Unsupervised Learning with Jacob Effron Ep 44: Co-Founder of Together.AI Percy Liang on What’s Next in Research, Reaction to o1 and How AI will Change Simulation
16 snips
Oct 3, 2024 Percy Liang, a Stanford professor and co-founder of Together AI, dives into the fascinating evolution of AI from simple token prediction to complex autonomous agents. He discusses the challenges of AI safety and interpretability, emphasizing a holistic approach to mitigating risks. The conversation also touches on generative agents in simulations, their potential in social dynamics, and the importance of evaluating AI systems with new methodologies. Plus, Liang shares insights into AI's future role in fields ranging from education to music, showcasing the technology's transformative power.
AI Snips
Chapters
Transcript
Episode notes
From Believable to Valid Simulations
- Current simulations are visually believable but lack validity to accurately reflect real-world behaviors.
- Valid simulations can unlock new applications like policy testing and scientific studies.
Challenges and Innovations in AI Evaluations
- Current AI evaluations are complicated by unknown training data overlap and superficial judgments.
- Using language models themselves for evaluation can improve coverage and depth.
Interpretability Challenges With Closed Models
- Interpretability is harder today since weights and training data are often unavailable.
- Mechanistic interpretability aids scientific curiosity; explanations help debugging and real-world trust.