A world model, in artificial intelligence, is an internal representation that an agent uses to predict how its environment will change over time, including in response to its own actions, allowing the agent to plan, imagine outcomes, and evaluate choices without needing to act in the real environment to observe the consequences. The term has roots in classical control theory and model-based Reinforcement learning, where an agent's model of environment dynamics is used explicitly for planning, but it has taken on broader significance in recent years as researchers have proposed that learning a predictive world model from raw sensory data, particularly video, could be a path toward more general and robust intelligence than training on text alone.
Origins in reinforcement learning
Model-based reinforcement learning has long used explicit or learned dynamics models to simulate future states before acting, in contrast to model-free approaches that learn a policy or value function directly from trial and error. A widely cited 2018 paper by David Ha and Jürgen Schmidhuber, titled simply "World Models," demonstrated that an agent could learn a compressed generative model of a video-game environment and then train a policy almost entirely inside that learned simulation, popularizing the term in its modern AI sense and building on earlier work on predictive, self-modeling agents.
Video prediction and generative world models
A newer line of research treats large-scale video prediction itself as a way to learn a world model implicitly, without a hand-designed simulator. Google DeepMind's Genie, introduced in 2024, learned to generate playable, interactive 2D environments purely from unlabeled internet video, without access to underlying game engine states or action labels, demonstrating that action-conditioned world dynamics could be learned from passive observation. Video-generation systems such as Sora have similarly been described by their developers as early, general-purpose "world simulators," on the argument that accurately predicting how a video should continue requires an implicit model of physics, object permanence, and cause and effect, though critics note that such models frequently violate basic physical consistency and that fluent video generation does not necessarily imply a reliable or generalizable internal physics model.
JEPA and LeCun's proposal
Yann LeCun has proposed the Joint Embedding Predictive Architecture, or JEPA, as an alternative to both purely generative video models and standard Autoregressive model language modeling. Rather than predicting raw pixels or tokens, JEPA-style models predict future states in an abstract, learned representation space, an approach LeCun argues is more efficient and more capable of ignoring irrelevant detail, such as texture noise, while still capturing the causal structure needed for planning. LeCun has framed world models built this way as a central missing ingredient for Embodied AI and for progress beyond the capabilities of today's Large language model systems, positioning the idea in direct contrast to labs betting primarily on scaling text-based Pretraining.
Significance
World models are increasingly discussed as relevant not only to Robotics and game-playing agents but also to broader questions about Artificial general intelligence, since an agent with an accurate internal model of its environment can, in principle, plan over long horizons and generalize to novel situations more efficiently than one that only reacts to immediate input. As of the mid-2020s, world-model research remains an active and unsettled area, with substantial disagreement over how much of it can be learned from passive video versus how much requires embodied interaction.