A world model is a learned, internal simulator of how an environment behaves: given the current situation and a proposed action, it predicts what happens next. Instead of reacting only to what it sees right now, an agent equipped with a world model can run that simulator forward in its own "imagination" — rolling out hypothetical futures, scoring them, and choosing actions by their predicted consequences rather than by trial and error in the real world. It is the machine-learning embodiment of the idea that intelligence rests on a predictive model of reality, and it has become the connective tissue linking reinforcement learning, video generation, and self-supervised representation learning into one research program.\n\nA world model learns the environment's dynamics so the agent can foresee the consequences of an action before committing to it. Formally it approximates the transition distribution p(s_{t+1} | s_t, a_t) — and usually a reward model too — turning a black-box environment into a differentiable, queryable predictor. This is the sharp line between model-based and model-free reinforcement learning: a model-free agent (DQN, PPO) learns only a policy or value function by directly interacting with the world, while a model-based agent first learns to simulate the world and then plans or trains inside that simulation. The payoff is sample efficiency — real interaction is slow, dangerous, or expensive (a robot arm, a fab tool, a car), whereas simulated rollouts are cheap and infinitely repeatable.\n\nModern world models predict in a compact latent space, not in raw pixels. Reconstructing every pixel of the future is wasteful and brittle, so the dominant designs (RSSM, Dreamer) use an encoder to compress each observation into a low-dimensional latent state, learn the dynamics between latents, and only decode back to observations when needed. Predicting in latent space is faster, generalizes better, and forces the model to keep the task-relevant structure while discarding noise like exact textures or lighting. The recurrent latent then carries a running belief about the world — including parts the agent cannot currently see — which is what lets it plan over long horizons from partial observations.\n\nThe signature trick is "learning in imagination": the agent trains on trajectories the model hallucinates, not on real experience. Once the latent dynamics are accurate, an agent like Dreamer generates thousands of imagined rollouts entirely inside the world model and optimizes its policy and value function against those dreamed futures, touching the real environment only to keep the model honest. This decouples policy learning from the cost of real interaction and is why world-model agents reach strong performance with dramatically fewer environment steps — the expensive real world is queried sparingly, and the cheap learned simulator does the heavy lifting.\n\nWorld models now span three fields that used to be separate. In reinforcement learning they are the planner's simulator (Dreamer, MuZero-style latent models). In generative AI they have become large video models — Sora, Genie, and their kin learn an implicit, controllable simulator of visual reality and can be driven by actions, producing playable or steerable environments. In self-supervised learning, joint-embedding predictive architectures (JEPA) take a different stance: rather than generating the future pixel-by-pixel, they predict the representation of the future in latent space, sidestepping the wasted capacity of pixel reconstruction. All three are the same bet — that predicting the world is the route to understanding it.\n\n| Approach | What it predicts | Prediction space | Primary use |\n|---|---|---|---|\n| Dreamer / RSSM | Next latent state + reward | Compact latent | Model-based RL, planning in imagination |\n| MuZero-style | Latent dynamics tuned for value | Value-relevant latent | Planning without a given simulator |\n| Sora / Genie | Future video frames, action-conditioned | Pixels / tokens | Generative, controllable environments |\n| JEPA | Representation of the future | Latent embedding | Self-supervised world understanding |\n\n``svg\n\n``\n\nThe unhelpful way to see a world model is as just another neural network bolted onto a reinforcement-learning agent. The useful way is to see it as a shift in where the intelligence lives: from a reactive policy that maps observations to actions, to a learned simulator the agent can query, plan inside, and dream with — reserving precious real-world interaction for keeping that simulator accurate. Compress perception into a latent, learn how latents evolve under actions, and you can train an agent almost entirely in imagination, generate controllable video environments, or learn representations by predicting the future without ever drawing a pixel. Read world models through a learned-simulator-you-plan-inside lens rather than a bigger-policy-network lens, and the encoder, the latent dynamics, the imagination rollout, and the JEPA-versus-generative split stop looking like separate tricks and resolve into a single idea: predict the world in order to act in it.