Training an AI Scientist with Feedback from Reality, w- Liam Fedus & Ekin Dogus Cubuk (from a16z)

Episode Summary

Executive Summary: This episode of the Cognitive Revolution features a cross-post from the A16Z podcast, where host Anjane Mitta interviews Liam Fettis (co-creator of ChatGPT) and Eken Dojes Chubuk (former head of materials science at Google DeepMind), co-founders of Periodic Labs. They discuss their $300 million seed round to build an AI-powered physical laboratory that connects AI-generated hypotheses to real-world experiments, using nature as the reinforcement learning signal. Their North Star is discovering a high-temperature superconductor, with sub-goals including autonomous synthesis and characterization. The conversation emphasizes the need for high-quality experimental data, a mission-driven team culture, and a path to commercializing an intelligence layer for advanced manufacturing.

Main Topics: The Need for Physical Experimentation in AI-Driven Science (Priority: 5/5): Periodic Labs argues that existing scientific literature lacks high-quality experimental data and negative results, making it insufficient for training foundation models in physics and chemistry. They advocate for automated labs that generate data through real-world experiments, using nature as the reward function. Periodic Labs' North Star: High-Temperature Superconductivity (Priority: 5/5): The company's primary goal is to discover a superconductor above 200 Kelvin, which would be a fundamental scientific breakthrough. This goal requires solving sub-problems like autonomous synthesis and characterization, making it a unifying challenge for the team. Team Composition and Culture (Priority: 4/5): Periodic Labs prioritizes intense curiosity and mission alignment over advanced degrees. They foster a 'no stupid questions' culture with weekly teaching sessions where ML researchers, physicists, and chemists learn from each other. The team of ~30 includes world-class talent across LLMs, experiments, and simulations. Commercialization Strategy: Intelligence Layer for Advanced Manufacturing (Priority: 4/5): Beyond the North Star, Periodic Labs plans to commercialize its technology as an intelligence layer for companies in space, defense, and semiconductors. They aim to accelerate R&D by automating simulations, design processes, and data integration, using a land-and-expand approach. Mid-Training and Scaling Laws (Priority: 3/5): The founders discuss mid-training (continuing pre-training on new data) to inject physics and chemistry knowledge into LLMs. They argue that scaling laws hold but require optimizing against the target distribution, which for physical sciences necessitates experimental data. Academic Partnerships and Grant Program (Priority: 3/5): Periodic Labs is establishing an advisory board and a grant program to support academic research in LLMs, AI agents, synthesis, and materials discovery. They recognize the importance of academia for developing simulation tools and novel synthesis methods.

Key Arguments: Existing scientific literature is insufficient for training AI models due to noisy data, lack of negative results, and the need for iterative experimentation. Scaling laws hold but require optimizing against the target distribution; for physical sciences, this means generating experimental data in the loop. High-temperature superconductivity is an ideal North Star because it requires solving many sub-problems and has fundamental scientific value. The team's culture of curiosity and cross-disciplinary learning is essential for bridging the gap between ML and physical sciences. Commercialization will focus on solving well-scoped problems for advanced manufacturing customers, using a land-and-expand strategy. Mid-training on simulation and experimental data can inject physics knowledge into LLMs, improving their performance on scientific tasks.

Data Points: Seed investment: $300 million - Led by Andreessen Horowitz for Periodic Labs. Current highest ambient pressure superconductor temperature: 135 Kelvin - Baseline for measuring progress towards high-temperature superconductivity. Team size: ~30 - Current size of Periodic Labs team. Years since founders met: 8 years - Liam and Doge met at Google Brain eight years ago.

Pivotal Quotes: "Ultimately, science is driven against experiment in the real world. And so that's what we're doing with Periodic Labs. We're taking these precursor technologies and we're saying, okay, if you care about advancing science, we need to have experiment in the loop." — Liam Fettis: Explaining the core thesis of Periodic Labs: using real-world experiments as the reward function for AI. "If we could find a 200 Kelvin superconductor, even before we make any product with it, to be able to see such quantum effects at such high temperatures, I think would be such an update to people's view of how they see the universe." — Eken Dojes Chubuk: Describing the fundamental scientific value of the North Star goal. "The amount that even our best physicist doesn't know about physics is much bigger than the amount that they know about physics." — Eken Dojes Chubuk: Explaining why advanced degrees are not required and why curiosity is valued over existing knowledge.

Implications: Periodic Labs' approach could revolutionize materials science and chemistry by creating AI systems that autonomously explore the physical world. Success would accelerate R&D in advanced manufacturing, defense, and semiconductors, potentially leading to breakthroughs like room-temperature superconductors. The emphasis on mission-driven, cross-disciplinary teams may influence how AI-for-science companies are built.