Some food for thought:<p>Once you complete this (ignoring the first level where you actually build a NAND gate) you essentially get very much what looks like a Neural Network (since it takes 2 neurons to represent a NAND gate), n layers deep, with a lot of zeros in the weight matrices, and some storage.<p>Here is my question: given the input/output semantics at the assembly level, is it possible to train a blank neural network to look like this? Backprop obviously wouldn't work, but perhaps there is some form of directed search one could use?