a16z Podcast Building an AI Physicist: ChatGPT Co-Creator’s Next Venture
306 snips
Sep 30, 2025 
Liam Fedus, co-creator of ChatGPT and now co-founder of Periodic Labs, teams up with Ekin Dogus Cubuk, former head of materials science at Google DeepMind, to discuss their innovative approach to automating physical science discovery. They dive into the importance of integrating LLMs with experimental data to tackle complex problems like superconductivity and magnetism. Their lab focuses on real-world experimentation, emphasizing the necessity of iteration and learning from failures. With a unique hiring philosophy and a commitment to bridging ML and experimental cultures, they aim to revolutionize how we engage with hard sciences.
AI Snips
Chapters
Transcript
Episode notes
Experiment As The Reward Function
- Periodic couples experiments, simulations, and LLMs to create a closed-loop scientific discovery process.
- They treat real-world experiments as the reward function that trains agents to design physical systems.
From Flipping Tires To Physics Talks
- Liam and Doge first met at Google Brain and bonded over physics conversations and a shared practical task flipping a large tire.
- Their ongoing chats about quantum mechanics and superconductivity later inspired them to build Periodic together.
Train Against Real Experiments
- Do train models with physically grounded reward signals tied to experiment rather than solely human preference proxies.
- Use lab-generated experimental outcomes to error-correct simulators and guide RL optimization.