I agree in the long run whitebox interpretability would be better than CoT monitoring but the technology is very much not ready for it today (and it's not clear we'd solve enough of interpretability before the AIs have actually scary capabilities).