896, we’ve known that they all lie strictly inside the critical strip \(0<\Re s<1\); that’s what it took to prove the prime number theorem. Beyond that, writing \(s=\sigma+it\), the zero-free regions we knew looked like de la Vallée Poussin’s classical
\[ \sigma>1-\frac{c}{\log t}, \]
hugging the line \(\Re s=1\) ever more tightly the further up you go. Over the decades, people improved the logarithm a bit, but nobody could rule out zeros creeping arbitrarily close to \(1\). A fixed zero-free half-plane \(\Re s>\theta\) for any \(\theta<1\), a so-called quasi-Riemann hypothesis, seemed hopeless.
Then, in early October, OpenAI announced exactly that: no zeros with real part greater than \(7/8\), formally verified in Lean. By the symmetry of the functional equation, all non-trivial zeros therefore lie in the strip \(1/8\le\Re s\le7/8\). Alongside it came a second, shorter argument for the weaker bound \(11/12\). That’s the one human experts actually sat down and read.4
Terence Tao again. The poor fellow has become the public face of understanding what all these AI-for-maths advances mean, so he’s heavily over-cited. But in this case it’s for the very good reason that he undoubtedly is one of the smartest people on earth who can assess the impact of all those announcements. He’s neither AI-pilled nor AI-doomed, but pragmatic and sharp-eyed. And above all, he’s thoughtful enough beyond mathematics to philosophise about the societal impact, too.
In the same interview, he called OpenAI’s result “the biggest advance on the subject in 50 years”. We’re “one quarter of the way to the Riemann Hypothesis”, although their method “cannot get beyond three quarters without another extremely clever idea”. And the technique itself:
It’s sort of a combination of […] three or four existing ideas arranged in a very unusual way. […] It’s not something any single mathematician alive, I think, would have come up with. I think if you got the right people to work together, lock them in a room, they might have come up with this after lots of trial and error.
I was curious what that means, so I asked Claude and ChatGPT to take me through it. Roughly, the shorter \(11/12\) proof, which shares the key ideas, decomposes into four steps:
The classical skeleton. Remember tossing the prime coin? The Möbius function \(\mu(n)\) looks like a sequence of random coin tosses, and \(\sum\mu(n)n^{-s}=1/\zeta(s)\). A zero of \(\zeta\) is a pole of \(1/\zeta\). So if you can show enough cancellation in the Möbius sums, say \(M(x)=O(x^{\theta+\varepsilon})\) for the Mertens function, there can’t be any zeros with real part greater than \(\theta\). For \(\theta=1/2\), that’s the Riemann Hypothesis itself. All of this is textbook analytic number theory. The hard part, of course, is proving the cancellation.
Hide the sum in a bigger family, many times over. Instead of the ordinary integers, the proof works with the Eisenstein integers, the integers of the number field \(\mathbb{Q}(\sqrt{-3})\). That costs nothing: the zeta function of that field is \(\zeta(s)\) times another \(L\)-function, so every zero of \(\zeta\) is a zero there, too. What you gain are the sixth roots of unity, and with them sixth-power residue symbols, the machinery behind cubic and sextic reciprocity. Twisting the Möbius sum by those symbols gives a whole family of sums \(A_u\). Now the trick: if \(u=p^6\) is the sixth power of a prime, the twist is trivial (except where \(p\) divides \(n\)), so \(A_{p^6}\) is essentially our original sum again. It hides in the family once for every prime small enough for its sixth power to fit in.
Here’s the numerology, as far as I understand it. Say the family has about \(D\) members, each a sum of about \(D\) terms. If they behave like random walks on average, each one has size about \(\sqrt{D}\), and their squares add up to about \(D^2\). Our original sum appears about \(D^{1/6}\) times in that total, once for every prime up to \(D^{1/6}\), so it can be at most about
\[ \sqrt{D^2/D^{1/6}}=D^{11/12}. \]
That’s where the \(11/12\) comes from: it’s the sixth power. The catch is that proving the “random walks on average” part needs a strong large sieve inequality for sixth-power or cubic characters, and the known ones are too weak. For cubic characters, Alexander Dunn and Maksym Radziwiłł had even shown (assuming GRH) that the best known bound can’t be improved. Dead end.
Turn the Möbius function into a modular form. Both AIs singled this out as the real spark. Expanding the squares and applying Poisson summation replaces the sixth-power characters with Gauss sums. Then a classical identity for Gauss sums absorbs the Möbius sign, and what’s left are cubic Gauss sums. These are something very special: the Fourier coefficients of Kubota’s cubic theta function, a so-called metaplectic form. That’s a cousin of modular forms, with an enormous amount of symmetry. So the Möbius function, our prime coin, the very thing that resists any structure, gets traded for the coefficients of a modular form! The analytic heavy lifting is then done by a summation formula from that same 2021 paper by Dunn and Radziwiłł, which was really about Patterson’s conjecture on the bias of cubic Gauss sums. These are obscure objects.
From sextic to quadratic. The theta function’s symmetry, its reflection formula, turns the sixth-power twist into the cube of a sixth-power character. And a sixth root of unity cubed is just \(\pm1\): the character has become quadratic. For quadratic characters, Roger Heath-Brown’s large sieve is essentially optimal, which is exactly what was missing in step 2. A recursion of Poisson steps that’s guaranteed to stop after a fixed number of rounds closes the argument. Note the irony: the proof deliberately starts with a more complicated family, sixth-power twists, just to get access to a simpler estimate after the transformation.
