LinkedIn quietly replaced its feed ranking system in 2026. Not with a tweaked version of the old one. With a single 150-billion-parameter language model called 360Brew, built on top of LLaMA 3 and fine-tuned on internal data. They published the technical details on their engineering blog. Most people did not read it. Here is what the paper actually says, with the parts that matter for anyone building on or analyzing the platform. What it replaced The old LinkedIn ranking pipeline was a chain of around thirty specialized models. Each one scored a numerical feature: dwell time, sender-receiver affinity, click-through rate, comment likelihood, and so on. The features were stitched together by a final ranking layer that picked which posts to show. This is the same pattern most social platforms have used since the 2010s. It is fast, cheap, well-understood, and easy to A/B test. The downside is that it can only see what the engineers thought to measure. Anything outside those features is invisible.…