Sep 30, 2025, 1:00 PM ETInterview

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

Periodic Labs was founded by Liam and Doge (co-creators of ChatGPT and DeepMind physicists) to create an "AI physicist" that advances physical science through real-world experimentation.
The founders met eight years ago at Google Brain while attempting to flip a massive tire together, bonding over shared interests in quantum mechanics and superconductivity.
The company's core thesis is that accelerating science requires an "experiment-in-the-loop" where real-world physics, rather than digital ground truth (like math graders), serves as the primary reward function for AI agents.
Periodic aims to replace current digital reward functions with a "physically grounded reward function" where Nature itself acts as the RL environment.
The founders identified a gap in current LLM capabilities: while models excel at logic and math, they lack the iterative scientific inquiry method required to discover new physical phenomena without direct experimental feedback.

Periodic operates a frontier AI research lab that tightly couples Large Language Models (LLMs), high-throughput simulations, and automated physical experiments.
The lab utilizes "mid-training," a process distinct from standard pre-training or post-training, to inject new, specific experimental and simulation data into model weights to address domain-specific gaps.
Current models are deemed insufficient for material discovery due to "epistemic uncertainty," noisy literature data, and a lack of published negative results, which the new lab will actively generate as learning signals.
Agents are trained to perform the full scientific loop: reading literature, running simulations, conducting physical experiments, and iterating based on real-world failure or success.
The team plans to leverage tools like neural nets, graph neural networks, and diffusion models as specific "tools" for agents to handle geometric reasoning and synthesis recipes.
Unlike scaling laws which predict performance on in-distribution data, Periodic argues that out-of-distribution generalization (e.g., physics from internet text) has a "power law" slope too shallow to be useful for discovery without direct domain-specific training.

The first scientific target is "high-temperature superconductivity," specifically aiming to surpass the current ambient pressure record of ~135 Kelvin to potentially reach 200 Kelvin.
Superconductivity was selected as the initial "north star" because it is a robust phase transition, unites the team across physics and chemistry, and offers immediate fundamental insight into quantum mechanics.
The initial technical pipeline focuses on "powder synthesis," a low-cost, automatable method where robots mix powders and heat them to discover new materials like superconductors and magnets.
Periodic views superconductivity as a gateway to a broader "AI scientist" that can generalize across solid-state physics, materials science, and chemistry.
The commercial strategy involves a "land and expand" approach, starting with "co-pilots for engineers" in space, defense, and advanced manufacturing to accelerate R&D workflows and reduce iteration time.
The company explicitly rejects the idea that current frontier labs will eventually "crack" physics through pure scaling of pre-trained models on internet data, necessitating a dedicated physical lab.

Periodic currently employs roughly 30 people, comprising a "fractal" mix of ML scientists, experimentalists, simulators, and operational staff.
The team is structured to bridge the "simplex" between pure ML, pure experimental physics, and pure simulation, utilizing "bridge people" who understand multiple domains.
Hiring prioritizes candidates with a deep sense of mission urgency, a desire to "make contact with reality," and world-class expertise in one of the three pillars (ML, experiment, simulation).
The culture emphasizes "no stupid questions," with weekly cross-disciplinary teaching sessions where ML researchers teach RL loops and physical scientists teach quantum mechanics and synthesis history.
Advanced degrees in physics or chemistry are not strictly required; the founders argue that any candidate, regardless of background, faces a similar learning curve regarding the specific complexities of the domain.

Deployment strategy targets high-R&D-budget industries (semiconductors, aerospace, defense) by solving specific, scoped problems like automating simulations and integrating design pipelines, rather than attempting immediate total transformation.
Periodic is launching an advisory board with experts like ZX Shen (Stanford), Steve Kelson, and others to align industry work with long-term academic research directions.
A grant program will be established to fund academic work that is too specialized for industry but essential for the broader community, particularly in LLMS, synthesis, and material discovery.
The company leverages open-source simulation tools and academic-developed code (often Fortran-based) while adding its own proprietary RL and automation layers.
Periodic acknowledges that current models are "terrible at scientific analysis" and views their approach as building upon base model advances through targeted, high-compute reinforcement learning and mid-training.