Plectis Atlas
Search

↑ and ↓ move, Enter opens, Esc closes.

372 families in 17 disciplines. Point at a dot to see its family; choose it to open the family. Choose a discipline’s name to see all its families.

Drawing the map…

Every family, by discipline

Each family is a group of related papers in OpenAI’s catalogue. Open its page to read those papers and the reports about them.

Number theory 31 families, 17 with Lean

Algebraic and complex geometry 36 families, 8 with Lean

Real and complex analysis 16 families, 10 with Lean

Convex and metric geometry 15 families, 13 with Lean

Theoretical computer science 40 families, 34 with Lean

Dynamical systems and ergodic theory 12 families, 9 with Lean

Combinatorics 37 families, 33 with Lean

Algebra 18 families, 10 with Lean

Probability and statistical mechanics 29 families, 19 with Lean

Mathematical logic 6 families, all with Lean

Group theory 14 families, 12 with Lean

Mathematical physics 25 families, 17 with Lean

Operator algebras 19 families, 15 with Lean

Topology 18 families, 3 with Lean

Functional analysis 11 families, 10 with Lean

Differential geometry 29 families, 15 with Lean

Partial differential equations 16 families, 11 with Lean

What OpenAI published

OpenAI labels 189 of the 416 Lean statement files as main results in its Lean catalogue.

The papers cite one another in 586 directed pairs and cite 11,019 works outside the release. OpenAI has withdrawn 3 papers and replaced 27 with later editions.

Atlas links public contributions credited to 371 names and places 671 reports beside the results they discuss.

Who examined the release, and how

These groups report different kinds of work: verifying a formal proof with a program, comparing a Lean statement with a paper, or testing a finite claim by calculation. Atlas links their original reports and has not repeated their work.

  • Carlos Toledo, cert-machine ran Comparator on OpenAI’s proofs with a second independent verifier, tested finite claims in some papers by exact calculation, and compared the Lean statements with OpenAI’s catalogue descriptions.
  • MathVet ran Comparator on OpenAI’s proofs in an isolated environment and compared the Lean statements with OpenAI’s catalogue descriptions under published assessment rules.
  • Neal McBurnett, openai-math-lean-tooling ran Comparator on OpenAI’s proofs using Lean’s own proof verifier and recorded the time and memory used.
  • Ben Knill, openai-math-hol-cards proved selected intermediate steps using HOL Light, a different proof assistant, and tested incorrect examples that the program should reject.
  • Thomas Bloom, erdosproblems.com identifies which Erdős problems the papers address. These links do not assess whether the claimed proofs are correct.

How to help

Each contribution covers one result or one page, and is credited by name on the pages it changes.

Without an AI account

Open a short GitHub form to explain a result, compare a Lean statement with its paper, or report a problem with a page.

Ways to contribute

With an AI account

Give your assistant the start-here guide from the repository. It helps you choose one useful note, drafts it privately and leaves the signing to you.

The start-here guide