Don't Worry About the Vase Podcast On Dwarkesh Patel's Second Interview With Ilya Sutskever
Dec 3, 2025
In this enlightening conversation, Ilya Sutskever, co-founder and chief scientist of OpenAI, delves into the intricate world of AI and deep learning. He shares insights on why models perform well on benchmarks yet struggle in real-world applications, framing emotions as key value signals. Sutskever discusses the importance of continual learning post-deployment and the challenges in aligning AI with human values. He even speculates on the timelines for achieving superhuman learners, painting a picture of both potential and uncertainty in our AI-driven future.
AI Snips
Chapters
Transcript
Episode notes
Eval Scores Don't Equal Economic Impact
- Models score well on benchmarks but underdeliver economically because they are brittle and reward-hacked.
- Expect an impact gap while engineers chase eval-driven metrics instead of robust real-world behavior.
Emotions As Dense Value Signals
- Ilya frames emotions as value functions that provide dense information and training signals.
- Emotions guide exploration and resolve uncertainty in ways scalar rewards alone cannot.
We're Entering An Age Of Research
- Scaling pre-training is becoming less productive and data is running out, so new research ideas are needed.
- We are returning to an 'age of research' where novel techniques matter more than raw scale.