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

Key Takeaways

Periodic Labs aims to automate scientific discovery, particularly in physics and chemistry, using AI.
Current AI models, trained primarily on digital data, lack the real-world experimentation necessary for scientific breakthroughs.
Periodic Labs integrates physical lab results as reward functions for AI agents, moving beyond theoretical models.
Scaling laws, successful for language models, are insufficient for complex scientific problems like cracking physics.
Interdisciplinary collaboration between AI and physical scientists is crucial for advancing hard sciences.

Liam Fedus and Ekin Dogus Cubuk, co-founders of Periodic Labs, combine backgrounds from OpenAI and Google DeepMind, respectively, to focus on scientific discovery.
Periodic Labs functions as a frontier AI research lab, utilizing Large Language Models (LLMs), simulations, and real-world experiments.
The company aims to create physically grounded reward functions for AI optimization by integrating experimentation into the AI learning loop.
AI agents at Periodic Labs will use quantum mechanics and solid-state physics tools to perform experiments in a physical lab.
The initial focus is on probing the quantum mechanical energy scale through powder synthesis to discover new materials, such as superconductors and magnets.

Early language models, including ChatGPT, were trained using supervised data and reinforcement learning against human preferences, leading to limitations in mathematical reasoning due to imprecise reward functions.
Current AI models struggle with scientific discovery because they are primarily trained on digital data and lack an iterative experimental process.
These models cannot collapse epistemic uncertainty or learn from noisy or negative experimental results, preventing deeper scientific understanding.
True scientific advancement requires acting in the real world and iterating through experimentation, a capability missing in today's AI systems.
The common AI concept of a 'lab' differs from Periodic Labs' physical laboratory, which explores physics at the quantum mechanical energy scale.

Periodic Labs' initial targets are superconductivity and magnetism, chosen for their impactful sub-goals and fundamental scientific interest.
The pursuit of high-temperature superconductivity is a 'North Star' with goals like autonomous synthesis and characterization, potentially exceeding the current 135 Kelvin record.
These domains are technically advantageous due to their robustness to simulation uncertainties related to microstructure.
Achieving AI scientists involves completing the full scientific loop—experimentation, simulation, data analysis—in these specific domains.
Measuring progress includes synthesizing high-temperature superconductors and directly measuring material properties like ductility and toughness.

The success of scaling laws in language models, from GPT-1 to GPT-5 Pro, predicts emergent capabilities with increased data and compute.
However, this scaling approach has limitations for cracking complex physics problems, requiring a different methodology.
The 'Y-axis' or performance metric for scientific discovery differs significantly from general internet data distributions.
General knowledge is insufficient for specialized scientific breakthroughs; AI must optimize against the specific data distribution of the problem.
Scaling laws in vision models show power-law gains in-domain, but out-of-domain performance often requires centuries of improvement.

Periodic Labs employs a team comprising roughly half Machine Learning (ML) scientists and half physical scientists.
Collaboration is fostered through weekly teaching sessions, encouraging an open culture where all questions are welcomed.
ML researchers learn scientific domains while physical scientists teach LLMs reasoning about physics and chemistry.
This approach bridges the knowledge gap, mirroring the necessity of interdisciplinary expertise for complex discoveries like superconductors.
No single human is likely to possess the combined expertise in physics, chemistry, and synthesis required for complex scientific breakthroughs.

The company culture emphasizes open learning, with computer scientists translating scientific concepts into API-like structures.
Individuals with cross-disciplinary experience serve as bridges between ML and physical science teams.
Joining Periodic Labs does not require an advanced degree in physics or chemistry, emphasizing curiosity and pragmatism.
The primary differentiator for researchers is a strong mission focus on accelerating science.
Periodic Labs seeks candidates who are deeply curious, pragmatic, solution-oriented, world-class in their fields, and possess a strong sense of urgency.

Periodic Labs focuses on deploying AI technology in mission-critical industries such as space, defense, and advanced manufacturing.
The deployment strategy involves identifying specific customer bottlenecks and mapping Periodic's capabilities to solve well-defined problems.
Customer needs include automating complex simulations for product development and integrating data and processes across design pipelines.
The concept of 'mid-training' injects new, real-time knowledge into AI models beyond their initial pre-training cutoff, enhancing performance in specific scientific domains.
Pre-training encodes knowledge directly into AI weights, offering deeper understanding than simple retrieval systems for complex materials.

The university ecosystem is crucial for advancing physical sciences, developing simulation tooling, and novel synthesis methods.
Periodic Labs is establishing an advisory board with experts in superconductivity, solid-state chemistry, and physics to align with long-term research.
Notable advisors include ZX Chen, Steve Kylson, Mercury Canadsides, Chris Wolverton, and Kostia Novoselov.
Periodic Labs is launching a grant program to support academic research relevant to LLMs, agents, synthesis, materials discovery, and physics modeling.
AI models can benefit from integrating geometric reasoning capabilities (e.g., equivariant graph neural networks, diffusion models) to complement linguistic strengths.