Why AI Scientists Must Experiment: Periodic Labs' Vision
Periodic Labs co-founders Liam Fedus and Ekin Dogus Cubuk argue that scientific discovery differs fundamentally from math or coding due to physical noise and uncertainty. They propose 'synthesis superintelligence,' using reinforcement learning in real labs to train models on failed experiments, overcoming the limits of internet-only data.
ImportanceStructuralEvidenceE3 inspectableWrite-upStandard
While large language models have made significant strides in mathematical proofs and code generation, they struggle in fields like materials science that require interaction with the physical world. Liam Fedus and Ekin Dogus Cubuk, co-founders of Periodic Labs, argue that this disparity stems not from a lack of intelligence, but from the nature of scientific discovery itself: new knowledge emerges from validating conjectures against reality, not just logical deduction.
Fedus emphasizes that "intelligence is necessary, but not sufficient." In math or coding, all context—axioms, APIs—is deterministic and accessible. In physics, systems contain Avogadro's number of atoms, making complete information inaccessible. Experiments are noisy: furnaces degrade, instruments vibrate, and labeling errors occur. Therefore, an AI scientist must make decisions under high uncertainty and missing information, a skill set distinct from optimizing known answers.
To address this, Periodic Labs proposes "synthesis superintelligence," a closed-loop system combining prediction, synthesis, and characterization. Unlike traditional RL environments, their reward functions are based on identifying phases from raw data like X-ray diffraction (XRD). If an AI accurately identifies the target crystal structure from noisy experimental data, it receives a reward. This mechanism forces models to learn how to extract valid signals from ambiguity, rather than relying on memorized answers.
Furthermore, Cubuk and Fedus argue that failed experiments are more valuable than successful ones. Existing literature suffers from survivorship bias, publishing only successful syntheses. Periodic Labs collects vast amounts of "negative results" through automated labs, using them as key training signals. This helps models distinguish between "un-synthesizable" materials and those requiring different conditions, while recording the entire exploration process—including errors—teaches AI strategies akin to human scientific intuition.