I think this points to future optimizing will be focused on making inference really inexpensive.