“The most effective way to use frontier models to improve a new model is to employ them as judges in a reinforcement learning setup.”