(Mind you, this is my reading of the AIs’ reading of expositions of an AI proof. Take the details with a pinch of salt.)
Once you know the trick, Radziwiłł reckons he can sketch the core argument on half a page. Half a page! Compare that to what we’re looking at now.
So why had nobody found it? That’s exactly what Tao considered so unusual: there are experts in all those techniques, but no working mathematician would have all of them in their tool belt. The metaplectic people don’t work on zero-free regions, and the zero-free-region people don’t use metaplectic theta functions. On top of that, the specialists believed their tools were too weak for the job. Radziwiłł himself said as much to Quanta. So nobody had any reason to look for this particular combination.
No more universal geniuses
There’s been long-standing moaning about how fractured mathematics has become. There are no more universal geniuses, just specialised experts with a deep knowledge of one field, who find it hard to communicate their ideas to colleagues, even when their areas of expertise are adjacent.
And some of the greatest breakthroughs came precisely from building bridges between those fields.
Wiles’ proof of Fermat’s Last Theorem was celebrated in public because it solved a centuries-old question. Within mathematics, it was celebrated much more for how he solved it: by proving the Taniyama–Shimura–Weil conjecture,5 now the modularity theorem, and thus building a bridge between the previously separate worlds of modular forms and elliptic curves.
Riemann’s genius itself was translating the discrete world of prime numbers into the seemingly far away world of complex analysis, via the zeta function, which encodes all the primes in Euler’s product. That allowed us to tackle problems in number theory with an entirely different set of tools and methods.
(Did I mention that I wrote my master’s thesis in Göttingen, where Riemann was a professor when he wrote his 1859 paper on the primes?)
This is where AI truly shines. It’s not just an expert in one field. It has read all the literature, and it’s an expert in every field. In a single chat about my undergrad thesis, Claude pulled in Linnik’s theorem, GRH, the parity problem from sieve theory, Tao’s work on Collatz and finite automata. That’s what enabled OpenAI’s model to come up with the unusual combination behind quasi-RH, which humans were unlikely to find.
Again, we don’t know yet what kind of clever ideas and unusual combinations there are in the RH proof. Alpöge’s announcement only offers glimpses: the argument descends from Stepanov’s elementary proof of RH over function fields, and finding it took “this rather harebrained tornado through the whole of number theory”. The overwhelming consensus amongst human experts was that some entirely new techniques, maybe even new fields of mathematics, would have to be developed before RH would fall. Maybe humans did in fact come up with all the required building blocks already. As Tao put it: “The RH proof uses a century of past literature on related topics.” Even the smartest humans just weren’t able to keep all of them in their heads, so we were waiting for computers big enough to do that.
Yes, the proof as (not quite) presented is and may forever be incomprehensible to humans. But there is immense value in knowing that there is a way to solve this problem. People who might have had the skills, but were deterred by the fear that their efforts could go to waste, can now confidently study Claude’s paths, make them clearer, find shortcuts and maybe entirely new mathematical worlds.
Utopia or extinction?
There’s an under-discussed angle to this whole development: whilst we see mathematical progress like we’ve never seen in all of human history, it comes at enormous costs. OpenAI, Anthropic and certainly other undisclosed organisations have thrown countless millions at those problems (at least they might recoup token haha millions from the Millennium Prizes). There was the joke that physicists are so expensive, whilst mathematicians only needed paper and a bin. Only philosophers were cheaper: they even saved the bin. But now maths seems to be the playground of the ultra-wealthy. That would be the saddest development…
I am even more torn than before. Do all those breakthroughs of the past years mean we’re entering some kind of utopia where knowledge and progress are abundant? Or is this humanity driving towards its own extinction?
One thing’s for sure: the world of mathematics, and maybe the world as a whole, won’t ever be quite the same after yesterday’s announcement…
Case in point: on the same day, a human mathematician’s claimed proof of the Generalised Riemann Hypothesis also got some attention, a self-confessed “hurried announcement” with some holes in the write-up. Several people have since reported problems in it – found by AI, of course. Even if those could be repaired, IMHO the fact that an AI proof dropped first wouldn’t take the slightest bit away from such an achievement. Its author would still be the first human to understand why RH is true. ↩︎
Known as Betteridge’s law of headlines. ↩︎
Alpöge himself writes that “it is natural to feel that this is an end, in my case a loss of a principal guiding light”, though what he saw “so amazed” him that he believes “it is one dawn across science of something far greater”. ↩︎
Around the same time, OpenAI also claimed that there are no Landau–Siegel zeros: hypothetical real zeros of Dirichlet \(L\)-functions sitting extremely close to \(1\), whose mere possibility makes many results about primes in arithmetic progressions frustratingly ineffective. Their absence follows from the quasi-Riemann hypothesis, too, since it covers all Dirichlet \(L\)-functions, but the standalone paper uses a completely different argument, borrowed from transcendence theory. William Banks and Bryce Kerr have since written human expositions of it. ↩︎
Strictly speaking, Wiles (with Taylor’s help) proved the semistable case, which was enough for Fermat. The full conjecture followed in 2001, by Christophe Breuil, Brian Conrad, Fred Diamond and Richard Taylor. ↩︎