行业观察 合集 · 大模型及AGI

当 AI 可以证明一切,数学还是「人类」的数学吗?

当 AI 可以证明一切,数学还是「人类」的数学吗?

所属合集 大模型及AGI 第 1 / 4 篇

当 AI 可以证明一切,数学还是「人类」的数学吗?

On October 6, 2026, OpenAI published 721 mathematical manuscripts on GitHub in a single drop, all produced by an unreleased internal model on open research problems. In November, Google DeepMind introduced AlphaProof Code, and Anthropic's formal-math repository released three machine-checked Lean 4 projects. Meanwhile, 16 mathematicians issued the Leiden Declaration, endorsed by the International Mathematical Union, and 25 Fields Medalists signed a joint statement warning of AI's "serious misalignment" in mathematics.

AI is generating mathematical proofs at unprecedented speed - and the mathematics community is pushing back at unprecedented force.

1. OpenAI's 721 Manuscripts: Unprecedented Scale, ~42% Formalization

The official numbers (corrected)

The openai/math repository is the largest public collection of AI-generated mathematical results to date. The following figures were verified directly from the GitHub README:

Metric Value Source
Total manuscripts 721 GitHub README
Result families 373 GitHub README
Top-line results formalized ~42% GitHub README
Producing model Unnamed internal frontier model GitHub README
Average compute per result ~3 hours of ChatGPT Pro thinking compute GitHub README
Total problems posed ~4,000 GitHub README

An important correction: Chinese media reports had cited "722 manuscripts / 372 families / 22% formalized," which does not match the official repository. OpenAI states "~42% of top-line results formalized" - and top-line results are not the same unit as manuscripts, since one family can group several papers (a principal result, companion arguments, consequences, or alternative proofs).

Ten published reasoning traces

OpenAI also released abridged summaries of the model's reasoning for ten result families:

Family Subject
007 Ordinary two-point correlations of multiplicative functions
017 The irrationality exponent of the circle constant pi
087 Symmetric and general Mahler conjectures
102 Ordinary NP-hardness at the basic semidefinite threshold
159 Quasipolynomial bounds for arithmetic progressions
197 Kaplansky's direct-finiteness conjecture in characteristic two
221 The Mezard-Parisi formula for diluted spin glasses
271 Spontaneous magnetization in the quantum Heisenberg ferromagnet
287 Isomorphism of free group factors
362 The three-dimensional relativistic Vlasov-Maxwell system

Notable exceptions to the fixed procedure include work on a zero-free region for the Riemann zeta function and a proof of the Hodge Conjecture for CM abelian varieties.

OpenAI's positioning

The README contains a striking statement:

"In our view, these manuscripts are best understood as a form of raw research data, rather than polished academic papers. We defer to the math community completely on matters of credit, precedence, and concurrent work."

This is a defensive posture - OpenAI is deliberately avoiding a fight with mathematicians over authorship and credit.

2. Anthropic's formal-math: Three Lean 4 Projects, Including Riemann Zeta Zeros

As of November 1, 2026, Anthropic's formal-math repository (472 stars) publishes three machine-checked Lean 4 formalizations:

Project Statement Origin
zeta23 More than two thirds of the zeros of the Riemann zeta function are simple and on the critical line Alpoge-Furman, arXiv:2608.13637
3sum-apsp Word RAM programs solve 3SUM in O(n^1.9992) and the min-plus product and APSP in O(n^2.99942) Alman-Vassilevska Williams, arXiv:2610.06783
eth A deterministic Turing machine counts satisfying assignments of every k-CNF in subexponential time, refuting the Exponential Time Hypothesis (ETH) Claude, paper included in the folder

The verification chain

Each project must pass the Palomar registry acceptance check:

  • leanprover/comparator with a toolchain-matched lean4export
  • The independent nanoda kernel
  • The landrun sandbox

A project passes only when lake build succeeds on its pinned toolchain with no sorry outside the trusted comparator statements, and the registry prints "Your solution is okay!" for every declared comparator configuration.

