AI design tools don't have bad taste. They have no taste. That's worse.

Everyone blames the model. "GPT can't design." Wrong. The model isn't the bottleneck. The pipeline is.

Here's what's actually happening when an AI tool spits out that same purple-gradient, rounded-card, centered-hero layout every single time:

1. They generate pixels, not systems. Most tools prompt a model to output a screen. One screen, in isolation. No shared token layer, no type scale, no spacing system carried across views. So the second screen doesn't know what the first one decided. Consistency is impossible by construction, not by accident.

2. They optimize for "renders without breaking," not "looks intentional." The reward signal is valid output. A layout that compiles. Taste isn't in the loss function anywhere, so the model regresses hard to the mean of its training data. The mean of all web UI is Tailwind defaults and a violet gradient. That's the slop.

3. They treat design as generation when it's actually constraint. Good design is mostly subtraction. Rules about what you don't do. Contrast ratios, optical alignment, a 4px grid you never break, restraint with color. A raw generative model has no constraints unless you build them in. Almost nobody does, because constraints are unglamorous engineering.

Our bet with Baroque was that the fix isn't a smarter model. It's encoding design laws as hard constraints, and generating against a shared system instead of one screen at a time. You take a reference you love, we extract the actual system behind it (tokens, scale, rhythm) and generate your whole product against that, consistently.

Is that the right call? Genuinely not sure it's the only way. Maybe the model does eventually swallow taste end to end and constraints become a crutch.

So tell me I'm wrong. Where's the real bottleneck? Model, pipeline, or data?