On August 1, OpenAI published ten new results in mathematics and theoretical computer science, all produced by an internal AI model S¹P⁵. The first pins down the exact strength of a decades-old method for packing spheres in high dimensions P⁵. Alongside the paper, OpenAI posted machine-checkable Lean proof certificates to GitHub, so anyone with the right software can verify the work without taking the company's word for it P⁷. Whether that is enough to convince mathematicians that an AI can do real research rather than pattern-match is the question this collection was built to answer.

My read: This is the first time I have seen a major lab pair AI-generated mathematical results with formal, machine-checkable Lean certificates on a public repo. The certificates are the whole story. A blog post saying "our model proved X" is just a claim. A Lean certificate is a file that a proof checker can run in seconds and return a yes or no. I am skeptical of the broader framing until independent mathematicians weigh in on whether these results are genuinely novel or known corollaries dressed up, but the verification infrastructure is the right move and the part worth watching.

One result pins down sphere packing in high dimensions

The collection spans geometry, cryptography, and complexity theory S¹. The result OpenAI describes in most detail in the abstract concerns high-dimensional sphere packing: the exact asymptotic strength of the Cohn-Elkies linear program, which yields an improved general packing bound in high dimensions P⁵.

Sphere packing asks how densely you can arrange identical spheres in space without overlap. In two and three dimensions, the answers are known and intuitive. In high dimensions, the problem becomes far harder and connects directly to coding theory and communications, where dense packings correspond to efficient error-correcting codes.

The remaining nine results are not detailed in the available material, but the abstract confirms they span the three named fields P⁵S¹.

The proof checker does not care who wrote the proof

OpenAI did more than publish a paper. It also released a GitHub repository called "openai/ten-proofs" containing Lean certificates for the proofs P⁷. Lean is a proof assistant, a programming language for writing mathematical proofs that a computer can verify with certainty. If a Lean certificate compiles, the proof is correct. There is no room for human error, hand-waving, or persuasive rhetoric.

The repository had attracted 135 stars within hours of its creation on August 1 P⁷. Its top contributor is listed as "balexeev-oai" P⁷, indicating an OpenAI researcher.

This matters because AI-generated proofs face a trust problem. When a language model claims to have solved a problem, you cannot tell whether it genuinely derived the result or produced something that looks right but contains a subtle error. Lean closes that gap. The proof checker does not care who wrote the proof or how. It checks the logic.

Benchmarks test known problems with known answers. OpenAI's collection attempts something harder: producing new results on open problems.

What to do about it

For a research group at a university or a cryptography lab, the practical step is simple. Clone the repository, install Lean, and run the proof checker on the certificates P⁷. If they compile, the proofs are formally correct. That is the floor. The ceiling is whether the results are mathematically interesting, which is a question for domain experts, not software.

Consider a postdoctoral researcher working on coding theory at a university. The sphere packing result, if novel and correct, could tighten bounds used in error-correcting codes. That researcher can now take the Lean certificate, verify it in minutes, and then read the proof to understand the technique. The verification is free. The understanding still costs human effort.

One thing to check this week: the openai/ten-proofs repository on GitHub P⁷. If you have Lean installed, try compiling the proofs. If you do not, read the README and the issue tracker to see whether other mathematicians have raised concerns.

What we don't know yet

The abstract names only one of the ten results in detail P⁵. The other nine are not described in the available material. We do not know which specific open problems were addressed, how hard they were, or whether the results are genuinely novel or refinements of existing work.

No independent academic verification is present. OpenAI's own blog and PDF are the only sources S¹P⁵. The Hacker News discussion, which drew roughly 97 points and 79 comments S⁴, is a signal of interest, not of correctness.

The Lean certificates verify logical correctness, but they do not verify mathematical significance or novelty. A proof can be correct and uninteresting. Whether these results have been submitted to peer-reviewed journals or conferences is unknown.

The next signal: whether independent mathematicians in sphere packing, cryptography, or complexity theory post reactions or refutations in the coming weeks. We will check the openai/ten-proofs issue tracker and arXiv for follow-up papers. If you want that follow-up in your inbox, subscribe and we will send it.

Sources: S1 — Ten advances in mathematics and theoretical computer science · S2 — Ten advances in mathematics and theoretical computer science - openai. · S3 — Ten advances in mathematics and theoretical computer science · S4 — Ten advances in mathematics and theoretical computer science · P5 — Ten Advances in Mathematics and Theoretical Computer Science · P6 — malkreide/hn-tech-signal-mcp · P7 — openai/ten-proofs · P8 — Abdrakib/ml-research-assistant

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.