Translated data: Yann LeCun points out that while large language models (LLMs) have their uses, they cannot accurately understand the physical world and lack support for fundamental intelligent features such as long-term memory, reasoning, and planning. He discusses the possibility of building models with a deep understanding of the world and introduces the advantages of the Joint Embedding Predictive Architecture (JEPA) over LLMs. JEPA is capable of better extracting abstract representations, enabling the system to fundamentally learn the abstract characteristics of the world.