Ask a chatbot to plan a route across a room and it will produce a fluent, confident description that may or may not survive contact with an actual wall. It has read about rooms; it has never been in one. A world model is the attempt to fix exactly that — to give an AI an internal, predictive model of an environment, so it can imagine what will happen next and plan against it, rather than guessing from text.

It is one of the most active ideas in AI right now, and one of the clearest dividing lines in the debate over what comes after the large language model.

The core idea: imagine, then act

A world model learns to answer a simple-sounding question: given the current situation and an action, what happens next? Once a system can predict that reliably, it can do something powerful — run the future forward in its head. Instead of trying an action in the real world to see if it works (slow, expensive, sometimes dangerous), it can simulate thousands of possibilities internally and pick the best one. That is why world models are often cited as delivering an order of magnitude or more — roughly 10-to-100× — better sample efficiency than traditional trial-and-error reinforcement learning: most of the trial happens in imagination.

The idea isn't new — it echoes how people think, running a quick mental simulation before we act. What's new is that it's starting to work at scale.

Two paths: understand, or generate

The field splits into two camps that disagree about what "predict the future" should even mean.

The generation path renders the future, frame by frame. DeepMind's Genie line learns an interactive, playable environment from ordinary video — Genie 3, shown in 2025, generates navigable worlds at 720p that stay coherent for minutes. OpenAI's Sora and NVIDIA's Cosmos platform sit here too: the model literally paints what comes next, and that rolling video is the simulator. It is visually stunning, and expensive — you are predicting every pixel, most of which don't matter for a decision.

The understanding path argues that is a waste. Yann LeCun's JEPA (Joint Embedding Predictive Architecture) predicts in an abstract latent space instead of in pixels — capturing that a ball will fall and roughly where, without rendering every frame of it falling. Danijar Hafner's Dreamer line works similarly for control: DreamerV3 famously used a single set of settings to learn Atari, robotics, 3D navigation and even to collect diamonds in Minecraft, with no per-task tuning — an early sign that one general world model was plausible.

Why it matters now: bodies, not just screens

The reason world models jumped from research curiosity to headline in the last year is robotics and physical AI. A model that only predicts text can't run a warehouse robot; a model that predicts consequences can. Meta's V-JEPA 2, trained on more than a million hours of video, was used to control a real robot arm on tasks it had never seen, without lab-specific retraining — the kind of zero-shot transfer that has been the field's holy grail. NVIDIA's Cosmos is explicitly pitched as a foundation layer for exactly this: teaching machines the physics before you put them in a body.

It is also why some researchers — LeCun among them — argue the world model, not the ever-larger language model, is the more likely road to general intelligence. Text captures what people have written about the world. A world model has to capture the world.

The honest limits

None of this is solved. Generated worlds still drift and hallucinate over long horizons — errors compound, and a simulated future that looks convincing for ten seconds can quietly become nonsense at thirty. Latent-space models are more efficient but harder to inspect: it is not always clear what they've learned to predict. And the two camps sharply disagree on which approach scales, which means a lot of confident roadmaps rest on an unsettled bet.

The short version: a language model learns to talk about the world. A world model tries to learn the world itself — and whether that is the next big step or a very expensive detour is, right now, one of the most interesting open questions in AI.