Integrating some non-computing science into reinforcement learning from human feedback (RLHF) can give us the models we want.