--- schema: 1 type: video title: 'Why LLM Recommenders Will Be AI''s Biggest Consumer App — Devansh Tandon, Meta' description: >- Four of the 10 most-used apps in the world are content feeds, and for an hour of engagement they can be up to 100x cheaper to run than AI chat apps. The reason is that a feed decodes pointers to existing content instead of generating every token itself. Devansh Tandon, who leads Meta Recommendations Research, argues that recommendation systems scale just like LLMs, that the field is still early on that curve, and that the LLM recommender will become one of the biggest consumer applications of AI. Tandon lays out four S-curves the industry is climbing: traditional recsys, LLM-inspired, LLM-native and agentic. He then walks through the recipe for building an LLM recommender. First, tokenize content with semantic IDs, which turns a three-minute Reel from about 10,000 tokens into about 10. Next, make the model bilingual in English and your catalog. Then post-train it to rank, with chain-of-thought reasoning you can read. He also shows how this makes feeds steerable, like Instagram's Your Algorithm, where you can see and edit what the algorithm thinks you like. Speaker info: X/Twitter: @devanshtandon_ (https://x.com/devanshtandon_) LinkedIn: https://www.linkedin.com/in/devanshtandon/ Related links: Meta AI: https://ai.meta.com Timestamps: 0:00 Intro: two big arguments 0:45 About Devansh 1:30 Semantic IDs and generative retrieval go mainstream 2:15 Scaling laws, from LLMs to recommenders 3:45 Real scaling curves at Meta 4:15 Instagram Reels: 30% more watch time 4:55 The tokens in, engagement out flywheel 5:40 Four S-curves of recommendation 6:35 LLM-native recommenders 6:50 Agentic recommenders 8:05 The three-step recipe for an LLM recommender 8:35 The five-layer cake 9:35 Semantic IDs: a Reel in 10 tokens 10:45 Pre-training on English and semantic IDs 11:25 Post-training: an LLM re-ranker with chain of thought 12:25 Steerable feeds: talking to your Instagram algorithm 13:15 Explaining recommendations 13:50 Prompted playlists, custom feeds and Ask DoorDash 14:05 Feeds vs. chat apps: the same flywheel 14:50 Why feeds are 100x cheaper per hour of engagement 16:25 Why LLM recsys is AI's biggest consumer app 17:40 Wrap-up --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/lIgdnF0s0kQ-youtube-122d1d966cb579cf6435 Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- canonical: 'https://liquidgold.cc/watch/lIgdnF0s0kQ' markdown: 'https://liquidgold.cc/watch/lIgdnF0s0kQ.md' llms: 'https://liquidgold.cc/watch/lIgdnF0s0kQ/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:20:49.017Z' language: null children: [] videoId: lIgdnF0s0kQ durationSeconds: 1081 publishedAt: '2026-09-25T14:00:27.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 24 endSeconds: 47 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=24' heldReason: rights - ordinal: 1 text: null startSeconds: 57 endSeconds: 101 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=57' heldReason: rights - ordinal: 2 text: null startSeconds: 146 endSeconds: 234 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=146' heldReason: rights - ordinal: 3 text: null startSeconds: 234 endSeconds: 246 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=234' heldReason: rights - ordinal: 4 text: null startSeconds: 257 endSeconds: 270 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=257' heldReason: rights - ordinal: 5 text: null startSeconds: 304 endSeconds: 316 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=304' heldReason: rights - ordinal: 6 text: null startSeconds: 362 endSeconds: 396 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=362' heldReason: rights - ordinal: 7 text: null startSeconds: 396 endSeconds: 407 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=396' heldReason: rights - ordinal: 8 text: null startSeconds: 418 endSeconds: 429 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=418' heldReason: rights - ordinal: 9 text: null startSeconds: 487 endSeconds: 498 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=487' heldReason: rights - ordinal: 10 text: null startSeconds: 524 endSeconds: 534 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=524' heldReason: rights - ordinal: 11 text: null startSeconds: 589 endSeconds: 648 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=589' heldReason: rights - ordinal: 12 text: null startSeconds: 658 endSeconds: 691 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=658' heldReason: rights - ordinal: 13 text: null startSeconds: 702 endSeconds: 750 