Liam Fedus and Ekin Dogus Cubuk, co-founders of Periodic Labs, discuss their mission to automate scientific discovery by grounding large language models in real-world physics and chemistry experiments. They emphasize using experimental outcomes as training signals, enabling AI to learn from direct physical feedback. The team targets high-temperature superconductivity as a grand challenge, while building near-term tools like AI co-pilots for scientists. Success depends on integrating noisy experimental data, including negative results, and combining expertise across AI, materials science, and quantum mechanics. Mid-training techniques allow models to absorb new scientific knowledge dynamically, and geometric reasoning is seen as essential for understanding physical systems. To accelerate progress, Periodic Labs is fostering deep collaborations with academic researchers through advisory boards and grants, aiming to bridge the gap between machine learning and hard science innovation.