X has updated its open-source 'For You' information flow algorithm code on GitHub as of May 15, according to Foresight News. The update includes a fully operational end-to-end inference pipeline. The system utilizes the Thunder component to fetch content from accounts users follow, while the Phoenix component employs machine learning to retrieve relevant content from a comprehensive corpus. Interaction behaviors such as likes, replies, retweets, and clicks are scored and ranked using a Transformer model modified from Grok-1.

The update introduces several new features, including an ad injection module, a content understanding classifier for spam detection and policy enforcement, and multiple hydrator data expansion components. X has stated that all manually designed features have been removed, relying entirely on the Transformer model for relevance assessment.