Facebook's AI Now Guesses Which Group Posts You'll Love
Meta is testing AI that picks your group posts — and it borrows clues from your main feed.
Meta researchers published a paper describing a new way to decide which posts you see in Facebook Forum, a standalone app built for people who live in Facebook Groups. Most recommenders today work by translating every post into a long list of numbers and finding the ones closest to what you liked before — think of shelving books by topic and handing you the nearest one. The new approach, called generative recommendation, skips the shelf. The AI simply writes out the label of the post you should see next, the way your phone autocompletes a sentence.
The catch they had to solve: Forum is new, so there isn't enough clicking data to teach an AI from scratch. Their fix was to borrow from two places. First, they trained on activity across Facebook Groups generally, not just Forum. Second, they reused "semantic IDs" — short numeric labels describing what a post is about — that were already learned from Facebook's main Feed. Those labels are the AI's vocabulary. They then tuned a 3-billion-parameter language model, a mid-sized AI brain comparable to models that run on a single powerful server, to spit out those labels based on your context.
Most of the paper is about testing which choices actually matter. They experimented with how the labels are built, how much of your past activity to feed the model, and whether to include details from your profile. Their headline finding: labels learned on one part of Facebook do carry over to a brand-new surface, which is genuinely useful news for any company launching a fresh app with no history. They also offer practical tips for teams doing this on real social platforms.
The honest limitation is that this is a research paper, not a product launch. There's no word on whether or when it reaches your phone, and no side-by-side proof that it beats the current system for real users. It also depends on mixing behavior data from one Facebook surface into another — exactly the kind of data-sharing that privacy regulators and advocates watch closely.
- Meta tested an AI that writes out the post you should see next, rather than matching similar posts — like autocomplete for your feed.
- Because the new Forum app had too little data, the AI reused labels already learned from Facebook's main Feed and Groups activity.
- The experiment used a 3-billion-parameter language model, and the paper says those borrowed labels did transfer to the new app.
Why It Matters
Better Recommendations could mean less scrolling and more posts you actually want — but it also means your Main Feed Behavior shapes other apps.