Prime Intellect's Will Brown argues three or four scaling axes in reinforcement learning are currently overlooked

i'll give one meta-framing which can be unpacked into several

we often want to simultaneously optimize two competing objectives, e.g. performance and reasoning efficiency, or task volume and generalization-per-task

these can all be minmax problems with lagrangian formulations

@willccbb maybe more

@willccbb Not exactly the same (i think?) but i've been thinking about the predictability vs diversity tension for test time scaling, lately.

@willccbb Chat which