I just read LeCun’s recent thoughts on world models. Thoughts on JEPA as a path forward? [D]

Explain Like I'm Normal

Meta's Chief AI Scientist Yann LeCun is doubling down on Joint-Embedding Predictive Architecture (JEPA) as the successor to current LLMs. He argues that today's models lack a 'world model' to understand physical reality, and that JEPA allows AI to learn by predicting missing parts of a scene rather than just guessing the next word. This approach aims to give machines the common sense and spatial reasoning that current generative models fundamentally lack.