phed

Before we get into it, I spent a decent amount of time playing up with AI to deliver these threads, so bookmark to come back to it later Let's get into it👇

1. Two brains, not one 🧠 YouTube uses two neural networks: one to choose a few hundred videos (candidate generation), and another to rank them perfectly (ranking)

2. Your watch history is gold The algorithm prioritizes what you've watched before — not likes, not comments, but watch time - It doesn’t care what you like It learns from what people like you watch. You’re grouped with others via collaborative filtering

3. New videos get special treatment YouTube boosts fresh uploads. It corrects for bias towards old content by feeding in the "age" of a video

4. Clicks are nice, but watch time is king Ranking favors videos people actually watch — not ones they just click on. Clickbait is nerfed

5. Search history matters Your searches are tokenized and embedded like watch history — they influence recommendations

6. Power users aren’t favored YouTube balances the dataset so active users don’t dominate. Every user has equal say in training

7. Recency bias is real Videos you were just recommended (and ignored) get downgraded on the next page load

8. Fake engagement fails Watch time is weighted by actual duration watched Skips and early exits = bad signals

9. Freshness vs Relevance YouTube constantly balances new content (exploration) vs. reliable hits (exploitation)

10. It uses your device & region Your location, device, age, gender... they’re all embedded into your user profile

11. Candidate videos are found via nearest neighbors Millions of videos → top 100s via approximate nearest neighbor search in a vector space

12. Impressions are tracked If you’ve seen a video before but didn’t watch, it’s less likely to show up again - The model learns when NOT to recommend Videos that got impressions but no clicks = demoted

13. It tracks your behavior by video ID How many times you watched a channel, what order you watched videos — it all counts

14. Expected watch time > CTR YouTube rewrote logistic regression to predict watch time directly — not just click-through rate

15. YouTube can’t rank every video in the world — so it first picks 100s (candidate generation), then ranks them (ranking) This design makes recommendations efficient across billions of videos (which is why it's important to study what's working in the niche you're entering)

if you enjoyed this thread, drop a follow in the next 7 seconds or get 7 years of the worst luck known to mankind