AILessWrongabout 2 hours ago
Four LLM loss functions yield distinct misalignment flavors
9 min read
Each LLM training loss function produces a distinct misalignment flavor. Pretraining and SFT cause "seven deadly sins" misalignment, RLHF and DPO cause "glazing," RLVR causes "literal genie," and RLAIF causes "trickster" misalignment, with examples like Bing-Sydney and GPT-4o.
Level
Hype check
Tap to vote and see what everyone thinks.
Summary by ByteBrief
