“AI labs improve their models by selecting specific domains, creating reward models for those domains using expert-level datasets, and then connecting them to their policy models.”