“Direct Preference Optimization (DPO) is disappearing from leading-edge post-training recipes, with most labs other than AI2 (for OLMO-3) having moved away from it.”