137 sats \ 3 replies \ @elvismercury OP 28 Mar \ parent \ on: Some thoughts on SN rewards meta
I love the KBC metaphor.
Casting the reward function in that light then pushes the design space down a level -- it becomes a different grounding problem -- you define beauty based on the consensus of the users you've attracted. But you're still attracting users. Those are still choices that the definition of beautify is grounded in.
You could find that you've made choices that result in a curious definition of beauty, for good or ill.
You could find that you've made choices that result in a curious definition of beauty, for good or ill.
Ah yes, it's all a bit semi-supervised in that I did a lot of data labelling early on to kick start the whole thing.
The way I've thought about this so far is that so long as the group of "people who have historically spent well" can change, ie the historical input to the algorithm is a recent window of time, the definition of beauty can progress.