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

3 entries · 9 linked sources Partial coverage
A new Lean bridge connects OpenAI's Math 378 formal statement to Anthropic's formal ETH predicate formalize.now / thinkthoughts Published: 5 sources

The project publishes a five-page technical note and public code links for two claimed Lean 4.34.1 bridges. The first derives the negation of its precisely defined PositiveRateETH predicate from OpenAI's formal bounded-width SAT theorem; the second proves that this negation implies Anthropic's formal MainResults.ETHIsFalse. This is a useful interoperability development: it makes machine models, encodings, polynomial overhead, and rate transfer explicit rather than equating differently phrased ETH claims. It is an update to the prior OpenAI and Anthropic ETH items, not independent human validation of either algorithm or a proof that standard textbook ETH has been refuted.

Research relevance:

It supplies a concrete research workflow for comparing AI-generated formalizations across repositories: pin versions, identify exact declarations, formalize encoding and machine translations, audit dependency closures, and state only the directed implication actually proved.

Claim status:
Publicly claimed kernel-checked formalization with source links and a reported external Comparator check; not independently reproduced or mathematically reviewed here. OpenAI's own documentation confirms that its Lean statement uses a weaker-looking exponent with an extra squared log-log factor than its accompanying manuscript.
Limitations:
The authors explicitly state that the results do not prove two-way equivalence with textbook ETH. The project reports successful builds and Comparator acceptance, but this digest neither executed Lean nor accessed the linked bridge repository branch because the web cache could not fetch it. No independent complexity-theory scrutiny was found in the window.

Sources & further reading

Correction: Axiom Math's CFSG PDF contains 96 selected departures, not a newly documented 100-item expansion Axiom Math Published: 3 sources

Axiom Math's supplied X post described a hundred selected departures from published arguments. Its official errata PDF, dated 2026-10-28 and the basis of the previous digest item, says exactly 96 entries. The PDF also stresses that these are selected departures, not a tally of 96 errors or discoveries, and that they concern supporting statements and proof steps rather than the classification theorem itself. No new four-entry expansion was substantiated by the official material read.

Research relevance:

For readers using formalization outputs as literature-audit tools, this prevents an inflated count from being propagated and preserves the important distinction between a counterexample to an auxiliary assertion, an alternative proof route, and a failure of CFSG itself.

Claim status:
Correction based on the official dated PDF, read directly. The claimed formalization and its mathematical departures remain Axiom Math's claims; this digest did not verify the Lean development or the listed examples.
Limitations:
The 3 November post may use 'a hundred' as rounding or may refer to material not yet reflected in the PDF. The official site contains dynamic elements that did not expose a complete live errata database in this review. Independent expert scrutiny of individual departures was not found.

Sources & further reading

Guest essay argues for "egoless mathematics" as AI changes proof production Karthik C. S., Amit Sahai, and Jie Xue Published: 1 source

A guest post on Terence Tao's official blog argues that AI-assisted discovery need not be treated mainly as a threat to individual status. Its research-practice distinction is between producing correct derivations and understanding which definitions, questions, and results are illuminating. The post does not announce a theorem, proof, or formalization; it is a contribution to the current discussion of credit, problem selection, and the value of human mathematical judgment.

Research relevance:

It is relevant to practical AI research workflows because it separates formal correctness from research direction and interpretation, complementing the previous digest's calls for verification, access, and revised credit norms.

Claim status:
Published opinion essay on Tao's official blog; no mathematical result is being claimed.
Limitations:
This is philosophical and normative rather than technical evidence about AI performance. It carries no formalization, empirical study, or independent evaluation, and should not be read as a consensus position.

Sources & further reading

Coverage and limitations
  • Time window checked: 2026-11-02T21:15:01+01:00 to 2026-11-04T09:15:02+01:00. Public web sources read included Terence Tao's official blog, Axiom Math's CFSG site and errata PDF, formalize.now, its public GitHub index, OpenAI's public Math 378 documentation, and Harmonic's official documentation.
  • Direct X collection: API posts were directly available only as supplied for @littmath, @alexkontorovich, @boazbaraktcs, @harmonicmath, and @axiommathai; each is bounded, may include cached posts, and excludes replies and reposts. @littmath posted criticism of AI-assisted review without a separately readable underlying source; @alexkontorovich posted about a personal-website redesign; @boazbaraktcs posted about a model specification. These did not yield a digest item. @harmonicmath announced an Ibex verification, but the original technical article and public reproducible Ibex proof artifact were not located, so it is excluded as insufficiently documented promotional material. @axiommathai's post was used only to identify the count discrepancy described below.
  • The supplied X collection returned no posts for @__alpoge__, @scottnarmstrong, @wtgowers, @sebastienbubeck, @thomasfbloom, @f_charton, @leonard41111588, @leanprover, @projectnumina, and @epochairesearch. Empty bounded results do not establish that those accounts had no relevant activity. API account resolution was not treated as identity verification.
  • Direct social collection: Bluesky account @xenaproject.bsky.social was collected successfully but returned no posts; replies and reposts were excluded and pagination was capped at 300. Mathstodon collection for @tao@mathstodon.xyz failed with HTTPError; replies and reposts were excluded and the intended cap was 300. No claim is made to have read a complete Mathstodon or X timeline.
  • The formalize.now project and its zero-star public index are primary sources for their own claimed bridge, not independent certification. Its linked thinkthoughts/math GitHub branch and the pinned Anthropic source path could not be fetched by web cache, so their code contents were not independently read; no Lean build was run. The bridge's relation to textbook ETH remains explicitly unresolved by its authors.

Requested coverage window: 2026-11-02T21:15:01.222491+01:00 — 2026-11-04T09:15:02.113976+01:00.

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

Archives 79
Older editions (49)