LinkedIn feed rebuilt with 360Brew
LinkedIn has reportedly overhauled its feed using a large 150 billion‑parameter model called 360Brew to prioritize credibility, profile alignment, and authentic engagement over simple like counts. The change signals that content relevance and professional fit may now matter more than hashtag tactics for distribution on the network. (x.com)
LinkedIn’s move to a single AI ranking system has already changed creator performance: an analysis of more than 3 million posts found median reach fell about 47% year‑over‑year, video reach dropped roughly 72%, and reliable text posts saw about a 34% decline after the overhaul. (authoredup.com) The same change put heavier weight on how well a person’s profile and past activity match a topic, and on deeper interactions like saves and substantive comments rather than raw like counts or hashtag tricks; LinkedIn’s new model now influences not just the main feed but job recommendations, search results, and people suggestions. (arxiv.org) (leaders.social) Technically, LinkedIn described 360Brew as a 150‑billion‑parameter “decoder‑only” foundation model — parameters are the internal values the model adjusts during training, a decoder‑only architecture is a type of language model that predicts the next token in sequence (useful for generating or scoring text), and “foundation model” means a large pre‑trained system that can be applied across many recommendation tasks. (arxiv.org) (alphaxiv.org) Under the hood LinkedIn runs a two‑stage pipeline: an LLM‑based retrieval step turns profiles and posts into numerical vectors (dense embeddings) to quickly narrow hundreds of millions of candidate posts to a few thousand, then 360Brew (the ranking model) scores those candidates for relevance to each member’s professional interests. (humai.blog) (arxiv.org) Practical signal shifts that matter for portfolio and content strategy: LinkedIn’s analyses and practitioner writeups flag profile fields (headline, About, Featured items) as primary classification signals, “saves” and reading/dwell metrics as far more influential than simple reactions (one account of the new weighting described a save as roughly five times the impact of a like), and the platform appears to deprioritize generic or template‑style posts. (krucho.ski) (authoredup.com) (yepads.com) Agencies and analysts recommending next steps point to tightening topic pillars (consistent, narrow subject areas shown across profile and posts), featuring two to three portfolio pieces that explicitly demonstrate domain outcomes, and documenting measurable results within post copy and Featured items so the model and human reviewers can associate concrete impact (metrics, client names, case outcomes) with a clear professional identity. (leaders.social) (upgrowth.in)
