rlhffactneutralDPO (Direct Preference Optimization) implementation in frontier models is messy, particularly on the data sideNathan Lambert29 Jul 2026https://x.com/natolambert/status/2082485850257207307