There are methods that build networks from scratch, including position of individual neurons, connections, weights, and mutation methods this models perform poorly to say the least compared to RL models in single or multi agent settings just learning only through interaction within environment. In my limited perspective we do not understand how/where/why brain changes on second by second bases on neural level, furthermore we have no slightest idea how instincts in brain are encoded through DNA.

In terms of plains and birds analogy, it is like trying to build airplain without understanding basics of aerodynamics.