A world model is an AI system’s internal picture of how its environment works, learned well enough that the system can predict what will happen next. Instead of reacting only to what it sees right now, an agent with a world model can run a small simulation in its head: if I take this action, the model says the world will look like that. This lets it plan and even practice inside its own imagination rather than only in the costly, slow real world.
The term was crystallized by David Ha and Jurgen Schmidhuber in their 2018 paper “World Models.” They trained an agent to build a compact internal model of a game environment, then showed it could learn a good strategy almost entirely by training inside that learned simulation, the dream, before being dropped back into the actual game. The demonstration made concrete an old idea: that an agent which understands the dynamics of its world can be far more sample-efficient than one that learns purely by trial and error.
Why business readers should care: world models are at the heart of the current frontier where robotics meets foundation models, often called embodied AI or physical AI. A warehouse robot, a self-driving car, or a humanoid that must act in the messy physical world cannot afford millions of real-world mistakes, so giving it a learned model of physics and consequences to plan against is a key to making such systems safe and practical. Several leading labs now treat large-scale world models as a route toward general-purpose robots and simulators for training them.
The honest limits are significant. A world model is only as good as its understanding of the environment, and small prediction errors compound over long imagined sequences, so an agent can confidently plan around a future that will not actually happen. Building models accurate enough for the open physical world, with its endless edge cases, remains hard and unsolved, which is exactly why this is described as a frontier rather than a finished capability.