Five tabs, five machines
If you’ve spent any time on your timeline lately, you’ll probably have seen a clip where somebody walks through a landscape with the arrow keys and the caption says it was generated rather than built. It looks like a video game that nobody made. The replies call it a world model, so you go and look the phrase up.
Ten minutes later you have five tabs open, and they describe five different machines. One generates video you steer with the arrow keys. One exports geometry into Blender. One is a paper about predicting embeddings — lists of numbers that stand in for pictures — and it never shows you a video at all.
The mess has a history, and the history is tidier than the tabs. Between about 2018 and 2024, four traditions arrived at the same phrase from four directions: reinforcement learning and control, self-supervised representation learning, generative video and spatial models, and interpretability — the study of what goes on inside a trained model. Each meant something real, and none of them had to check what the others meant.
The five definitions
Treat these categories as contracts rather than species. Each one describes what a system promises to hand you. Modern work crosses between them all the time.
The Renderer predicts observations, usually pixels. GameNGen, from Dani Valevski and colleagues at Google (2024), draws frames of the video game DOOM — each frame comes from the frames before it and the player’s inputs. It runs at about twenty frames a second on a single TPU chip, fast enough to play. Genie 3, from DeepMind, generates interactive video at 720p and 24 frames per second. Sora belongs in this category too, but without the action input: it predicts observations you cannot steer.
The Simulator predicts structure you can query, meaning 3D shapes you can ask questions of. Marble from World Labs (co-founded by Fei-Fei Li) takes text, an image or a rough 3D layout and hands back Gaussian splats for the look and triangle meshes for the structure. NVIDIA open-sourced Cosmos in January 2025, generating physically-aware video built as training data for robots and self-driving cars.
The Dynamics Model predicts compact state under actions. Compact means a short list of numbers rather than a picture. MuZero, from Julian Schrittwieser and colleagues at DeepMind, is the proof: its model predicted only reward, value and likely moves — never the board. With that it matched AlphaZero at Go, chess and shogi, having never been told the rules. PlaNet and Dreamer, from Danijar Hafner and colleagues, use it in different ways: PlaNet plans inside a model at every step, Dreamer learns behaviour from futures imagined through one.
The Representation predicts embeddings — a list of numbers standing in for an image or clip, with similar things close together — then throws the prediction away. I-JEPA (2023), from Mahmoud Assran and colleagues, hides part of an image and predicts the embedding of the missing piece; it never redraws the pixels. V-JEPA 2 (2025) pre-trains on more than a million hours of video and drives a Franka robot arm in a lab it never saw during training.
The Implicit Model is a different kind of claim altogether. There is nothing to run: the claim is about the inside of a network. Othello-GPT is a small language model trained only to predict legal moves in Othello. It was never shown a board. Kenneth Li and colleagues (2022) found the board was there inside it anyway — a probe could read it out, and when they changed that inner board by hand, the model’s next moves changed too. Neel Nanda and colleagues (2023) showed the board was stored plainly enough for the simplest probe to read, once you asked for “mine” and “theirs” instead of black and white.
One equation, sixty years
The term started somewhere specific: control theory, the engineering maths of keeping a machine on course. A state and an action go in, and the next state comes out. The oldest ancestor is the Kalman filter, from 1960. Rudolf Kalman set it out in A New Approach to Linear Filtering and Prediction Problems. Within a decade it was flying in Apollo’s guidance computer. It is still the maths inside a GPS receiver.
The other half took thirty years. By 1990 three groups reached for the same design within a year of each other. Jürgen Schmidhuber, then in Munich, built a system in two parts: one network learned to predict what the world would do next, a second chose actions, and the first told it what each choice would lead to. He called the first network the world model, and that is where the phrase comes from.
In the same period Richard Sutton built Dyna — a learning agent that practised partly on real experience and partly on experience its own model made up. Sebastian Thrun, Knut Möller and Alexander Linden made the third, in Planning with an Adaptive World Model. All three had the shape of the modern idea: a learned model says what happens next, and something else uses it to decide.
