Objectives

Pre-training has one objective and post-training has many. These entries cover the alignment losses — the policy-gradient methods and the closed-form preference objectives that replaced them — and the arguments about which parts of the reinforcement-learning apparatus were ever load-bearing.