[ben_w]

> The problem though is that LLMs write code to represent designs as well, which means you get consistent designs. You end up aligning on a generic mean because of it, which is often why you see these repeated aesthetics.

I think this is more due to temperature (while the LLM is producing a design) rather than consistency. Consistency means it'll be yes-and-ing all the way once it starts, but that initial seed? You meet one instance of Claude 5, you've met them all. Ditto ChatGPT-x or Gemini-y.

Because each blank context window starts the same, the only initial seed of variation is the user prompt. If the users don't mention style, then 5000 people all prompting for "single page web app to do X" will get very close to the same result.

Most humans keep asking for the same, and the output is therefore analogous to Simon Willison's pelican on a bike, where the output is overwhelmingly facing to our right, and to the extent that the bike is functional and would actually e.g. steer, it's a standard bike rather than recumbent, cargo, or tandem.