RT @MatthieuWyart: LLMs learn by predicting tokens. World models (JEPA, data2vec) learn by predicting their own abstractions. Which needs m…