Dwarkesh Podcast Ilya Sutskever – We're moving from the age of scaling to the age of research
3919 snips
Nov 25, 2025 
Join Ilya Sutskever, co-founder of OpenAI and Stability AI, as he dives into the world of AI and machine learning. He discusses the intriguing concept of model jaggedness, explaining why AI sometimes behaves inconsistently. Ilya contrasts pre-training and reinforcement learning, emphasizing the importance of generalization and the barriers faced by AI. He also explores how emotions can serve as value functions and proposes new strategies for ensuring AGI is aligned with human values. Insights on continual learning and the future of superintelligence add depth to this fascinating conversation.
AI Snips
Chapters
Transcript
Episode notes
Eval-Focused RL Produces Brittle Models
- Current RL fine-tuning can overfit to evals and create brittle behaviors that oscillate between mistakes.
- Pre-training's broad data avoided that manual reward-targeting and may explain superior robustness on many tasks.
Specialist Training Undermines Generalization
- Training a model solely to excel at a narrow benchmark produces specialists who generalize poorly.
- Broad pre-training supplies vast, natural data that acts like many hours of practice and helps generality in ways RL often doesn't.
Value Functions Improve RL Efficiency
- Value functions give intermediate feedback so agents learn faster without waiting for terminal rewards.
- They can short-circuit credit assignment and accelerate RL in long-horizon tasks like coding or math.