◇Interpretability Pipelines Get Automated

Models are increasingly used to analyze other models' internals, converting interpretability from a labor-limited craft into a compute-limited pipeline and concentrating auditing capability with whoever holds the weights.