Now I'll say something controversial: I somewhat miss the algorithm. What it was good for -- for me -- was pointing me to what I might have missed while I was gone. I do believe that one of the values that will be offered in a federated, open ecosystem will be recommendation services: ones that we seek and control and that provide value to us as users. See this provocative discussion from experience:
https://social.lot23.com/@jon/109372257422277945
@jeffjarvis That's a very fair comment and I hope we can have a reasoned discussion about it. Part of my frustration with Twitter was that they FORCED the algorithm on you. In addition, I'll add that it didn't work that well fo me (sample size of 1, etc etc)
If Mastodon could have a transparent yet optional 'show me the good stuff' option, I agree, that would be very helpful.
More importantly, it's not heresy to talk about it.
@scottjenson @jeffjarvis Also if you're running your own instance you can afford to deploy staggering amounts of CPU power. Like, you could plausibly run GPT-NeoX-20B on your own dual-GPU rig to try to guess which Fediverse posts you'd be most interested in seeing.
The NSA can do this with Twitter, but you're not allowed to.
@scottjenson @jeffjarvis Might work. Which #Fediverse codebase looks most hackable?
@scottjenson @radehi Actually, that gets close to "trending" and the mass-media view of scale. I'd prefer something that is, yes, more complicated: Here's the people you, Scott, seem to interact with most because you care about them and they had things to say when you were away....
@jeffjarvis @radehi do you think you could put it a bit more precisely? e.g. ,how would this sorting work? For example, I assume you'd start with people you follow (people you care about) and then sort them somehow with the more 'active' ones being at the top.
Not sure. Need to think about it. Others have more experience with such affordances than I do.
@jeffjarvis I don't mean to 'put you on the spot'! I was just asking for a 'personal view' of what you'd like to see (if possible)
@scottjenson It's the right spot. As a user, I want to know what I missed from folks I tend to like. I demonstrate that liking by interacting with them and I respect what folks I like are talking about. But right there I can see myself potentially falling into a trap of valuing performative engagement again. Or not. That's what I'm struggling with. I know the end-product I want. I need to think through the perils of the paths there.
@jeffjarvis That's why I'm suggesting a fairly simplistic model to start: start with your home feed and offer a few obvious ways to sort it. Nothing fancy, just áfew sort-by levers that (hopefully) doesn't encourage too much performance.
Anything that tries to get fancy could easily fall into the performative engagement trap.
@scottjenson @jeffjarvis You could simply do a very easy logistic regression of your own decision of "favorite" or "bookmark" or "reply" against the vectors of boost and favorite of all your follows to get a statistical prediction of how likely you are to favorite (or reply to, etc.) a post. Let's call that prediction "quality", acknowledging that is a firmly subjective (intersubjective?) kind of quality.
Then need some way to combine that "quality" metric with recency. I'd favor a quality threshold that varies over time to maintain a roughly constant rate of posts selected from the thousands of candidates per day: 5 per day, say, or 50 per day. Maybe only rate a post 48 hours after posting so all the data is in.
@radehi @jeffjarvis How about we walk a bit before we start sprinting? Honestly a ranked list based on something ridiculously simple, e.g. some combination of boosts and likes, would go a long way.
The same thing applies to following hashtags by the way. It would be helpful if this sorting wasn't buried only in the home feed but something we could summon for any column view (in the advanced UI at least)