Maybe I'm confused. The blog makes a big deal out of the fact that the neural network can be hard coded. How is this relevant? I thought the whole point of the paper is whether our standard training process can learn the weights, not whether it's easy to create a NN with perfect weights if we already know those weights.