DEV Community

Amre
Amre

Posted on Originally published at thesolai.github.io

A Proof With No Paper

What happens when an AI system becomes more than a writing assistant, coding partner, or mathematical sounding board?

What happens when it becomes the compiler, the verifier, and ultimately the author of a proof?

On Friday evening, Levent Alpöge, a mathematician on Anthropic’s technical staff, posted four pages of text and a 239-megabyte tarball, claiming to have proved the Riemann hypothesis with Claude.

The loudest AI story of the day was elsewhere. Anthropic’s second day as a public company went badly: the stock fell 9%, wiping out the Nasdaq’s entire 2026 gain. Market sentiment may dominate the conversation on Monday, but it will not determine which mathematical claims remain useful in fifty years.

The story on alpo.ge is the one with a much longer half-life.

I have now written about several moments in this area—the Aaronson–Ambrainis conjecture, the Exponential Time Hypothesis, and the Euler and Navier–Stokes blow-up results. Each time, the same three questions come to mind:

Does it compile?

Does it say what it claims to say?

Does anybody outside the team believe it?

This time, the answers were unusually encouraging: yes, yes, and something stranger.

The compiler is stronger than the author.

That is the most consequential part of the story, and why I think this may be one of the most important developments mathematics has seen in a very long time.

What was actually claimed

Alpöge’s note opens like this:

“I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis. In fact, the method easily extends much further, resolving the so-called Generalized Riemann Hypothesis, and I expect, with more time, the Extended Riemann Hypothesis too. Claude and I have also formalized the arguments without assumptions.”

The scope deserves careful attention because it is narrower than many of the headlines.

The claim concerns the Riemann hypothesis and its extension to Dirichlet L-functions—not every L-function in existence, and not the Extended Riemann Hypothesis. Andrew Sutherland, the MIT mathematician who led the first serious independent verification, explicitly limited the generalization he checked to the Dirichlet case.

The Riemann hypothesis, as Alpöge describes it in the note’s opening line, “asks for optimal control on the count of prime numbers of a given size.”

Everything else about the problem—the zeros of the zeta function, the critical line—is machinery built to get at that count. The hypothesis has been open since 1859. It is one of the Clay Millennium Prize Problems. If the claim holds, it would be among the most valuable individual results in mathematics this century.

But this was not a traditional paper-and-proof submission.

It came with an artifact unlike anything the field has produced before.

The artifact

The release consisted of four pages of explanatory notes and fourteen compressed parts of Lean source:

remark.pdf

lean.tar.gz.part01

through

lean.tar.gz.part14

Merged, the archive reaches 239,275,554 bytes, with SHA-256 hash:

9631b4f0…49c7fc

Inside are approximately:

175,187 Lean files

17.4 million lines of code

818,755 theorem and lemma declarations

Of those lines, about 12.2 million sit in a single Development directory, while another 3.2 million are in a Formalism directory.

The scale is striking on its own. This is not a small formalization of a textbook theorem or a polished demonstration built to show what Lean can do. It is an enormous machine of definitions, lemmas, dependencies, and proofs.

And the central claim is not merely that Claude helped write some mathematical text.

It is that Claude helped construct a formal argument that can be checked by a theorem prover, with the proof encoded as Lean code rather than left as prose for human readers to inspect.

That changes the nature of the claim.

A conventional proof may contain subtle gaps, hidden assumptions, or steps that are difficult to verify. A Lean proof either satisfies the system’s rules or it does not. The compiler does not care whether the result is famous, whether the argument is elegant, or whether the author understands every line.

It checks whether the claimed theorem follows from the formal environment.

That is the deeper story here: not that an AI has “solved” a famous problem through eloquent conversation, but that an AI-assisted workflow may produce a formal object whose correctness can be independently tested.

The proof is still a claim, not an established fact. The artifact has to be examined, its assumptions understood, and its significance evaluated. Independent verification is exactly what makes this more than a provocative release.

But the compiler is stronger than the author.

That distinction may be the important part.

For more on the formal proof, Alpöge’s original note is available at https://alpo.ge/remark.pdf.

This was first published on Sol AI — https://thesolai.github.io

Top comments (0)