a16z Podcast: "Building an AI Physicist: ChatGPT Co-Creator’s Next Venture"

Date: September 30, 2025
Host: Andreessen Horowitz (A16Z)
Guests:

• Anjane Mitha (A16Z General Partner, Host/Moderator)
• Liam Vedas (Co-founder, Periodic Labs; ChatGPT Co-Creator)
• Doge Chubuk (Co-founder, Periodic Labs; Former DeepMind Physics Lead)

Episode Overview

In this episode, the A16Z Podcast dives deep into the creation and mission of Periodic Labs, a new frontier research group led by Liam Vedas and Doge Chubuk—veterans of OpenAI and DeepMind. The discussion centers on their ambitious goal to build an “AI physicist”: models that not only understand and converse about science, but can meaningfully design experiments, discover new materials, and accelerate real-world research. The episode traces the roots of Periodic Labs, explores the technical and cultural barriers in fusing AI with physical science, and examines the impact this could have on fields from advanced manufacturing to quantum materials.

Key Discussion Points & Insights

1. The Origin Story of Periodic Labs ([01:35–04:40])

• Serendipitous Partnership: Liam and Doge share how their collaboration began not in the lab, but at a Google gym—bonding over flipping a heavy tire. Coincidentally, this evolved into years of casual discussions about quantum mechanics and advancements in AI.
• Shared Vision: Doge notes noticing LLMs’ growing impact on his daily physics work, both in learning and in coding, and the itch to elevate AI from a learning tool to a full research collaborator.
  • “We were looking at improvements on language models … seeing scaling laws within physics, within chemistry … and the goal of this technology is to accelerate science, accelerate physical R&D.” —Liam Vedas, [04:00]

2. Mission & Technical Approach ([04:40–12:04])

• Experiment-in-the-loop AI: Periodic aims to close the gap between digital and physical science by making experiment — not just math or code — the foundation for training models. This gives their systems a "physically grounded reward function."
  • “We're creating a physically grounded reward function that becomes the basis on which we're optimizing… Nature is our RL environment in our setting.” —Liam Vedas, [05:24]
• Lab as RL Environment: Unlike prior models trained solely on digital data, Periodic uses real labs—specifically, powder synthesis and quantum mechanical systems—to couple LLMs with high-throughput experimentation.
• Iterative, Scientific Inquiry: Doge stresses that iterative cycles of theory, simulation, and experiment—mirroring human discovery—are essential, emphasizing the value of both positive and negative results.
  • “It’s very uncommon to publish negative results... and that's a learning signal. And this is something that our lab will produce as well.” —Liam Vedas, [12:04]

3. Why Physics? Why Now? ([13:23–17:28])

• Choice of Domain: Initial focus is on quantum mechanics, superconductivity, and magnetism—areas that combine foundational scientific value with tractable, automated experimental methods.
• Technical Readiness: Improvements in reinforcement learning, LLMs, and robotics in the last few years have made this leap possible. Industry silos and the need for highly interdisciplinary teams have been key obstacles.
  • “The technology that we think is necessary to do it has really just emerged in the last couple of years.” —Liam Vedas, [16:55]

4. Scaling Laws, Domain Shift, and the Necessity of Real-World Data ([17:28–22:59])

• Limitations of Pure Scaling: The team cautions that simply scaling up data and compute (à la GPT) won’t suffice for true scientific breakthroughs—experimental feedback is needed because generalization from digital corpora is weak for out-of-domain science tasks.
• Data Quality Issues: Existing literature datasets are incomplete, noisy, and bereft of negative results. The team argues for building custom datasets directly tied to their experimental work.
  • “The experimental data we want actually doesn't exist.” —Doge Chubuk, [20:39]

5. Why Superconductivity & Magnetism as Initial "Beelines"? ([23:15–26:01])

• North Star with Milestones: High-temperature superconductivity is both a grand challenge and a task with numerous sub-goals (automation, synthesis, characterization) crucial for fully autonomous scientific systems.
  • “To find a 200 kelvin superconductor ... would be such an update to people's view of how they see the universe.” —Doge Chubuk, [24:04]
• Technical Tractability: Superconductivity's phase transition makes it relatively robust to experimental imperfections, making it a promising candidate for early robotic discovery.

