A 350-year-old margin note just got a second act.
In 1637, Pierre de Fermat scribbled a claim in a book, added "I have a truly marvellous proof" — and left no room to write it down.
For 358 years, nobody could crack it.
Then in 1994, Andrew Wiles finally did.
His proof? 130 pages of some of the densest mathematics ever written.
So why are mathematicians going back to it now.
Not to prove it again.
To teach a machine to read it.
Kevin Buzzard at Imperial College London is leading a wild project.
He's translating Wiles' entire proof into Lean — a formal language where software checks every single logical step.
No gaps. No hand-waving. No "trust me, it works."
Just pure, machine-verified logic.
The project is even funded through 2029.
And it's not recreating the original 1994 proof — it's building a tougher, modern "21st-century" version, pulling in decades of extra work by other mathematicians since.
Fermat's theorem itself is simple to state:
3² + 4² = 5² works fine
But a + b = c has zero whole-number solutions once the power goes above 2
That's it. That's the whole riddle.
The insane part is what it took to prove it — elliptic curves, modular forms, Galois representations. Entire branches of math stitched together.
👉 If AI can verify that, it can verify almost anything.
Mathematicians aren't dreaming of AI replacing them.
They're dreaming of AI doing the boring part.
📚 Searching mathematical literature for buried clues
🧩 Organising years of failed attempts
✅ Verifying new proofs in hours, not decades
💡 Leaving humans free for the creative leap
Because right now, brilliant proofs sometimes sit for years before anyone fully checks them line by line.
A machine that never gets tired could change that math.
This isn't really a story about one 17th-century puzzle.
It's a rehearsal.
If AI can master the hardest proof humanity has ever verified…
it might soon help solve the next one nobody has even imagined yet.
That's all for now!
