Meta Runs Weekly AI Over Trillions of Friendships to Pick Your Reels

What happened

Meta's Friend Bubbles feature for Facebook Reels runs weekly machine learning inference across trillions of person-to-person connections to surface videos your friends liked. The system fuses two signals: survey-based closeness scores and real-time interaction data. A conservative threshold gates which friends' activity actually appears as bubbles on your feed, cutting clutter while boosting watch time. The technical constraint? Prefetch friend metadata alongside video content so scrolling never stutters, even on budget phones. The ranking model weighs P(video engagement | bubble impression) against pure content quality, letting engineers tune how much friendship trumps algorithmically perfect suggestions.

Why it matters

Social context is no longer a nice-to-have overlay but a core ML problem requiring its own retrieval pipeline and continuous retraining loop.

Sources