
Futurology Today's AI Is Too Dumb to Be Dangerous (with Yann LeCun and Dawn Nakagawa)
13 snips
Nov 17, 2026 Yann LeCun, French computer scientist, Turing Award winner, and AMI Labs chair, explores why today’s language models lack genuine world understanding. He discusses JEPA, machine learning through video, embodied intelligence, planning, and the limits of chain-of-thought reasoning. The conversation also tackles AI safety, regulation, open models, Project Tapestry, and the geopolitical risks of concentrated control.
AI Snips
Chapters
Books
Transcript
Episode notes
Intelligence Means Adapting When You Do Not Know
- Yann LeCun defines intelligence as rapidly adapting to unfamiliar situations, not storing facts or skills.
- A teenager learns to drive in roughly 20 hours, despite machines having billions of hours of driving data.
Why JEPA Predicts Meaning Instead Of Pixels
- LLMs predict discrete symbols effectively but struggle with unpredictable pixel-level video details.
- JEPA learns abstract representations that preserve predictable structure, such as branches and leaves, while discarding each leaf’s exact position.
World Models Turn Prediction Into Planning
- A world model predicts how actions change the world, enabling planning toward goals.
- Babies gradually learn gravity and object dynamics through observation and interaction, knowledge machines still struggle to reproduce.



