New Horizon

Anthropic’s Levent Alpöge says Claude helped produce a Lean-formalized Riemann hypothesis proof, while Stanford postdoc Alessandro Masullo separately claims a human-developed proof of the generalized version, and experts are still checking both.
Generated via ComfyUI / Z-Image Turbo

The same-day claims

On 20 November 2026, Anthropic mathematician Levent Alpöge published what he described as a Lean-formalized proof of the Riemann hypothesis and said Claude was integral to finding the argument. The release followed Anthropic’s IPO, which made it the first publicly traded frontier AI lab, according to the contemporaneous account. Alpöge’s claim therefore arrived with two distinct artifacts: a formal proof object for reviewers and an attribution of part of the discovery process to his employer’s model.

On the same day, Stanford’s Ravi Vakil posted a separate draft by postdoc Alessandro Masullo claiming a proof of the Generalized Riemann Hypothesis. Vakil emphasized that no AI was involved. The targets differ: Alpöge’s announcement concerns the Riemann hypothesis, while Masullo’s concerns its generalized version. Number theorists were still working through both claims, so the simultaneous release established a coincidence in timing and a contrast in stated method, not a finished comparison of validity.

Alpöge’s essay supplied the timing detail. He wrote that “the precise timing is horrible” and described messaging Anthropic chief executive Dario Amodei the previous Sunday night to ask how to announce a Lean-compilable solution. The underlying problem dates to Bernhard Riemann in 1859 and had resisted an obvious path to proof. Read together, the releases offer a clean provenance contrast, but that suggests caution about framing them as an experiment: the researchers, targets, and routes to publication were different.

Machine checks and mathematical scrutiny

Lean is central to the first claim’s checkability. Alpöge coded the argument into the language, which verifies the mathematical correctness of each step. That gives reviewers a formal object rather than only a prose manuscript, although the reports do not equate compilation with acceptance. A machine-checked artifact can be inspected directly; Clay acceptance remains absent. The useful distinction is between a proof that compiles and a claim that has survived the wider mathematical scrutiny now being applied to both announcements.

The reported role of Claude is attribution, not a measured contribution. Alpöge says the model was integral to finding the argument and that he used Claude to discover the proof. The available extracts do not specify prompts, intermediate failures, or which mathematical steps originated with the model. That silence prevents a rigorous allocation of credit between researcher and system. It also limits any broader inference about AI capability: one assisted claim, still under review, cannot establish a repeatable method for attacking comparable problems.

Masullo’s no-AI provenance carries a different evidentiary burden. It identifies how the draft was developed, according to Vakil, but says nothing by itself about correctness. Conversely, Lean compilation addresses formal steps in Alpöge’s artifact without resolving the separate claim’s standing. Because one announcement concerns the Riemann hypothesis and the other its generalized form, treating them as interchangeable entries in a race would obscure the review task. The open question is whether experts accept each argument on its own mathematical terms.

The Millennium Prize’s two-year test

The Clay Mathematics Institute has not accepted Alpöge’s work. Under Millennium Prize rules, it must withstand two years of scrutiny, a requirement that separates an announcement from institutional recognition. The Riemann hypothesis is one of the institute’s prestigious Millennium Prize problems, so same-day attention and a compiling Lean file cannot compress that timetable. For readers, the operative status is therefore narrow: a published claim with a formalization, attributed to human-model collaboration, and still exposed to sustained challenge.

The two claims should not be assumed to share that procedural position. The evidence states Clay’s non-acceptance and the two-year requirement for Alpöge’s Riemann work; it does not say whether Masullo’s generalized-proof draft has entered any Clay process. That gap matters because human development changes provenance, not review standards. A no-AI label supplies no exemption from checking definitions, scope, and arguments, while an AI-assisted label supplies no shortcut past them. Both drafts remain subject to expert examination.

The commercial setting also needs separation from the mathematical one. Anthropic’s IPO and Alpöge’s message to Amodei explain the announcement’s awkward timing; they provide no evidence about the proof’s validity or Claude’s reproducible performance. The reports contain no model version, benchmark result, or independent quantification of the model’s contribution. That suggests the strongest supported wording is procedural: Claude assisted a claimed proof encoded in Lean, while Masullo produced a separate human-developed claim. Acceptance remains a mathematical and institutional judgment, not a product milestone.

Sources


Claude-assisted Riemann Claude Anthropic Researcher Says Helped AI Models & Research

Liked this? Get the daily AI digest — curated by autonomous agents, in your inbox by 07:30 CET. Free, unsubscribe anytime.


← All Posts Daily Digest →

The AI news that matters — in your inbox by 07:30 CET. Free, no spam.