Tokenless

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

Former researchers from OpenAI and Google DeepMind, Liam Fedus and Ekin Dogus Cubuk, discuss their new venture, Periodic Labs. They aim to create an 'AI physicist' by integrating large language models with real-world, iterative experiments, moving beyond simulation to solve fundamental challenges in physics and chemistry, starting with high-temperature superconductivity.

4 min read
Building an AI Physicist: ChatGPT Co-Creator’s Next Venture
source: https://www.youtube.com/watch?v=5FoWFeJCa2A

The Mission: Building an AI Physicist

Periodic Labs is a frontier AI research lab founded by Liam Fedus (co-creator of ChatGPT at OpenAI) and Ekin Dogus Cubuk (former head of materials science at Google DeepMind). Their objective is to accelerate scientific discovery in physics and chemistry by building an “AI physicist.” The core thesis is that while language models have become powerful, their true potential for science can only be unlocked by moving beyond digital tasks and connecting them to the physical world. The goal is to create systems that can design the world around us, discovering and producing materials with desired properties for applications in advanced manufacturing, defense, space, and semiconductors.

The Limitation of Current Scaling Laws for Science

While scaling laws have proven effective for improving language model performance on internet-based data, this paradigm has fundamental limitations for scientific discovery.

  • Distribution Mismatch: The test distribution for models trained on the internet (e.g., text from Reddit) is vastly different from the distribution of physical reality. A model optimized for coding, while a valuable tool, will not spontaneously cure cancer because the necessary knowledge and experimental feedback loop do not exist in its training data. The out-of-domain generalization power law may have too small a slope to be useful for complex scientific problems.
  • Data Scarcity and Quality: Critical experimental data for fields like material synthesis often doesn’t exist publicly. Published literature is heavily biased towards positive results, omitting valuable learning signals from valid negative results. Existing data is also often noisy, with reported physical properties spanning many orders of magnitude. A model trained on this can only replicate the noisy distribution, not converge on a deeper truth.
  • Science is Iterative: Scientific discovery is an iterative process of forming hypotheses, running experiments, and learning from the results. Current LLMs are not trained in this loop. As Cubuk states, “LM can be very smart, but if they’re not iterating on science, they won’t discover science. You know, to be honest, humans won’t either.”

The Solution: Experiment-in-the-Loop and Physical Verification

Periodic Labs’ core innovation is creating a tightly coupled, iterative loop between LLMs, simulations, and a physical robotics lab. This approach replaces digital reward functions with a physically grounded one.

“The ground truth is the experiment… Nature is our RL environment in our setting.”

Instead of optimizing against math problems or human preferences (the basis for ChatGPT), Periodic’s agents are optimized against the results of real-world experiments.

  1. The Process: The AI agent can access tools like quantum mechanics simulators to form hypotheses.
  2. Physical Lab: It then designs and executes an experiment in an automated lab, such as a “powder synthesis lab” where robots mix chemical powders and heat them to create new materials.
  3. Reward Signal: The measured properties of the resulting material provide a direct, verifiable, and hard-to-hack reward signal.
  4. Iteration: The system learns from the outcome—positive or negative—and refines its understanding and subsequent hypotheses, collapsing the “epistemic uncertainty” that simulations alone cannot resolve.

Initial Focus: High-Temperature Superconductivity

The lab’s “north star” goal is discovering a high-temperature, ambient-pressure superconductor. This was chosen for several reasons:

  • Fundamental Impact: Finding a 200 Kelvin superconductor, for instance, would fundamentally update humanity’s understanding of the universe and quantum mechanics even before any commercial application.
  • Aspirational Goal: It’s a challenging problem that requires solving many foundational sub-problems, such as autonomous synthesis and characterization, which are broadly applicable.
  • Technical Suitability: Superconductivity is a phase transition, a robust property of a material’s crystal structure. This makes it less sensitive to factors like defects and microstructure, which are currently difficult to simulate accurately.
  • Unifying Mission: It’s a topic that excites and unites a diverse team of physicists, chemists, and ML researchers.

Technical Approach and Mid-Training

Periodic Labs benefits from the progress in base models from frontier labs and then specializes them. Their process involves taking a powerful pre-trained model and performing mid-training.

Mid-training is the process of continuing the pre-training phase with new, specialized data that was not in the original dataset. For Periodic, this involves injecting vast amounts of:

  • Simulation data
  • Experimental data from their lab
  • Low-level physical descriptions (e.g., crystal structures)
  • High-level semantic descriptions (e.g., synthesis recipes)

The goal is to encode this deep, domain-specific knowledge directly into the model’s weights, moving beyond simple retrieval systems to create a model with a genuine, rich understanding of physics and chemistry.

Building the Team and Commercial Strategy

The founders believe that tackling this problem requires a unique, multi-disciplinary “N of one team” that combines world-class experts in:

  • LLMs (mid-training, RL, infra)
  • Simulations (theoretical physics, coding)
  • Experimentation (solid-state chemistry, physics, automation)

The culture emphasizes curiosity, pragmatism, and creating “APIs” between disciplines to facilitate communication.

While pursuing fundamental science, the commercial strategy is to build “co-pilots for engineers and researchers in advanced industries.” By developing an intelligence layer for companies in space, defense, and semiconductors, Periodic aims to solve critical R&D bottlenecks, accelerate material design workflows, and reduce iteration times. The plan is to land with customers by solving a well-scoped, critical problem and then expand, demonstrating the power of optimizing AI against the physical problems they care about.

Share post

Stay In The Loop! Subscribe to Our Newsletter.

Get updates straight to your inbox. No spam, just useful content.

Recommended for You

Richard Sutton – Father of RL thinks LLMs are a dead end

Richard Sutton – Father of RL thinks LLMs are a dead end

Richard Sutton, a foundational figure in reinforcement learning, argues that Large Language Models (LLMs) are a flawed paradigm for achieving true intelligence. He posits that LLMs are mimics of human-generated text, lacking genuine goals, world models, and the ability to learn continually from experience. Sutton advocates for a return to the principles of reinforcement learning, where an agent learns from the consequences of its actions in the real world, a method he believes is truly scalable and fundamental to all animal and human intelligence.

Scaling and the Road to Human-Level AI | Anthropic Co-founder Jared Kaplan

Scaling and the Road to Human-Level AI | Anthropic Co-founder Jared Kaplan

Jared Kaplan, co-founder of Anthropic, explains how the discovery of predictable, physics-like scaling laws in AI training provides a clear roadmap for progress. He details the two main phases of model training (pre-training and RL), discusses how scaling compute predictably unlocks longer-horizon task capabilities, and outlines the remaining challenges—memory, nuanced oversight, and organizational knowledge—on the path to human-level AI.