A latent ocean. Tens of thousands of particles drift through a curl-noise field on the GPU.
Every few seconds, a language model is asked for a single sensory fragment — a texture, a weather, a piece of geology, a marine or neural word. The fragment is hashed into a 16-dimensional dream vector that slowly reshapes the simulation: flow direction, viscosity, palette, depth of focus.
You never see the words. Only the weather they make.