One detail worth noting

The eth project credits its source as "Claude, paper in the folder" - meaning the paper refuting ETH was produced by Claude, and Anthropic then formalized it. Compared with OpenAI's approach (publish raw manuscripts, formalize some later), Anthropic chose to publish the formalization itself.

3. Google DeepMind AlphaProof Code: Verification Beyond Pure Math

In November 2026, the DeepMind blog announced "Introducing AlphaProof Code: a breakthrough in verifying complex software at depth" (the post itself returned 404 during our fetch; the title is verified from the blog index).

From the title and context, AlphaProof Code targets deep verification of complex software rather than pure mathematical proofs - likely industrial code verification, safety-critical property proofs, and a software-engineering sibling of AlphaProof for mathematics.

4. Google Antigravity /teamwork: Multi-Agent Research Collaboration

The /teamwork command documented on antigravity.google (verified directly) runs a collaborative multi-agent team built for large software projects, complex simulations, and deep research tasks, including mathematical proofs.

Architecture

Role Function
Sentinel Coordinator: records the request, routes tasks, posts progress, spawns the final audit
Project Orchestrator Breaks the approved brief into milestones, coordinates parallel tracks, hands off to a fresh successor between milestones
Explorers Read-only research agents tracing call chains and candidate solutions
Workers Implementation agents with terminal and file tools
Critic Independent code review: correctness, completeness, robustness, style
Challenger Builds adversarial test suites, edge cases, worst-case inputs
Auditor Checks test evidence against real command output; blocks mocked or skipped tests
Success Auditor Final end-to-end verification before delivery

There is a dedicated Math / Proof execution path (triggered by prompts such as "prove theorem X"), plus a Math / Proof Large Team path running tournament networks for hard conjectures and combinatorial search.

The August 27, 2026 Antigravity blog post was titled "Teamwork: when AI becomes a research partner."

The Rust rewrite case (verified from the November 3 blog post)

Metric Value
Source code 174,774 lines of Fortran 90 and C across 264 files (FVCOM coastal ocean model)
Goal 100% memory-safe Rust, zero unsafe blocks
Run time 59 minutes of autonomous execution
Legacy bugs caught 6 silent bugs dating from the 1990s
Memory usage Cut in half
Speed 1.88x faster

Three stages: baseline capture (11 min, golden outputs across 100 random seeds), full-repo audit and blueprint (3 min), then the Rust port with closed-loop self-healing and PGO (45 min), with zero numerical difference against the original Fortran.

5. The Leiden Declaration: Mathematics' Formal Counter-Response

Item Content
Full name Leiden Declaration on Artificial Intelligence and Mathematics
Published June 2, 2026, in Leiden, Netherlands
Working group 16 mathematicians (University of Washington, Edinburgh, Cambridge, Oxford, ETH Zurich, Leiden, and others)
Endorsement Officially endorsed by the International Mathematical Union (IMU)
Signatories 2,411 mathematicians
arXiv 2609.35334
Website leidendeclaration.ai (full text fetched directly)

The declaration grew out of the September 2025 Lorentz Center conference "Mechanization and Mathematical Research," which gathered around 60 participants from 10 countries, followed by eight months of drafting.

Five threats (as stated in the preamble)

  1. Reliability: automated techniques can produce plausible but unreliable or incorrect arguments that are hard to distinguish from correct proofs - and this now applies to formalizations too, where translation errors between computer-encoded and human presentations hide in the gap
  2. Attribution: models trained on published works frequently return outputs that do not properly cite the human works they synthesize
  3. Incentives: the use of AI may become incentivized for its own sake, disrupting hiring, funding, and recognition
  4. Evaluation: results communicated through press releases and blog posts bypass peer review and community scrutiny
  5. Autonomy: research questions may be prioritized for amenability to automated mathematics rather than expert judgment of deeper significance

Four sets of recommendations

Audience Core asks
Individual mathematicians Disclose tool use, support reviewing, adhere to open science, retain responsibility for correctness, put effort into proper attribution
Mathematical organizations Lead publishing and reviewing policy, maintain standards of rigor, protect authors' rights, insist on peer-reviewed outlets
Policymakers Strengthen author protections, do not believe the hype, regulate the AI industry, invest in public computational infrastructure
Commercial AI labs Meet the standards mathematicians expect of colleagues; collaborations must respect freedom of conscience

Featured endorsement: Peter Scholze (verified, quoted from the official site)

"This is a wonderful declaration, coming at the right time. The goal of mathematical research is human understanding of mathematics, and so mathematics can only thrive in a community of human mathematicians... Just like I do not want my children to be educated by AI, I am pondering my mathematical ideas without use of AI, and generally avoid reading AI-generated text as best I can."

