So we made a learning network out of transistors instead! For instance, here's a piece which does softmax. The only problem is that pulling the output no longer perturbs the whole thing... which we solve by using a clever "dyadic" construction which won't fit in a tweet!
