Distillation allows compute-starved AI labs to achieve results equivalent to having more compute by using the outputs of stronger models