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

Episode Summary

Executive Summary: The episode centers on Periodic Labs’ thesis that AI can only advance science if it is trained and optimized against real-world experiments, not just digital labels. Co-founders Liam Vettus and Doge Chubuk explain why physics and chemistry are ideal starting points, how labs can create physically grounded reward functions for RL, and why superconductivity and magnetism are the first ambitious targets. They also discuss team-building, academia partnerships, and near-term enterprise copilots for advanced industries.

Main Topics: Experiment-in-the-loop AI for science (Priority: 5/5): Periodic Labs is building an AI research lab where experiments, simulations, and LLMs are tightly coupled so models learn from the physical world, not just static text or internet data. Why physics and chemistry are the right frontier (Priority: 5/5): The founders argue that physics offers fast feedback, strong verifiability, and clear reward signals, making it the best entry point for an AI physicist that can eventually generalize to chemistry and materials discovery. Reward functions, RL, and mid-training (Priority: 5/5): They contrast current AI training with Periodic’s approach: using experiments as the reward function, adding domain knowledge through mid-training, and leveraging high-compute RL to teach scientific reasoning. Superconductivity and magnetism as North Stars (Priority: 4/5): High-temperature superconductivity is both a scientific milestone and a practical benchmark. It serves as a motivating target that forces progress in autonomous synthesis, characterization, simulation, and iteration. Data quality, noisy labels, and negative results (Priority: 4/5): The conversation emphasizes that scientific datasets are noisy, often incomplete, and biased toward positive results. Periodic aims to generate better data and capture valid negative results as learning signals. Commercial applications and deployment (Priority: 4/5): Beyond discovery, Periodic wants to provide copilots for advanced industries such as semiconductors, space, defense, and manufacturing, helping companies accelerate design and simulation workflows. Team, culture, and academia partnerships (Priority: 3/5): The company is building a multidisciplinary team spanning ML, simulation, physics, and chemistry, while also partnering with academia via advisory boards and grant programs to stay close to frontier science.

Key Arguments: AI progress in science requires physical experimentation in the loop; digital-only training is insufficient for discovering new scientific truths. Physics and chemistry are especially suitable because they are verifiable, iterative, and often supported by simulators and measurable outcomes. Current frontier models are good at logic and math because they were trained on verifiable rewards, but they are not yet trained to reason and act effectively in the physical sciences. The existing scientific literature is too noisy and incomplete for reliable model training; many valid negative results are missing, and some properties vary by orders of magnitude. Scaling laws still matter, but they must be applied to the right distribution; if the target task is too far from the training data, the slope may be too weak to be useful. Periodic’s approach is to change the training environment itself so the target distribution becomes closer to what the model sees during training. Superconductivity is an ideal first benchmark because it is scientifically profound, technically robust to some simulation gaps, and provides a clear measurable target. The startup’s near-term business value comes from copilots and intelligence layers for advanced industries that need better simulation, design, and materials workflows. Building the right team requires deep collaboration between ML researchers and physical scientists, with weekly cross-training and “no stupid questions” culture. Academia remains essential because it produces much of the foundational simulation tooling and scientific insight that Periodic wants to operationalize.

Data Points: Periodic Labs team size: ~30 - Founders said the company is roughly 30 people at the time of the interview. Ambient-pressure superconductivity benchmark: ~135 Kelvin - They cited this as the current best ambient-pressure superconductor temperature to beat. Hypothetical superconductor target: 200 Kelvin - A 200 K superconductor was used as an illustrative breakthrough goal. LLM knowledge cutoff: mid-training use case - Mid-training was described as inserting new knowledge into a model after pre-training when freshness or missing domain knowledge matters. Early ChatGPT training paradigm: RLHF - Liam described ChatGPT as originally built with supervised data plus reinforcement learning from human preferences. Functional team split: roughly half ML / half physical scientists - The discussion described a team composition balanced between machine learning and physics/chemistry backgrounds.

Pivotal Quotes: "Ultimately, science is driven against experiment in the real world." — Liam Vettus: Explaining why Periodic Labs centers real-world experiments as the ground truth reward function. "Nature is our RL environment in our setting." — Liam Vettus: Describing how experiments replace internet text or human preference as the optimization target. "If we could find a 200 Kelvin superconductor... that in itself says so much about the universe that we didn't know yet." — Doge Chubuk: Justifying superconductivity as both a scientific and inspirational North Star.

Implications: If Periodic succeeds, AI could shift from analyzing science to actively discovering it, accelerating materials, chemistry, and industrial R&D. The broader effect could be faster innovation loops and a new class of real-world scientific copilots.