WORLD MODELS

World models are AI systems that learn a function from a (current world state, action) pair to the next world state, letting an agent predict how its environment will evolve and simulate the outcomes of actions before taking them. The term spans two paradigms: internal world models that predict the future at a high, semantic level (cognitive sense, in the spirit of LeCun's JEPA family) and external world models that aim to simulate reality at full visual fidelity (e.g. Genie-2, GAIA-1). Both are foundational for embodied AI: agents can plan and act using internal models while learning inside external simulators.