Can't wait for when I can vibe code a production recommender system. Until then, here's some system designs: • Retrieval vs. Ranking: https://t.co/zTsfElM3R7 • Real-time retrieval: https://t.co/hSfRr3Ibwm • Personalization: https://t.co/Qm3kScI7dI • Bandits:... See more

A central challenge for any recommendation algorithm is the tradeoff between safe but somewhat boring recommendations that are similar to recommendations that worked well in the past (“exploitation”), and risky recommendations that are unlikely to be good but have a high payoff if they do turn out to be good (“exploration”). Exploration lets the

... See more

Like a completions-based product, the advantage of a feed is that users don’t have to interact with a chatbot. The inputs to the model come from how the user interacts with the feed (likes, scrolling speed, time spent looking at an item, and so on). Users can experience the benefits of an LLM-generated feed (if any) without having to change their... See more

The two-tower era: efficient but static

The two-tower architecture has been the workhorse of modern recommendation systems, and for good reason: it’s elegant, scalable, and remarkably effective at learning patterns from massive amounts of data. Here’s how it worked for us once we deployed it in early 2024.

The system had two neural networks working... See more