Comment by 8note

making the specifications is still hard, and checking how well results match against specifications is still hard.

i dont think the model will figure that out on its own, because the human in the loop is the verification method for saying if its doing better or not, and more importantly, defining better