x
Training On Interpretability Probes Is Bad In Proportion To How Contingent The Features They Rely On Are — LessWrong