Yann LeCun: LLMs won't lead to AGI

Why it matters

Turing Award winner and Meta AI chief Yann LeCun in a spirited exchange with Kara Swisher. In his signature contrarian style, LeCun explains why autoregressive LLMs won't lead to AGI — 'A baby learns more by watching the world than any LLM learns from trillions of tokens'. He proposes JEPA (Joint Embedding Predictive Architecture) as the alternative and explains why open-source AI isn't dangerous but essential for safety.

Key takeaways

• 'LLMs will never get us to AGI' — pure text training lacks physical world perception and interaction; this is the fundamental bottleneck.
• A baby learns more about the world through vision in 9 months than any LLM from trillions of tokens — vision + interaction is the foundation of intelligence.
• JEPA architecture: not predicting the next token, but predicting abstract representations of the world — the path to world models.
• Open-source AI is essential for safety: 'closed source = a few people controlling super-powerful AI' is the biggest risk.
• The four AI players (OpenAI, Google, Anthropic, Meta) shouldn't become an oligopoly — diversity itself is protection.