But this is part of the author's point: where are all the production reinforcement learning models?