LLMs accumulate a lot of declarative knowledge.
That's how they can answer a large number of questions. That's also why they require so many parameters and hence so much memory. But one can replace the accumulation of knowledge by *understanding*
That what JEPA world models

→ View original post on X — @ylecun