Meta's Devansh Tandon: LLM recommenders are AI's biggest consumer app, 100x cheaper than chat
AI Engineer · youtube · 2026-09-25
Meta Recommendations Research lead Devansh Tandon argues LLM recommenders will be AI's biggest consumer app: 4 of the world's 10 most-used apps are feeds, and feeds are up to 100x cheaper per engagement hour than chat apps since they decode pointers to existing content rather than generating tokens.
Key points
- Four S-curves: traditional recsys, LLM-inspired, LLM-native, agentic
- Recipe: semantic IDs (a 3-min Reel from 10,000 tokens to 10), pre-train bilingual on English + catalog, post-train a chain-of-thought re-ranker
- Meta reports 30% more watch time on Instagram Reels from this approach
- Steerable feeds: Instagram's Your Algorithm lets users inspect and edit what the model thinks they like; prompted playlists, custom feeds, Ask DoorDash
More from Apps
- Microsoft rebuilds Copilot into a super app with persistent Autopilot agents — rohanpaul_ai · 2026-09-25
- From technical drawings alone, Claude builds interactive smelting furnace in Blender — EricBuess · 2026-09-25
- invideo launches Effects: build custom timeline effects by describing them in chat — Uncanny_Harry · 2026-09-25
- Solo dev builds rolefate, a data-driven tracker of AI's impact on jobs — DependentCodd · 2026-09-25
- User turns to AI voice cloning to preserve late father's voice from tiny audio snippets — Pure-Cover4679 · 2026-09-25
- Lila Sciences' AI lab screens 2,942 catalysts in 3 months, cracks green hydrogen's OER problem — tw_killian · 2026-09-25