马斯克罕见低头：开源𝕏推荐算法，自嘲“很烂”不过未来月更

Core Viewpoint - GitHub has made Elon Musk's open-source recommendation algorithm system fully visible, which is primarily driven by AI models [1][2] Group 1: Algorithm Transparency and Community Reaction - The open-source announcement has generated significant excitement within the community, with many praising the transparency of the system [2] - Musk acknowledged the algorithm's shortcomings, stating it is "dumb" and requires substantial improvements, but emphasized the importance of transparency in the improvement process [4][5] - Musk has consistently criticized the previous platform's lack of openness and has followed through on his promise to publicly share Twitter's core recommendation algorithm since the acquisition [6][7] Group 2: Algorithm Mechanism - The recommendation system is built on a Transformer architecture similar to Grok-1, which learns from users' historical interactions (likes, replies, retweets) to recommend content [9] - The system begins by identifying the user and their recent activities, aiming to create a "real-time user profile" without pre-set assumptions [12][13] - Two types of user information are collected: Action Sequences (direct interest signals) and Features (long-term attributes) [14] Group 3: Content Filtering and Scoring - The algorithm filters through a vast amount of tweets to select a few thousand potentially relevant ones, using both familiar and external sources [16][17] - The system employs a Hydration module to complete candidate tweet information and a Filtering module to eliminate unwanted content [21][22] - The final scoring is done by a Phoenix ranking model, which predicts various user interactions and assigns scores based on weighted combinations of these predictions [25][26] Group 4: Key Features of the System - The system is purely data-driven, rejecting manual rules and allowing AI models to learn directly from raw user data [33] - It utilizes a candidate isolation mechanism to ensure independent scoring of each piece of content [34] - The algorithm predicts multiple user behaviors rather than providing a single recommendation score [36] - The modular design of the system supports rapid iteration and development [37] Group 5: Acknowledgment of Limitations - Despite the praise for transparency, the algorithm has been criticized for certain flaws, such as the lack of a time decay mechanism for "block" signals, which could negatively impact account recommendations [39][41] - Musk himself acknowledged the algorithm's deficiencies, indicating a commitment to ongoing improvements and updates every four weeks [42][44]