Yann LeCun told Futurology host Don Nakagawa that intelligence is learning new skills fast with little interaction. He pointed to a camera moving up a tree trunk as proof that pixel prediction cannot work.
You know leaves are coming, but you cannot predict every leaf in wind.
LeCun said an LLM works because it scores every word in a finite dictionary and samples the next one, a loop he called autoregressive prediction. Video has no finite dictionary, he said, and most pixel detail is unpredictable, so the same trick trains the model to do the impossible.
His fix is JEPA, joint embedding predictive architecture, which learns an abstract representation that drops unpredictable detail. LeCun, founder and chairman of AMI Labs, said the system predicts that representation, enough to expect branches and leaves without fixing each leaf position.
A world model, he said, uses that kind of prediction to forecast the result of an action and then plan, the way babies learn that unsupported objects fall after about nine months of observation. He contrasted that with models that rebuild 3D scenes to generate consistent video, work associated with Fei-Fei Li, which he said aims at pixels for media rather than action.
LeCun conceded he spent about 15 years failing at pixel-level video prediction himself, so JEPA is still a research bet. Production chat keeps context by refeeding history because the models themselves are stateless with no built-in memory of prior requests, per Onely, which is why memorization scales but robots still cannot match a cat jumping to a shelf.




