The X platform has open-source the algorithm weights of the entire platform, which is rare in history. Grok usually doesn't read the content of a post, but predicts what everyone will do after seeing it. For example, whether you will like, reply, forward, copy links, or follow the author can increase your weight score significantly. However, reporting, blocking, and clicking are not of interest, and each behavior carries negative weight. The system combines these predicted results to determine the final score of this tweet for the user. After calculating all its weight scores (as shown in the figure below), we obtained several key results, namely that scores cannot determine everything. 1. Punishment measures for repeated brushing If your tweet has already appeared in the information flow of the same reader, then your second tweet will only retain a score of 62.5%, and your third tweet will only have about 44% remaining, with a minimum score of no less than 25%. That is to say, the algorithm explicitly restricts the same user from repeatedly brushing through your tweets, and this behavior will lower the weight and score. 2. Original tweets will not receive any discounts X clearly prefers original tweets and is less willing to recommend replies and shares. If readers do not pay attention to you, your replies or reposts will hardly enter the other party's timeline. Even if readers have already followed you, reply and forward content should still be multiplied by 0.75. Only original tweets from accounts that have already been followed can retain full weight. Algorithms intentionally consider replies and forwards as relatively weak content. 3. Similar content filtering After completing all ratings and weights, the Grok system will also use a reordering model to disperse similar posts. Even if your score is good, if the content is too similar to the previous tweet, your ranking may still drop a few places. What things should not be done Don't brush likes, the like weight is only 0.5, which is almost the worst in the entire Grok algorithm. Similarly, personal homepage visits are almost useless. The reason why long tweets have become less important than before is mainly because the score for the duration of stay is no longer as high as before. The weight of whether or not to stay is 0, and the weight score for actual stay time is also very low. Don't treat replying and forwarding everywhere as a way to reach strangers. The replies and reposts from unfamiliar accounts will be filtered before rating, even if the other party has already followed you, you are in a mutual following status. Do not intentionally post content that is easy to trigger reports, block, or be clicked on without interest. As long as the system predicts a significant report, the account will be severely downgraded. X does not encourage tearing and unhealthy content. This may also be the reason why pornographic, borderline, junk content, or abusive tweets have significantly exited the timeline, and the system may be preventing these contents from entering everyone's timeline. The final result, what X truly rewards is not how lively it looks, but whether users are willing to actively participate in discussions, share, and follow the author. The real punishment is not that no one likes it, but that the readers who come across it clearly express that they don't want to see this kind of content again. Finally, thank you to Musk for being willing to open source the X platform algorithm, which has shown us the operation behind X. AB KUAI DONG