The Flaw in the YouTube Recommendation Algorithm

YouTube's recommendation algorithm is just a state machine with two states: 'guessing what you want' and 'aggressively reminding you of that one time you clicked on a sea shanty video by mistake.'

In our recsys post-mortem we discovered one accidental rickroll skewed the embeddings so hard that the principal component became “80s nostalgia” - turns out the only thing with higher cosine similarity than Rick Astley is our tech debt

After 15 years of building sophisticated deep learning models with billions of parameters, YouTube's recommendation system still hasn't figured out the difference between 'user clicked on a rickroll link as a joke' and 'user wants to relive their entire 1987 summer camp experience' - proving that even transformers can't transform correlation into causation

This perfectly captures the classic cold-start problem in reverse: YouTube's recommendation system demonstrates impressive recall but catastrophic precision when it decides that one rickroll means you're ready to pivot your entire music taste to the Reagan era. It's like a recommendation engine that learned from a single training example and decided to overfit so hard it turned your feed into a time machine - proof that even with petabytes of behavioral data, sometimes the algorithm just really wants you to never give up on 80s synth-pop

Apparently their bandit runs with epsilon=0 and watch-time as the only reward, so one rickroll overfits my feed for a fortnight

One rickroll from social and YouTube’s recsys pins your latent vector to 80s synthpop - then the bandit sets epsilon=0 and “Never Gonna Give You Up” becomes your entire personality for two weeks

Overfitting 101: one rickroll training sample, and your recsys deploys eternal 80s pop exploitation