Notes · 01

Why AI design
all looks the same.

Ask any model for a landing page and you can predict the result before it streams: the gradient blob, the three-tier pricing, the purple on white. People call this a failure of creativity. It isn't. It's the system working perfectly. This note is about why, and about the three things we've found that actually move it.

01 · The mass

The cliché is the correct answer.

A language model is a probability machine. It learned what design usually looks like, and "usually" is dominated by the most common, most repeated, most agreed-upon work in its training data. When it generates, it samples near the peak of that distribution.

what generation actually is
the question you asked"design something good"
the question it answers"what is most likely?"
The cliché isn't a malfunction. It's regression to the mean, performed flawlessly.

Instruction tuning makes it stronger, not weaker: models are further trained toward what most people rate well, which drags output even closer to consensus taste. The average of everyone's preferences is, by construction, nobody's taste.

02 · The valley

Adjectives are locations, not directions.

Picture design space as a landscape with one massive, dense valley in the middle. Every prompt drops you somewhere on it, and the sampler rolls downhill. The trap: the words designers reach for are themselves places inside the valley.

what you write
  • "clean and modern"
  • "minimal, elegant"
  • "be creative, unique"
  • "premium feel"
every adjective has a center of mass
where it lands
  • the same white card, again
  • the same 8px grid of gray
  • gradient orbs
  • dark mode with gold serif
the name of the valley, not a path out

Every generation rolls downhill into the pile — the layouts the mass prefers. Distinctive ones sit at the edges, almost never sampled.

A warped field with a pile of common AI-generated layout wireframes at the centre labelled valley. Dashed rails lead inward; pink samples accelerate along them into the pile — gravity pulling generation toward the cliché. At the edges, four distinctive layouts in dashed outline, labelled as never sampled.

03 · The trap

The mass cannot criticize itself.

The obvious fixes fail for the same reason. Raising temperature adds noise, not direction: you scatter randomly around the same center. And asking the model to judge — "generate two, pick the better one" — consults the same probability mass that produced the options. In our early tests, a model judge asked "which is better?" picked the cliché almost every time. There is no minority opinion inside a distribution.

04 · What moves it

Three things that actually work.

Steer in values, not adjectives
Don't ask for "distinctive." Inject specifics: exact tokens, spacing logic, rules about what a color is allowed to mean, the patterns a person never uses. Concrete values relocate the starting point before sampling begins — words about style just rename the valley.
A deviation budget
Distinctiveness is a delta, not a distance. Deviate deliberately on a few decisions and hold competent defaults everywhere else. Deviate on everything and you get noise; deviate on nothing and you're back in the valley. Coherence is carried by the defaults you keep.
An anchored judge
Never ask a model "which is better?" — that consults the mass. Ask "which would this specific person save?", anchored to their actual saved work. In our development tests that one reframe took the judge's cliché-picking from nearly always to nearly never. The judge stops consulting the crowd and starts consulting a person.

05 · The consequence

The escape vector comes from outside.

Gravity can't be deleted. It's what the model is: the mass supplies the competence, the fluency, the defaults. It can only be escaped locally, and the direction has to come from somewhere the model doesn't have — a particular person's recorded eye.

That's the bet uistash is built on: the model stays frozen, and the person's taste becomes a file. The references you keep, the patterns you never touch, the rules your eye enforces without you noticing — written down, cited, versioned, and injected as the steering force. Whether it works isn't something we'll claim. It's something we're measuring, blind, in the open.

The same valley, now with an escape path: orange rails pull samples from the centre labelled valley toward an off-centre orange cluster labelled you. Three ticks name the mechanism — values, budget, anchored judge. The direction comes from outside the model. valley

the direction has to come from outside the model.

honesty notefield notes
these findings come frombuilding uistash
sample sizessmall
peer reviewedno
the blind studyrunning now
notes, not conclusions — numbers land when the study does

Read how we measure it.

The lab behind these notes runs sealed blind experiments with controls that test the experimenter, and refuses to display numbers it can't defend.

How the Lab works Back to the product