OpenAI just published 722 AI-generated math papers solving 90 of the top 500 open problems — here's what the Lean formalization dump means

OpenAI published 372 AI-generated mathematical results on GitHub yesterday. Each proof consumed roughly three hours of ChatGPT Pro compute. All 372 include Lean formalizations, which means they're machine-verifiable — you can run the proof through Lean's type checker and confirm it's valid.

The repository claims 90 of these results solve problems from the top 500 open problems in mathematics. One of those is allegedly Barnette's Conjecture, a graph theory problem that's been open since 1969. Jake Boggan, who spent 24 years working on that conjecture on and off, found his life's work in problem 180 of the OpenAI dump. He posted on HN: "I don't know what to think exactly."

Twenty-five Fields Medal winners signed an open letter warning that mass-producing mathematical truths could destroy the field rather than advance it. Their argument: mathematics isn't just about cranking out true statements. The process of working through a hard problem — the dead ends, the insights, the connections to other domains — is where the real value lives. Dumping 372 proofs into the commons short-circuits that.

What the Lean formalization actually guarantees

Lean is a proof assistant and programming language. When you write a proof in Lean, the type checker verifies every step. If it compiles, the proof is correct by construction. This is different from a traditional math paper, where correctness depends on peer review and human judgment.

OpenAI's 372 proofs all ship with Lean code. That means anyone can clone the repo, install Lean 4, and run the type checker to confirm the result. No trust required.

But Lean formalizations don't answer the "so what" question. A proof might be technically valid and also completely uninteresting — a minor variation on a known result, or a brute-force argument that doesn't generalize. The Lean code tells you the proof is sound. It doesn't tell you whether the theorem matters.

The Fields medalists' concern is that flooding the field with machine-generated results buries the signal. If OpenAI can generate 372 proofs in a few weeks, what happens when that number is 3,720? Or 37,200? The community's attention is finite. Sorting valuable results from derivative ones already takes work; scaling that filter to handle LLM output at scale is a different problem.

The compute budget tells you what OpenAI is optimizing for

Three hours of ChatGPT Pro compute per proof works out to roughly $150–$180 per result, assuming Pro tier pricing (~$50/hour of heavy o3 usage). Multiply that by 372 proofs and you're looking at $55k–$67k in compute spend for the entire dump.

That's a rounding error for OpenAI. The real cost is researcher time — building the search infrastructure, curating the problem list, validating the Lean formalizations, writing the repo documentation. If this project took a small team two months, the fully-loaded cost is probably $500k–$1M.

What does that buy? A proof-of-concept that o3 (or whatever internal model ran this) can formalize non-trivial math at scale. The Github repo is a credential play: "We can do this, and here's the machine-checkable evidence." It's also a recruiting signal — mathematicians who care about formal verification now have a dataset to benchmark against.

But it's not clear OpenAI expects the math community to use these results directly. The repo has minimal documentation. There's no narrative explanation of which problems are significant or why. It reads like a data dump, not a contribution to the literature.

What this means for formal verification adoption

Lean adoption in pure mathematics has been growing slowly. Major results like the liquid tensor experiment got formalized in Lean, and that work took years of collaboration between mathematicians and proof engineers. The Lean community has been careful about building social norms — formalizing a result is a collaboration, not a replacement for human insight.

OpenAI's dump doesn't follow those norms. It's automated formalization at scale, with no human mathematician signing off on which problems are worth solving or how the proofs connect to the broader field.

If this becomes the standard pattern — LLMs churning out Lean code faster than mathematicians can evaluate it — formal verification stops being a tool for collaboration and becomes a credibility arms race. Labs compete on proof throughput. The bottleneck shifts from "can we formalize this?" to "should we care about this?"

The Fields medalists are worried about exactly that shift. Their letter argues that mathematics needs human judgment at the problem selection stage, not just at the verification stage. Lean can tell you a proof is correct. It can't tell you whether the theorem advances understanding.

The question is whether the math community can build evaluation infrastructure that scales with LLM output. If not, the 372-proof dump is just the beginning. OpenAI has shown the capability exists. The next move is figuring out what to do with it.