writing code of higher complexity (we know from CICERO that longer time spent on inference is worth orders of magnitude more than the equivalent in training when it comes to improving end performance), or doing real world tasks with unknown fractal depth (aka yak shave)