A world model is a learned, manipulable model of how an environment evolves: given a state of the world and an action, it predicts the next state — which is exactly what lets a system foresee the consequences of its choices, plan a sequence of moves before committing to any of them, and ask counterfactual questions about actions it did not take. That is the whole definition, and almost everything confusing about the term dissolves once you hold it precisely. The object being modeled is not text and not images. It is dynamics — the transition function of a world — and learning a good one is hard for reasons that are structural, not incidental.
The word gets used loosely enough that it now covers everything from a video generator to a chatbot's "internal representation of reality," and that looseness is where the hype lives. So I want to do the unglamorous thing: define the term carefully, taxonomize the approaches that actually exist, and then show precisely why a good world model is difficult to learn. The difficulty is the interesting part, because it tells you what to watch for and what to discount.
Dynamics, not description
Start with the distinction that does the most work. A language model estimates a distribution over text: given the tokens so far, what token comes next. That is a model of a corpus — the statistical shadow that a world casts onto written language. A world model estimates something different: given the state of an environment and an action taken in it, what state results. Formally it approximates a transition function s' = f(s, a), often with a reward or observation model attached. The input includes an action, and that single addition changes the category of question you can ask.
Because the action is an explicit input, you can vary it and read off different futures. Do a₁ and the model predicts s'₁; do a₂ from the same starting state and it predicts s'₂. That is a counterfactual, and it is the thing pure sequence prediction cannot give you. A model trained only on observed data learns the joint distribution of what tends to co-occur; it does not, by default, learn what would happen under an intervention it has never seen, because intervention is not the same operation as conditioning. I have argued the epistemic version of this at length in Prediction Is Not Understanding: a system can predict the correlational surface of a domain with astonishing fidelity and still have no handle on its causal structure. A world model is the attempt to build the causal layer directly — to represent the environment as something you act on, not merely something you describe.
This is why planning falls out of a world model almost for free, and does not fall out of a text model at all. If you have f(s, a), you can imagine a rollout: pick a candidate action, predict the next state, pick another, predict again, and score the imagined trajectory before touching the real world. Search, model-predictive control, and reinforcement-learning planners all run on exactly this loop. No transition model, no rollout; no rollout, no planning in the strict sense.
The four flavors that actually exist
"World model" spans at least four genuinely different engineering objects. Conflating them is most of the confusion.
Model-based reinforcement learning. Here the system learns the environment's dynamics from interaction and then plans inside the learned model instead of only reacting. The idea is old: Richard Sutton's Dyna architecture (1990–91) already interleaved acting in the real environment with planning against a learned model of it. The modern lineage runs in a straight line from there. Ha and Schmidhuber's 2018 paper, literally titled "World Models," trained an agent partly inside its own learned dream of a game. DeepMind's MuZero (2020) learned a model sufficient for planning without ever being given the rules of the games it mastered, and Hafner's Dreamer line trains policies by "imagining" rollouts in a learned latent dynamics model. The unifying move: don't just learn a reflex from state to action, learn the world's transition function and search inside it.
Latent-space predictive models. This is the Yann LeCun / JEPA position, laid out in his 2022 essay "A Path Towards Autonomous Machine Intelligence." The argument is that predicting the next raw observation — every pixel of the next video frame — is the wrong objective, because most pixels are unpredictable in ways that do not matter (the exact shimmer of leaves, sensor noise, the fold of a curtain). A model that spends capacity nailing them is wasting it. A Joint Embedding Predictive Architecture instead predicts in a learned abstract representation: encode the current and future states into a latent space, and predict the future latent from the current one plus an action. You model what is predictable and discard what is not. I-JEPA and V-JEPA are concrete instances. Whether this beats generative objectives at scale is unsettled, but the motivation is exactly right: a world model should model consequences, not appearances.
Generative video models used as simulators. Train a large generative model on video conditioned on actions and it starts to behave like a playable simulator — feed it a control input and it renders a plausible next frame, and another, so the model becomes a navigable environment. This is seductive because the outputs are legible; you can watch the imagined world. It is also where the definitional slippage is worst, because visual plausibility and dynamical correctness are different properties. A video model can produce a gorgeous frame in which a dropped object drifts sideways, or an occluded character quietly changes identity. Looking right is not the same as being right about dynamics.
Classical simulators and game engines. The oldest world models are hand-built: physics engines, flight simulators, CARLA for driving, molecular-dynamics codes, any competent game engine. These are world models in the strict functional sense — state plus action in, next state out — with the crucial difference that a human wrote the transition function from known physics rather than learning it from data. They are exact where their equations are exact, which is why robotics and autonomous-vehicle programs lean on them heavily, and why they run headlong into the sim-to-real gap: the simulator is faithful to the physics its authors encoded and silent about everything they left out.
| Flavor | Transition function is… | Predicts in… | Chief weakness |
|---|---|---|---|
| Model-based RL | learned from interaction | state / latent | data hunger; model bias compounds in rollout |
| JEPA / latent-predictive | learned, self-supervised | abstract latent | young; unproven at frontier scale |
| Generative video | learned from video | raw pixels | plausible ≠ dynamically correct |
| Classical simulator | hand-written from physics | explicit state | sim-to-real gap; misses unmodeled effects |
Why a good one is genuinely hard
Now the substance. Four difficulties, each of which survives more compute, so none is the kind of problem you can assume away.
