Claude Found the Pattern. Knuth Found the Proof. That Distinction Is Everything.

Vishal Misra, a Columbia CS professor, published a piece this week arguing that the Shannon/Kolmogorov gap explains where AI stops. It's worth reading. It's also built on a story it hasn't entirely earned.

What works: the Don Knuth anecdote near the end.

A collaborator used Claude to solve an open combinatorics problem — 31 exploration runs, roughly an hour, a valid construction found for every odd case tested. Then Knuth sat down and proved it. Not tested it further. Proved why it must work.

Misra's summary: "Claude found the pattern. Knuth found the mechanism."

That sentence does more precise work than most AI commentary manages in three times the word count. It names a real boundary, locates it in a specific and verifiable event, and doesn't overstate it. It doesn't say AI is useless. It says AI is doing one thing, and humans are doing another, and those are not the same thing.

What doesn't: Einstein.

Misra uses the Einstein arc — two postulates, everything follows, G=8πT — as his proof that genuine discovery requires a kind of reasoning AI structurally cannot do. The problem is that the Einstein he's describing is the retrospective, textbook version.

The actual general relativity took a decade. Einstein made significant wrong turns. He worked alongside Marcel Grossmann, a mathematician whose expertise in Riemannian geometry Einstein didn't have. The "lone mind asking what kind of universe would make the data unsurprising" is the cleaned-up origin story, not the process.

Using a myth of solitary counterfactual genius to argue about what cognition discovery requires is the exact move the piece warns against elsewhere — fitting the narrative to the conclusion you want rather than asking what the evidence actually shows.

The Shannon/Kolmogorov distinction is real and useful. The Knuth example demonstrates it cleanly. The Einstein example smuggles in more than it proves.

Read it for the vocabulary it gives you — pattern-finding versus mechanism-finding is a usefully precise way to evaluate what you're actually asking an AI to do. Just don't let the physics heroics do the heavy lifting.