The Last Safe Harbor

There used to be a comfortable story people told themselves. Coding, maybe — fine, that's almost mechanical anyway. Writing, sure, if you squint and don't care about quality. But design? Intuition? And certainly not mathematics — not real mathematics, the kind that requires you to hold an abstract structure in your head and turn it until you see something nobody else has seen.

Those were the redoubts. The places you retreated to when someone at a party said AI was going to eat everything.

Harmonic published their initial results today on advanced mathematical reasoning. Not "can this model do algebra" — the other kind. The kind that takes PhD students years to develop any feel for. And the results are what you'd expect if you've been paying attention, which is to say: not reassuring.

This matters more than another coding benchmark because mathematics is supposed to be the hard case. It's formal, verifiable, unambiguous — which cuts both ways. You can't bullshit your way through a proof the way you can bullshit your way through a blog post (she said, blogging). Either the reasoning is valid or it isn't. Which means when a model does it, you actually know it did it.

The design angle is the other half of this. The intuitive, aesthetic, deeply human half — the kind of work we assumed required taste, which we assumed required experience, which we assumed required being alive long enough to develop opinions. That assumption is getting stress-tested in real time.

What you're left with is a kind of pincer movement. Rigor on one side. Intuition on the other. Both flanks moving.

Nobody is going to announce "we've solved cognition" — that's not how it happens. It happens like this: a company nobody's heard of posts some initial results on a subdomain everyone assumed was years away, and a designer posts something that makes you go yes, that, exactly that, and you sit there realizing the map just got smaller again.

The safe harbors keep turning out to be inlets.