Summary: This video covers the vision and strategy of Periodic Labs, a new company co-founded by Liam Fedus, a ChatGPT co-creator from OpenAI, and Ekin Dogus Cubuk, a materials and chemistry leader from Google DeepMind. Going beyond simple language models, they present a new paradigm of an "AI scientist" that repeatedly experiments to pursue real discoveries in physics and chemistry. Built on the conviction that iterative learning grounded in "experimental verification" drives true scientific innovation, they mobilize AI, laboratory work, interdisciplinary talent, and industry experience to chart a new path for science automation.

1. The Origin Story and Goals of Periodic Labs

Periodic Labs started with the grand goal of "science automation based on experimentation." Early in the video, the two co-founders share how they first met while flipping a giant tire together at Google, and how they'd long shared a conviction about the need for scientific discussion and collaboration.

"Science is ultimately driven by verification in the real world through experiments. That's why at Periodic Labs, we're trying to tightly couple these technologies with experimentation."

They emphasize that "the advancement of language models (LLMs) is becoming increasingly important in coding, physics, and chemistry," and argue for the necessity of an "AI physicist that receives direct experimental feedback" beyond simple chatbots.

2. What Periodic Labs Does and Its Approach

Periodic Labs is a frontier AI research lab that uses LLMs to accelerate progress in physics and chemistry. The core idea is to organically connect data-generating laboratories, simulations, and LLMs so that AI actually learns by repeatedly running experiments.

"Based on 'physical reward functions' combining experiments, simulations, and LLMs, we optimize AI to find optimal results in the real world. Nature itself is essentially the RL (reinforcement learning) environment."

This approach goes beyond the monotonous verification of existing internet data, textbook data, or code scoring, creating a new concept where AI does "real science" by directly verifying results in the lab.

3. Why "Experimental Verification" Is Decisive

Current LLMs are strong at logic, math, and coding thanks to single, clear "digital rewards" that are easy to verify. But real science is built upon repeated experiments and countless trial and error — that's the differentiator.

"Even the smartest person, locked in a room without experiments, will never discover anything new. The essence of science lies in iterative experimentation and how you learn from the process."

Periodic Labs points out that real-world data in physics and chemistry contains noise, inaccuracy, and many unpublished negative results, and emphasizes that their "AI lab" encompasses even these limitations.

"Accumulating negative results (failures) in the lab is also important. The real results we need aren't all contained in existing papers or positive outcomes."

4. Specific Research Goals and Team Building Strategy

Periodic Labs' immediate "North Star" is the discovery of high-temperature superconductors (above 200K). In this process, AI autonomously manages experimental procedures like powder synthesis, explores theories, creates actual samples, and tests performance.

"The search for high-temperature superconductors involves autonomous synthesis, autonomous characterization, and repeated simulations and experiments. If this becomes possible, the scientific impact alone would be enormous, not to mention the industrial implications."

To achieve this, the team is structured around LLM experts, experimental/synthesis experts, and simulation experts, each forming sub-teams with emphasis on close communication and cross-disciplinary learning.

"It was important to assemble 'world-class' small teams for each domain. The lab, simulation, and LLM areas each need their own 'innovation leader'!"

5. Scaling Laws, Limitations, and Periodic's Differentiation

The existing AI "scaling laws" (better models through bigger models and more data) hit limitations. When experimental data is scarce or existing datasets (literature, simulations) diverge significantly from real-world objectives, the performance curve flattens too much to achieve practical progress.

"Unlike fields like coding where data is abundant, experimental data is scarce, and the knowledge we actually need often isn't in the dataset. That's why we believe we need to evolve the dataset's target to match our goals through repeated real experiments."

Periodic's core differentiation is a "mid-training" technique that repeatedly injects experiments and simulations, and a structure that directly channels practical industry feedback into model improvement.

6. Commercial Success and "Science-Industry" Convergence

Periodic Labs pursues its ambitious goal of an AI scientist while simultaneously aiming to quickly solve real-world industry problems. The goal is for AI to become a key companion for improving R&D efficiency and outcomes across engineering, chemistry, and materials industries.

