Epistemic status: the idea it’s not fully fleshed out (there are a bunch of problems that I’m skipping over, the post is meant more as: “has anyone been seriously thinking about this?” and throwing out a starting point, than: “here is a proposal that would actually work”) and I wouldn’t be surprised if either it’s unfeasible or it has already been implemented.

 

I was reading Neutrality and the part about social media struck me: “Or, they use ‘impartializing tactics’ chosen in the early-Web-2.0 days when people were more naive and utopian, like ‘allow everyone to trivially make a user account, give all user accounts initially identical affordances, prioritize user-upvoted content.’ […] — with LLMs, zero-knowledge proofs, formal verification, prediction markets, and the like — can make a better stab at these supposedly-subjective virtues than ‘one guy’s opinion’ or ‘a committee’s report’ or ‘upvotes’?”. I thought “Seriously, why aren’t social media companies doing this already?” (and they might be). The thing that stood out to me was that we could try to use prediction markets, or just markets, to improve the “ALGORITHM”.

The easiest possible way to do it (and that almost definitely wouldn’t work) is to use likes (or upvotes, etc.) as currency, we could give people a certain amount of likes and treat them as “bids” one could place on a given post, if the post does well, you earn more likes. Another approach, that solves the obvious problem of “But what if I run out of likes?”, is to use a reputational system: likes are weighted by the reputation you have, infinite likes finite reputation. There are a couple other problems that need to be solved. First, how do we measure a successful prediction? Second, how can we avoid Goodhart without overcomplicating the system? (Another issue is setting up the right incentives, a paid model probably works better than one based on ads but I’m not sure what would be best here).

Let’s start from Goodhart: the solution is just don’t tell the users what you’re doing. Don’t put the reputation anywhere on the site and you should mostly be fine.

Now, how do you measure a good prediction? First we need to figure out what good is. Personally, I think avoiding to promote all the memetic slop, “like to see the animation” or the random celebrities' news or “funny” videos, would be a decent goal. We want to promote good posts for the right users and not everyone will agree on what is good for them, but we assume companies already know how to solve that (I imagine something like “cluster together people with similar internet tastes”). You can then proceed to measure a good like if it’s a good prediction that other users in the same “tpot” will like the post too.

And assuming everything goes right the result is a social media that can distinguish between users that randomly like most things and users that can act as curators and spread their good tastes.

New Comment
11 comments, sorted by Click to highlight new comments since:

I am surprised why social media companies don't create individualized bubbles using the following algorithm: First find people who upvoted the things you upvoted, and downvoted the things you downvoted. Then show you the content those people upvoted that you haven't seen yet. (Authors automatically upvote the content they submitted; which is how the new content gets its first upvote ) Add a little noise and sometimes show someone a completely random video; also show random videos if you run out of things to recommend.

Maybe trying to determine the best content is a mistake, and instead you should try to predict the best content for given user.

I guess the problem is with users who are not logged in, what should be showed to them? The simple option is to show the things with most upvotes; the second most simple option is to have special admin accounts who are displayed the most upvoted things and can upvote or downvote them, and then the anonymous user gets displayed the content that got most upvotes by the admin accounts. Or some combination of these two things.

How does that differ from standard recommender systems?

They absolutely do. This phenomenon is called filter bubble.

I've yet to see a personalized filter bubble, I've certainly been part of some nebulous other's filter bubble, but I am constantly and consciously struggling to get social media websites to serve up feeds of content I WANT to interact with.

Now while I am in someone else's bubble - there is certainly a progressive slant to most of the external or reshared content I see on Meta platforms (and it seems to be American-centric despite the fact I'm not in the United States), it took me an incredibly long time to curate my 'for you/Explore' page in Instagram, to remove all 'meme' videos, basically anything with a white-background and text like "wait for it" or "watch until the end" or intentionally impractical craft projects. It is only through aggressively using the "not interested" option on almost every item in the For You page, it now shows me predominantly High Fashion photography.

Alas, one stray follow (i.e., following a friend of a friend) or like and Voompf the bubble bursts and I'll start seeing stuff I have no interest in like "top 5 almond croissants" or "homemade vicks" - I have never EVER liked any local cuisine content ever and I'm really against the idea of getting medicinal craft projects from popular reels, sorry.

