On X, not all engagement is equal.
Elon Musk’s social media company on Thursday released the source code that determines how posts are recommended to users, the latest development in Musk’s quest to champion transparency on the platform.
“The goal is straightforward — we want people to be able to answer for themselves whether a platform is limiting their reach, whether the system is fair, and why they see particular content,” the company said in an X post.
Musk said on X that the decision to make X open source was to “improve fairness” and seek feedback for improvement.
Here’s a breakdown of what was revealed in the code.
Positive weights
Posts are ranked for a user’s For You timeline in part based on how the algorithm predicts that user will interact with them.
The algorithm, which X calls “Phoenix,” examines a user’s recent viewing history to predict how that user will interact with a post from an account they don’t follow.
If Phoenix predicts a user will share the post by copying its URL, the post is awarded a weight about 40 times as large as if Phoenix predicts a post will receive a “like.”
Replies, quotes, and shares by direct message each carry a weight of 5, which is 10 times the 0.5 weight assigned to a predicted like. For follows, the weight is equivalent to eight likes. For reposts, the weight is two likes.
Phoenix combines those predicted probabilities with their respective weights to produce a ranking score, which X then uses to decide the order of posts in the For You feed.
Negative weights
If Phoenix predicts that a user will demonstrate displeasure with a post, negative weights will severely limit a post’s visibility.
Phoenix treats a predicted “report” as 468 times — in the opposite direction — as a predicted like. A predicted mute carries a negative weight about 118 times the size of a predicted like. A ‘not interested’ interaction is about 86 times as large in the negative direction, and a block is about 62 times as large.
The system suggests that “ragebaiting,” or posting inflammatory content meant to spur angry replies, might not be a recipe for growing a following, which is maybe surprising given X’s reputation under Musk.
Looking under the hood
Also included in X’s Thursday announcement is a pilot feature called “Under the Hood,” which lets participating users see aggregate information about labels applied to their account and posts that can affect visibility.
X’s source code shows that its systems classify content and accounts across categories, including spam and adult material. Those labels can then be used by a separate system when determining whether a post is shown.
X said the feature has only been activated for a “randomized test group of eligible accounts,” and that a wider rollout will depend on the feedback the company receives.
Making X open source
X’s Thursday release of its recommendation algorithm source code is the latest push to open-source parts of the platform.
In January, Musk called the algorithm “dumb” and pledged to make the system more transparent.
“At least you can see us struggle to make it better in real-time and with transparency,” Musk wrote on X. “No other social media companies do this.”
