Limitations and Potential Applications of Mechanistic Interpretability

In this chapter, the speakers discuss the limitations and potential applications of mechanistic interpretability (Mac Interp). They mention that even if Mac Interp could provide a comprehensive explanation of a model's behavior, it may still fail to detect unforeseen interactions and deception. They explore the idea of focusing on safety-relevant components or tasks with a large performance gap between different models to make Mac Interp more feasible. Additionally, they consider the possibility of ruling out problematic behavior without fully understanding the model's internals. They express the need for testing Mac Interp soon and having progress metrics to ensure its effectiveness.