a16zBuilding an AI Physicist: ChatGPT Co-Creator’s Next Venture
CHAPTERS
From ChatGPT and DeepMind physics to a new goal: an “AI physicist”
The conversation frames Periodic Labs’ ambition: building AI systems that can design and improve real-world physical systems, not just generate text. The hosts set context on the founders’ backgrounds and why this direction matters for science and industry.
The founders’ origin story—and why LLMs felt ready for physics
Fedus and Cubuk recount meeting at Google Brain and reconnecting through shared interests like quantum mechanics and superconductivity. They describe how LLMs became increasingly useful for physics work (learning, coding, simulations), motivating the leap to make LLMs first-class tools for research.
What Periodic Labs is building: experiments + simulations + LLMs as one system
Periodic is positioned as a frontier AI lab focused on physics and chemistry, emphasizing tightly coupled experimentation, simulation, and model training. The central idea is to generate high-throughput, high-quality data and use it to iteratively improve models and scientific outcomes.
Replacing digital reward functions with “nature” as the reward signal
Fedus explains how current LLM progress relies on verifiable digital rewards (math/code graders) and human preference models (RLHF). Periodic aims to ground reinforcement learning in physical reality, where experiments provide hard-to-game feedback and correct simulator deficiencies.
Why a physical lab is essential (and what kind of lab Periodic means)
Cubuk contrasts “AI lab” with a literal experimental lab and argues physics is the next frontier of verifiable reasoning beyond math/logic. Periodic targets quantum-mechanical regimes relevant to materials and chemistry, starting with automatable powder synthesis that robots can run at scale.
Why existing models can’t do discovery: iteration, uncertainty, and missing negatives
The founders argue discovery requires iterative action and learning from outcomes, including failures—something current models lack. They highlight noisy literature data, broad uncertainty in reported properties, and the scarcity of published negative results as key blockers to training reliable scientific models.
How to measure progress: superconductors, properties, and ‘designing the world’
They propose concrete, physics-grounded metrics of success—especially raising superconducting critical temperatures. Applied metrics include measurable improvements to material properties (strength, ductility, toughness) and demonstrating the ability to design materials for specified targets.
Why now: frontier AI enables a rare “N-of-1” cross-disciplinary team
Fedus explains the difficulty is not only technical but organizational: combining top-tier ML, experiment, and simulation talent in one coordinated effort. The needed AI techniques are recent, and the relevant data is siloed across industries, requiring new data creation and integration.
Scaling laws aren’t enough: the Y-axis, domain shift, and absent datasets
They affirm belief in scaling laws but argue performance depends on what’s measured and the target distribution. Out-of-domain improvements can be too slow (small slopes), and for many materials tasks, the necessary experimental data either doesn’t exist or is too noisy to learn from—making physical verification and targeted data generation essential.
Why superconductivity first: a North Star with many enabling sub-goals
Superconductivity is chosen because it is scientifically profound, motivational, and technically suited to current limits (phase transition robustness). Pursuing higher-temperature superconductors forces development of a full autonomous loop: synthesis, characterization, simulation, and agentic reasoning—creating reusable infrastructure for broader domains.
Startup reality: commercial path via ‘copilots’ for advanced-industry R&D
They address the tension between moonshot science and building a viable business. Periodic’s near- and mid-term product direction is AI copilots for engineers and researchers in industries with massive R&D budgets (semiconductors, defense, space, manufacturing), reducing iteration time in physical workflows.
Integrating cultures: shared teaching, ‘API thinking,’ and training physics reasoning
The team blends ML researchers and physical scientists through weekly cross-training and a culture of asking basic questions. Physicists help specify reasoning steps for mid-training/RL (how to reason about quantum mechanics), while ML researchers learn domain tools and objectives—using “API-like” mappings between inputs, outputs, and evaluations.
Deployment into traditional industries: start scoped, go beyond retrieval, use mid-training
They outline a pragmatic deployment approach for slower-moving industries: solve a well-scoped critical problem with clear evaluations and expand from there. Discussions include limits of retrieval-only approaches, challenges of permissioned data when moving to training, and how mid-training injects fresh, domain-specific knowledge into models.
Academia partnerships, research tooling, and what makes a great Periodic researcher
They emphasize strong ties to academia for simulation tools, scientific task definitions, and reasoning strategies (e.g., thinking in symmetries). Periodic plans an advisory board and a grant program, and they close with hiring criteria: mission alignment, curiosity, pragmatism, world-class strength in a pillar, and urgency.
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome