What Makes TikTok's Algorithms So Effective?

That article misses that TikTok gets effective negative feedback: they know when you don't like a video. Contrast that to the interface that YouTube cribbed from [1]. When you show people a list of ten videos on the side you can't come to any conclusion of whether or not people liked the videos they didn't click on, in fact, I'm frequently frustrated because I also wanted to watch another video that was recommended and now it's lost. If Facebook had a "dislike" button that worked that would fix many of the things wrong with it. [2]

I read many papers on "negative sampling" which is used in western recommender algorithms to guess at what you didn't like because there's no dislike button: I always thought "I can't believe it works" and fact is, it doesn't or it doesn't work very well.

My RSS reader gets a thumbs up/thumbs down and gets AU-ROC of 0.79, TikTok gets 0.84 or so and has a lot more data and 1000x the budget.

I'd almost say "throw the western recommendation literature in the trash and treat it as a classification problem [3]" except that when you have less than <1000 data points for a user there is a cold start problem. You could throw AI and the computational power of a Dyson sphere at the problem and it won't do as much as a dislike button that speaks with authority.

[1] https://montevallotimetravel.wordpress.com/2012/05/03/idiocr...

[2] would be good for people who don't like hysterical posts about politics but it's an existential threat to the advertising economy, influencers, etc...

[3] calibrated, even

I've always suspected that the secret sauce was that there are plenty of humans in the loop tagging and classifying, that's a cheap option in China. It's from a censorship regime after all. Today they likely have enough data to automate most of it.