"The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis Training an AI Scientist with Feedback from Reality, w- Liam Fedus & Ekin Dogus Cubuk (from a16z)
149 snips
Oct 2, 2025 Liam Fedus, a former OpenAI research leader who co-created ChatGPT, and Ekin Doğuş Çubuk, an ex-Google DeepMind head of materials research, discuss their groundbreaking venture, Periodic Labs. They aim to create AI-driven physical labs that connect real-world experiments with AI insights. Key topics include 'experiment-in-the-loop' strategies, measuring robotic success through superconductivity metrics, and the importance of cross-disciplinary teams. They highlight a unique culture prioritizing curiosity and collaboration, aiming to reshape scientific discovery through AI.
AI Snips
Chapters
Transcript
Episode notes
Experiment As The Ultimate Reward
- Periodic Labs ties AI reasoning directly to real-world experiments to create a physically grounded reward function.
- They treat nature as the RL environment so experiments become the ultimate correctness signal for models.
From Tire-Flipping To Talking Quantum
- Liam and Doge first connected years ago at Google Brain, even flipping a giant tire together.
- Their casual chats about quantum mechanics and superconductivity seeded the idea for Periodic.
Optimize Models With Verifiable Rewards
- Improve model reliability by optimizing against precise, verifiable reward functions rather than vague assistant goals.
- Build reward signals that check physical correctness, not just helpfulness or friendliness.