· via dev.to (home feed)
Claimed 17.4M-line Lean proof of the Riemann Hypothesis awaits its first full build
Levent Alpöge says he and Claude produced a 17.4-million-line Lean formalization of the Generalized Riemann Hypothesis with zero sorry — and no one has completed a full build yet.

The claim
On November 20, 2026, Levent Alpöge, a mathematician at Anthropic, announced that he and Claude had discovered a proof of the Riemann Hypothesis, that the method extends to the Generalized Riemann Hypothesis for Dirichlet L-functions, and that the arguments had been formalized in Lean without extra assumptions. The announcement is a four-page personal note; the proof itself was published as a fourteen-part archive of Lean source. In that note, as quoted by Fernando Borretti in a blog post, Alpöge describes Claude pulling far ahead of him near the end and says he expects the Extended Riemann Hypothesis to yield to the same approach with more time.
The day brought a cluster of Millennium-scale claims. Alessandro Masullo posted a second, independent GRH claim. Alpöge and Tristan Buckmaster, again with heavy AI assistance, claimed finite-time blow-up for the unforced incompressible Navier–Stokes equations in three dimensions. Terence Tao blogged about the GRH proof, the Navier–Stokes blow-up and a separate Euler-equations blow-up, while Scott Aaronson wrote that his bet is that the GRH proof will stand. The Hacker News discussion reached 2,304 points.
What the archive contains
A dev.to write-up that cloned a GitHub mirror and measured it reports 175,187 Lean files totalling 17,374,215 lines, roughly 1.4 GB. There is no .lake directory, so none of those lines are vendored Mathlib; the count is the project's own source. It includes blank lines and comments, making it a size measure rather than a count of proof steps, but by that write-up's assessment it is still the largest formalized proof ever made public, by a wide margin.
The checks that matter most came back clean: searches for sorry, admit, sorryAx and top-level axiom declarations all returned nothing across the project. In Lean, sorry is the placeholder that lets an unfinished proof typecheck, and a formalization containing one proves nothing, so its absence is a necessary first gate rather than a final verdict.
The target statement is also anchored to Mathlib's standard definitions of the zeta function and Dirichlet characters, closing off the classic trick of proving a vacuous homemade variant. The entry point is twelve lines long, and everything reduces through a single step to one claim: that the Dirichlet L-function statements hold for every character. According to the write-up's summary of Alpöge's note, the argument descends from Stepanov's proof of RH over function fields, using chained theta lifts whose chain length is √X for a von Mangoldt sum of length X.
What nobody has verified yet
As of the write-up's publication, no one had publicly completed a full build of the project. The mirror's own README states that a partial independent replay confirmed the statement definitions and several conditional reductions, and that the full proof build remains unverified. A zero sorry count is necessary but not sufficient; the completed build is what would make it sufficient.
Aaronson adds two caveats. Lean's kernel has historically contained bugs that admitted proofs of false statements, so a system could in principle exploit a soundness hole rather than prove something true, though he sees no sign of that here and notes that no AI-assisted result carrying a Lean certificate has needed retraction so far, while several without certificates have. Separately, mathematicians will not treat 17 million lines of Lean as an answer until they can explain and distill the ideas themselves, which he expects to take years. Tao's initial read is that the argument appears to draw on the full strength of modern work in the Langlands program and arithmetic geometry. Verification efforts are being followed on the Lean Zulip.
The explanation debt
Alpöge says the formalization is finished and that producing a human-readable document from it would be routine. According to a second dev.to post, he generated more than 250,000 pages, a pile Claude reportedly called a book, before deciding that publishing such a dump would be inappropriate and releasing only the Lean. By his own account he has not yet mastered the mathematics surrounding the argument and expects a much simpler version to emerge. That post frames the gap as the real story: a green build is permission to start explaining, not permission to stop. Borretti's essay reads the announcement's elegiac tone as a marker of where discovery now happens, with the AI at the center and the human guiding from the edges.
The Navier–Stokes repository, checked the same way, holds 2,271 files and 616,060 lines behind a 79-page paper. Its 24 sorries sit in the top-level statement files exactly where its README says they belong, and the proof library itself contains none.
Why it matters
The Riemann Hypothesis is a Clay Millennium Prize problem and a standing assumption across analytic number theory; GRH is the generalization much of the field quotes conditionally. If the build passes independent replay, mathematics gains the result and formal methods gain the largest checked artifact in history, produced with an AI doing much of the technical work. The event also relocates the burden of proof: acceptance would rest not on reading an argument but on running one, followed by the slower human work of understanding it. And it sharpens a new question for AI-assisted research, namely what it means to hold a proof whose certified correctness outruns its author's comprehension.
- #lean
- #formal-verification
- #mathematics
- #riemann-hypothesis
- #ai