New AI Trick Makes Recommendations Scarily Accurate
Why your feed finally knows what you actually want
Deep Dive
Instead of compressing an entire user history into one generic summary, a new approach tokenizes each individual interaction—user, item, context, and outcome—into a single compact Event Token. This gives LLM-based recommenders richer signals at every position, and because the tokens are precomputed and cached, higher snapshot resolution doesn't add real-time serving cost. The method improves ranking and retrieval quality and can even boost a non-LLM ranker when added as historical features.
Key Points
- AMBER compresses a full user action (clicks, time, context) into a single token, letting AI see more without slowing down.
- On industrial-scale tests, it improved recommendation quality while using the same computing power — and even helped older ranking systems.
- The same tokenizer works across different content types, hinting at one AI system powering many apps.
Why It Matters
Smarter, faster recommendations mean less time scrolling and more time enjoying what you actually like.