World models are already used in domains like spatial simulation and autonomous driving, and can complement LLMs where prediction and causality are crucial.

While acknowledging practical challenges and safety considerations, the author argues world models may be the most promising path to physical intelligence if matured.

World models produce embeddings that represent the predicted state of the environment after an action, enabling accurate prediction and potential causal reasoning rather than mere sentence generation.

LLMs and world models tackle different problem classes: LLMs excel at token-based problem solving, while world models forecast environment states and causality after actions to support future planning.

Looking ahead, world models could enable billions of years of simulated learning in a short real-world timeframe, reducing real-world exploration risk but facing issues like sim-to-real gaps, limited action-labeled data, and partial observability.