a16z Podcast• Sep 30, 2025• 51:48Interview

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

From a16z Podcast

Doge(Guest)•null(Guest)•Liam and Doge(Guest)

Executive Summary

Periodic Labs, a frontier AI lab founded by a co-creator of ChatGPT and a former DeepMind physics lead, aims to accelerate scientific discovery in physics and chemistry.
Their core strategy involves a tight feedback loop between Large Language Models (LLMs), simulations, and a physical laboratory to create a "physically grounded reward function" from real-world experimental data.
The company's long-term "North Star" is the discovery of a high-temperature superconductor, a grand challenge that drives foundational research in automated synthesis and characterization.
Their near-term commercial strategy is to build and sell AI "co-pilots" to accelerate R&D for engineers and researchers in advanced industries like aerospace, defense, and semiconductors.

8 quotes

Concerns Raised

Existing LLMs are incapable of novel scientific discovery without being grounded in new experimental data.
The necessary high-quality experimental data for training AI in their target domains does not currently exist in public literature.
There is a risk of getting siloed in a specific domain (e.g., superconductivity) without developing a generalizable 'AI scientist' platform.

Opportunities Identified

Discovering a high-temperature superconductor, which would be a monumental scientific and technological breakthrough.
Building and commercializing AI 'co-pilots' to accelerate R&D in multi-trillion dollar advanced industries.
Creating a repeatable, AI-driven process for materials discovery that can be applied across various domains beyond their initial focus.

Key Themes

Physically Grounded AI for Scientific Discovery

The central thesis is that current AI models, trained on existing digital data, cannot make novel physical discoveries. Periodic Labs addresses this by building a physical lab to generate experimental data, which serves as a real-world reward function to train and validate their AI agents.

This approach aims to overcome the limitations of simulation and historical data, potentially enabling AI to explore and discover new materials and physical phenomena that are outside the current human knowledge base.

The AI Scientist Flywheel

Periodic Labs is creating a closed-loop, iterative system where an LLM-based agent proposes hypotheses, runs simulations, directs automated experiments in a physical lab, and then learns from the real-world results. This flywheel of hypothesis, experiment, and feedback is designed to rapidly accelerate the scientific process.

If successful, this flywheel could exponentially increase the rate of materials discovery and physical R&D, compressing research timelines from years to weeks or days.

Balancing Moonshots with Commercial Viability

The company pursues the ambitious, long-term goal of discovering a high-temperature superconductor while simultaneously developing a practical, near-term business model. They plan to commercialize AI "co-pilot" tools for R&D-intensive industries to generate revenue and fund their foundational research.

This dual strategy de-risks the capital-intensive nature of frontier research, creating a sustainable path to achieving a major scientific breakthrough by leveraging commercial applications along the way.

Overcoming Data Scarcity in Science

The founders argue that existing scientific literature is insufficient for training powerful AI models due to high noise, a lack of published negative results, and domain-specific context. Their in-house automated lab is critical for generating the high-quality, targeted, and proprietary data needed to train their systems effectively.

This highlights a key bottleneck for applying AI to the physical sciences and justifies their strategy of vertically integrating data generation, which could become a significant competitive advantage.

Fusing ML and Physical Science Cultures

The lab's success depends on the deep integration of machine learning experts with physicists and chemists. This involves physical scientists teaching AI models how to reason about quantum mechanics and ML scientists building the systems to execute and learn from physical experiments.

This cross-disciplinary collaboration is essential for translating abstract scientific principles into effective AI training methodologies and building a system that can genuinely contribute to scientific inquiry.

Get started free

Topics

AI for Science Periodic Labs Large Language Models (LLMs)Reinforcement Learning (RL)Reward Functions Physical AI Materials Science Superconductivity Magnetism Automated Synthesis Scientific Discovery R&D Acceleration AI Co-pilots Frontier AI Industry-Academia Collaboration

Processed Apr 6, 2026 yt-dlp + mlx-whisper + Gemini