Training an AI Scientist with Feedback from Reality, w- Liam Fedus & Ekin Dogus Cubuk (from a16z)
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
Periodic Labs, co-founded by former OpenAI and Google DeepMind researchers Liam Fetest and Ecken Dojeschubuk, has raised $300 million in seed funding to build AI systems that can autonomously conduct physics and chemistry experiments. (01:35) The company aims to create an "AI physicist" that can iterate directly against physical reality, using experiment results as the reinforcement learning signal rather than relying solely on digital rewards. Their North Star goal is discovering a high-temperature superconductor, which they believe will require achieving countless sub-goals including autonomous synthesis and characterization. (32:36) Unlike traditional AI labs that optimize against math and code graders, Periodic Labs is building physical laboratories where nature itself serves as the ground truth for training their models.
- Main Theme: Building AI systems that can advance scientific discovery by conducting real-world physics and chemistry experiments, moving beyond digital-only AI applications to create systems that iterate directly against nature's ground truth
Speakers
Liam Fetest
Former VP of Post-Training Research at OpenAI and co-creator of ChatGPT. Fetest was instrumental in developing the reinforcement learning from human feedback pipeline that transformed raw language models into useful assistants. He has deep expertise in scaling laws, reasoning systems, and creating effective reward functions for training AI systems.
Ecken Dojeschubuk
Former Head of Material Science and Chemistry Research at Google DeepMind. Dojeschubuk led physics teams focused on quantum mechanics, solid-state physics, and materials discovery. He has extensive experience in computational physics simulations and understanding the intersection of machine learning with physical sciences.
Anjaney Mida
General Partner at Andreessen Horowitz (a16z) who led the $300 million seed investment in Periodic Labs. Mida focuses on frontier AI investments and has deep expertise in evaluating and supporting companies at the intersection of AI and scientific research.
Key Takeaways
Physical Experiments Are Essential for AI Scientific Discovery
Current AI models trained on existing scientific literature have fundamental limitations because published research lacks negative results and contains noisy data spanning orders of magnitude. (15:41) As Dojeschubuk explained, experimental data for materials properties can vary so wildly that training on it produces models with no predictive power. The only way to collapse this "epistemic uncertainty" is through actual experimentation. This means AI systems need to iterate against physical reality, not just digital simulations, to develop true scientific intuition and make breakthrough discoveries.
Scaling Laws Apply to Physics But Require Domain-Specific Training
While scaling laws continue to hold for language models, they don't guarantee performance on out-of-domain tasks like physics and chemistry. (25:47) The team discovered that out-of-domain performance also improves as a power law, but the slope may be so small that it's practically useless. This explains why general models struggle with scientific analysis - the knowledge simply doesn't exist in their training distribution. The solution is to make your target domain as close to your training distribution as possible by generating high-quality experimental data directly.
Mission Alignment Trumps Advanced Degrees
Periodic Labs prioritizes intense curiosity and mission alignment over formal credentials, not requiring advanced degrees for team members. (44:00) As Dojeschubuk noted, even their best physicist doesn't know vastly more about the total scope of physics than someone starting from zero, because the field has become so vast. They maintain a "no stupid questions" culture with weekly teaching sessions where ML researchers, physicists, and chemists learn from each other. This approach recognizes that advancing science requires collaboration across disciplines rather than siloed expertise.
Real-World Applications Drive Sustainable Progress
While high-temperature superconductivity serves as their North Star, Periodic Labs plans to commercialize progress by serving as an intelligence layer for advanced manufacturing companies. (38:36) Industries like space, defense, and semiconductors spend massive R&D budgets on materials iteration but lack good AI tools for their workflows. By solving immediate commercial problems while building toward their ultimate scientific goals, the company ensures sustainable funding and creates a feedback loop where commercial success accelerates fundamental research.
Academic-Industry Collaboration Is Critical
The company is launching both an advisory board with leading academics and a grant program to support university research. (59:22) They recognize that much of the foundational simulation tooling they use was developed in academia, and that industry alone is blind to important analysis methods and thinking strategies used by premier scientists. This collaboration ensures they benefit from academic advances while contributing back to the broader scientific ecosystem, creating a virtuous cycle of knowledge sharing and innovation.
Statistics & Facts
- Periodic Labs raised $300 million in seed funding led by Andreessen Horowitz, making it one of the largest seed rounds in AI history. (00:32) This massive funding reflects investor confidence in their approach to AI-driven scientific discovery.
- The current record for ambient pressure superconductivity is 135 Kelvin, which serves as Periodic Labs' benchmark to beat. (16:43) Breaking this record would represent a significant breakthrough in materials science with transformative applications.
- The team consists of roughly 30 people, with about half being ML scientists and half physical scientists. (46:03) This balanced composition reflects their strategy of deeply integrating machine learning expertise with domain knowledge in physics and chemistry.
