Skip to main content

Two Proofs of the Riemann Hypothesis Dropped on the Same Day. One Used AI.

An Anthropic mathematician says Claude helped him crack the most famous problem in math. Hours later, Stanford released a twelve-year human proof it wasn’t planning to share yet.
Josef Waples
Josef Waples

Create Your Free Account

Practice and track your progress.

Continue with GoogleBy continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.

Today was a shocking day: Two separate proofs of the Riemann hypothesis have landed. One was found by an Anthropic mathematician working with Claude. The other took a Stanford postdoc twelve years, and no AI touched it.

Let's start with the AI version: Levent Alpöge, a number theorist at Anthropic, posted a note this morning saying that he and Claude have proved the Riemann hypothesis, along with its generalization to Dirichlet L-functions, and released a Lean formalization to back it up.

Now for the non-AI version: Only several hours later, Ravi Vakil, president of the American Mathematical Society, put out a draft by Alessandro Masullo that claims the Generalized Riemann Hypothesis by an entirely human route, one that has been under private review for months.

Back in August, I wrote about Claude’s first swing at this problem, which fell well short of a proof but produced a real, Lean-verified improvement on a 50-year-old bound. Nobody I spoke to then expected the actual hypothesis to fall this year. Now there are two candidate proofs, built on different methods by people who, as far as anyone has said, were not working together.

If either holds, it is the biggest result in mathematics, period. If both hold, number theory just got two independent routes to solving the most famous open math problem out there.

Quick answer: The Riemann hypothesis has not been confirmed as proved, but there are now two serious, independent claims. Alpöge’s is a formal Lean proof that outside experts have not yet built or matched against the standard statement. Masullo’s is a human-written draft that Stanford’s Brian Conrad has been vetting for months and that the community is only now seeing. The first checks on both could come within days or weeks. A readable, accepted proof and the $1M Clay prize are each much further away.

What Is the Riemann Hypothesis?

The prime numbers, 2, 3, 5, 7, 11 and so on, look random as you walk up the number line. In 1859, Bernhard Riemann showed that their spacing is secretly controlled by a function, now called the Riemann zeta function, and specifically by where that function equals zero. His hypothesis is that every one of the "interesting" zeros (a mathematician would say "non-trivial") sits on a single vertical line in the complex plane, with real part exactly 1/2.

Riemann hypothesis solved ai

That is the whole claim. It has been checked numerically for the first ten trillion zeros. It has never been proved for all of them.

visualizing riemann zeta function with help of ai

Why is the Riemann Hypothesis Worth $1 Million?

The Riemann hypothesis is one of the Clay Mathematics Institute's seven Millennium Prize Problems, each carrying a $1M bounty. Before this year, only one had ever been solved: the Poincaré Conjecture, in 2003, by Grigori Perelman.

Progress on Riemann was so rare that most top number theorists spent essentially none of their time on it, not because it didn't matter, but because there was no visible path. As recently as August, when Claude improved a related bound, Oxford's James Maynard told Scientific American there was no route from those techniques to the actual hypothesis.

Why we care about this particular problem: Navier-Stokes mattered because weather models and aircraft design depend on it behaving well. Riemann is different. What it would do is instantly unlock hundreds of results that mathematicians have proved "conditionally," with the hypothesis as an assumption, over the last century and a half. The Generalized Riemann Hypothesis, which Alpöge also says he has settled, does the same for an even larger family of functions and an even larger pile of waiting theorems. That is why Scientific American calls it the North Star of number theory.

Does Proving the Riemann Hypothesis Break Encryption?

This is a first question I had, too.

If the Riemann hypothesis is about prime numbers, and RSA encryption is built on prime numbers, doesn’t proving it put online banking at risk? The answer is no. RSA depends on one specific hard task: take a huge number that’s the product of two big primes and recover those two primes. That’s a factoring problem, and nothing about the Riemann hypothesis makes it easier.