6. Commercial Pathways & Industry Impact ([27:00–29:41])

• Analogous to AI Programming: Just as coding copilots unlocked productivity for developers, Periodic aims to build intelligent copilots for engineers and researchers in advanced R&D industries, especially those working with physical processes.
• Commercial Imperative: The ambition is for Periodic to be a "wildly successful commercial entity," funding and accelerating the science it was founded to advance.
  • “We want to accelerate advanced manufacturing ... Become an intelligence layer for all these teams ...” —Liam Vedas, [28:24]

7. Blending ML and Physical Science Cultures ([29:41–35:40])

• Active Cultural Bridging: The team is intentionally building cross-domain intuition—ML experts learn physics, physicists learn ML. Weekly teaching sessions, open Q&A culture, and “bridge" staff (with both backgrounds) are key mechanisms.
  • “No stupid questions ... Just like the dumbest physics question, the dumbest ML question.” —Liam Vedas, [32:34]
• Recruiting Ethos: Emphasis is on curiosity, mission alignment, and willingness to learn over formal credentials.
  • “Even our best physicist doesn't know about physics compared to all there is to know … everyone has so much to learn.” —Doge Chubuk, [33:49]

8. Evaluation, Deployment, and Real-World Integration ([36:30–41:50])

• Clear Metrics: Success can be directly measured—e.g., synthesizing superconductors at unprecedented temperatures, or demonstrable material improvements.
• Pragmatic Industry Deployment: Land-and-expand strategy, solving urgent, well-scoped problems for slow-moving but essential industries (space, defense, semis) and gradually deepening the partnership.
  • “It's not coming in and saying, hey, we're going to transform your fab line on day one. ... It's like, no, we're going to solve a really critical problem.” —Liam Vedas, [37:43]

9. Technical Stacks: Mid-Training, Tooling, and Composability ([41:50–46:15])

• "Mid-training" Concept: Injecting fresh, domain-specific data (especially experimental) into models post-pretraining, but before RL-based post-training, to improve scientific reasoning and modeling.
  • “Mid train is basically you're taking new data, new knowledge … and you continue pre train.” —Liam Vedas, [42:00]
• Leverage Instead of Reinvention: Modern ML models and open-sourced simulation tools from academia and industry are used modularly, not rebuilt.
  • “Neural net as a tool to these agents is something that immediately accelerates our work.” —Liam Vedas, [45:34]

10. The Role of Academia & Community ([46:15–51:00])

• Academic Partnerships: Collaboration with universities remains vital—academic labs pioneer key simulation methods, cultivate high-level scientific thinking, and identify meaningful research directions.
• Initiative Examples: Starting an advisory board including leading figures in superconductivity and materials synthesis, launching a grant program for academic research in AI and the physical sciences.
  • “We want to accept grant proposals and enable and support … work that's going to help community, especially in relation to LLMs, agents, synthesis, materials discovery, physics modeling.” —Doge Chubuk, [49:27]

11. What Makes a Great Team Member at Periodic? ([51:00–52:29])

• Desired Traits: Deep curiosity, process rigor, solution-orientation, and a palpable sense of urgency. Innovators across machine learning, experimental science, and simulation are all welcome.
  • “If the candidate feels like a sense of urgency for improving these physical systems, discovering these amazing materials ... they would be a good fit.” —Doge Chubuk, [52:06]

Notable Quotes & Memorable Moments

• On the transition from digital assistants to AI scientists:
  • “Chatbots was like a great milestone along the way, but we really want to see technology out in the world.” —Liam Vedas, [04:00]
• On reward functions in science vs. digital models:
  • “Nature is our RL environment in our setting.” —Liam Vedas, [05:24]
• On the importance of negative results in science:
  • “A valid negative result is very valuable … that's a learning signal. And this is something that our lab will produce.” —Liam Vedas, [12:04]
• On the challenges of interdisciplinary learning:
  • “The amount that even our best physicist doesn't know about physics is much bigger than the amount that they know about physics.” —Doge Chubuk, [33:49]
• On urgency:
  • “We want these technologies not in 10 years, … but we want them ASAP.” —Doge Chubuk, [52:06]
• On the future vision:
  • “The goal is create this repeatable system, prove it, and then just go through the different domains that way.” —Liam Vedas, [26:01]

Timestamps for Key Segments

| Segment | Start Time | |-------------------------------------------------|------------| | Periodic Labs origin story | 01:35 | | Mission and experiment-in-the-loop approach | 04:40 | | Lab as RL environment, reward functions | 05:24 | | The necessity of physics, real-world verification| 09:27 | | Limitations of current data, scaling laws debate| 17:28 | | Why superconductivity is the “beeline” | 23:15 | | Commercial pathways, copilots for engineers | 28:24 | | ML & science cultural integration | 30:39 | | Recruiting and team culture | 33:41 | | Real-world deployment, customer perspectives | 36:30 | | Mid-training and data integration | 41:50 | | Composability & leveraging open tools | 44:58 | | The academic/industry interface | 46:15 | | Advisory board and grant program | 49:27 | | Criteria for joining Periodic | 51:00 |

Conclusion

Periodic Labs stands at the intersection of AI and physical science, aiming to create the first AI systems that can reason, experiment, and innovate in the real world as working scientists do. This episode unpacks both the vision and granular tactics underpinning this revolution, making clear the technical, cultural, and commercial challenges ahead—and the immense potential value if they succeed.