How LinkedIn rebuilt search with AI: 2-tower embeddings replace keyword matching, dwell time over likes
_jaydeepkarale · x · 2026-09-16
A thread by jaydeepkarale breaks down LinkedIn's AI-powered search overhaul. The old keyword engine either returned nothing when queries didn't match all keywords, or matched keywords while missing the intent. The new two-layer design: a retrieval layer combining token-based and embedding-based retrievers (a 2-tower architecture ranking by cosine similarity, with post embeddings precomputed offline in batch), followed by a multi-stage ranking layer. Post quality is scored on on-topic rate and long-dwell time — likes and comments are deliberately excluded since they're easy to farm.
More from Research
- Three Years On: GPT-4's Eureka Taught a Robot Hand to Spin Pens in Isaac Gym — JasonMa2020 · 2026-09-17
- Q-Planning Enables Self-Improving Robot Policies: LIBERO-10 Hits 99% With Frozen BC — chris_j_paxton · 2026-09-17
- GPT-6 Astra Autonomously Trains a Pen-Spinning RL Policy in 36 Hours — JasonMa2020 · 2026-09-17
- First Thermodynamically Favored DNA Computer Demonstrated on 100-Bit Addition — EricTopol · 2026-09-17
- OmniHarness: symbolic policy learning boosts generalizable visual generation — Xu Xu · 2026-09-17
- TokenRhythm Launches NeoHorse-1: 4B/9B Models Post-Trained on Agent Execution Traces — rohanpaul_ai · 2026-09-17