Interview, Fireside Chat
Fireside Chat with Ekin Dogus Cubuk, Co-CEO of Periodic Labs | RAISE Summit 2026
- LLMs are expected to revolutionize programming and mathematics while machine learning continues to play a significant role in physics, leveraging advances in simulations, RL, and high-throughput experimentation to automate scientific discovery at scale.
- Current limitations in computer context storage and the absence of negative results in publications create data gaps, necessitating a dynamic laboratory capable of iteratively testing with the universe rather than relying solely on static datasets.
- A new Bay Area laboratory will deploy robots to perform chemical experiments and transport samples alongside traditional tools like microscopy and diffraction, controlled by LLMs to increase throughput and consistency beyond human capabilities.
- The company aims to reach a throughput of 1,000 experiments per day by the end of this summer, a target that matches the organization's original goal and is expected to transform understanding of solid-state chemistry and powder synthesis.
- Future operations will focus on learning precise ingredients, ratios, and temperatures for synthesizing unknown chemicals, with a specific short-term plan to improve semiconductor process engineering in areas such as deposition, etching, lithography, and interface materials.
- Deep engagement in the semiconductor industry is projected to yield results by the end of this year, with the facility capable of performing all inorganic synthesis within its constraints, explicitly excluding radioactive or toxic elements like uranium.
- Long-term objectives include establishing a deeper recursion of intelligence where improved chips enable AI advancements that further improve chips, potentially creating a new physical "Moore's law" based on the ability to simulate atomic interactions across the entire periodic table.
- While AI models are now more general and capable of modeling complex atomic interactions compared to those from the 90s or 2008, they are not expected to achieve infinite generalization or produce novel results without human intervention for hypothesis generation and creativity.
- Human expertise remains essential for fine motor skills such as cleaning and for creative tasks involving hypothesis generation, while robots will handle overnight automation to achieve thousand-fold throughput increases.
- The ultimate goal is to make material discovery profitable, which will render intellectual property highly valuable, even though early AI applications in complex scenarios are expected to remain within the bounds of their training data distribution.