If 2/3 of AI progress is due to software improvements, doesn't 2x overall AI progress acceleration require (2 - (1 - 2/3))/(2/3) = 2.5x faster software progress, if training compute growth stayed the same?