Threads is running an AI system that turns feed complaints like less politics into actual changes to what you see next.
A new arXiv paper describes Dear Algo, an agentic intent layer deployed on Threads that interprets open-ended requests - explicit, inferred, negative, or compound - and compiles them into an executable retrieval plan, rather than a one-off search result. The system then hands off to existing retrieval and optional semantic or multimodal reranking, letting search and recommendation share one intent-processing pipeline instead of running as separate products. In a blinded audit of 300 public request-item pairs, an LLM-based judge measured 94.4% exact-relevant precision. A serving-path study over the reranker's first 72 eligible hours found the irrelevant share among judged results fell from 4.78% to 2.80% compared with an off-path baseline.
Feed algorithms are usually a black box you nudge by scrolling past things you dislike. Dear Algo treats a plain-language request as a real signal, effectively letting users edit their own ranking function without touching a settings menu. That is a different pitch than X's not-interested button or TikTok's implicit-signal-only model, and it puts Meta ahead of rivals in treating natural language as a first-class recommendation input, not just a search query.
The precision numbers come from an LLM judge grading the system's own output, not from the people typing less politics - so how this holds up outside a controlled audit is still an open question.