World models train more efficiently by predicting patterns, not raw data

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

Source: x.com ↗