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=702' heldReason: rights - ordinal: 14 text: null startSeconds: 761 endSeconds: 803 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=761' heldReason: rights - ordinal: 15 text: null startSeconds: 839 endSeconds: 849 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=839' heldReason: rights - ordinal: 16 text: null startSeconds: 894 endSeconds: 905 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=894' heldReason: rights - ordinal: 17 text: null startSeconds: 905 endSeconds: 990 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=905' heldReason: rights - ordinal: 18 text: null startSeconds: 990 endSeconds: 1002 watch: 'https://liquidgold.cc/watch/lIgdnF0s0kQ?t=990' heldReason: rights --- # Why LLM Recommenders Will Be AI's Biggest Consumer App — Devansh Tandon, Meta Four of the 10 most-used apps in the world are content feeds, and for an hour of engagement they can be up to 100x cheaper to run than AI chat apps. The reason is that a feed decodes pointers to existing content instead of generating every token itself. Devansh Tandon, who leads Meta Recommendations Research, argues that recommendation systems scale just like LLMs, that the field is still early on that curve, and that the LLM recommender will become one of the biggest consumer applications of AI. Tandon lays out four S-curves the industry is climbing: traditional recsys, LLM-inspired, LLM-native and agentic. He then walks through the recipe for building an LLM recommender. First, tokenize content with semantic IDs, which turns a three-minute Reel from about 10,000 tokens into about 10. Next, make the model bilingual in English and your catalog. Then post-train it to rank, with chain-of-thought reasoning you can read. He also shows how this makes feeds steerable, like Instagram's Your Algorithm, where you can see and edit what the algorithm thinks you like. Speaker info: X/Twitter: @devanshtandon\_ (https://x.com/devanshtandon\_) LinkedIn: https://www.linkedin.com/in/devanshtandon/ Related links: Meta AI: https://ai.meta.com Timestamps: 0:00 Intro: two big arguments 0:45 About Devansh 1:30 Semantic IDs and generative retrieval go mainstream 2:15 Scaling laws, from LLMs to recommenders 3:45 Real scaling curves at Meta 4:15 Instagram Reels: 30% more watch time 4:55 The tokens in, engagement out flywheel 5:40 Four S-curves of recommendation 6:35 LLM-native recommenders 6:50 Agentic recommenders 8:05 The three-step recipe for an LLM recommender 8:35 The five-layer cake 9:35 Semantic IDs: a Reel in 10 tokens 10:45 Pre-training on English and semantic IDs 11:25 Post-training: an LLM re-ranker with chain of thought 12:25 Steerable feeds: talking to your Instagram algorithm 13:15 Explaining recommendations 13:50 Prompted playlists, custom feeds and Ask DoorDash 14:05 Feeds vs. chat apps: the same flywheel 14:50 Why feeds are 100x cheaper per hour of engagement 16:25 Why LLM recsys is AI's biggest consumer app 17:40 Wrap-up --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/lIgdnF0s0kQ-youtube-122d1d966cb579cf6435 Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- Canonical: https://liquidgold.cc/watch/lIgdnF0s0kQ · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:12 to 1:41) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:41 to 2:15) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (2:15 to 3:54) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (3:54 to 5:04) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (5:04 to 5:50) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (5:50 to 6:47) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (6:47 to 8:07) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (8:07 to 8:44) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 8: Document excerpt (8:44 to 9:38) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 9: Document excerpt (9:38 to 10:48) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 10: Document excerpt (10:48 to 11:31) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 11: Document excerpt (11:31 to 12:30) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 12: Document excerpt (12:30 to 13:23) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 13: Document excerpt (13:23 to 14:09) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 14: Document excerpt (14:09 to 14:54) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 15: Document excerpt (14:54 to 16:30) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 16: Document excerpt (16:30 to 18:01) > Held: transcript rights do not permit this chunk's text.