On Friday, November 20, a mathematician at Anthropic put a short PDF on his personal website. The file is called remark.pdf, which is a funny name for what sits behind it. Levent Alpöge wrote that he believed he had, working with Claude, found a proof of the Riemann hypothesis. That’s the problem mathematicians have been stuck on since 1859.

My first thought was that this was another rumour. Those show up every few months, usually from some retired engineer with a 12-page “proof” and a lot of confidence, and they’re dead in a week. This one didn’t die. By Saturday, Scientific American was reporting that independent mathematicians had compiled the proof on their own machines without problems, and Nature, the Washington Post and Al Jazeera all had full stories up.
So the short answer to “is it really about to solve it” is that it’s already been claimed, and the claim is holding up so far. The code compiles, the checks are mostly done, and the slow human part is starting now. What nobody can tell you yet is whether any human understands the proof.
What the problem is, roughly
Primes are numbers like 2, 3, 5 and 7 that divide only by one and themselves. Along the number line they look random. In 1859 Bernhard Riemann tied them to a function now called the zeta function, and every place where that function gives zero adds a little more detail to the pattern of the primes. The Riemann hypothesis says all the important zeros sit on one single straight line.
Computers have checked more than 10 trillion of them, and every one was on the line. That sounds like proof but it isn’t, because one zero off the line, way out somewhere, would change the whole picture of how primes are spread. I’m oversimplifying (the real thing lives in complex numbers, and I won’t pretend I can do that maths), but that’s the shape of it.
The problem itself is easy to state and it beat everybody for 167 years.
It matters outside pure maths as well. Abhishek Saha at Queen Mary University of London told Al Jazeera that hundreds of published results are proved only by assuming the hypothesis is true. Kevin Buzzard at Imperial College London gave a more practical example, which is that some fast algorithms are known to work only if the generalised version holds. Cryptographers have been assuming it for years because the evidence is so strong, and Buzzard said we can all sleep slightly safer now that it’s (probably) proven. The Clay Mathematics Institute listed it as one of seven Millennium Prize Problems back in 2000, with $1 million attached. That money hasn’t moved. Clay still shows the problem as unsolved, and as I understand their rules a solution has to be published in a journal and then sit for two years before anyone gets a cheque, so nobody is collecting this winter.
How we got here in one year
This didn’t come out of nowhere. Early this year OpenAI’s models proved the quasi-Riemann hypothesis, a weaker cousin that says no zeros hide in the strip between seven-eighths and one. I mixed up the quasi version and the real thing for most of an hour while reading about it, so I’ll keep them separate from here on. The weaker result did kill off the “Siegel zeros” problem that had annoyed number theorists for almost a century.

