Post

there’s a joke to be made here about doing it for the golden ratio, I can sense it but it’s out of reach

hmm you'd probably know better than i would, but i believe the other paper was doing this roughly to enable cheaper models to be used for tasks by latger supervising models