I use AI to make AI-written maths and code easier to understand.
I’m Will Cook. I do maths research and build software with AI. Most projects here can be followed at several levels of detail, from a short video down to the proof or the code.
I build the tools so they get better as AI models get better.

The projects
Twenty-one projects in five groups. Each has its own page. Fifteen open with a video, and the six newest open on the website or the code.
Three places to start
These three are my examples in the introduction, one each from maths, code and working with AI coding assistants.
All twenty-one, by group
1 Maths
Research on Erdős problems and on two of OpenAI’s results, as papers, websites and proofs in Lean, and explanations of Troy Lee’s Lean proof of the PCP theorem.
Lean is a program that checks mathematical proofs.
The funnel: How a Lean proof becomes papers and a video
Maths Map: Eight Erdős problems on one map
Erdős #257: A case from 1968, proved and checked in Lean
Quartic separation: OpenAI’s fourth-power result, proved again in LeanResult by OpenAI
Spin glass: Two nearly best answers, half the arrows differTheorem and Lean by OpenAI
PCP theorem: Troy Lee’s Lean proof, explained at three lengthsLean proof by Troy Lee
Reader websites: Click a linked sentence, open the Lean beside itLean Exposition, Quartic Reader, Proof Reader
2 What a check covers
What a proof, a report, a benchmark’s grader or a specification covers, and what it leaves out.
Atlas: A map of every result in OpenAI’s maths releasePapers and Lean by OpenAI
Verification map: What one proof covers, in a toolkit of 88 components
MedAgentBench study: A script that reads one page of results passes all 30 released casesBenchmark by Jiang and coauthors, Stanford
FaithformBench Reader: A wrong step adds a number. In two thirds of false positives, the Lean omits it.Benchmark by Cornish, Ghinassi and coauthors
Spec change map: What three corrected clauses changed in a login contractContract by PDD Kit’s authors
3 Code
Code turned into a map and into a database that an AI model can search, and a test of which descriptions of code go out of date.
4 AI coding assistants at work
Tools for working with AI coding assistants, and experiments on what they do.
One Checkout: Commit one fix while another AI keeps editing
Safety Core: A written rule let 6 of 10 AI runs commit a made-up value. A check let 0.
Edit View: An edit can remove text that its monitor is never shownTested on code Anthropic released
Agent Trace: Read what your AI coding assistant ran
Zenith: Showing an AI tester the bugs it missed
5 How I make these videos
The plan I write before AI draws any animation.
About
I’m looking for a job, internship, fellowship or funding, and I’m open to working with other people on their problems. If there’s a problem you think I should work on, please email me.
If something on these pages is unclear or wrong, please tell me. If your feedback leads to a change, I’d be happy to credit you, with your permission.