OpenAI Claims Progress on 372 Math Problems, Including the Unique Games Conjecture — Here's the Catch

OpenAI has published results covering 372 open mathematical problems, including a claimed proof of the Unique Games Conjecture (UGC), a long-standing question in theoretical computer science. Only some of these results come with Lean-verified certificates, meaning the rest rely on natural-language argument rather than machine-checked logic. Human mathematicians have barely begun the work of reading, understanding, and vetting what's been produced. The headline number is striking, but it obscures a slower, harder process still ahead.

What OpenAI Actually Released

The announcement lists 372 problems where an OpenAI model produced a solution or proof attempt. Among them is the Unique Games Conjecture, a conjecture from complexity theory that has shaped how researchers think about the limits of approximation algorithms since the late 1990s. A confirmed resolution — proof or disproof — would be a genuine landmark for the field.

But "372 results" is not a uniform category. Some of these come with Lean certificates — machine-checked formal proofs that guarantee logical consistency within the Lean proof assistant. Others are presented as natural-language arguments without that formal backing. A Lean certificate tells you the logic doesn't contain a gap a computer can detect; it does not tell you that a human has understood why the proof works, or whether its techniques generalize.

The Gap Between Formal Verification and Mathematical Understanding

This is the real story here, more than the raw count. Mathematics as a discipline doesn't stop at "no logical errors found." A proof becomes accepted knowledge when the community can explain why it works, identify its key ideas, and connect it to other open questions. That process takes time, peer scrutiny, and often reformulation — none of which a Lean certificate can shortcut.

For UGC specifically, that review has only just started. Complexity theorists will need to go through the claimed proof structure, check whether it matches known partial results and lower bounds, and see whether the techniques used actually resolve the conjecture as historically framed — or whether they prove something adjacent.

This echoes a pattern we covered in Inside the AI Math Wars, when OpenAI, Anthropic, and other labs announced progress on Millennium Prize-level problems. The friction back then wasn't really about whether the models produced correct math — it was about the mismatch between the pace of AI labs publishing results and the slower, consensus-driven pace at which mathematicians validate them. The 372-problem release looks like that same tension playing out at a larger scale.

Why It Matters Beyond the Headline Number

For researchers in theoretical computer science and combinatorial optimization, whether UGC actually falls matters directly — a large amount of hardness-of-approximation literature is built on assuming it's true. If the proof holds up, parts of that literature get a firmer foundation; if it doesn't survive scrutiny, the conjecture remains open and the claim becomes a cautionary tale about verification speed.

For the broader AI industry, the more durable signal isn't the count of problems "solved" — it's how fast formal verification tooling like Lean is becoming load-bearing infrastructure for trusting these claims at all. Given how quickly labs are generating proof attempts, expect more reliance on automated checkers rather than purely human review, simply because human review can't scale at the same rate.

The Takeaway

A confirmed proof of the Unique Games Conjecture would be one of the bigger results in theoretical computer science in years. Right now, what exists is a claim with partial formal backing and a mathematical community that hasn't finished checking it. That verification process — not the headline count — is the thing worth watching next.

Reference: https://scottaaronson.blog/?p=10169