The usefulness of mechanistic interpretability

The podcast episode explores the concept of mechanistic interpretability and its utility. One speaker expresses skepticism about its current usefulness, mentioning that it fails to explain much of the performance of models. However, they acknowledge the potential for future advancements. While there are doubts about mechanistic interpretability allowing for solving core problems like auditing for deception, there is still interest in investing resources into the field. The discussion also touches on the importance of consensus and identification of concrete projects to advance understanding and application of mechanistic interpretability.