Otherwise I can't see how they could auto-decide to push out content to large audiences without something they don't like getting out.
Like, they'd only need to sift through a handful of videos to reward random publishers with thousands of views, perhaps even with fake numbers, and people would probably gobble it up.
The recommendations and people's video interactions allow you to classify both people and videos.
First time I hear of TikTok in those terms. Could you expand on what that means?
Take Youtube, a video site that added suggestions. If you're logged in, you get dumped on a homepage with an endlessly scrolling list of recommendations. You scroll until a thumbnail or video title catches your eye, you watch that video, might like or comment, then click one of ~10 recommendations. The algorithm gets a positive signal based on what you click and your watch-time, but the negative signal on anything you don't click is incredibly weak. Because you choose what you watch, positive signals are mostly driven by titles, thumbnails and familiarity, and negative signals are rare. As a result, most recommendations are stuff you know or things adjacent to it (people who watch this also watch this), and quickly get boring. The recommendations can't inject novelty or purposefully expand the algorithm's knowledge.
Now on Tiktok: You open the app, you get a single video in full screen. Either you scroll on (negative signal) or you watch it until the end (positive), or you let it repeat (very positive). Or you might like or comment. After scrolling on, you get another video in full screen, same game again. The cost of a bad recommendation is low because videos are short and quick to judge, so the algorithm can occasionally give you videos it's uncertain about or that are selected completely randomly. Your reaction will give a signal, no matter what (either you watch till the end or you don't) and as a result the algorithm knows more about you and about the video.
To put it another way, TikTok is Ice Creme, Youtube is dinner. A popular Ice cream stand has a lot more throughput than most restaurants. It only takes ~1 minute to serve a customer some ice cream.
YouTube used to be full with 3-4 minute videos back in the day, with a ten minute maximum. Longer than TikTok, but not much. But those short videos all but died out when the 10-minute videos got another ad spot and thus more money for the creator.
Vary rarely do I see a < 10 second clip, most are 30 seconds - 1 minute.
Maybe I’m too old for TikTok.
Sounds like this would heavily incentivize items that are immediately catchy. Slow buildup stuff would die in this environment.
My guess is every single video is marked by a human for how much exposure it should get, and if they let something get out that shouldn't, that person is reprimanded.
… Wait what? Why would that be your guess? How is this easier than, you know, anything else?
The whole point of TikTok is that you don’t have to follow people or engage with your friends on the platform. You just open it up and it’ll show you videos that you’ll like. Obviously, there is a little bit of “training” involved. Within about 20-40 minutes it really started to zero in on things I liked.
If you are really interested, check the yc article in 2016. The algorithm is their bread and butter. They have been doing it for 10 years. What's more amazing to me is that not many ppl are aware of this.
If someone post a clip playing guitar and singing for example, would it determine this is a performance, and it should show it to cohorts who have liked viewing performances in the past?
Or would the AI model work more like… show (unknown type) clip to random people, identify similarities in the people who like or dislike the video based on those user’s viewing habits and then try to identify the video type based on that knowledge? I guess the second method would be very inefficient unless combined with some feature extraction and metadata / tagging from the source.
Not an ML person, just a curious dev.
Don't have any illusions, they can easily control what goes viral if it matters to the mothership.
I don't think they are successful at keeping off political content though, in the sense that denying political content is a political statement itself in favor of the status quo.
The manual shadowbanning fight is one they will never win.