The Next Big Idea HUMAN COMPATIBLE: Can We Control Artificial Intelligence?
29 snips
Sep 28, 2023 
Stuart Russell, an expert in artificial intelligence, discusses the rapid development of AI, the potential termination of the human species, AGI, language models, specialized processes in the human brain, the potential of super intelligent AI, AI in education and science, and regulating dangerous technologies.
AI Snips
Chapters
Transcript
Episode notes
LLMs Are A Preview Not AGI
- Large language models give people a foretaste of what general-purpose intelligence will feel like.
- They are not AGI yet and experts don't claim they are.
Scaling Alone Won't Reach Real-World Intelligence
- Scaling compute alone doesn't solve core AI challenges because weak methods hit combinatorial limits.
- More hardware raised step depth modestly but can't reach the thousands of steps needed for complex tasks.
How Transformers Learn Language Patterns
- Transformers predict next tokens by learning sequence probabilities from vast text windows.
- Training tweaks quintillions of parameters to optimize next-word prediction, producing surprising abilities.