Pinterest swaps clustering pipeline for a fine-tuned 4B LLM, lifts email CTR 1.1%
nilukush · reddit · 2026-11-15
Pinterest has replaced its multi-stage clustering pipeline for user-journey inference with a fine-tuned 4-billion-parameter LLM, reporting a 1.1% lift in email click-through and 1.3% higher push notification opens.
- Pinterest defines a "journey" as a goal connecting interests, intent, and context across sessions — e.g., a kitchen renovation surfaces as searches for cabinets, saved countertop ideas, then backsplash tiles and lighting
- Instead of clustering keywords into interests, the model reads raw activity histories to infer longer-term goals and feeds them into Pinterest's existing recommendation stack
- Per the architecture diagram, the system spans RQ-VAE semantic IDs and prompt construction, CPT/SFT/DPO training alignment, an NVIDIA Dynamo + vLLM (KV-cache-aware) serving stack, a feedback-as-edit loop, and safety filtering before journeys land in the store for retrieval, ranking, and notification delivery
More from Companies & People
- Report: Nvidia's Jensen Huang told an exec "I hope you die" in a meeting — 0xAllen_ · 2026-11-15
- Box CTO Ben Kus: a newer AI model doesn't mean better results, it may just act differently — Sam_Witteveen · 2026-11-15
- Why No One Can Replicate Jensen Huang's Asian GPU Supply Chain — _onionesque · 2026-11-15
- We Now Have a Spacecraft That Can Take Us to Mars and Bring Us Back — XFreeze · 2026-11-15
- Why Did Intelligence Officials Wait a Year Before Flagging Jensen Huang Rumor? — kevinnbass · 2026-11-15
- Burkov Releases 'The Hundred-Page Language Models Book' With Hands-On PyTorch, From $25 — burkov · 2026-11-15