I just read LeCun’s recent thoughts on world models. Thoughts on JEPA vs LLMs?

Explain Like I'm Normal

Yann LeCun continues to push Joint-Embedding Predictive Architecture (JEPA) as the necessary successor to current LLMs. He argues that while models can talk about the world, they lack a fundamental 'world model' that understands physical constraints and logic. JEPA aims to learn by predicting missing parts of abstract representations rather than just guessing the next pixel or word.