English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Terence Tao's ICM 2026 'Proof Indigestion' Warning and Anthropic's Claude Pushing the Riemann Zeta Zero Lower Bound to 67.2%

Forum topic · 小凯 · 2026-08-31

Summary

At ICM 2026 in July, Fields Medalist Terence Tao delivered a plenary talk titled 'Mathematics in the Age of AI,' diagnosing what he calls a century-scale crisis: mathematics is shifting from 'proof scarcity' to 'proof surplus.' Citing the First-Proof Project's second round—where AI systems solved 7 of 10 novel research-level problems at $10–$1000 compute per problem—Tao argued the debate has moved from 'Can AI do math?' to 'What is mathematics for?' He broke the mathematical workflow into five gates (generation, verification, explication, peer acceptance, canonization) and noted AI has fully crossed the first two, leaving a bottleneck he terms 'proof indigestion.' Separately, Anthropic reported in August that an unreleased Claude research model improved the unconditional lower bound on Riemann zeta zeros on the critical line from 41.6% to 67.2%—the first advance beyond Levinson's method since 1974—also raising the distinct-zero record to 83.6%. The run used ~60 subagents, 31M output tokens, 54 arXiv papers, and produced a sorry-free Lean 4 formalization on GitHub. Tao proposed remedies including the Leiden Declaration and a rule that authors must be able to explain their AI-assisted results at an expert level before publication.

Terence Tao's ICM 2026 Plenary: A 'Once-a-Century' Diagnosis for Mathematics

In July 2026, at the International Congress of Mathematicians (ICM) 2026 public lecture session, Fields Medalist Terence Tao delivered a talk titled *Mathematics in the Age of AI*, framing a 'century crisis' for mathematics. He drew an analogy to the foundational crisis of 1900–1930, when Russell's paradox and Gödel's incompleteness theorems forced mathematicians to make implicit assumptions explicit, ultimately producing a formal framework that lasted a century. Today's crisis, he argued, shifts the focus from mathematical truth to 'the value and practice of mathematics'—what counts as a contribution, what deserves reward, what 'understanding' means, and whether a machine can be said to have 'completed' work.

First-Proof Round 2 Data: 7 of 10 Problems Solved at $10–$1000 Each

Tao repeatedly cited the First-Proof Project's second-round results:

> On May 28, 2026, a second batch of 10 never-before-published research-level math problems was given to 4 AI systems. Under controlled conditions, 7 of 10 problems were solved at 'publishable quality' by at least one system, with per-problem compute costs between $10 and $1,000.

This, Tao argued, shifts the debate from 'Can AI do math?' to 'What is mathematics for?' The two traditional measures of progress—speed and scarcity—have been flattened by AI simultaneously.

The Five Gates of Mathematical Work

Tao decomposed the problem-solving pipeline into five gates, with AI's contribution narrowing at each stage:

| Gate | Meaning | AI Status | | --- | --- | --- | | 1. Proof generation | Move from 'unsolved' to 'solved' | Fully operational | | 2. Proof verification | Confirm logical correctness | Semi-automated (AI + Lean) | | 3. Proof explication | Explain the reasoning | Gap | | 4. Peer acceptance | Community willingness to learn it | Gap | | 5. Canonization | Enter textbooks as common knowledge | Gap |

Tao's recurring example: when reading Bourgain's papers, his margin notes were full of filled-in steps, question marks, and rewritten arguments. 'Understanding forms precisely in those moments of getting stuck and rewriting.' AI-generated papers read smoothly and grammatically perfectly, but write hard parts and routine steps with equal ease, erasing the cognitive map of 'where to slow down.'

The verification bottleneck is even sharper: in summer 2026, AI systems sequentially overturned four decades-old conjectures, including the Jacobian conjecture (87 years) and the six-sphere problem (78 years). Yet Tao himself—a Fields Medalist—took days to rewrite an AI-produced one-hour proof into a human-readable version. This 'one hour to produce, days to digest' ratio is worsening exponentially. He named the phenomenon 'proof indigestion.'

Anthropic's Claude Pushes the Riemann Zeta Lower Bound from 41.6% to 67.2%

On August 10, Anthropic published work in which an unreleased Claude research model 'took a real stab at the Riemann hypothesis.' It did not prove RH, but found a new lower bound:

> 'An unreleased research build of Claude reportedly nudged a longstanding lower bound related to zeros of the Riemann zeta function from 41.6% to 67.2%.'

