In the summer of 1637, a French magistrate named Pierre de Fermat picked up his copy of Diophantus's ancient Greek text Arithmetica and scrawled a quiet provocation in the margin. He had discovered a truly marvelous demonstration that no three positive integers can satisfy the equation $a^n + b^n = c^n$ for any integer value of $n$ greater than two. Then came the infamous taunt: the margin was simply too narrow to contain it.
For more than three centuries, that single marginal note became a ghost haunting the attic of human thought. Amateurs mailed frantic proofs to academies. Brilliant minds broke their professional hearts against it. When Sir Andrew Wiles finally cornered the beast in 1995, working in secret for seven years in an attic study at Princeton, it took a 129-page manuscript and an armada of modern mathematical machinery to do it. Wiles stepped out of the fog of proof victorious, but the mathematical community was left with a terrifying hangover. How do you actually verify a modern mathematical argument that spans hundreds of pages of esoteric abstraction? Checking a proof by hand can take peer reviewers years. Sometimes, quietly, they just trust the author.
Then came eleven days in August.
A cluster of artificial intelligence agents running on a research architecture developed by Anthropic did not invent a new path through number theory. Instead, they undertook a task most human experts believed would consume years of tedious, skull-splitting labor: they translated the sprawling architecture of Wiles's intellectual monument into the cold, uncompromising syntax of Lean, a formal proof assistant.
Imagine trying to build a suspension bridge where every single atom of steel must be individually certified by an inspector who accepts nothing on faith. That is formalization. Every logical leap, every hidden dependency, every subtle shift from algebraic geometry to modular forms must be spelled out until nothing is left to intuition.
Working collaboratively through a shared dependency graph—a persistent digital to-do list that prevented individual agents from repeating past errors or forgetting what their digital peers had already solved—dozens of Claude instances tore into the problem. They did not sleep. They did not lose heart when a logical branch collapsed. When early runs stalled because agents drifted off-task, the researchers shifted the architecture toward structured state management. The machine remembered.
Thirteen million lines of Lean code later, the work was finished.
The resulting machine-checked artifact stands as a strange monument. It contains nearly thirty thousand intermediate theorems, meticulously chained upward from foundational axioms to the grand conclusion. It includes custom-built formalizations of algebra, harmonic analysis, and geometry. For mathematicians accustomed to the agonizingly slow pace of human verification, the speed feels almost illegal.
Yet the true weight of this milestone lies less in Fermat's ghost finally being laid to rest inside a silicon server and more in what happens next. Mathematics is expanding faster than the human infrastructure designed to police it. Thousands of papers flood repositories every month, and the pool of qualified referees is drowning. When the guardians of truth can no longer keep up with the volume of creation, the foundation begins to soften.
Machines like this offer a scaffolding against that erosion. They do not replace the chaotic, beautiful intuition of human genius that first dreamed up the proof; they act as a universal translation engine for certainty. They take the wild, tangled paths of human thought and pave them into roads that anyone can travel without fear of falling through the dark.
Fermat was wrong about his margin. It was never wide enough. But deep inside a modern data center, the machines have finally found all the room they need.
