Anthropic released the full 9,842‑line Lean script that reproduces Wiles’s proof, inviting the community to audit the AI‑generated proof.
ANTHROPIC'S AI FORMALIZES FERMAT'S LAST THEOREM, SHAKING MATHEMATICS TO ITS CORE
*Anthropic’s Claude‑3 model has turned a 350‑year‑old mathematical puzzle into a machine‑checked proof. The move threatens to upend how mathematicians validate results, while exposing the opacity of AI‑generated logic.*
Anthropic’s AI research team announced on Sept. 4 that its Claude‑3 model has produced a fully formalized proof of Fermat’s Last Theorem (FLT) in the Lean proof assistant. The claim follows a month‑long race with independent researchers who posted a similar effort on the Xena Project blog. Anthropic’s paper, posted on its research site, says the model generated the entire Lean script, checked it line‑by‑line, and passed the Lean verifier without human edits. Anthropic claims the proof required 2.3 million token calls and completed in under three hours on a single A100 GPU.
The achievement is more than a brag‑sheet for AI labs; it signals that large language models can navigate the rigor of formal mathematics, a domain traditionally guarded by specialist mathematicians. If AI can routinely translate informal arguments into machine‑checkable code, the bottleneck of proof verification could evaporate, reshaping peer review, grant funding, and the very tempo of mathematical discovery. Critics warn that the black‑box nature of the model leaves the logical chain opaque, demanding new standards for auditability.
Anthropic’s Breakthrough
Anthropic’s internal team, led by research scientist Dr. Maya Patel, fed the informal proof from Andrew Wiles into Claude‑3, prompting the model to decompose the argument into Lean’s type‑theoretic primitives. Within 12 iterations, the model produced a 9,842‑line Lean file that reproduced every lemma of Wiles’s original proof, including the modularity lifting theorems. The team logged 1,024 GPU‑hours, a fraction of the estimated 10,000 hours a human team would need for a comparable formalization. Anthropic released the full repository under an MIT license, inviting scrutiny from the Lean community. Early audits by the Lean prover’s core developers flagged zero verification errors, a rare clean pass for a proof of this scale.
The Technical Playbook
The pipeline hinges on Claude‑3’s chain‑of‑thought prompting, where the model is asked to “explain each step in formal language before coding it.” Researchers built a custom wrapper that translates natural‑language explanations into Lean syntax, then feeds the output back for self‑verification. The system leverages the “auto‑rewrite” feature of Lean 4, allowing the model to iteratively refine failed proof attempts. In total, 2.3 million tokens were consumed, a cost of roughly $5,800 at Anthropic’s published rates. The team also integrated a SAT‑solver to prune impossible lemma branches, cutting the search space by 87 %. The approach is reproducible: the codebase, prompts, and hyper‑parameters are all documented in the accompanying GitHub repo.
Community Reaction
Mathematicians on the Zulip Lean channel reacted with a mix of awe and caution. Dr. Elena García, a senior researcher at the Institute for Advanced Study, posted, “If a model can reproduce Wiles’s proof without error, we must rethink what constitutes a ‘human‑verified’ theorem.” Meanwhile, the formal methods community flagged the lack of a human‑readable audit trail as a security risk. The Xena Project, which posted a rival formalization two weeks earlier, issued a statement that Anthropic’s “black‑box” approach undermines the collaborative ethos of open mathematics. Social media saw a surge of #AIProofs memes, but also a coordinated petition on Change.org demanding that journals require explicit model provenance for any AI‑generated proof.
Implications for Society
Beyond academia, the proof raises policy questions about AI‑generated knowledge. Governments drafting AI governance frameworks now face a concrete example: an algorithm that can produce legally binding mathematical statements without human oversight. Insurance firms are already consulting legal teams on whether a formally verified AI proof could be used as evidence in patent disputes. Education sectors fear a future where students submit AI‑written proofs, forcing curricula to shift toward verification skills rather than derivation. At the same time, the speed of formalization could accelerate breakthroughs in cryptography, climate modeling, and drug design, where formal guarantees are critical. The trade‑off between rapid innovation and opaque methodology will shape funding priorities for the next decade.
Anthropic’s formalized FLT proof is a watershed moment that forces every stakeholder to confront a new reality: AI can now produce verifiable mathematics at scale. The immediate challenge is to embed transparency, provenance, and accountability into the proof pipeline before the technology outpaces the institutions meant to govern it. Failure to do so will hand unchecked logical power to a handful of corporations, eroding the public trust that underpins scientific progress. The next weeks will determine whether the community can harness this capability responsibly or watch it become a silent, unchallengeable arbiter of truth.
Sources: Anthropic research page, Xena Project blog, Hacker News discussion, Lean community Zulip channel, statements from Dr. Maya Patel and Dr. Elena García.