“It seems difficult to successfully use AI control as the primary method for safely deploying models that are not sufficiently aligned.”