“Interpretability science indicates that AIs are developing increasingly sophisticated internal world models.”