"Large language models predict the next token. That is not intelligence — that is statistical autocomplete at scale. Real intelligence predicts the state of the world. The model must have a world model, not just a word model. That third layer — the world — is what everyone keeps skipping."
Invented convolutional neural networks and proved that hierarchical feature learning could match human perception on structured tasks. Now leads Meta AI's fundamental research agenda, arguing that current LLMs are architecturally incapable of genuine reasoning and proposing JEPA — Joint Embedding Predictive Architecture — as a path toward world-modeling intelligence.