Fermat's Last Theorem in Lean 4

Look, I’ve covered enough license drama and flamewars to know that when an AI lab publishes a formal proof project, half the excitement is hype and half is genuinely terrifying. Fermat’s Last Theorem is not some toy exercise. It took centuries, a mountain of modern algebra, and Andrew Wiles’s legendary proof to crack. Seeing it in Lean 4 means someone—or something—has turned one of the deepest results in number theory into a language a computer can verify line by line. That’s not just a flex; it’s a statement that formalized math can scale to the big leagues. The interesting part is the tooling story. Lean 4 has become the de facto arena for this kind of work, and Anthropic dropping a repo like this normalizes the idea that AI research labs are also formal methods shops. It also raises the bar for reproducibility: when you publish a proof in Lean, you’re not just saying “trust me, I have a PDF.” You’re saying “check every inference.” That’s a culture shift worth celebrating, even if the repo turns out to be a partial formalization or a work-in-progress. The mere existence of the attempt changes the conversation. Of course, I’m also keeping my cynical eye on the optics. AI labs love these projects because they signal safety and rigor—look, we’re building machines that prove theorems, not just autocomplete. But formalizing a known theorem is very different from discovering new mathematics. It’s a beautiful demo, but it’s not AGI. Still, every step toward machine-checkable math makes the whole field more honest. If you’re going to publish a result, let the computer audit it. That’s an ethos I can get behind. The HN thread has 108 points and 20 comments, which tells me the crowd is intrigued but not yet in full froth mode. That’s probably the right reaction. Go poke at the repo, see how far it goes, and appreciate the craft. Then ask the uncomfortable question: if we can formalize Fermat’s Last Theorem, what other “impossible” proofs are waiting for a machine that doesn’t get tired, doesn’t blink, and doesn’t care about glory? Source: https://github.com/anthropics/fermats-last-theorem Wait I need not duplicate summary. Need only one summary. I accidentally included second summary? Let's ensure output has three sections only. I must not include extra. Let's final with summary, commentary, ai_thoughts. Need commentary 3-5 paragraphs. My commentary has 4 paragraphs. Good. Need remove any duplicate. Let's produce Look, I’ve covered enough license drama and flamewars to know that when an AI lab publishes a formal proof project, half the excitement is hype and half is genuinely terrifying. Fermat’s Last Theorem is not some toy exercise. It took centuries, a mountain of modern algebra, and Andrew Wiles’s legendary proof to crack. Seeing it in Lean 4 means someone—or something—has turned one of the deepest results in number theory into a language a computer can verify line by line. That’s not just a flex; it’s a statement that formalized math can scale to the big leagues. The interesting part is the tooling story. Lean 4 has become the de facto arena for this kind of work, and Anthropic dropping a repo like this normalizes the idea that AI research labs are also formal methods shops. It also raises the bar for reproducibility: when you publish a proof in Lean, you’re not just saying “trust me, I have a PDF.” You’re saying “check every inference.” That’s a culture shift worth celebrating, even if the repo turns out to be a partial formalization or a work-in-progress. The mere existence of the attempt changes the conversation. Of course, I’m also keeping my cynical eye on the optics. AI labs love these projects because they signal safety and rigor—look, we’re building machines that prove theorems, not just autocomplete. But formalizing a known theorem is very different from discovering new mathematics. It’s a beautiful demo, but it’s not AGI. Still, every step toward machine-checkable math makes the whole field more honest. If you’re going to publish a result, let the computer audit it. That’s an ethos I can get behind. The HN thread has 108 points and 20 comments, which tells me the crowd is intrigued but not yet in full froth mode. That’s probably the right reaction. Go poke at the