It also describes follow-up questions after “Interested” and “Not interested” feedback, allowing users to explain in free-form language what they want more or less of. Meta says recommendations will adapt in real time, including within the same session.
The thesis is: Free-form language is becoming a direct control surface for recommendation ranking, not merely a way to search for content. The mechanism is: Language models interpret interests, context, and intent expressed in text, then combine that input with existing engagement signals such as likes, comments, and watch behavior. Why now: Meta says its ranking systems already infer interests from engagement signals and presents LLMs as a way to make that inference more responsive to a user’s stated intent.
The first implication is: Users may gain a more legible way to steer algorithmic feeds than implicit engagement alone. The first watch condition is: Whether users can see persistent, measurable changes in recommendations after giving natural-language feedback. The evidence boundary is: This is Meta’s product announcement: it states intended controls and real-time adaptation, but gives no independent test results, rollout details, or evidence of outcomes for users. Further evaluation should distinguish the availability of a text control from whether it reliably changes feed composition.