Yeah, definitely. Combination of expert-system gating (some requests probably get routed to weaker models), distillation (for performance/cost), and RLHF lobotomization.