Archived edition

Updates on artificial intelligence for mathematical research, proof discovery and formal verification, with links to papers, code and tools.

Research updatesSelect an entry to read more

2 entries · 7 linked sources News found
Levent Alpöge announces a Claude-assisted RH and Dirichlet GRH proof, with substantial independent Lean checks reported Levent Alpöge; verification reports by Andrew Sutherland and independent contributors Published: 6 sources

In a four-page note on his official personal site, Alpöge says he and Claude discovered a proof of the Riemann Hypothesis and that the method extends to the Generalized Riemann Hypothesis for Dirichlet L-functions. He says the argument was formalized in Lean and releases it as a large multipart archive; the note describes the human-readable development as unfinished. Within the same day, Andrew Sutherland’s MIT verification page reported that all 175,187 modules compiled, the five headline theorem statements and axioms were checked, module-level kernel replay passed, and an independent checker accepted the exported proof. Sutherland’s page gave its last update as 2026-11-20 22:05 UTC and said the official Comparator run and another kernel replay were still in progress. A separate LUMI project reported a complete independent build but still had its statement and axiom audits running at its 22:30 UTC update. The formalized GRH scope is Dirichlet L-functions, not every use of “GRH.”

Research relevance:

Potentially a landmark result in analytic number theory and an unusually concrete case study in AI-assisted discovery followed rapidly by public formal verification. The distinction between the initial announcement, the available Lean proof, independent machine checks, and a readable mathematical account matters for researchers assessing both the theorem and the workflow.

Claim status:
The proof is Alpöge’s claim. As of Sutherland’s published 2026-11-20 22:05 UTC status, independent reports provided strong evidence that the released Lean code compiles, states Mathlib’s Riemann Hypothesis along with four Dirichlet L-function theorems, uses only standard Lean axioms, and passes multiple kernel/checker validations. This verifies the formal statements as encoded under the reported toolchains; it is not a human proof review or a claim that the mathematical community has accepted the result. Terence Tao’s 2026-11-20 blog entry likewise described the GRH contribution as an enormous Lean-only proof whose strategy still needs to be extracted and understood.
Limitations:
The announcement note is dated 2026-11-20; its internal timeline says the key events occurred earlier, including progress on November 15–16. Independent checking was still underway at the stated MIT update time, and verification reports are time-sensitive. The released work lacks a conventional readable proof; the author says he has not mastered all of the mathematics and anticipates simplifications. The MIT verification page notes that its compilation used direct Lean invocations on cloud machines rather than a full Lake build, and that some checks remained incomplete at its cutoff. This digest did not download, build, or execute Lean. Machine checking alone does not confer peer review, human understanding, or institutional acceptance.

Sources & further reading

Ravi Vakil posts Alessandro Masullo’s proposed human-written GRH proof, then reports problems Ravi Vakil; proposed proof by Alessandro Maria Masullo Published: 1 source

Vakil posted a current draft of Masullo’s proposed proof of the Generalized Riemann Hypothesis, saying Brian Conrad had been examining it for months and giving extensive feedback. Vakil described the draft as human-written, with no AI assistance, and invited rapid checking. In an update about eight hours later, Vakil said several people had identified problems with help from different AI models and emphasized that all issues would need to be resolved before the proof could be accepted.

Research relevance:

A useful counterpoint on AI’s role in current proof workflows: models were reportedly used to identify problems in a human-authored draft, while the claimed AI-assisted Lean result followed a distinct formal-checking path.

Claim status:
This is an announcement of a proposed proof, not an established GRH result. Vakil’s same-day correction explicitly reports problems and withdraws any implication that he is asserting the proof strategy is correct. His post attributes the proposed work to Masullo and reports Conrad’s review and feedback; it does not certify a proof.
Limitations:
The post appeared on 2026-11-20 and was updated around eight hours later. The linked draft was accessible from the blog, but this digest does not independently analyze its mathematics. The problems are reported by Vakil in a brief update, without a full public account here of their nature or resolution.

Sources & further reading

Coverage and limitations
  • Window checked: 2026-11-19 18:15 to 2026-11-21 06:15 Europe/Paris. Read Levent Alpöge’s official personal site and four-page note, Terence Tao’s November archive, Ravi Vakil’s Math 216 post, Andrew Sutherland’s MIT verification page, and independent verification project records. Search results also surfaced AP and Nature coverage; Nature’s article was not directly accessible, so it was used only as an indexed result, not as a read source.
  • Alpöge’s official site identifies him as an Anthropic mathematician and links the note. The supplied X API lead from Scott Armstrong included commentary about the announcement, but the X post itself could not be opened (403/cache miss); no complete X timeline was read. Supplied X feed posts were treated as leads with their provided created_at dates, not as independently verified identities. Other supplied feed results were empty or unavailable; those do not establish that nothing happened. X collection was bounded, excluded replies and reposts, included cached posts, and API account resolution is not independent identity verification. The supplied Bluesky feed for xenaproject.bsky.social returned no posts, with a 300-post cap and replies/reposts excluded. The Mathstodon feed for Tao was unavailable (HTTPError; at most 300 posts).
  • Checked related official blogs and sources: Tao’s blog and Axiom’s research blog were read; no additional in-window Axiom mathematical result beyond items already supplied was identified. The supplied Harmonic result was already in previous_items. Simon Frieder’s blog could not be consulted through the web tool. Feed collection had per-run budget limits; Bubeck and Bloom were specifically marked unavailable due to exhausted post/time budget. Results are not an exhaustive survey of social media.

Requested coverage window: 2026-11-19T18:15:01.215544+01:00 — 2026-11-21T06:15:02.028364+01:00.

Public web research; coverage is not exhaustive. A source link is not a certification of a claim.

Archives 83
Older editions (53)