Anthropic, Claude의 사고 구조 해독: 100만 개 피처 발견
Anthropic이 Claude 3 Sonnet에서 100만 개의 내부 피처를 해독해 개념이 형성되는 방식을 매핑했습니다. 이는 AI 투명성을 한 걸음 앞당기며, 안전 엔지니어가 근원 수준에서 모델 행동을 제어할 수 있게 해줄 잠재력을 갖고 있습니다.
anthropic released findings mapping approximately 1 million internal features—or concepts—within its claude 3 sonnet model, visualizing how specific combinations of artificial neurons fire for distinct ideas like fairness or biological structures. the research allows engineers to identify and deactivate unsafe traits, such as bias or manipulation, by targeting these specific features rather than filtering outputs post-generation. while this represents a significant step toward mechanistic interpretability, it remains an open question whether this specific mapping applies to other models or architectures. successful mechanistic interpretability is increasingly viewed as a critical safety requirement for future autonomous systems to prevent unforeseen behaviors.