So a couple years back I’m at a friend’s wedding, and this guy I barely know finds out I used to work at Facebook. First question, dead serious: “Okay, be honest with me. Is my phone listening to me? Because I swear I talked about hiking boots out loud and then I saw an ad for hiking boots.”
I get this question a lot. And the answer is almost always no, your phone isn’t secretly recording you. What’s actually happening is way less spooky and honestly way more interesting.
Here’s the one mental shift that unlocks the whole thing: stop picturing the algorithm as a mind-reader, or as some censor that “decides” what you deserve to see. It’s a prediction engine. That’s it. Every single post in your feed got there because a model guessed “this person is X percent likely to like, comment on, watch, or linger on this,” and that guess beat out everything else fighting for that slot. The hiking boots thing? You just fit a pattern with thousands of other people who looked at boots and then bought a pair. No microphone required.
Once that clicks, the weird stuff starts to make sense. Why does your feed keep feeding you rage-bait? Because the system optimizes for some metric, usually watch time or “meaningful interactions,” and anger keeps you scrolling. The algorithm isn’t evil. It’s just really, really good at hitting the number somebody told it to hit. Which, you know, is its own kind of problem. But that’s a different coffee.
For a resource: when I was starting out, the thing that made it real for me was a paper called “Deep Neural Networks for YouTube Recommendations” (Covington and co, 2016). It’s a real product team explaining how they actually do it, in two stages, candidate generation and then ranking. Not theory. Production. Go read it.
And if you want something newer and kind of wild, in 2023 Twitter open-sourced a big chunk of its recommendation algorithm on GitHub. You can literally read the real ranking logic. That basically didn’t exist when I was learning this stuff.
But the fastest way to get it? Build a dumb little recommender yourself. Grab a public dataset, write a hundred lines of Python, watch it suggest things. The mystery dies the second you’ve built one. It’s not magic. It’s just math, plus a metric somebody chose.
