It _can_ be an analog to your million lines of code. But often times a given architecture won't scale to greater capabilities just by adding more parameters in

It _can_ be an analog to your million lines of code. But often times a given architecture won't scale to greater capabilities just by adding more parameters in the same pattern as before. (Oversimplifying, but) the signal from the training data gets weaker and weaker the "further away" the parameters are from it in the model. It can take actual ingenuity to figure out how to get those further layers in a network to contribute anything useful.