I used stumbleupon in the 0s and that was primarily blogposts and writings. I don’t see why an algorithm couldn’t recreate the stumbleupon experience.
1. RSS doesn't help me find new interesting blogs and writers. One of TikTok's few virtues as a platform is giving algorithmic reach to creators who have no social reach.
2. My RSS collections have always been one of either I'm following a highly curated list that leaves me hungering for more content or I'm following a firehose that makes me miss the content I prefer.
I stumbled across a website that does something like what the article describes, but for just tweets[0]. It's surprisingly compelling and gives me the feel of using StumbleUpon. I am eager to see more things like it.
[0]: https://mood.surf (no affiliation)
For example, if you submit this link: https://stratechery.com/2023/netflixs-new-chapter/ then you will get connected to two RSS feeds that posted that link: https://stratechery.com/feed/ and https://hnrss.org/newest?points=100 (ie, the HN feed of items with >100 points)
As you rate content you get connected to more and more sources.
The way it solves the oversubscription problem is - whenever a user/feed posts something - your connection to them goes down slightly. As a result you see content from sources with the highest signal-to-noise ratio first.
The differences from Artifact is:
- Artifact uses implicit signals such as read time; LinkLonk uses your explicit upvotes. I think implicit signals are fine for entertainment content, but not for informational content. Time spent is not equal to becoming better informed. I think only you can be a judge of that.
- Artifact uses an opaque AI algorithm optimized to do what they want; LinkLonk uses a transparent algorithm, where each recommendation comes from a feed or a user that you have co-liked items in common.
Filtering helps but "i just like some type of content from that person" seems to be forever problem on every site. Like on youtube, you watched whole three videos of a guy cooking ? I will bomb your feed with random crap from his channel for next month or two!
I don’t know how much of figuring out what users want transfers from many short videos to text. For instance, dwell time seems an obvious metric but I find I ‘dwell’ longest on either things I find enjoyable or on things I find terribly boring. I hope it does transfer well – exposing people to more of the kind of writing they want seems generally good to me.
I don't like user-generated video content so I rarely use YouTube and I've never touched TikTok. But I don't think algorithms really help. I really prefer the user selection here on hacker news.
People underestimate this. TikTok's algorithm makes the app enjoyable to use while most other algorithmic feeds make their respective service less enjoyable. I imagine that is because TikTok factors in other criteria beyond active engagement. At this point we all know it is easier to trigger active engagement through negative emotions. By simply prioritizing any engagement, algorithms like the ones used at Twitter and Facebook end up prioritizing content that delivers us those negative emotions in order to get us to engage. TikTok's algorithm also factors in passive engagement like the dwell time you mentioned. That ends up creating a more positive user experience because it actually learns what we like instead of just what makes us reply angrily.