Riemann is about how primes are spread out along the number line, how closely their count tracks a smooth formula, and a proof would tighten the error bars on that estimate. It wouldn’t hand anyone a shortcut to a particular key. There’s another point: cryptographers have been assuming the hypothesis is true for decades, and plenty of security arguments are already proved “under GRH” (assuming the generalized Riemann hypothesis is true).

The actual threat to RSA is a working quantum computer running Shor’s algorithm.

What Alpöge Is Actually Claiming to Have Proven

This morning, Levent Alpöge posted a four-page note on his personal website saying he believes he's solved the Riemann hypothesis:

working with Claude, discovered a proof of the Riemann Hypothesis.

Levent Alpöge

He says the same method also settles the Generalized Riemann Hypothesis.

With his message, Alpöge released Lean code with a formal proof split into a 14-part archive, and a timeline that he says Claude assembled for him. There is no preprint, no journal submission, no independent verification.

The timeline is worth mentioning: Anthropic's shares started trading on Nasdaq only yesterday, while Alpöge's notes show months of open-ended work. Since October 6, he's been working with Claude across a range of mathematical areas, then steering different efforts toward combining ideas and pushing them toward the Riemann problem. Several of the key steps are dated November 15.

There are three distinct things being claimed, and they are not equally strong:

  1. The Riemann hypothesis itself, formalized in Lean. This is the headline claim, and the Lean files are the entire evidence for it.
  2. The Generalized Riemann Hypothesis, which extends the same statement to Dirichlet L-functions. Alpöge says the method resolves this too, and that the formalization covers it "without assumptions."
  3. The Extended Riemann Hypothesis, a further generalization (Dedekind zeta-functions). Here Alpöge is more careful: he describes it only as an expectation for future work, not a result.

In a footnote, Alpöge reports that a full natural-language write-up he generated ran past 250,000 pages! He chose not to publish it, calling it "slop." So what the world has today is Lean code and a note. 

He is also unusually candid about his own position. He writes that he is presenting a complete argument before having mastered the mathematics around it, and specifically acknowledges limited command of one area the proof relies on. His stated plan is to understand the argument, give talks on it, and write it up properly. So, Alpöge directed the approach throughout but, near the end, Claude moved faster than he could follow, and he's catching up.

Can You Trust a Proof Nobody Can Read? 

If you aren't familiar with Lean, know it's a formal proof assistant that checks whether each logical step follows from the previous one. A Lean proof can't be talked into looking correct: if it compiles, the chain holds.

That said, what Lean doesn't do is tell you whether the theorem you proved is the theorem you meant.

If you aren’t familiar with Lean, know it’s a formal proof assistant that checks whether each logical step follows from the previous one. A Lean proof can’t be talked into looking correct: if it compiles, the chain holds.

Harmonic CEO Tudor Achim made the broader case for this on DataFramed back in August: machine-checked proofs remove the bottleneck where mathematics has depended on a trusted reviewer signing off, and that same tooling is starting to spread into software. He also said AI’s edge in math is less insight than stamina, grinding through structures where a human would stop for coffee.

Verification runs in three stages:

  1. Does it build, and what does it rest on? Anyone can download the 14 files and ask Lean to check them. A project this size can take a while to compile, but it is mechanical work. Lean can also print the axioms a finished theorem depends on, which exposes any unproved assumption smuggled in.
  2. Is the final theorem really the Riemann hypothesis? This is the decisive test, and it needs human number theorists, not a compiler. Checkers have to confirm that the statement at the end of the code matches the standard one, and that every definition underneath it, starting with the zeta function itself, is the real thing and not a convenient lookalike. Until this is done, posts saying the hypothesis is "solved" are ahead of the evidence.
  3. Can a human understand it? Even a proof that passes both checks would be a black box (although it may forever be a black box for most of us). Mathematical results become knowledge when the community can read them, judge whether the ideas are new, and build on them. The OpenAI Navier-Stokes proof, more than two months on, is still being digested, and I go back to that artilce from time to time to update it.

