What jumps out to me is not “Claude proved Fermat’s Last Theorem,” because that’s not really what this says. The theorem was already proved by Wiles decades ago. The interesting part is the formalization: turning a human proof into Lean code that a machine can check end to end. That’s a very different task, and in some ways a nastier one.
And honestly, 13 million lines is the part that makes me squint. Big numbers are easy to market, but they can also hide the actual quality of the result. If a formal proof grows to more than five times Mathlib, my first question is not “wow, how smart is the model?” It’s “how much of this is elegant formalization, and how much is just brute-force scaffolding?” The article suggests the proof is multi-layered and built from lots of sub-results, which is plausible. But “13 million lines” feels like the kind of metric that sounds impressive while saying very little about whether this is a repeatable engineering breakthrough or a spectacular one-off.
The other thing I’d be careful about is the “11 days” framing. Continuous, largely unsupervised work is real progress, but it’s also not the same as a model cleanly doing a research task in a human sense. You still have an open-source tool in the loop, high-level guidance, retries, and the usual hidden human labor that always sits behind these demos. I’m not dismissing it — I think this is genuinely a milestone — but the article reads a little too eager to collapse “formalized a theorem” into “AI can now do serious mathematics.” That leap is not earned yet.
Still, if I were building with Claude, I’d pay attention to the workflow lesson more than the headline. Prove2Me, the repeated attempts, the theorem-by-theorem accumulation, the ability to keep going for days: that’s the interesting bit. It suggests these models are becoming useful as long-horizon search-and-verification engines, not just chatbots that occasionally stumble into a clever derivation. That’s a much more believable direction than magical theorem-solving.
The part I’d actually want to see next is simpler and stricter: can this approach formalize smaller, messier math results without ballooning into absurd code size? If the answer is yes, then this starts to look like a real tool for mathematicians. If not, it’s still a striking demo, but one that may be too expensive and too unwieldy to matter outside Anthropic’s own showcase.
Reference: A famous 129-page proof became 13 million lines of code — thanks to Claude
