An AI touched ten open math problems — and a Fields Medalist nodded
OpenAI says its internal Astra model solved or advanced ten open problems, publishing formal Lean proofs anyone can verify
The three lines
- OpenAI's internal model Astra claims to have solved or advanced ten open problems
- Results include a non-sofic group construction and sphere-packing bounds — Lean proofs on GitHub
- Fields Medalist Timothy Gowers: would recommend one proof to a top journal 'without hesitation'
Key questions
- What makes this different from AI acing math tests
- The problems had no known answers. OpenAI says its internal model Astra solved or substantially advanced ten open problems in mathematics and theoretical computer science — including proving the existence of non-sofic groups, a long-standing group-theory question, and new sphere-packing bounds. That is knowledge production, not knowledge retrieval.
- Why should the claims be believed
- Because of the verification stack. Part of the results were written as formal proofs in Lean — code a computer checks step by step, making plausible-but-wrong reasoning structurally impossible — and published on GitHub for anyone to verify. On top of that, Fields Medalist Timothy Gowers said he would recommend one of the proofs to a top journal without hesitation.
- What does it mean beyond mathematics
- It marks a change in what AI capability is. Mathematics became the first arena for machine-generated new knowledge precisely because its verification is airtight. The heavier question is when this crosses into science and engineering, where answers cannot be machine-checked as cleanly — that is the real weight of this announcement.
The stock sentence of AI news has been "it scored X on a benchmark." This week the sentence changed: "it solved problems nobody had solved." OpenAI announced that an internal model called Astra solved or substantially advanced ten open problems in mathematics and theoretical computer science. On its face, that is a self-serving claim like any other. Two things make this one different — machine-checked proofs were published, and a Fields Medalist nodded.
1. What was claimed — unsolved, not untested
The headline result is a construction proving the existence of non-sofic groups — a question in group theory that stood open for decades as "probably true, proven by no one." The list also includes improved sphere-packing bounds in high dimensions, binary code limits, and arithmetic circuit complexity results.
The distinction that matters: AI's math record until now was measured on competition problems, where answers are known and graded. Open problems have no answer key. If the advances hold, this is not recombination of existing knowledge — it is production of new knowledge, in the one field where that claim can be checked to the letter.
2. The verification stack — why this claim travels
| Layer | What it is | Why it matters |
|---|---|---|
| Formal proofs (Lean) | Proofs as machine-checkable code, on GitHub | Plausible-but-wrong reasoning cannot pass |
| External authority | Gowers: top-journal quality "without hesitation" | Recognition beyond the press release |
| Open verification | Anyone can run the checker | Reproducibility without trust |
| Open caveat | Full list and solved/advanced split under review | "Solved" vs "progress" is for mathematicians |
| Model status | Astra is internal, unreleased | A capability demo, not a product |
The formal-proof layer is the load-bearing one. Proofs written for a proof assistant like Lean are verified mechanically, step by step — the one gate that AI's chronic confident-nonsense problem cannot slip through. How that machinery works is today's companion reference, "Formal proofs explained."
3. What remains
Two questions stay open. Whether academia certifies all ten — the line between "solved" and "meaningfully advanced" belongs to mathematicians, and review is underway. And when this capability crosses out of mathematics — into sciences where verification is messier — which is where the consequences stop being academic. The parallel AI race tracks run in "ChatGPT's free tier goes unlimited" and "Five models in eight weeks"; the verification machinery in "Formal proofs explained."