And about that $1M prize: Clay's rules require a solution to appear in a qualifying publication, wait at least two years, and win general acceptance in the mathematics community. The earliest Clay could consider awarding a prize 2028, more likely 2029.

One thing is certain: AI is moving fast and keeps surprising us. Terence Tao reportedly said only weeks ago that AI might reach Riemann sometime before the 2050s. 

What Does This Mean for Anthropic's IPO?

Anthropic began trading on Nasdaq yesterday, November 19, and this announcement has come one day later. 

Alpöge addresses this: he says he consulted the company's leadership about whether he could announce, that there was a constraint of some kind he had to respect, and that he alone decided not to wait any longer. He also says that any cynical reading regarding announcing all this to help stock prices is false.

Now For the Second Proof, With No AI at All

While number theorists were still downloading Alpöge’s Lean files, Ravi Vakil, the Stanford algebraic geometer who currently serves as president of the American Mathematical Society, posted something on his own blog. He wrote it, by his own account, in a rush from an AMS board meeting, minutes before chairing a session that would last hours. The subject was a proposed proof of the Generalized Riemann Hypothesis by Alessandro Masullo, a postdoc at Stanford.

The key facts, as Vakil lays them out:

  • Brian Conrad, one of the most respected and most skeptical number theorists in the field, has been going through the argument closely for months and giving detailed feedback, which Masullo has been incorporating. Vakil says it has held up so far.
  • It is entirely human work. No AI was involved at any stage.
  • The original plan was to bring Masullo to Stanford for a year, polish the write-up, circulate it quietly among enough experts, and only then announce. Alpöge’s post made that plan untenable, so Vakil released the current draft as-is.
  • Vakil is explicitly asking the community to check it, formalize it, and patch any small gaps they find.

The draft is titled The Music of the Primes, which is a very pretty title, and in his introduction Masullo describes it as the product of twelve years of work. He also addresses, directly, the mood among mathematicians watching AI systems pick off problems that have stood for decades: his message to anyone feeling discouraged by the current climate is that human mathematicians are still here, and not going anywhere.

Three things are worth being careful about here.

  • First, Masullo’s claim is GRH, not RH alone. Since GRH implies RH (the Riemann zeta function is the simplest Dirichlet L-function), a correct proof of GRH settles the original hypothesis too. On paper, both claims cover the same ground.
  • Second, these are not the same proof. From what has been posted so far, Alpöge’s argument runs through heavy automorphic machinery (the Zulip thread, where number theorists are picking through the Lean code in real time, mentions Galois representations, automorphy lifting, and base change). Masullo’s draft is a long human-written manuscript with a different lineage. If both survive scrutiny, mathematicians will have two independent proofs to compare, which is far more valuable than one. If only one survives, the other becomes a very interesting near-miss.
  • Third, the two proofs are at different stages of review. Alpöge’s has been seen by essentially nobody outside Anthropic; its credibility rests on the Lean compiler and on Alpöge’s reputation. Masullo’s has had months of attention.

So, Is the Riemann Hypothesis Solved or Not?

Possibly. Possibly twice.

Here is what we have: two independent claims on the same day. One is a Lean formalization from a serious mathematician at Anthropic, built with Claude, that anyone can download and compile. The other is a twelve-year human manuscript from a Stanford postdoc that one of the field’s toughest reviewers has been checking for months. Behind both is a year in which AI-assisted results in mathematics have mostly held up under scrutiny.

What we don’t have is outside confirmation of either. Nobody has publicly reported building Alpöge’s files or matching the final theorem to the standard statement. Masullo’s draft has been public for a matter of hours. There is no paper in either case. Anthropic, Stanford, and the Clay Institute haven’t commented.

I’ll keep updating this piece as the information comes in.

Josef Waples

Author

Josef Waples

I'm a data science editor with contributions to research articles in scientific journals. I'm especially interested in linear algebra, statistics, R, and the like.

Topics

Related

blog

GPT-5.6 and the Dinitz-Garg-Goemans Conjecture