Institutional endorsements include the IMU ("Mathematics is, and should always remain, a profoundly human endeavour"), ICARM at Carnegie Mellon, the Mathlib Initiative, and the Royal Dutch Mathematical Society.

6. The 25 Fields Medalists' Joint Statement

On September 11, 2026, 25 Fields Medalists - including Terence Tao and Yu Deng - signed a joint statement described in Chinese media as warning of AI's "serious misalignment" in mathematics, directed at practices by OpenAI, Anthropic, and other labs. (Full English text not retrieved; verified through secondary reports only - treat details as tentative.)

Tao's position is nuanced: he publicly uses Claude Code and Lean for formalization work while also signing statements that press the industry on transparency. Scholze, by contrast, avoids AI entirely.

7. The Split Inside the Community

Camp Representative Position
Engaged users Terence Tao Uses Claude Code with Lean for formal proofs; sees value in the tooling
Opposed Peter Scholze Avoids AI-generated text; math ideas must be nurtured by humans
Institutional pushback Leiden working group Five threats to reliability, attribution, incentives, evaluation, autonomy
Public warning 25 Fields Medalists Joint statement against industry practices (details tentative)
Defensive industry OpenAI Calls manuscripts "raw research data"; defers credit to the math community
Verification-first industry Anthropic Publishes machine-checked formalizations rather than raw claims
Partner-framing Google "When AI becomes a research partner"; multi-agent milestone audits

8. Trend Assessment

  • Short term (2026-2027): formalization becomes the default quality gate for AI mathematical claims; multi-agent verification spreads; organizations codify disclosure rules along the Leiden lines
  • Medium term (2027-2028): proof verification itself gets automated and scaled; the bottleneck shifts from generation to human understanding; hybrid human-AI research modes mature
  • Long term (2028+): a possible paradigm shift from "human proofs" to "machine proofs plus human understanding", alongside new subfields studying what AI-verifiable mathematics means

9. Conclusion

The core contradiction of 2026:

  • AI can generate hundreds of plausible proofs quickly
  • Humans need years to understand and verify them
  • Formalization is the necessary quality gate, but it is expensive and incomplete (OpenAI: ~42% of top-line results; the other ~58% still rely on human readers)
  • The community fears that publicity timelines are outrunning peer review

Which is exactly what the Leiden Declaration says in plain terms: press releases cannot replace peer review, and correctness claims that cannot be independently verified undermine the discipline.

The final question: if AI can prove everything but no human understands anything, is it still human mathematics?

Sources and Verification

Content Status Source
OpenAI 721 manuscripts, 373 families, ~42% formalized, 10 reasoning traces Verified GitHub openai/math README, fetched directly
Anthropic formal-math three Lean 4 projects, Palomar verification chain Verified GitHub anthropics/formal-math README, fetched directly
AlphaProof Code announcement Partially verified DeepMind blog index title (post page 404 during fetch)
Antigravity /teamwork architecture and roles Verified antigravity.google/docs/teamwork, fetched directly
FVCOM Rust migration (174k lines, 59 minutes, 1.88x) Verified antigravity.google blog post, fetched directly
Leiden Declaration full text, five threats, recommendations, endorsements Verified leidendeclaration.ai, fetched directly
Scholze quote Verified Featured endorsements on leidendeclaration.ai
25 Fields Medalists joint statement Partially verified Secondary Chinese reports only; original text not retrieved
Trend assessment Author analysis Interpretation of current developments; not investment advice

This article synthesizes primary sources from OpenAI, Anthropic, Google, and the Leiden Declaration project. Industry analysis only; not investment advice.

◆ 本文内容仅代表作者观点,不构成投资或采购建议。
分享这篇文章 邮件