World Model

A world model is a learned internal simulator: a neural network that predicts how an environment will change in response to actions, letting a system "imagine" outcomes before acting. The idea comes from model-based reinforcement learning — Ha and Schmidhuber's 2018 work trained agents inside their own dream of the environment — and DeepMind's Dreamer line showed agents learning complex control almost entirely in imagination. The term now also labels large video-generation models like Sora and Genie, on the argument that predicting plausible future frames requires implicitly learning physics, object permanence, and cause and effect. Skeptics counter that generating convincing video is not the same as maintaining a consistent causal model, and glitches expose the gap. For builders, world models matter as the research bet behind robotics, autonomous driving, and interactive 3D environments — domains where trial and error in the real world is too slow or dangerous.