Internal Interfaces Are a High-Priority Interpretability Target — AI Alignment Forum