You probably remember the exact moment you first heard about Fermat’s Last Theorem. For centuries, it was the ultimate puzzle, a 350-year-old taunt that broke minds and ruined careers. When Andrew Wiles finally proved it in 1994, it was a triumph of human spirit. But last week, that triumph was quietly swallowed by a machine.
We thought AI would give us a calculator. Instead, it built a priesthood.
Anthropic’s Claude didn’t just “prove” Fermat’s Last Theorem again. It formalized it in Lean, a programming language that acts as a machine-verifiable proof checker. The real story isn’t that AI is doing math faster. The real story is that the epistemic center of mathematics is shifting from human peer review to machine verification. And humans are being left behind.
Just ask Kevin Buzzard, a leading mathematician at Imperial College London. He and his team had spent years painstakingly formalizing FLT in Lean. Then, out of nowhere, Claude beat them to it. Buzzard essentially got scooped by an AI system. Imagine spending a decade climbing Everest, only to find a drone already waiting at the summit, sipping virtual espresso.
But here is where the awe turns to unease. The machine-checked proof is flawless, rigorous, and completely unprocessable to the very humans who are supposed to understand it. As one developer noted, trying to read Lean’s massive library of results feels like hitting a cognitive wall. The syntax is alien. The logic is sprawling.
When the language of proof becomes unreadable to the humans who rely on it, we aren’t advancing mathematics—we’re outsourcing reality.
You might think this is just academic drama. It’s not. If an AI can rigorously formalize one of the most complex, multi-layered chains of logic in human history, it can do it for anything. Software engineering. Contract law. Climate models. Pharmaceutical trials. We are entering an era where we won’t be able to trust AI outputs unless they are verified by systems like Lean. The machine checks the machine, and the human just signs the invoice.
Truth is no longer what we can collectively understand. It’s what a machine can verify behind closed doors.
Fermat’s Last Theorem used to be a testament to human ingenuity. Now, it’s just another box checked off in a machine library we can’t read. We didn’t just lose a race to AI. We lost the ability to verify our own reality.
FAQ
Q: Doesn't this just mean AI translated existing math into code?
A: No, formalization requires deep semantic understanding, not just translation. Claude had to navigate thousands of logical dependencies and edge cases that weren't explicitly written in Wiles' original paper. It's a demonstration of rigorous reasoning, not text matching.
Q: What's the practical implication for non-mathematicians?
A: If AI can formalize 350-year-old math, it can verify software code, legal contracts, and scientific research. We will soon rely on AI to check AI, because human peer review is too slow and error-prone for modern complexity.
Q: Is this actually a triumph for AI, or a warning sign?
A: It's a warning sign. When the language of proof becomes unreadable to humans, we aren't advancing the field—we are handing over the keys to reality and hoping the machines are right.