A month ago, Donald Knuth did something that should have cracked the internet in half. The 87-year-old author of The Art of Computer Programming — a multi-volume algorithmic bible that computer scientists have been genuflecting before since the 1960s — published a mathematics paper titled "Claude's Cycles." Named after an AI model. Because that model found what he could not.
The internet, predictably, was too busy benchmarking chatbots on coding puzzles to notice.
Here's why this matters more than whatever leaderboard war you're following this week.
Knuth had been stuck for weeks on a problem involving Hamiltonian decompositions of directed graphs. Translation for humans: imagine a network of one-way streets. A Hamiltonian cycle is a route that visits every intersection exactly once and loops back to the start — the world's most efficient pub crawl, no repeats, end where you began. Knuth needed a general construction that works for any odd-dimensional cube of these connections larger than 2×2×2. He'd cracked the small cases by hand. Computers had verified solutions up to 16×16×16. The elegant general rule? Weeks of grinding. Nothing.
On February 28, Filip Stappers — a colleague — fed the problem to Claude Opus 4.6. Over 31 guided conversations spanning roughly one hour, Claude attacked the conjecture from every angle. Linear formulas. Brute-force searches. Geometric frameworks. Simulated annealing — a technique stolen from metallurgy where you "heat" and "cool" a solution to escape local dead ends. It hit walls. Pivoted. Kept going.
Then it cracked it. Claude independently identified the underlying structure as a Cayley digraph — a specific network type built from mathematical group operations — and reformulated the problem accordingly. The resulting "serpentine" pattern turned out to correspond to the classical Gray code, a combinatorial sequence where consecutive entries differ by exactly one digit. Claude derived this from scratch, without knowing the classical version existed. An AI reinvented a known mathematical structure while solving a problem a living legend couldn't touch.
The raw numbers for the pedantic: 11,502 Hamiltonian cycles for the 3×3×3 case. Of those, 1,012 generalize to 5×5×5, and 996 work for both 5×5×5 and 7×7×7. Exactly 760 "Claude-like" decompositions hold for all odd dimensions greater than 1.
Now for the part where I ruin the celebration.
Claude didn't do this alone. Not even close. Stappers steered every one of those 31 conversations. Knuth wrote the formal proof himself. The even-dimension case — literally the other half of the problem — remains unsolved, and Claude made exactly zero useful progress on it. So we're talking about a model that, with a skilled human holding the wheel, solved half a problem in a highly structured domain. Not exactly Skynet.
But here's what the skeptics need to sit with: the "skilled human holding the wheel" had already tried. And failed. The tool made the difference. Not a marginal difference — a qualitative one. Weeks of effort from one of history's finest algorithmic minds, cracked in an hour by a model that doesn't understand what a graph is.
Knuth called it "a joy to learn not only that my conjecture has a nice solution but also to celebrate this dramatic advance in automatic deduction." And in the revised version (March 16), he dropped this: "It seems I'll have to revise my opinions about 'generative AI' one of these days."
From a man who has forgotten more about algorithms than most CS departments collectively know — that sentence detonates quietly. This isn't some LinkedIn influencer declaring victory for artificial intelligence. This is the person who defined computational complexity standards admitting, in a published academic paper, that his priors about AI were wrong.
And we collectively responded by scrolling past it to argue about whether Gemini's context window is really 2 million tokens.
The lesson isn't that AI is coming for mathematicians. It's darker and simpler than that. The people best positioned to use these tools — domain experts with hard, well-defined problems — are getting results that look like magic. Everyone else uses the same models to generate LinkedIn posts and argue about benchmarks. The gap between "AI as a sharp tool in expert hands" and "AI as a party trick" isn't closing. It's widening.
Knuth didn't name a paper after a chatbot because he's gone soft at 87. He named it that because intellectual honesty demanded it. The AI found the answer. He had the integrity to say so.
Most of us won't. We'll keep pretending the tools are either useless or omnipotent, because the uncomfortable truth — that they're powerful but only in competent hands — doesn't fit anyone's narrative.
