Pacing would let interpretability research catch up to model capability