“The success of transferring legibility training from small models to humans suggests that using small models for scalable oversight experiments is a more promising and faster research path than using humans.”