Navigating LLM embedding spaces using archetype-based directions — LessWrong