Mosseri corrects a common misconception: rather than 'knowing' that a user likes surfing, the algorithm relies on massive, non-human-readable embedding vectors, and only now with LLMs can that be translated into semantic descriptions. ✦ AI generated
Adam Mosseri · Lenny's Podcast · 2026-07-09 · original ↗
starts at this moment · 34:23
“What's something that the Instagram algorithm knows about human behavior that people may not realize?”
One of the most common misconceptions is actually in the opposite direction. I think people assume that there's a much more detailed semantic understanding of everybody's interests and preferences in the algorithm than there is. Most of what's really driven the progress in the world of recommenders... have been these large embedding models... that basically produce artifacts that cannot be read by people.
verbatim transcript · starts at 34:23
34:23>> I want to transition talk to talk about Instagram the product the platform things you guys have learned there. Let me start with this question. What's something that the Instagram algorithm knows about human behavior that people may not realize? >> One of the most common misconceptions is actually in the opposite direction. I think people assume that there's a much more detailed semantic understanding of everybody's interests and preferences in
34:48the algorithm than there is. Most of what's really driven the progress in the world of recommenders over the last 5 10 years have been, you know, these large embedding models and these other techniques that basically produce artifacts that cannot be read by people. They're not legible. They're like giant vectors. It's like, sure, I can show you the vector, but it's just going to be a bunch of numbers
35:11in like a seven dimensional space. It's like, and so when when we talk about does the algorithm know something, usually we think in these more semantic terms. It knows I like surfing and it's like it doesn't. It just has this big ass number that happens to correlate with surfing. Um, that said, I think that is starting to change, right? I think that what one of the things that LLMs are enabling
35:40is they can describe in you know words you know English for or whatever language you prefer what some of those previously illeible artifacts um are at least approximate to if not mean directly right so this is like the thing I've been really I posted about this this week this thing called your algorithm basically the idea is we take a look at all of the stuff that you've
36:05interacted with. And then you know we all of that is in an embedding space. You can think of embedding space as a map. You can map a bunch of videos into the same map. And so videos that are close or similar. And now we can just have an LLM just be like describe that part of the map. And it can be like oh that is like deep pourover coffee