Skip to the stories
Robotics & Automation

Anthropic Researcher Claims Claude Helped Prove the Riemann Hypothesis

  • Aamir Khollam
  • 713 reads

Anthropic mathematician Levent Alpöge says Claude helped him prove the Riemann hypothesis and its generalized form. Independent checking is still underway.

Anthropic Researcher Claims Claude Helped Prove the Riemann Hypothesis

The claim is large enough that it deserves a careful read before anyone celebrates or dismisses it. Levent Alpöge, a mathematician at Anthropic, announced on Friday, November 20, that he has a proposed proof of the Riemann hypothesis, and he credits Claude with rapid progress during the final stages. If the argument survives scrutiny, a conjecture that has stood since 1859 would finally be settled.

What's actually being claimed

The Riemann hypothesis is a statement about the distribution of prime numbers. It concerns the nontrivial zeros of the Riemann zeta function, a complex-valued function that encodes deep information about how primes thin out as numbers grow. Bernhard Riemann proposed that every nontrivial zero lies on one specific vertical line in the complex plane, the line where the real part equals exactly one-half.

Alpöge says his work goes beyond the original conjecture. He claims to establish the Generalized Riemann Hypothesis (GRH), which extends the same idea to a broader family of functions called Dirichlet L-functions. Those functions show up across number theory, and many results have been proven only on the condition that GRH or a related statement holds. A proof of GRH would turn a large body of conditional results into unconditional ones, which is why the generalized version carries as much weight as the original.

What Alpöge released

Alpöge published a four-page explanation alongside a large set of Lean verification files. The archive reportedly totals about 239 megabytes across 175,187 source files. That scale is unusual for a mathematical announcement, and it's the first thing that separates this from a typical AI-assisted sketch.

Lean is a formal proof assistant. It checks each step of an argument against strict logical rules, so a proof either compiles or it doesn't. Nothing gets waved through on the strength of a plausible-sounding paragraph. Mathlib, Lean's standard mathematics library, supplies the shared definitions and lemmas that formal proofs build on.

The method Alpöge describes relies on weighted prime sums, known as von Mangoldt sums, and on mathematical constructions called theta lifts. Those are established tools in analytic number theory, so the interesting question is how they're combined and whether that combination delivers the zero-location result it needs.

Alpöge also described the personal side of the work. He said that resolving a problem this influential would mean losing a guiding light for his research, and in the same breath he expressed excitement about what the field might do next. Researchers who have spent careers around a single open problem often describe that mix of loss and momentum.

Why a compiled proof isn't the finish line

A Lean archive is stronger evidence than an ordinary written argument, but compiling is not the same as proving the right thing. The code can be internally consistent and still formalize a slightly different statement from the one mathematicians care about. That gap, often called statement fidelity, is where formal projects most commonly go wrong.

An independent verification effort has reportedly confirmed that the downloaded files match Alpöge's published archive. Its early checks also found that the main theorem statements compile against Mathlib. That's a useful first step: it rules out the most basic kind of mismatch between what was posted and what was run.

Still, the hard part comes next. Reviewers need to confirm that the formalized theorem addresses the conventional Riemann hypothesis, check that every definition matches its standard meaning, and identify any assumptions that the code relies on without proving. A single unsupported axiom buried in a large library can invalidate a result that otherwise looks clean. With 175,187 files to read, that review will take real time.

The clock on a million-dollar problem

The Clay Mathematics Institute lists the Riemann hypothesis among its seven Millennium Prize Problems, each carrying a $1 million award. Alpöge's announcement doesn't qualify for immediate consideration. Clay requires publication in an approved peer-reviewed outlet, and the proposed solution must then survive at least two years of scrutiny and gain broad acceptance in the mathematical community.

That timeline is the point. Peer review in pure mathematics is slow by design, and a result of this size tends to attract many competing checks, not one. The community will probably look for the most fragile step first: the places where a large argument leans on an earlier result that was stated in a slightly different form.

So the immediate story isn't whether the problem is solved. It's whether a formal, machine-checkable proof of one of the field's central questions can hold up under expert examination. If it does, the more lasting change may be in how mathematicians use AI tools and proof assistants together, with the verification layer doing the job of catching errors that human readers miss. For now, the Riemann hypothesis remains officially unsolved, and the next few months of checking will decide whether that changes.

Source: Aamir Khollam · interestingengineering.com

From the same set