Yann LeCun says LLMs are limited for reaching AGI because text/next-token prediction can't capture the continuous, high-dimensional physical world. He advocates world models — joint embedding predictive architectures (e.g., VJepa) that learn abstra...