Constitutional AI vs. RLHF vs. Deliberative Alignment
Compares RLHF, Constitutional AI and Deliberative Alignment through a persona-emotion-behavior framework, argues CAI yields more stable personas, and cautions against activation steering that models could learn to evade.