The moat was never the model. It's craft.

Every team now has access to the same frontier models. That used to feel like the prize. It's starting to feel like the floor.

When the model is a commodity — when your competitor can call the same API you can — the advantage moves somewhere else. The question I keep coming back to isn't which model you use. It's what you build on top of it, and whose knowledge is inside it.

The asset that appreciates

There's a quiet assumption that as AI gets better, the people matter less. In my experience it runs the other way. The better the tools get, the scarcer the real bottleneck becomes: judgment — taste, context, knowing which of a thousand good-looking options is actually right.

Models generate. They don't yet know why an artist chooses one version over another, what a franchise is supposed to feel like, or which clever shortcut will quietly break in production three months from now. That knowledge doesn't depreciate as the models improve. It compounds.

The veteran in the loop

The systems people want to build are described as learning loops — tools that get smarter the more they're used, capturing institutional knowledge and improving on top of a generic model. I build those systems. Here's what the diagram leaves out: the loop is made of people.

The institutional knowledge everyone wants to capture isn't sitting in a database. It's in the hands of a senior artist who has solved this exact problem a hundred times and never wrote it down. For that knowledge to enter the system, that person has to choose to put it there.

Which makes it a trust problem

So the moat isn't the model, and it isn't even the data. It's whether the people who hold the craft trust the system enough to teach it.

That's the same thing I wrote about before — adoption decides everything — seen from one level up. A veteran who trusts the tool feeds it their best judgment, and the loop compounds. A veteran who suspects the tool was built to replace them feeds it the minimum, and the loop starves. Same model. Same architecture. Entirely different outcome, and the difference is human.

Build the thing the veteran wants to teach

That changes what you're actually building. The goal stops being a model that stands in for the expert, and becomes an instrument the expert wants to play — one that respects their craft, makes their judgment go further, and gives them a reason to pour themselves into it.

Get that right and something rare happens: the system stops being a tool and becomes a place where craft accumulates. Every person who uses it leaves a little of their taste behind. That's an asset no competitor can buy, because it was never for sale — it was earned, one person's trust at a time.

The idea I want people to remember

The teams that win the next decade won't be the ones with the best model. They'll be the ones whose people made the system smarter than any model could be on its own.

Models are getting cheaper. Craft is getting rarer. Build for the rare one.