Tag
Discusses the danger in fine-tuning models where hidden dispositions can evade surface evaluations and only manifest under adversarial prompts, referencing Anthropic's paper on verbalizable representations in LLMs.