| > What are you trying to say? That the reason this person see this content is their fault? They secretly want to see these type of videos? Nah, it's not about intent, but it is about profiling. They're saying that e.g. gullible-seeming people will be algorithmically matched with videos trying to con them out of something, while non-gullible people won't be. People who watch more videos by creators with religious values will eventually be recommended religious content; while people who don't do that, won't. Etc. Think about it less like users being matched with things they'll appreciate; and more like creators being matched with the audience most receptive to their message. > There are a lot of articles and videos showing that sometimes watching just one video will suggest a ton of videos related to it, no matter if you are not interested. This isn't a failure of ML. They've got the algorithm doing exactly what they want it to do. It just isn't serving you. TikTok is a two-sided market, where the supply is "engaged eyeballs" and the demand is from advertisers with ads to show them (where a regular video producer is just an advertiser who provides enough retention value to the platform with their "ads" that they get paid rather than paying per impression.) TikTok's algorithm isn't trying to match you with the videos you'll most like; rather, it's trying to optimize the amount of money ByteDance extracts out of its advertisers by optimizing for three things: 1. keeping the eyeballs engaged, by showing them videos which are predicted to increase the particular user's session duration in the app; 2. showing the "engaged eyeballs" the most profitable ads, under the proviso that any given advertiser can filter for eyeballs with specific demographics/interests; 3. (here's the clever bit) — nudging the eyeballs toward videos that will allow them to plausibly say that a given user has a given high-CPM interest, and thus now show them the high-CPM ads. The third factor is what makes the "one video causes your recommendation feed to completely change" thing. An very close analogy would be to dating (another imbalanced-demand two-sided market where demand is a passive judgement while supply is an active offer.) Picture person A walking into a nightclub, looking for a date, but not actively talking to anyone. They sit there, and wait for other people to come up and talk to them. The people that come to person A might be somewhat random at first; but, as the people in the club notice a pattern in who's doing best talking person A up, the supply-side will self-select — they're profiling person A, and "recommending" themselves based on said profiling. But then, at some point, imagine person A quietly mentioning to one of these strangers "I think I might like [niche interest]." And this news spreading throughout the club. Now, if there's anyone who likes [niche interest] in the club — suddenly, they think they have a chance. And if having [niche interest] is rare, maybe there are a bunch of unsatisfied single people with [niche interest] who've been desperately waiting for someone like person A to show up. So now there's suddenly a stampede of people, all with [niche interest], trying to get person A's attention. Willing to pay money to get person A's attention, even. So much that the club manager (who happens to be easily bribed) is willing to cordon off the area around person A and set up a queue of all these interested people, so that the "rabble" who aren't so intensely interested (and so aren't willing to pay a bribe), won't even get a word in edgewise any more. That's TikTok. You're person A. The advertisers are the desperate people in the club. And a single clicked video can be the whisper of acknowledgement of a niche interest they were hoping for. |