All News
mathematicsriemann-hypothesisleanformal-verificationanthropicclaude

Alpöge's Riemann proof is 17.4 million lines of Lean and no paper

Alpöge and Claude claim the Riemann Hypothesis and the GRH, delivered as 17.4m lines of Lean with no readable paper. Here is what is established and not.

Vlad MakarovVlad Makarovreviewed and published
7 min read
Alpöge's Riemann proof is 17.4 million lines of Lean and no paper

On the evening of 20 November, Levent Alpöge published a four-page note and a fourteen-part tarball. The note says he believes he has, working with Claude, discovered a proof of the Riemann Hypothesis, that the method extends to the Generalized Riemann Hypothesis, and that the argument is formalized in Lean without assumptions. The tarball is the argument. There is no paper behind it, no referee report, and no page a number theorist can read the way they read a proof.

What was published, and what was not

Alpöge anticipated the objection in the note itself, and his answer is why this is not a normal priority dispute. Attempting some massive unreadable travesty of "slop", he writes, would have been obviously inappropriate, so he decided "to only present Lean". He is candid that this is a stopgap: he intends to understand and impart understanding of the argument and its simplifications, to speak about it, to write it properly.

The note is therefore two things at once. It is an announcement in the genre of a priority claim, and it is a pointer to an artifact. It also carries a rough timeline, a sketch of how a chain of theta lifts, each mimicking a square root of multiplicity, might recover the divisor multiplicity that Stepanov's function-field argument gets for free, and a list of people he thanks, among them David Dohan, Ben Michel, Ralph Furman, Eric Price, Cora Vinther, Peter Sarnak and Manjul Bhargava. What it does not carry is a proof anyone can follow. Alpöge says so himself: where Claude was the expert and not him, he is paraphrasing a summary "which might make zero sense".

Two readable proofs, one Lean-only

The best calibration came the same day from Terence Tao, who listed three results in decreasing order of how digested they currently are. Two are blowup results, for the unforced incompressible Euler and Navier-Stokes equations, both with Alpöge among the authors, and both ship with a readable argument alongside the Lean formalization. Of the Navier-Stokes paper Tao writes that the strategy is "broadly clear", while more work is needed for a professionally written proof.

The third entry is one sentence. Alpöge, "using an internal AI model", has generated an "(enormous) Lean-only proof of the Generalized Riemann Hypothesis". Same author, same day, and the difference is the shape of the artifact. Tao adds that early analysis suggests the argument uses the full power of modern work on the Langlands program and arithmetic geometry, and that digestion "will take some time, though in this case at least we have a human author to communicate with".

What the archive contains

A mirror on GitHub reproduces the fourteen parts unmodified and reports the merged archive at 239,275,554 bytes, SHA-256 9631b4f0. Its layout table counts 175,187 Lean files and roughly 17.4 million lines, with 12.2 million of them in a single Development directory and 3.2 million in a Formalism directory.

The front door is small. Main.lean states five theorems: the Riemann Hypothesis, the zeros-on-the-line form, nonvanishing for Dirichlet L-functions to the right of one half, and the two GRH statements. Each is one line applied to a proof term defined deep in the supplements. RiemannHypothesis itself is Mathlib's own definition rather than something written for this archive, which removes one obvious way to fake a result. Statements.lean is 16 lines and imports only Mathlib. The toolchain is pinned to lean4 v4.33.1 with Mathlib at a fixed commit.

The mirror's static scan found no sorry, no admit, no axiom declarations, no native_decide and no opaque. It did find one addDecl, a command that copies existing declarations under new names, and 22 custom macros described as tactic shorthands. Those are precisely where an auditor looks, and none of them is a proof by itself. The README is blunt: as of 20 November, "no independent build or review of this code has been published. Treat it as a claimed proof until it has been compiled and checked."

The kernel is now load-bearing

Scott Aaronson, writing that afternoon, put the structural risk where it belongs. Lean, he notes, has had kernel bugs that in the past "allowed proofs of 0=1", and its underlying type theory has uncertain soundness. Until those are fixed, "a sufficiently motivated AI could almost certainly exploit them to prove a false statement". On this instance he sees "zero indications that that's happening right now".

That moves the question rather than answering it. A formal proof shifts the enquiry from whether the mathematics is correct to whether the kernel implements its logic correctly, and whether the statement means what the field means by it. The second half is not hypothetical: a proof of a mis-stated theorem compiles perfectly and is worth nothing. Here the target predates the proof, since the headline proposition is imported from Mathlib, which is the strongest available defence against that failure mode.

A base rate, not a verification

Aaronson's confidence deserves its precise weight. He bets the proof will stand because no AI result carrying a Lean certificate has so far needed retraction, while several without one have. That is a track-record argument, published within hours of his waking to the news, and it is not an inspection of twelve million lines. He is explicit that mathematicians will not be satisfied by the artifact: they will be satisfied "when, and only when, they're able to explain and distill and reconstruct the ideas for themselves", which "will probably take years".

His own expectations the night before are the cleanest measure of the speed. Asked at dinner how long until an AI proves the Riemann Hypothesis, with several more years suggested, he had answered that one should not be sure of such things anymore. He cited the earlier result bounding nontrivial zeros to within 3/8 of the critical line, the rung below this one, which we covered in October and which he notes already carries most of the consequences for number theory. Five weeks later there is a claim to the full statement.

Attention arrived immediately: the Hacker News submission stood at 2,142 points and 1,442 comments when I checked it. Attention is not verification, and most of that thread reacts to the size of the number rather than to the code.

What would settle it

There is a cheap first step and an expensive second one. The cheap step is a build. Fetch the archive, verify the hash, run the dependency cache, compile Statements and Main, print the axioms for the five theorems, and replay the resulting environment through an external checker such as lean4checker so that the result does not depend on the repository's own metaprograms. That answers whether the kernel accepts the proof. Tao points readers to the Lean Zulip thread as the place where the work is being done.

The expensive step is comprehension, and Alpöge's note is unusually direct about the gap. He describes publishing while "presenting the world with a complete argument before having mastered a massive neighbourhood around its mathematics", says he may have "baked lots of strange suboptimal choices into the argument", and expects shorter arguments to follow fairly quickly. He also discloses the timing: the work landed late on a Sunday night, he messaged Dario Amodei about whether he could announce, and writes that "the precise timing is horrible".

So what is established tonight is narrower than the headlines. A public artifact exists and matches its author's description; a prominent complexity theorist read the announcement and expects it to hold; and the author says the readable mathematics is still ahead of him. What is not established is that seventeen million lines of Lean state the Generalized Riemann Hypothesis and prove it under a sound kernel, checked by someone with no stake in the answer, and nobody has yet explained the argument well enough for a number theorist to reconstruct it. Until both of those happen, the accurate description is Alpöge's own: a claim, with a certificate attached.

Related Articles

Scroll down

to load the next article