X Releases Timeline Algorithm – Transparency with Gaps

The platform X has published the source code of its “For You” recommendation algorithm on GitHub. Company owner Elon Musk announced that the code will be updated every four weeks in the future. In a post on his network, he candidly admitted: “We know this algorithm is dumb and needs massive improvements, but at least you can watch us fight to make it better in real-time.”

Phoenix instead of Heavy Ranker

The new algorithm first collects the candidates that are eligible for the timeline. A component called Thunder delivers “In-Network” candidates.

The component central to ranking also bears the name Phoenix. Phoenix is based on the AI model Grok from Musk's company xAI. The model analyzes each user's behavioral history: likes, replies, time spent, profile visits, but also blocks and mutes.

Based on this, it predicts possible user reactions for each post: positive actions such as liking, replying, retweeting, quoting, clicking, profile visits, video playback over 50 percent, image enlargement, sharing, long dwell time, and following – as well as negative ones such as “Not interested,” blocking, muting, and reporting. Each predicted action is multiplied by its probability and assigned a weighting factor. The sum results in the final score. The higher the score, the more prominently the post appears in the feed.

Phoenix thus fundamentally differs from the version released in 2023 called “Heavy Ranker.” At that time, Twitter still relied on classic machine learning: engineers manually defined hundreds of features – whether a post contained images, how many followers the author had, when it was posted – and weighted them manually.

The current documentation clarifies: “We have removed all manually created features and most heuristics from the system. The Grok-based Transformer does all the work by analyzing your interaction history (what you liked, what you replied to, what you shared, etc.) and using this information to determine which content is relevant to you.”

Key Information Missing

The central problem with the released algorithm: While X publishes the architecture and logic, it does not publish the specific weighting factors. Users and researchers can therefore understand that a block is rated negatively and a reply positively – but not how strongly.

In addition to the weightings, other key information is missing: the internal parameters of the Phoenix model are not public. The same applies to the training data – which user interactions were used for training and in what period remains in the dark. All in all, one learns that Phoenix calculates a weighted sum of behavioral predictions – but not what exactly happens in the black box of the neural network.

The release does not seem to be coincidental, as X is under considerable pressure from European regulatory authorities. Several investigations and proceedings are already underway based on European digital laws. Just at the beginning of December, the EU had imposed a fine of 120 million euros on Musk's online platform due to transparency deficiencies. The open-source strategy can be seen in this context as an attempt to proactively address the accusation of a lack of transparency and to accommodate the regulatory authorities.

With the release, X has disclosed more than other social networks. Researchers and regulators can understand the fundamental architecture of the network. However, this is not sufficient for genuine verifiability of algorithmic decisions – for example, whether certain political content is systematically favored or disadvantaged.

(jo)