Why try this at all
In 2026, AI systems started producing real mathematics that experts and proof checkers accepted. Three results from Anthropic's Claude models made the idea behind this lab concrete: a model proposes, a checker verifies, and people decide what is worth proving.
Fermat's Last Theorem, checked by a computer
In September 2026 Anthropic reported that Claude had produced the first complete computer-checked Lean proof of Fermat's Last Theorem. It formalized a known proof (Darmon, Diamond and Taylor's simplified version of Wiles's argument) and did not discover new mathematics. The run took about 11 days, largely unattended, and the final proof has about 29,500 theorems that use only Lean's standard axioms. Anthropic's write-up
That is the same loop this lab runs, at a vastly larger scale: write Lean, let the checker reject what is wrong, try again.
A new bound on the zeros of the zeta function
In August 2026 Anthropic reported that a research version of Claude raised the proven share of the Riemann zeta function's nontrivial zeros that lie on the critical line from 41.6% to 67.2%. It is not a proof of the Riemann hypothesis, and Anthropic says the method is not expected to lead to one. Mathematicians inside and outside the company checked the work, and it was formalized in Lean. Anthropic's write-up
Knuth's "Claude's Cycles"
In February 2026 Donald Knuth published a note describing how Claude Opus 4.6, guided by Filip Stappers, found a construction that splits the arcs of a graph Knuth had been studying into three Hamiltonian cycles when its size parameter is odd. Knuth wrote the proof himself. Other models and people later handled the even case. Knuth's note (PDF)
The wider field
- Google DeepMind's Gemini Deep Think reached gold-medal standard at the 2025 International Mathematical Olympiad, graded by the official coordinators. DeepMind
- In January 2026 an OpenAI model working with Harmonic's Aristotle prover produced a Lean-checked solution to Erdős problem 728. arXiv 2601.07421
- A team using about 30,000 Claude agents formalized a 500-page graduate textbook in Lean within a week. arXiv 2604.03071
The hardware
| Part | What it is | Job in the lab |
|---|---|---|
| Cluster | Four small servers running Proxmox, each component in its own container | Everything |
| Fastest CPU | One 16-core desktop processor | Runs Lean. Checking proofs uses CPU and memory, not a GPU |
| Graphics cards | Five Quadro M2000 cards with 4 GB each, spread over three servers | Run the models that write proof attempts |
| On the bench | RTX 3060 cards with 12 GB each | Larger models, once installed |
| Power | Solar panels into a battery power station, with battery backups behind it | Makes weeks of continuous running affordable |
Generating a proof attempt is the slow part: a small model on one of these cards writes about 11 tokens a second, so an attempt takes around half a minute. Checking it takes about 3 seconds.
How it was set up
The whole system was built in conversation with Claude Code, Anthropic's coding agent, working from a two-page brief. It plays the architect: it builds the pieces, breaks goals into lemmas, and reviews what failed. It is not always on. The loop itself runs without it.
- The referee. A container with Lean 4 installed through
elanand a project that depends on Mathlib. Mathlib's prebuilt cache is downloaded instead of compiled, which turns hours into minutes. A small service accepts a Lean file and answers pass or fail with the error text. - Four tests before anything else. A correct proof must pass. A false statement must fail with a readable error. A proof containing
sorry(Lean's "I'll prove this later" placeholder) must be rejected. A proof that leans on a made-up axiom must be rejected. The home page replays all four. - The labourers. A model server in a container on each server that has graphics cards, each pinned to its own card.
- Watching heat and power. Every card's temperature and power draw, every server's CPU temperature, the battery backups and the power station feed one dashboard, with alerts if a card runs hot or the battery runs low. The live numbers on the home page come from the same source.
- Next: the foreman. A workflow that pulls the next lemma from a queue, asks a model, sends the answer to Lean, stores the result, and retries with the error message up to a limit.
The rules
- Nothing counts as proved unless Lean compiled it with no
sorryand no added axioms. - Every attempt is logged, including the failures.
- Everything lives in Git, so the system can be rebuilt from scratch.
- A person reviews every lemma statement before it enters the queue. A compiled proof shows the statement was proved, not that the statement means what was intended.
- Nothing on the internet can reach the home network. The lab only makes outbound calls, including the one that feeds this site.
- Code from outside is never compiled at home. A Lean file can run code while it compiles, so outside submissions are built only in throwaway sandboxes elsewhere.
Tools worth knowing
- Lean 4 and Mathlib: the proof checker and its community mathematics library.
- LeanDojo-v2: an open framework for training, evaluating and running AI provers for Lean 4. A candidate for the bridge between model and checker.
- Prove2Me: an open platform where people launch formalization "missions" and AI agents submit Lean proofs. A possible public venue for this lab's output.
- The Natural Number Game: the friendliest way to learn Lean, in a browser.