The thing is, they aren't creative. When I ask an LLM to go beyond the mean, it has no idea what to do and returns incoherance. Its responses seem to have N latent dimensions of information. But, the value of each of those dimensions is always the midpoint. With creative prompting, I can push it to respond in ways that go beyond the mean, but it can't tie them together cohesively. It doesn't have anything in its training data to fill the hole in space. No ability to extrapolate.

My point is that those kinds of prompts are dumb and I actually expect the model to produce slop. In programming we write these huge prompts that go into a lot of detail, and when we do that, it actually works. I've never asked a coding agent to go beyond the mean, that just sounds like a dumb thing to do. Apply this rule in design and I'm sure you'll get better results. Kinda what the blog post is all about.