Yann LeCun: LLMs won't lead to AGI, bets on world models instead

Yann LeCun, Turing Award winner and former chief AI scientist at Meta, has once again publicly dismissed the idea that large language models (LLMs) alone can lead to artificial general intelligence (AGI). In a post on X, he noted that people are increasingly realizing AI systems are far from human-level intelligence and learning abilities. LLMs, he claims, compensate for their lack of common sense, understanding of reality, and limited reasoning and planning skills by accumulating enormous amounts of declarative knowledge. They excel in language-based domains like math and programming, but LeCun insists they are not creative mathematicians, software architects, or scientists. Instead, he sees them as tools to assist human creation, not autonomous thinkers.

LeCun proposes a different path: hierarchical "world models" that learn by observing the real world—through video and environmental interaction—rather than just predicting the next word. He believes a general method for training such systems could emerge within 12–18 months, enabling deeper understanding of causality and long-term planning. Since 2025, LeCun has led his own startup, Advanced Machine Intelligence Labs (AMIL), based in Paris, which has already raised $1 billion for foundational research with no immediate commercial product. This positions him against much of the AI industry that pursues AGI by scaling up LLMs and compute clusters.