A field survey of learned world models — how machines build small imagined copies of reality to predict, plan, and decide.
Each territory below is explained three times over: once in plain bearings anyone can follow, once in a surveyor's technical field notes, and once in a brief written for the person deciding whether to fund the expedition.
What a "world model" actually is, and why prediction, planning, and control all live on the same map.
Where "world model" ideas came from before the name existed — and the probability math that makes them precise.
The interface between a learned model and the environment it is trying to predict — and what a "rollout" actually is.
How a model carries a compressed memory of the past forward through time.
How a model decides what is worth remembering from raw pixels, sound, or text — and what to throw away.
The four major families of model that learn to generate: VAEs, GANs, autoregressive models, and diffusion/flow methods.
How an agent uses a learned world model to decide what to do next, by imagining consequences before acting.
The field of learned world models moves fast and cites itself in shorthand — "JEPA," "MPC," "the ELBO" — that can gatekeep as much as it clarifies. This atlas exists to chart the territory properly: what each idea actually is, how it fits next to its neighbors, and why it matters to three very different readers at once.
New territories are surveyed and added over time. Read more about the project →