X Open Source Algorithm Weights: Originality and Interaction Determine Recommendations, Like Weight Only 0.5, Homepage Visits and Stay Duration Weights Extremely Low
Details of X's open source For You recommendation algorithm weights show that the system primarily scores based on predicting user behavior.
Grok predicts whether readers will like, reply, share, copy links, or follow, with positive actions increasing scores; reporting, blocking, and "not interested" significantly deduct points.
In the same reader's information stream, the score for the second post drops to 62.5%, the third post to about 44%, and the lowest is no less than 25%, clearly limiting repeated exposure.
Original posts receive no discounts, and replies and shares from accounts that are not followed generally do not enter the timeline; even for followed accounts, replies and shares are multiplied by 0.75, while original content from followed accounts retains full scores.
The reordering model after scoring will disperse similar content, meaning that even high-scoring posts may move back due to similarity.
Like weight is only 0.5, and homepage visit and stay duration weights are extremely low; unfamiliar replies and shares are filtered out before scoring, and reporting predictions can lead to severe downgrades.
In terms of market mechanisms, after the algorithm is made public, creators are more inclined towards high-quality originality and genuine interaction, with funds and traffic flowing towards content that can spark discussion and sharing; the beneficiaries are high-quality original accounts, while those under pressure are accounts relying on inflated metrics, spam replies, and borderline content.
Source: Public Information
ABAB AI Insight
X's open sourcing fully demonstrates a scoring and multi-discount mechanism based on behavior prediction, continuing previous transparency efforts, transforming "traffic limits" from black box speculation into rules that can be referenced in code.
In terms of capital pathways, by publicly disclosing weights, it guides creator behavior towards the types of interactions the platform expects, reducing the distribution costs of low-quality content; the motivation is to enhance timeline quality and user retention, strategically shifting the algorithm from a controversial focus to an optimizable public system.
Similar cases can be seen where other platforms have limited public disclosure of recommendation logic leading to adjustments in creator strategies; current social recommendations are transitioning from implicit optimization to explicit rule guidance.
Structural judgment belongs to technical substitution: behavior prediction and multiple penalties replace simple popularity ranking, with the mechanism being that when the system can accurately predict negative feedback, it can filter out low-quality content in advance and reward posts that genuinely provoke active participation.
ABAB News · Cognitive Laws
The algorithm rewards active sharing, not passive liking.
Repeated exposure penalties can determine visibility more than content quality.
After publicizing weights, inflated metrics strategies become ineffective first.