Yann Lecun: Meta AI, Open Source, Limits of LLMs, AGI & the Future of AI | Lex Fridman Podcast #416

Yann LeCun argues autoregressive LLMs cannot reach human-level intelligence and that joint-embedding predictive architectures (JEPA) are the more promising path forward.

• A 4-year-old receives ~10^15 bytes through vision vs. ~2×10^13 bytes in all publicly available text LLMs train on — sensory data dwarfs language data.
• LeCun says LLMs fundamentally cannot plan, reason, maintain persistent memory, or model the physical world — four traits he considers essential to intelligence.
• Generative video prediction has failed for 10 years at Meta FAIR; predicting pixel-level video futures is computationally intractable unlike discrete text tokens.
• JEPA (Joint-Embedding Predictive Architecture) learns abstract representations rather than reconstructing raw pixels, allowing systems to discard unpredictable noise like blowing leaves.
• V-JEPA (video JEPA) is the first Meta system to learn good video representations and can detect physically impossible events in video, a primitive form of intuitive physics.
• Hierarchical planning — breaking a trip from NYC to Paris into nested sub-goals down to muscle control — is completely unsolved in AI; no system learns appropriate abstraction levels.
• LeCun dismisses AI doom scenarios as requiring false assumptions: superintelligence will not be a single event, guard rails improve iteratively like turbojet reliability, and open-source AI prevents power concentration.
• He compares banning open-source AI to the Ottoman Empire banning the Arabic printing press for 200 years, calling proprietary AI control a bigger danger than misuse.

Guests: Yann LeCun, Chief AI Scientist at Meta, professor at NYU, Turing Award winner · 2024-03-07 · Watch on YouTube