World Models: An Overview

Imagine trying to play a video game with your eyes closed, using only a friend who whispers what's happening. To do well, you would build a picture in your head: where things are, how they move, what happens if you press a button. That picture is a world model — a small imagined copy of the real thing, kept inside a mind so it can guess what comes next before it actually happens.

People do this constantly without noticing. You know that if you let go of a cup, it falls. You know a friend will probably laugh at a joke before you finish telling it. A world model is what lets a machine do the same: hold a compressed sense of "how things work" so it can imagine forward, instead of only reacting to what is in front of it right now.

Formally, a world model learns a function over three ingredients: a state (a compressed summary of everything relevant so far), an action (something an agent does), and a transition (how the state changes in response). Given (s_t, a_t), the model predicts s_{t+1}, and often a reward or observation alongside it. Everything downstream — planning, control, imagination-based training — is built on top of this one predictive object.

The interesting engineering tension is what to compress and what to throw away. A pixel-perfect simulator is a world model with zero compression and enormous cost; a hand-written physics equation is a world model with total compression and zero flexibility. Learned world models sit between these poles, trading fidelity for tractability, and the entire history of the field — covered territory by territory across this atlas — is really a history of where that trade-off has been renegotiated.

Strategically, a world model is a bet that simulation is cheaper than reality. Every domain — robotics, logistics, customer operations, energy grids — has a version of "we can't test this in production." A world model gives you a sandboxed, differentiable stand-in for the environment, so a policy or a plan can be rehearsed thousands of times before it touches anything real.

The organizations that will benefit fastest are the ones already sitting on sequential, decision-heavy data — operational logs, control loops, customer journeys — because that data is exactly the fuel a world model needs. The capability to watch is not "does it generate realistic video," but "does it let you ask what-if questions cheaply and trust the answer."