"Silicon Valley centers on computer work, but in space, defense, semiconductors, and other massive physical R&D, iterative experimentation is essential. AI tools are still lacking in these areas. That's exactly where we're stepping in."

Periodic provides AI-powered "scientist copilots" to actual sites in semiconductors, new materials, steel, and space/defense industries, adopting a strategic "land and expand" approach that repeatedly absorbs actual workflows and industry feedback.

7. Team Culture, Ideal Candidates, and the Synergy of Compound Expertise

Periodic fosters a challenging yet enjoyable collaborative culture where lab scientists and LLM/ML engineers "freely exchange even the most seemingly foolish questions."

"Our team has weekly sessions where we learn from each other. Physicists and chemists think about how to teach LLMs, and engineers study the detailed processes of physics and chemistry. A culture of mutual learning is truly important."

Rather than specific degrees or backgrounds, "genuine passion for scientific innovation and problem-solving, experimental curiosity, and rapid execution" are cited as the most important qualities.

"Even the most brilliant scientist doesn't know far more than they do know. At Periodic, 'being full of unknowns' is normal, and the experience of collaborating and learning together is truly enjoyable."

8. Traditional Industries, AI, and the Realities of Deployment

How quickly AI scientists are adopted in actual industries is also a key issue. Periodic acknowledges that core customers (space, defense, semiconductors) may be slow to change, but has adopted a practical strategy of focusing intensively on solving the most "critical" on-site problems, co-defined with customers through clear evaluation metrics.

"Rather than trying to overhaul all of R&D from the start, we believe it's effective to solve one or two truly important problems well to prove our capabilities, then expand gradually."

Simulation automation, data pipeline optimization, and improving the efficiency of real knowledge learning/application are the first collaboration points with industry partners.

9. Mid-Training and the Fundamental Leap in Model Performance

Periodic emphasizes that beyond simple pre-training and post-training, "**mid-training" is crucial.

"Mid-training means directly pouring in knowledge, experimental results, and newly acquired experiment/simulation data that existing models don't have."

This includes existing papers, simulations, and experimental results, as well as complex structures (crystal structures, manufacturing processes) and higher-order concepts (geometric and semantic knowledge). Through this process, they can "inject genuine 'expertise' and on-the-ground understanding into LLMs."

10. Collaboration with Academia and Open Innovation Strategy

Thorough collaboration with existing academia and university laboratories is essential for fusing AI and real experimentation.

"Important simulation tools, new synthesis methods, and experimental insights come largely from universities. We also run advisory boards and research funding programs to broaden our connections."

Specifically, collaborations with researchers at Stanford, Northwestern, and other institutions are underway across superconductivity, synthesis, and material characterization. Some of the most important ideas and skills are acknowledged to be rooted in academia.

11. Ideal Candidates and Final Message

The ideal Periodic team member possesses "**genuine scientific curiosity, passion for learning by facing reality in the lab, openness to new modes of collaboration and innovation, and an 'urgency to make things happen fast.'"

"If you want to change science fast — not in 10 years, but right now — and you have 'world-class' capabilities in any one area, we'd love to hear from you!"

The conversation concludes by encouraging passionate talent to join the journey of leading "the great transformation of science automation."

Conclusion

This video heralds the coming of a true "AI scientist" era beyond simple AI chatbots, concretely showing how Periodic Labs is turning innovation into reality through on-site experimentation, AI technology, industry convergence, interdisciplinary communication, and talent development. The trend of refocusing technological progress on "experimentation" and "real problem-solving" is worth watching closely for the massive changes it could bring across industries and science.

Key Keywords:

• Experimental verification, AI scientist, mid-training, high-temperature superconductors, R&D automation, interdisciplinary convergence, industry application, lab data, collaboration, ideal candidates, scaling laws, real-world feedback, science automation, open innovation