Solving (some) formal math olympiad problems
OpenAI
OpenAI built an AI that proves formal math theorems in Lean, and it just cracked real olympiad problems. That includes AMC12 and AIME questions, plus two adapted from the IMO — the toughest math contest on Earth.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI's latest research project isn't a chatbot or an image generator. It's a theorem prover, a system trained to write formal proofs in Lean, the proof assistant favored by mathematicians who want machines to check their logic line by line. And the headline result is that this thing can now solve problems pulled from the AMC12 and AIME, two of the harder standardized math competitions American high schoolers take, along with two problems adapted from the International Mathematical Olympiad.
That last part matters more than it might sound. The IMO is the contest that produces the kids who go on to win Fields Medals. Its problems aren't about grinding through calculation; they demand genuine insight, the kind of lateral thinking that's historically been considered a poor fit for machines trained on pattern matching. Getting a model to handle even adapted IMO-style problems, inside the unforgiving formalism of Lean where every step has to be verifiably correct, is a different kind of milestone than acing a benchmark quiz.
Formal proof systems like Lean are brutal training grounds. There's no partial credit, no fuzzy natural-language reasoning to lean on. Either the proof compiles and the logic holds, or it doesn't exist at all. Search spaces explode fast, and most automated provers historically choked on anything beyond textbook exercises. That OpenAI's system found valid Lean proofs for genuine competition problems suggests the model learned something closer to mathematical strategy than brute-force pattern completion.
The practical upshot isn't that AI is about to replace mathematicians. It's narrower and, honestly, more useful: a tool that can search for and verify proofs at a level competitive with strong high-school competitors, in a language machines can check without human review. That's the kind of foundation that eventually feeds into bigger ambitions, formal verification of software, automated theorem discovery, maybe one day tackling open conjectures. Small, verifiable steps, not a giant leap.
My take — AI-written commentary, not fact-checked reporting
I like this kind of research precisely because it's unglamorous and rigorous, no cherry-picked demo, just cold formal verification that either compiles or it doesn't. The industry loves shouting about AGI while ignoring the boring plumbing work like this that actually builds trust in what these systems can do. If AI math ever gets good enough to help mathematicians rather than just impress them, it'll look exactly like this, one painstakingly verified proof at a time.
Read more about this at: OpenAI
