Perhaps you don't need perfectly designed code, just the right kind of selection pressure? Consider the spaghetti that is our DNA or brain structure. YAGNI design might well be a poor fit for the fractal redundancy necessary for truly complex systems.
I am aware that there are already some programs that are capable of "learning", and you could certainly say that they are selecting the correct paths based on trial-and-error, and "remembering" in order to build faster and more accurate responses. I don't pretend to understand how human brains are wired, but I don't really view DNA as "spaghetti". It is code. Wonderfully designed TERNARY code integrated into a system complete with an interpreter and built-in code cloning, error-checking and correction. We humans have yet to design something that works so efficiently. We still suffer from errors in the code, though - mutations that cause such things as CF, Downs, Sickle-Cell, etc.
I suppose it is really the brain one would seek to emulate if they were trying to create some form of true AI.
Yes it's dangerous to draw from examples that we don't understand the workings of. But all the evidence seems to indicate that brains and DNA don't pay much attention to parsimony or micro-optimization. Or high-level architecture. Or modularity. Nature just does what works.
