Interview · 12 min read

The Proof Compiles or It Doesn't: On Astra, Lean Certificates, and What Changes When Math Gets Machine-Checked

Reconstructed from OpenAI's August 1 release and public statements by Sebastien Bubeck and Noam Brown: ten decade-open problems, Lean 4 certificates, and why machine-checkable proofs shift the verification burden.

By Classy AI News Staff — Research Desk · August 4, 2026

OpenAI's August 1 announcement landed with an unusual artifact attached: not just manuscripts, but Lean 4 certificates for ten problems in mathematics and theoretical computer science that had sat open for at least a decade. The work came from an internal evaluation of Astra, described as the company's next major model family — still unreleased — and the release included a 249-page manuscript collection, reasoning walkthroughs, and the openai/ten-proofs repository under Apache 2.0.

This piece is not a sit-down interview. It reconstructs the public record: OpenAI's blog post, the GitHub repository instructions, and posts on X from OpenAI research leadership. Where quotes appear, they are attributed to documented public statements.

What OpenAI actually shipped

On August 1, 2026, OpenAI published "Ten advances in mathematics and theoretical computer science", listing results spanning high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics.

The company said an internal Astra system generated the underlying mathematical arguments during development-time evaluation, estimating total token cost at roughly $2,000 at Sol API rates. Human researchers then prepared manuscripts and formalized each argument in Lean; OpenAI stated it takes responsibility for correctness while crediting the model for the arguments themselves.

Among the headline results:

  • An explicit construction of a non-sofic group, addressing a question open since Mikhail Gromov's 1999 formulation of soficity.
  • A disproof of Connes's rigidity conjecture on von Neumann algebras.
  • Progress on Ehrhart's volume conjecture.
  • Results touching Paul Erdős problems 146, 180, and 183, including multicolor Ramsey numbers.

Mathematical formulas written on a whiteboard

Why Lean changes the conversation

The GitHub repository documents a standard verification path: with elan installed, lake exe cache get followed by lake build All compiles all ten formalizations. Each result lives in a named module (NonSoficGroup.lean, ConnesRigidity.lean, and others). OpenAI reported a "sorry" count of zero across the certificates — meaning no step was left unproven in the formalization layer.

That distinction matters. Peer review evaluates whether a human-readable proof is correct and significant. Lean's kernel returns a binary verdict: the proof compiles, or it does not. Independent verification of the formal layer can happen in hours, not months — though mathematicians still must confirm that each formal statement corresponds to the intended open problem and judge whether the result advances the field.

OpenAI also noted alignment with the Leiden declaration on AI and Mathematics, emphasizing that claiming sole human authorship for AI-generated proofs would misrepresent how the results were produced.

Public voices from OpenAI research

Sebastien Bubeck, OpenAI's head of mathematics research, confirmed the results on X, calling them "beautiful," according to reporting from SiliconANGLE and Quartz.

Noam Brown, an OpenAI research scientist associated with test-time reasoning work, posted on X that the results represent "a major step for scientific reasoning," as cited in the same reports.

Neither researcher claimed the problems were "solved" in the peer-reviewed sense. OpenAI explicitly stated that none of the ten results had been through peer review at release. The Lean certificates address logical correctness of the formalization; the mathematical community still must engage with significance, novelty, and whether the formal statements match the historical problems.

Research papers and notes spread across a desk

What Astra is — and is not

OpenAI describes Astra as a model family built to run long tasks by coordinating multiple agents over extended periods. The ten-proofs release is an evaluation artifact, not a product launch. Astra itself remains unreleased.

The company released reasoning walkthrough PDFs narrating the model's thinking process for each solution, plus the full paper at cdn.openai.com/pdf/ten-proofs-oai.pdf. That transparency is double-edged: it invites scrutiny of both the mathematics and the formalization pipeline.

Community response on Hacker News and math forums has split along predictable lines. Enthusiasts point to the scale — ten unrelated problems, eight fields, all formalized — as unprecedented for AI-produced mathematics. Skeptics note vendor-reported results, the absence of peer review, and the need for independent mathematicians to validate that the Lean statements capture the intended problems.

The verification checklist going forward

For readers evaluating this release without taking sides, a practical checklist emerges from the public materials:

  1. Compile the certificates locally using the repository's Lake build instructions.
  2. Read the formal statements in each .lean module and compare them to the classical problem formulations in the manuscript.
  3. Wait for peer review and independent mathematical assessment of significance — compilation does not substitute for that step.
  4. Track whether Astra ships publicly with reproducible evaluation protocols for scientific reasoning tasks.

OpenAI framed the release as an invitation: "We hope the mathematical community will engage deeply with these results, place them in context, and bring the ideas behind them to life through new research and discovery."

Close-up of handwritten equations in a notebook

Why this desk cares

The Astra proofs sit at the intersection of two threads Classy has tracked all summer: test-time compute scaling for hard reasoning tasks, and formal verification as a trust layer when model outputs outpace human checking bandwidth. Lean certificates do not end the debate over AI and mathematics. They move one part of it — logical validity of the formalization — onto ground where a machine, not a reputation, delivers the first pass.

That is a different bar than a press release claiming "AGI solved math." It is also a higher bar than a PDF proof alone. The proof compiles, or it does not.

### Sources