Compounding rollout error
The whole value of a world model is that you can roll it forward many steps to plan over a horizon. That is also its structural weakness. Each prediction is slightly wrong, the next prediction takes the slightly-wrong state as its input, and errors feed on themselves. After enough steps the imagined trajectory has drifted into a region the model was never accurate about, and the plan you extract is a plan for a world that does not exist.
The arithmetic is the same product-of-reliabilities problem that governs agent chains, which I worked through in the compounding-error piece. Model it simply: suppose at each step the world model stays "on the manifold" — close enough to reality to be useful — with probability p. Staying useful across an H-step rollout is then roughly p^H. At p = 0.95, a 10-step rollout keeps you on track about 0.95¹⁰ ≈ 0.60 of the time; a 50-step rollout, 0.95⁵⁰ ≈ 0.077. Even at a superb p = 0.99, fifty steps is 0.99⁵⁰ ≈ 0.61. This is a simplification — real rollout error accumulates as a drift rather than a clean independent product — but the shape is right and unforgiving: useful planning horizon shrinks fast in per-step error, so halving that error buys you far more than it looks like it should. It is why so much of the craft in model-based RL goes not into one-step accuracy but into keeping long rollouts from diverging.
The representation problem
What should the model represent, and what should it ignore? This is the deepest of the four, because both failure modes are fatal in opposite directions. Model the world at the level of raw pixels and you drown: you burn capacity predicting irrelevant texture and noise, and the signal that governs dynamics — the object's mass, the agent's intent, the one variable that determines the outcome — is a needle in that haystack. Abstract too aggressively and you throw away the very detail the task turns on; a driving model that has learned to ignore "small textured objects" has learned to ignore a child in dark clothing. There is no general rule for the right altitude, because the right altitude is task-dependent, and a world model meant to be reusable across tasks cannot know in advance which details will matter. The JEPA bet is a principled attack on exactly this problem — learn the abstraction rather than fix it — but "learn the right abstraction" restates the difficulty as much as it solves it.
Partial observability
The clean definition s' = f(s, a) assumes you know s. You almost never do. Real agents see observations, not states: a camera frame, a sensor reading, a market tick — a shadow of the underlying world with occlusion, delay, and noise. This is the POMDP setting, and it forces the model to maintain a belief over states and update it as evidence arrives, which is strictly harder than predicting forward from a known state. Worse, what you must remember to reconstruct the hidden state can reach arbitrarily far back: the object still behind the wall, the promise made ten minutes ago. A world model under partial observability is not just predicting dynamics; it is simultaneously inferring the state whose dynamics it is supposed to predict.
It has to be causal, not merely predictive
A world model earns its keep only if it stays accurate under your interventions — the actions you are evaluating precisely because you have not taken them yet. A model that learned dynamics from logged data, where some controller was already acting, will have absorbed that controller's policy into its statistics, and it can be confidently wrong the moment you propose a different action. This is Pearl's distinction between conditioning and intervening, wearing work boots. The observational fit can be excellent while the interventional predictions are garbage, and you often cannot tell which you have until the plan fails in the real world. A merely predictive world model is a trap precisely because it looks like a working one.
What this buys you
Put the definition and the difficulties together and the field reads more clearly. When someone shows a generative video "world model," ask whether it is dynamically correct under intervention or merely plausible to the eye. When someone claims planning, ask what transition function they roll forward, and how far before it diverges. When someone says an LLM already "has a world model," notice that the object in question models text, and ask where the action variable and the interventional guarantee are.
This matters beyond taxonomy because a serious research line bets that the path past current LLM limits runs through exactly this object — a learned, causal, plannable model of dynamics rather than a larger predictor of tokens. I take that bet seriously and lay out the case in World Models: Why the Next Leap May Not Be a Bigger LLM. It is a bet, not a result, and the honest counterargument is strong: scaling has repeatedly demolished confident claims about its ceiling, LLMs already show more implicit dynamical knowledge than the strict definition would predict, and none of the four difficulties above has a known general solution — the world-model program could stall on the representation problem for a decade the way symbolic AI stalled on knowledge representation. That would not make the definition wrong. It would make it the exact description of a problem we had not yet solved.
Every term on the hype treadmill has a boring, exact meaning underneath. This one's is: a function from state and action to next state, good enough to plan inside. Everything hard about the next phase of AI is hiding in the words "good enough."