i think there's probably a whole family of improvements hiding wherever we softmax over bounded quantities without thinking about it. this paper makes me want to audit every loss function that does…