Comment by PaulHoule
4 hours ago
To praise TikTok it has a highly effective recommendation engine precisely because it is showing you one piece of content at a time and registering your engagement on that.
YouTube's interface gives people a feeling of agency because it tempts you with 10 or so videos on the side and you can choose one, it also means YouTube does not get information about the 9 you didn't click, maybe you would have liked 5 of them and hated 4 of them but it can at best guess about that. I read about negative sampling in the recommender literature to address this issue and never felt I understood it or believed in it -- the literature clearly indicates that it sorta-kinda works but I think it does not work very well.
So far as hating on algorithmic feeds it is not the algorithms themselves that are bad but how they are chosen. If there is any characteristic of the content that can be quantified or evaluated a feed can be designed to privilege that. A feed could be designed to be highly prosocial, calming and such or designed to irritate you as much as possible. It's possible that people get bored with the first.
My own reader works like TikTok in that it shows one content piece of the time but it is basically the stuff that I submit to HN and it is scientific papers and articles about LLMs and programming languages and social psychology and political science and sports and and advanced manufacturing and biotech and such. You might say my world view is unusual or something but it is certainly not mindless lowest common denominator stuff or outrage (e.g. to be fair I post a few things to HN because YOShInOn thinks they are spicy -- YOShInOn has a model that can predict if y'all are going to comment on an article or not and I felt it was a problem that my comments/submission ratio was low before I had YOShInOn)
> To praise TikTok it has a highly effective recommendation engine precisely because it is showing you one piece of content at a time and registering your engagement on that.
I'm a bit divided on TikToks efficiency. It's a well working doom-scrooling-machine, better than any other platform, but from my personal experience, it's not actually that good at recommending the content I actually want. And I think it's largely because it has the wrong focus, namely the attention. High attention-content is not always what I want and need, but TikTok has barely any way to realize this, exactly because of how It works.
> YouTube's interface gives people a feeling of agency because it tempts you with 10 or so videos on the side
Interesting, never used that side-thing.
> it also means YouTube does not get information about the 9 you didn't click,
Yes, and that's OK. The not-clicked entries can still give me relevant information. And yes, the system can't act on this, but that's the whole point of RSS Readers, to make your own choice, on the spot, and switch it constantly as necessary. No system can react to this. "Smart" algorithmic solutions are doomed to stay mediocre because of this.
Well...
Personally I can't stand TikTok or Youtube Shorts or the videos on Instagram. I just can't stand the meaningless motion to get attention, it makes my skin crawl, it makes the bottom drop out of my stomach, etc. One time YT Shorts showed me an AI generation video of a pretty girl transforming into a fox on America's got talent, which is a good choice for me but then I got saturation videos of Chinese girls transforming into just about everything on AGT with the same music and reaction shots and it was more than I could use and not looking cool anymore but rather like AI slop. That said, I enjoy classic YouTube with relish.
My RSS reader gives recommendations based on explicit up/down and it has an AUC of maybe 0.78 or so, I saw a paper where TikTok is getting 0.83 so I feel like I'm doing OK.
I haven't done anything to change it in the last year except increase the number of random articles it inserts a little because making the recs worse actually can make them better, see [1] TikTok is famous for doing this. I think I could tune it up so for a given batch it could have a target "thumbs up" percentage or something more systematic but really I am very happy with the recs so it is not clear to me what "better" really is.
There is the problem with it that the system has a lot of latency which does not really matter for articles on most subjects because news about software or science or political science or engineering is usually OK if it is delayed a few days or a few weeks but it is a problem with sports where you really look like a dumbass if you post about something that happened on week 2 during week 4. It's a toughie though because I'd have to rework the thing to take out latency in 5+ stages of the system and then think systematically how to balance "urgent" vs "interesting" so I don't face the problem that urgent but interesting sports articles don't crowd other things out. [2]
[1] https://en.wikipedia.org/wiki/Multi-armed_bandit
[2] personally I don't mind the old articles for myself because I'm a weird kind of sports fan. Two years ago I used to follow the NFL but since I started doing sports photography I might go to 5 games on one weekend and if I am doing that the NFL is a lot less interesting than, say, Arknights so I am a little embarrassed to say I have no idea how the Bills are doing this year. But if I'm going to post sports articles to Bluesky or something it's a problem.