Is there any discussion somewhere about adding in the data that makes the x.com/twitter recommend/ranker so functional?
The "Grok-based Transformer"[0] that uses P(click/dwell/not_interested/photo_expand/video_view) seems pretty important and I can't tell how atproto is capturing it. I use @spacecowboy17.bsky.social's For You and from what I understand that feed wouldn't get that data?
(I also struggle with the omni-purpose likes - endorsement, approval, discover-algorithm-input. Maybe a more prominent more/less button addresses this, but then provides less network signal.)
Personally, I really like that my feeds aren't getting that level of granular detail. I prefer the explicit control I have with 'Show more like this' and 'Show less like this'.
I generally think that. But letting dwell time/clicks/open-rates expand the recommender and then (bound to swipe) 'disinterested'/'show less like this' to cull has been pretty efficient. I used to feel dumped into simclusters and now I see a more specific subset of posts I prefer (while still casting what feels like a wide net).
I really liked when bsky introduced the 'show more/less' and then expanded it to custom feeds. But I'm afraid the recommender systems work better with more data. And I think the feed operator alone gets sent a limited set of interactions?
I'm not exactly sure how it would work in atproto but I could imagine an enriched 'graph-interactivity' where you can turn on and off which/how much signal/privacy you want.
