In this post I go over training LPs using cvxpylayers, but I think the more interesting discussion lies at the end, where I argue that current AI models lack the ability to reason counterfactually in discrete settings & that LPs can bridge this gap.