-
is there a chewy piece out there on the feeling of the tiktok recommendation algorithm?
-
like, this covers some basics vice.com/en_us/article/kzdwn9/tiktok-cant-save-us-from-algorithmic-content-hell but the really wild feeling is that the main way i train the algorithm is not by liking shit or whatever. it's just by watching some stuff for slightly longer than other stuff
-
which turns looking at the feed into this active participation in the algorithm. i'm actively steering through the latent space of possible tiktok videos, and i'm doing this on a moment-to-moment basis, based on exactly how long i watch each video for.
-
and the posters know this too - people posting stuff with a "wait for it..." bit, they're trying to get more viewers by getting longer watch times on their content. so i'm trading off "is this payoff going to be worth being recommended more stuff like this?"
-
Goodhart's Law says something like: if you change a system by optimizing against something that's connected to the thing you care about, then the two things will stop being connected. here, that's stuff you wanna watch & how long you watch a video for.
-
also, the thing that's kinda got to be said: one characteristic of YouTube's content is that it picks up on politics & obsession and magnifies that. TikTok picks up on horniness & magnifies that.