GenRec: Towards LLM-Native Recommendation at Netflix

Impressive results in a massive A/B test, replacing hand-tuned ranker famous for how much engineering hours went into it with a "simple" post trained LLM.

I've worked with ranking algorithms at scale (5k reqs/s) and hand tuned, manually crafting features, searching through user logs to see what went wrong. And now we can just pop an LLM on top of that pile of data and it's just better... End of an era, to me.