Why LLM Recommenders Will Be AI's Biggest Consumer App — Devansh Tandon, Meta

Recommendation systems follow power-law scaling curves similar to large language models, positioning the field at an early stage of development. The integration of LLMs into recommendation engines—specifically through semantic IDs and generative retrieval—enables more efficient, steerable, and interactive user experiences. This transition moves systems from traditional, black-box architectures toward LLM-native and agentic paradigms where models reason over user history and content. LLM-based recommenders are structurally more token-efficient than generative chat applications because they decode pointers to existing content rather than generating new text, making them a highly scalable consumer application. By leveraging a "tokens in, engagement out" flywheel, companies can optimize model training and inference to drive significant improvements in engagement and monetization across major platforms like Instagram and YouTube.