So they are reading and analyzing the weights and probabaly run some experiments to see how modifying them will make some answers better. makes sense they can do that.

But how do they know which answers are "better"?

if they don't loop?