LeCun defines world model in latent space

While many AI companies promote generative video models or LLMs as world models, LeCun's definition exposes their core limitation: predicting pixels or tokens is not the same as understanding physical dynamics. True machine intelligence requires predictive planning, not just realistic-looking generation.

* Pixel-Level vs. Latent Space: Generative video simulators are computationally inefficient and focus on irrelevant details (like background noise), whereas JEPA models capture essential abstract physics.

* Action-Conditioning is Crucial: A world model must predict the consequences of specific actions to enable planning and control, rather than just passively predicting the next frame.

* The Path to Common Sense: By learning from raw sensory data (such as video and physics simulations) without reconstructing pixels, world models offer a far more promising path to human-level common sense and reasoning than text-only LLMs.