I suspect the filter bubble is grossly exaggerated - (Counter theory: I am not profitable enough for the algorithm to build such a bubble) -  I suspect that much the fear and shock when you see Google ads delivered to you that match a conversation you had just 3 hours ago (although there is a lot of truth to the idea that smart devices are spying on you). Likely this is a combination of good old fashioned Clustering Illusion and the ego-deflating idea that the topic of that conversation you had just isn't that unique or uncommon (particularly). I've had three different people DM me the same OK GO Reel in one week recently. I never interacted with it directly - clearly the algorithm was aggressively pumping that to EVERYONE.

Spotify ads suggest I'm a parent who plays golf.
Instgaram Ads are slightly better these days - I just flicked through and most had to do with music and photography. But recently it was serving me a lot of obvious scam ads, with a photoshop image of a prominent political figure with a black eye. (again - politics. I don't follow or interact with ANY political content.)

It's someone else's bubble, I'm just living in it.

P.S. do not allow Spotify to add suggested tracks to the end of your playlist if you're not a fan of top 50 pop music, it will start playing Taylor Swift et. al., and you won't be able to remove it from your 'taste profile' rendering certain mixes unlistanble as it won't just be her music but every other top 50 artist.

On YouTube I keep getting Russian propaganda videos recommended all the time, and I really don't remember upvoting anything of that kind. So either YouTube does not use this algorithm (or maybe gives to "upvoted by people similar to you" much lower weight than "upvoted in general"), or there is some clever trick the authors of those videos use, but I am not sure what it could be.

On Facebook, I think people used this trick that they posted hundred cat videos, people upvoted them, and then they deleted the cat videos and posted some advertisement on their channel, and now the advertisement was shown to everyone, including "people you know liked this page". But I think this should not be possible on YouTube, because you cannot change the videos once uploaded.

So... I don't know. I wish YouTube created a bubble for me, but it seemingly does the opposite.

I think their metric might be click and not upvote (or at least, clicking has a heavy weight). Are you more likely to click on a video that pushes an argument you oppose?

As a quick test, you can launch a vpn and open private browsing to see how your recommendations change after a few videos

As I understand it the reason is that it is too computationally expensive to tailor a feed for each user, I remember seeing somewhere on the Facebook developer blog that that for any given user they take a batch, let's say 1 million posts/piece-of-content from their wider network, this will likely overlap if not be the same as many other users. From that it personalizes and prunes to whatever number of items they elect to show in your feed (say 100).

Out of the millions of pieces of content posted every day, it couldn't possibly prune it down for every user a-new every 24 hours, if there's a 200 million people using a platform at least once every 24 hours, that quickly rockets up. So they use restricted pools which go through successive filters. The problem is if you're not in the right pool to begin with, or your interests exist across multiple pools - then there is no hope of getting content that is tailored to you.

This is a half-remembered explanation of a much more intricate and multi-layered process as described by Facebook, and I don't know how similar/different Youtube, TikTok and other platforms are.

Computational costs make sense. I kinda assumed that if Google can calculate PageRank (a numeric value for every URL, depending on the content of documents in other URLs), calculating this would be a piece of cake. But I guess it's the opposite.

PageRank can be calculated iteratively -- you assign the same starting value 1 to all pages, and then just keep updating, and in theory if you keep doing it infinitely long, you will get the theoretically correct value, but if you check at any moment after finite time, at least the value will be something, and quite often it will be close enough to the theoretically correct value. It scales well.

On the other hand, similarity between users X and Y probably needs to be calculated individually for each pair of users. (I am not sure about this, maybe there also is a smart way to do that iteratively.)

If YouTube just puts me into a generic bag like "people from Slovakia", or even a narrower one like "people from Slovakia who watch discussion videos", that would explain a lot.

I think it's unlikely that computation is the bottleneck here, even if it used large batches I would still expect it to provide a better feed than what I usually see, I think the problem is more likely to be that they're incentivized to give us addictive content rather than what we actually like, and my hope is that something like X or other socials with different incentives could do a lot better.

I don't believe it's doing a good job of delivering me 'addictive' content. Am I in denial or am I a fringe case and for the most part it is good?

From my experience they're getting pretty good, depends on the social but IG reels or YT can keep me entertained with nothing-content for hours