one of the worst things about working with LLMs is that you have to read the slop to find out if it's slop. since you don't know if the output is slop or something worthy of attention, you have to read all of it. but you neither skim nor pay full attention, because you you either miss a detail or slopify your mind. so i spend a lot of time in that state of half-reading half-supervising -- correcting mistakes and giving handouts between chats -- until the model finally has the right context and direction by the time you're getting a good direction, you've already wasted most of your attention just trying to get there.