Instagram Does Not Know You Like Surfing. It Just Knows Your Numbers Line Up.
The app ran on unreadable math for years. Now language tools can translate the math into topics, and let you change them.
Lenny's Podcast: Product | Growth | Career · The rise of taste, human authenticity and judgment in an AI world | Adam Mosseri (Head of IG)People imagine Instagram keeps a neat file on them. It would say you like surfing, love sourdough, and are planning a trip to Hawaii. Adam Mosseri, the head of Instagram, says that file does not really exist.
For most of the last decade, progress in recommendations came from what are called embedding models. Every photo, video, and person gets turned into a giant string of numbers, called a vector, and placed on a kind of map. Things that behave similarly end up close together on that map. Those strings are not legible to people. Displayed on a screen, they would look like rows of figures spread across many dimensions.
Mosseri says people assume a much more detailed understanding of their interests and preferences than the system actually has. In his simplified version, the logic was more like, you enjoyed certain pictures, and because other accounts with overlapping tastes also enjoyed additional pictures, the system guessed you would enjoy those extras. The match is numerical rather than verbal, linking your activity pattern with patterns tied to surfing content without naming the interest itself.
That is starting to change. New AI language tools, the same kind of technology behind chatbots, are good at putting words to patterns. Instagram can now point one at a cluster on the map and ask it to describe that neighborhood in plain English. As an example of the kind of label such a tool might produce, Mosseri offered deep pourover coffee snobbery.
That translation powers a new feature called Your Algorithm, which shows you the topics Instagram thinks you are into and lets you add or remove them. Mosseri says Instagram hopes to allow broader requests in the future, like asking for more posts from friends or less of a certain format.
Questions people ask
What is an embedding, exactly?
An embedding is a list of hundreds or thousands of numbers that represents something, like a video or a user, as a point in space. Points that end up near each other have similar patterns of behavior, even if no person ever labeled what they have in common.
Does Instagram need to know my age or location to do this matching?
No. The recommendation system learns from what you tap, watch, and linger on, and from what similar behaving accounts do. It does not need to know your age or job to notice that your pattern matches the pattern of surfing fans.
Can I correct Instagram if it gets my interests wrong?
Yes, in a limited way. Your Algorithm lets you see the plain English topics the system inferred and add or remove them. Mosseri says Instagram hopes to allow broader requests in the future, like more posts from friends or less of a certain format.