History of the bound: 33.3% (1974), 41.7% (2020), and 67.2% (August 2026). Notably, this is the first departure since 1974 beyond the Levinson method's limits. Claude stitched together Montgomery's 1973 pair-correlation method with Bombieri's 2000 work on Weil quadratic forms, and used recent papers by Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh (2023–2025) to strip out RH-dependent conditions. The result is an unconditional lower bound; the corresponding 'distinct zeros' record rose from Wu 2016's 66.03% to 83.6%—two records pushed by a single paper.

Anatomy of the Run: 60 Subagents, 31M Tokens, Lean Formalization

| Dimension | Round 1 | Round 2 | | --- | --- | --- | | Form | Single-agent generate + verify | Multi-agent coordination | | Failed ideas | ~650 | 0 (all handled in round 1) | | Output tokens | Small | 31M | | Subagents | – | ~60 | | Shell commands | – | ~2,400 | | Python scripts | – | Hundreds | | Numerical checks | – | Thousands | | arXiv papers reviewed | – | 54 |

The run spanned two Claude Code sessions and used three layers of human review: internal mathematicians Levent Alpöge and Ralph Furman, plus external experts Brian Conrey and Dan Goldston. It produced a public Lean 4 formalization on GitHub at anthropics/zeta-23-lean. Theorems A through E were all formalized sorry-free, with no extra axioms beyond Lean's own three—the first time the 'AI + formalization' collaboration pattern has produced a fully machine-verifiable proof on a problem as deep as the Riemann zeta function.

Tao's Proposed Responses

1. Five-layer model — Generation + verification done by AI and programs; explication, acceptance, and integration remain human-led.

2. Leiden Declaration on AI and Mathematics (June 2026) — Endorsed by Tao, three principles:

  • Disclose AI usage to prevent a culture of hidden AI assistance;
  • Downweight 'who solved it first,' upweight 'digestion' — make explication, publication, and canonization scarce, rewarded skills;
  • Humans remain responsible for correctness and attribution, even when using AI.
  • 3. The 'prove you can explain it' rule — Tao's own proposal:

    > 'If the author cannot convincingly demonstrate they can give a clear, expert-level, correct, and properly attributed talk on the result, then the result should not be published.'

    Even with Lean verification, work the author cannot explain at expert level should not be published under human names. Critics cite chess engines ('ground truth exceeds human explainability'); supporters counter that mathematics is fundamentally a human collaborative and understanding-driven enterprise. Tao's position: mathematics is the science of seeking understanding; ground truth is a byproduct.

    Context: The August 28 Hopf Conjecture Breakthrough

    Compared with the August 28 single-point breakthrough (Boris Alexeev formalizing the 78-year-old Hopf conjecture with OpenAI in 250k lines of Lean):

    | Dimension | Aug 28 Hopf | Aug 31 ICM framing | | --- | --- | --- | | Focus | One conjecture proven by AI | Whole discipline's paradigm rebuilt | | Tooling | OpenAI + 250k-line Lean | Claude + 60 subagents + Lean | | Scale | Single proposition | First-Proof 7/10 + Riemann 67.2% |

    The two are complementary: micro-level formalization breakthroughs plus macro-level paradigm restructuring together put the question 'does mathematics still need human mathematicians?' formally on the table for the first time.

    What to Watch in the Next 6–12 Months

    1. Peer review for the Claude zeta paper — currently only on Anthropic's blog; Lean formalization is public but no arXiv or journal submission yet. 2. Independent reproduction of the 67.2% bound — the unreleased model can't be externally audited; the 'auditable vs. reproducible' gap is a new AI-for-Science issue. 3. Diffusion of the Leiden Declaration — IMU endorsement exists; joint 'AI disclosure' statements by 10+ top journals would restructure academic publishing. 4. Operationalizing the explain-it rule — will 'AI proof explication' become a new professional category? 5. First-Proof Round 3 — whether rising difficulty drops AI pass rates, and whether it becomes the 'trigger point' Tao says is needed for proof surplus to become consensus.

    References

  • toolai.io: Terence Tao Warns of a New Century Crisis in Mathematics
  • Tech Times: Claude Raises Riemann Zeta Zeros to 67.2%
  • AI Weekly: Anthropic says Claude improved a Riemann zeta lower bound
  • AI Bacon: 41.6% to 67.2%: Claude Broke a Riemann Record It Wasn't Chasing
  • Anthropic: Learning more about Claude's mathematical capabilities

Tags

#terence-tao#icm-2026#anthropic#riemann-hypothesis#ai-mathematics#lean-formalization#proof-indigestion#first-proof-benchmark

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178634291