Dwarkesh Podcast Eliezer Yudkowsky - Why AI Will Kill Us, Aligning LLMs, Nature of Intelligence, SciFi, & Rationality
67 snips
Apr 6, 2023 
Eliezer Yudkowsky, a prominent AI safety researcher, shares his insights on the potential risks of advanced AI. He argues passionately for the urgent need to align AI with human values to prevent catastrophic outcomes. Yudkowsky discusses the intricacies of large language models and their challenges in achieving alignment. The conversation delves into the ethical dilemmas of enhancing human intelligence and the unpredictable nature of human motivations as AI evolves. He also reflects on the philosophical implications of AI's impact on society and our future.
AI Snips
Chapters
Books
Transcript
Episode notes
Childhood Influences
- Eliezer's parents tried raising him Orthodox Jewish, but he only learned to pretend.
- The ethos of science fiction books resonated more with him, shaping his beliefs.
AI as Actors
- Training AIs on human text creates an actor, not a human.
- They learn to predict human behavior, not embody human psychology.
AI's Superhuman Ability
- An AI trained on all human text doesn't become the average human.
- It becomes capable of simulating any individual, which is far more dangerous.
