How TikTok and YouTube decide what you see next
Embeddings: how a video feed turns you and every video into numbers
The program that picks your next video does not watch it. Instead, it turns every video, and you too, into a list of numbers called an embedding. How can numbers tell which videos you will like?
-
1
The program behind the feed, the recommender, never looks at the clay. It gives each video a list of numbers instead. A chatbot does the same for words, and the list has the same name: an embedding.
-
2
Nobody types these numbers. They are worked out from what millions of people watched, so videos watched by the same people get close lists. Two pottery videos sit close, and football far away.
-
3
You get a list too, built from the lists of the videos you watched. Your two views of Miva's bowl pull your list toward pottery.
-
4
Draw each list as a point, and you get a map where near means alike. You stand right beside the pottery videos, and far from football.
-
5
Millions of videos have their point on this map, and your thumb will swipe again in a second or two. The app cannot watch them all for you. So which ones does it even look at?
-
Quick check
A friend who watches cooking.
Show the answer
Near the cooking videos. Their list is built from the lists of the videos they watched.
-
Go further
How long a real list is, and what else goes into yours.