Cat videos on YouTube

In My Humble Opinion
1

I watch WAY too much YouTube, and of course it feeds me things it thinks I’ll like. A specific kind of cat video is popping up a lot. A little kitten is being intimidated by a big dog, and the mom cat comes out of nowhere and just looks at the dog, who backs down. That’s the whole plot, and I keep getting these and it makes me wonder how they differentiate that kind of vid from any other cat video.

2

The exact recommendation algorithms used by TikTok (and competitors like YouTube Shorts) are closely guarded trade secrets, but we can speculate…

  • If one person watches that clip for a long time, it’s interesting to that one person. If ten million people watch that clip, it’s probably interesting to everyone. But if only 1000 people watch it fully and everyone else skips, then that’s a strong signal that it’s appealing only to a certain niche. Cluster enough of these niches and audiences together and you’ll start to get a graph of “videos liked by other people who generally like the same things I do”, and do that enough times and you can get pretty hyper-specific recommendations like “people who like cat videos but only if they include a big angry dog”.
  • They can also use AI (or cheap labor) to label these videos one by one, especially if they’re short
  • They can also cluster by creator and upload time, e.g. if this one video got really popular over a few days and then a bunch of other related creators also had videos popular a few days after, they might be response videos or copycats (copydogs?)
  • If the videos have sound/dialogue, you can also fingerprint those sounds (“meow meow bark”) or transcribe the dialogue and make clusters out of those too.

With enough videos and viewers with different content and habits, the statistical patterns alone can be pretty powerful.

3

Has anyone else seen the vids where a cat or bird watches their owner slice up a cat or bird cake?

4

Be aware that those cat rescue videos (there are also vids of cats rescuing small children) are AI-generated slop, often by an outfit called Sora, which is aiming to become the AI TikTok. If you see a small “Sora” logo, it’s fake. They pop up in Facebook reels too.

5

Question: is there a way to include tags, or the AI prompts, as data embedded in the video? (Similar to the data in still photos whuch record place, time, shutter speed, etc).

That would make it very easy to push videos which match your viewing history.

6

The term of art for this is Collaborative Filtering. Netflix famously had a contest for making these a while back.

Nowadays I’d expect that there are semantic embedding models that might be sophisticated enough to put “cat protecting kitten from dog” videos all into a similar space, and the recommendation system could just be “here are some videos that embed to a point near the embedding for the video you like”.