Then in August, Anthropic said an unreleased version of Claude had pushed up the share of zeros proven to sit on the line, from 41.6 percent to 67.2 percent. The way it started is almost funny. A staff member who isn’t a mathematician, Jarred Sumner, told Claude to take “a real stab” at the hypothesis. It tried 650 ideas, none of them worked, and then it drifted into a related problem and found the improvement there. Anthropic said itself that it didn’t expect those techniques to lead to a full proof. The number theorist James Maynard said that even being very optimistic, he saw no path from there to the actual hypothesis.
Anthropic had also been building a track record with Lean before this. Its research page mentions a computer-checked proof of Fermat’s Last Theorem that Claude wrote largely on its own over 11 days. And the August result was looked at by two experts in the area, Brian Conrey and Dan Goldston, who read the paper on short notice before it went out. That doesn’t prove the November claim is right. It does mean the process (formal code, outside readers, a public write-up) wasn’t invented for this one announcement.
Three months later Alpöge posted remark.pdf.
Another Millennium problem, Navier-Stokes, was also claimed with AI help this year. Al Jazeera says OpenAI announced in September that an unreleased model had solved it. I haven’t seen that one confirmed the way this one is being confirmed, so I’m treating it as a claim and leaving it there.
What the proof actually is
It isn’t a normal paper. Alpöge wrote a four-page explanation, but the proof itself is written in Lean, a programming language where the computer checks every logical step. If the code compiles, each step follows from the one before it, and nobody has to take the author’s word for anything. Buzzard estimates it’s around 17 million lines of Lean, which could be tens of thousands of pages if you wrote it out as normal maths. Science reported that it’s larger than some complete operating systems, and mathematicians guess the computing bill ran into millions of dollars. Alpöge was open about who did what. He said Claude moved far faster than him near the finish line.
Buzzard wrote that before you trust any Lean proof you have to ask two boring questions. Does the theorem written in the code really match the Riemann hypothesis, and does the code actually run for other people? Researchers including Andrew Sutherland at MIT have now compiled it separately with no trouble, and nobody found a hidden shortcut or a changed definition nobody mentioned. Sutherland put the chance of anything being wrong at “approximately zero”. One last check was still running in the report I read. Normal peer review hasn’t started, though, and a journal could take a long time.
And there’s the obvious conflict of interest. Anthropic makes Claude and Anthropic announced the result, so I wouldn’t take the announcement alone. That’s exactly why the independent compile matters more than any press release.
What mathematicians are saying
The mood isn’t only celebration. Saha said he’d expected the proof to need new theories built over decades, and then privately guessed it was still several years off. He told Al Jazeera the claimed proof “shatters my model of reality”. He described the usual approach as building roads to reach a deep mineral deposit, and this proof as a long tour through roads built for a hundred other deposits that somehow ends up at the right place. He also argued the credit should go to Alpöge and not to Claude, because others with the same AI access didn’t manage it, and he thinks Alpöge should get a Fields Medal.
Buzzard is more careful. Right now, he says, nobody knows why the theorem is true. It could be that humans were missing no ideas at all and nobody was smart enough to join up the ingredients. It could also hide a thousand pages of deep new theory. People around the world are translating the Lean code back into English, and he expects a much better picture in a week or two. He also pointed out that stronger versions, like the extended and grand Riemann hypotheses, still sit above this one. In his words we’re at the foot of the mountain and this is where the climb begins.
Then there’s the group that is bothered by something else. Maksym Radziwiłł at NYU told Science that the point of maths and science is understanding why things happen, and a pile of checked code doesn’t give you that on its own. Maynard said the result does very little for him right now, because he never doubted the hypothesis was true. I think that second reaction is underrated. If you already believed it, a yes from a machine tells you nothing new, and what you wanted was the reason.
Buzzard sees the grief and the bruised egos as real, but he doesn’t buy the idea that maths is finished. He compares AI to the telescope. It has no taste, he says, you tell it what to do and it does that, and mathematics is infinitely big so there’s plenty left. He also noted that right now only the tech companies with unreleased models are producing these results, and he expects that to change as better models become public.
Who gets the name on it
There’s a side argument going on that has nothing to do with whether the maths is right. In June, a group of well-known mathematicians signed a public declaration saying a proof should be tied to specific authors who take credit for it and take responsibility if it’s wrong. Their worry was that AI breaks that link. Tim Gowers took a looser view, as TechCrunch reported it. He said that if theorems stop carrying mathematicians’ names, that might be no worse than the fact that most stars aren’t named after astronomers.
The Washington Post described the mood among mathematicians as a mix of wonder and loss, and I think that’s accurate. This result makes the split harder to ignore. Alpöge signed his own name to the claim, and Saha, like I said, wants the credit to go to him and not to the model. But a person who signs a 17-million-line proof is signing something he can’t fully read yet. Who’s responsible if a hole turns up in line 9,000,000? I don’t have a clean answer.
There’s a money problem too. Mathematicians estimate this one cost millions of dollars in computing, and only a few companies have the unreleased models and the budget to run something like it. A graduate student in Pune or Lagos can’t try this next month. Buzzard thinks that gap will close as stronger models reach the public, and he may be right. I’d just add that “will close” is a prediction and not a fact, and the people making it have been wrong about timing all year.
The other proof, with 940 pages
The same day, Stanford’s Ravi Vakil, who is also president of the American Mathematical Society, put out a draft by Alessandro Masullo claiming a proof of the generalised Riemann hypothesis. Vakil called it all human, with no AI. Masullo says it took him 12 years. It runs to 940 pages, and Brian Conrad had been reading it for months.
Vakil released it early. The plan had been a year of editing and review at Stanford, and he said the Anthropic news made it better to just get the draft out. That has a cost. Several mathematicians have already raised problems with it, some of them found with help from AI models, and Vakil said those would need fixing before anyone could accept the proof.
It’s a strange situation. One proof is huge, machine-checked and not yet understood by any person. The other is human-written, and humans are still working out whether it’s right. The old worry that an AI will make up a step doesn’t apply when a made-up step simply won’t compile. What can go wrong is the statement, and Buzzard and others went after exactly that. Masullo’s draft is much earlier in the process, so I’d wait.
Does this mean P vs NP is next?
No, not from this.

I should say that I didn’t find any mathematician commenting directly on P vs NP after the Riemann news, so what follows is my own reading and not a quote from anyone. The first reason is plain. They’re different problems from different corners of maths. The Riemann hypothesis lives in number theory, in the behaviour of primes, while P vs NP is about how fast computers can solve certain problems, which is complexity theory. Proving the hypothesis gives you no step toward the other one, and neither answer to the Riemann question would settle it.
The second reason is why the Riemann problem was even reachable. It had a staircase. Mathematicians had spent decades pushing the share of zeros proven to be on the line up from small numbers, then came the quasi-Riemann result, then Claude’s 41.6 to 67.2 percent jump in August. Each step gave the AI more to work from. P vs NP has no such staircase, and it has famous walls. Baker, Gill and Solovay showed in 1975 that a whole family of standard proof techniques can’t work on it, and Razborov and Rudich added another wall in the 1990s. I’m not sure that anyone has a map of what the right kind of proof even looks like.
Lean helps with one half of the job. If someone, human or AI, ever claims a P vs NP proof and writes it in Lean, a fake would get caught quickly. Finding the real thing is the hard half, and Lean does nothing for that.
I should be careful though, because Saha thought the hypothesis was years away, and he knows this field far better than I do. He was wrong, and so was almost everyone else. So I’m not saying P vs NP is safe for decades. I’m saying about 90-10 that it doesn’t fall in the next couple of years, and that the 30 comes from not trusting any of our guesses this month.
As for “all the other problems”, Buzzard’s answer is simpler than mine. Mathematics is infinitely big, you can’t solve all of it, and one good tool doesn’t change that.
What to watch over the next few weeks
The first thing is that final Lean check, which was still running. After that, Clay’s response. They haven’t changed the status yet.
But the bigger one is the plain-English version. If Buzzard is right, humans will have translated enough of the Lean code in a week or two to say what’s inside, and it’s still unclear what they’ll find. It could be a thousand pages of new ideas or a very clever arrangement of old ones. I’d guess some of both.
Also keep an eye on the Masullo draft, because Conrad is still reading it and the first corrections are already in.