Taste doesn't come from the model
A model can draw fifty interfaces. It can't tell you which one is right.
Ask a model for a landing page and you get the same landing page everyone gets. Gradient hero, three feature cards, rounded everything. It isn't bad. It's the average of every landing page it has seen, which is exactly what it was trained to produce.
That's not a bug. A model predicts the likely next thing. Taste is knowing when the likely thing is wrong.
Generating is free now
The cost of producing an interface went to roughly zero. When something costs nothing, it stops being the valuable part. What's left is choosing.
This site is a small argument for that. Almost nothing on it is hard to generate. What took the time was deciding how each piece should feel.
Taste is a list
Taste sounds like a vibe. Most of it is a list of small, specific opinions, each one earned by noticing that something felt off:
- A highlight that travels between items feels like one control. One that blinks in and out feels like four buttons.
- Zoom should stay under your cursor, not the centre of the screen, or the thing you were reading slides away.
- A hit target should be bigger than the icon. The padding is the part people actually hit.
- An animation you can't interrupt is a small lie about who's in control.
Every one of these is invisible in a screenshot. Every one is the difference between something that feels made and something that feels generated.
Where AI engineering gets this wrong
The conversation is all about generation — bigger models, faster tokens, longer context. Almost nobody talks about selection, because selection doesn't benchmark.
When a model can produce a thousand options, the scarce thing is the person who knows which one to keep, and can say why.
Taste can be written down, a little. An eval is taste in a file. If you can't say what good looks like, no model is going to find it for you.
The model is the best intern I've ever had. It still needs someone to say no.