The paper that made the label popular
David Ha and Jürgen Schmidhuber’s World Models (2018) is the paper that made the label popular. Ha was then at Google Brain. The paper showed the 1990 design working on a car-racing game and a level of DOOM, in three pieces: an encoder (squeezes a video frame down to 32 numbers), a dynamics model (predicts where those numbers go next), and a controller (picks actions). The controller was trained almost entirely inside the model’s own imagined rollouts, then moved back into the real game, and it still worked.
Then the term travelled in two directions at once. In 2022 Yann LeCun published A Path Towards Autonomous Machine Intelligence, arguing that predicting every pixel is the wrong job because most pixels are detail nobody can predict. The JEPA family that followed predicts a summary of the next frame instead. The other direction went toward worlds you can see and walk through — OpenAI’s Sora in February 2024, Genie from DeepMind that same month, then Cosmos and Marble in 2025.
The test that was supposed to settle it
One test ought to sort all five out, and it takes about four seconds. Turn around and walk away, then turn back: is it the same room? Something that holds a world inside it keeps the furniture where you left it. I call it the turn-around test, and it does not work the way you’d hope. A large model trained on a lot of video can get very good at drawing a room that matches the one it drew a moment ago. Continuing consistently is part of that. From the outside, the two ways of passing the test look the same.
So the turn-around test cannot sort them, but it was reaching for the right distinction. When you turn away from the chair, it stops being visible, but it does not stop existing. What is there whether or not you can see it is called the state. The part of it you can see right now is called the observation. Working out the rest from the part you can see is called partial observability.
Craik’s small-scale model
None of this is new. A Cambridge psychologist, Kenneth Craik, described the fix in 1943, in a book called The Nature of Explanation. He was twenty-nine when it came out, and two years later he was dead. That was seventeen years before Kalman, and long before there was anything to build it out of.
If the organism carries a ‘small-scale model’ of external reality and of its own possible actions within its head, it is able to try out various alternatives, conclude which is the best of them, react to future situations before they arise… and in every way to react in a much fuller, safer, and more competent manner to the emergencies which face it. — Kenneth Craik, 1943
If you carry a small model of the world in your head, you can test an action before you take it. Every definition on the map is a different answer to what Craik’s small-scale model should hold.
The three questions underneath
One object sits underneath all five definitions. An environment has a state. An agent receives observations that reveal part of it, and from observations and memory it builds an internal state. A model predicts how things could change. The agent acts, the world changes, and a new observation arrives.
Three problems fall out of that loop. How sure should it be? A model that hands back one confident coordinate has thrown away the fact that it was guessing. PlaNet carries a deterministic part (whatever reliably follows from what came before) and a stochastic part (whatever could still go either way) — the model admitting what it does not know.
Do actions change the prediction? What happens next and what happens if I push this are two different questions. Only the second lets you compare options before choosing. A model that answers it is called action-conditioned.
How far can it roll forward? One prediction is rarely enough. To plan, a model has to feed its own answer back in and predict again. Errors pile up along that stretch — nothing dramatic happens at any single step, which is what makes the problem hard to notice, but get step one slightly wrong and step twenty can be somewhere else entirely.
Why the naming is confusing
One thing that can mislead you in the papers is the naming. Ha and Schmidhuber call their third module the controller, meaning the small policy that picks actions. The world model is the module beside it, the one the controller plans inside. If you called the whole thing “the controller” you would name it after the one piece that is plainly not a world model.
A platform is not a category. One product can expose a pixel predictor, a 3D simulator and an action-conditioned model all at once. Ask which one you are being sold and which interface you are about to build against.
The four traditions only noticed they were neighbours when their outputs started to look alike. By then each of them had been using the term for years. By June 2026 World Labs had published A Functional Taxonomy of World Models, and the dictionaries had begun. The next clip that goes past your timeline will be called a world model too. The question to ask is which of the five it is, and whether the person posting it could tell you.
Recap
- The phrase “world model” is used in five senses: Renderer, Simulator, Dynamics Model, Representation, and Implicit Model.
- Four traditions arrived at the phrase independently: reinforcement learning, self-supervised learning, generative video, and interpretability.
- The term was coined by Schmidhuber in 1990 for the prediction network in a two-part architecture; Ha and Schmidhuber’s 2018 paper made the label fashionable.
- All five definitions are different answers to what Craik’s 1943 “small-scale model” should hold.
- Three open questions underlie all of them: uncertainty, action-conditioning, and rollout horizon.