Periodic Labs：AI 科学家的下一步，不是多想一点，而是把真实实验接进来

Scientific intelligence requires an iterative loop of conjectures and physical experiments, because reasoning over existing knowledge cannot anticipate every outcome of a complex universe.

Physical-world reinforcement learning must handle noisy labels, hidden variables, and scarce samples rather than assume that every measurement provides a deterministic answer.

Decomposing discovery into candidate selection, synthesis, and characterization creates more tractable learning tasks than rewarding only the final discovery of a desired material.

Combining calibrated simulations with multiple measurement modalities can reduce ambiguity, while experimental evidence remains necessary to test structures and properties that simplified models miss.

Instrument-level agents can improve data capture by using experimental intent and prior evidence to choose informative observations, provided their latency matches the physical process.

Pragmatic automation should target bottlenecks and recurring errors to increase data quality and diversity, rather than treat complete autonomy as an independent objective.

Negative results become useful training evidence when their conditions and subsequent iterations are preserved, because failure to synthesize a material does not prove that it is impossible to make.

Recording the full lineage of scientific work enables training on decisions and reasoning processes, while newly generated data reduces the risk of rewarding memorized answers instead of genuine problem solving.

Forward-deployed engineering can translate internally tested scientific tools into customer-specific systems through secure local inference, tailored training, and close integration with industrial workflows.

Commercial adoption may create a reinforcing cycle of capital, talent, and scientific capability, while pricing outcomes requires more reliable autonomy than selling research assistance alone.