A bandgap circuit subtracts log(A.x) and log(B.x) to get a ratio log(A/B) that is independent of x (x>0), and to achieve this topology by evolution is a striking result. The question is whether one would trust the resultant topology in the absence of a known pattern. This is not casting a doubt on the evolutionary algorithm, it is more about the quality of the device compact models and the ability of the circuit equation solver to capture subtle instabilities and starting problems.

My guess is that today's compact models and solvers are both good enough for us to contemplate a 'Turing test' for analog circuit design.