Musk drops a bombshell: open-sourcing X platform's latest algorithm! Integrating the Grok AI model architecture and directly releasing a 3GB pre-trained model for you to play with

📄Full Article· Automatically extracted by trafilaturaGemini 翻譯1583 words

Musk has open-sourced it again! The X platform (formerly Twitter) today made a major release of the latest "For You" recommendation algorithm source code. This major overhaul abandons traditional hand-engineered rules and fully adopts the "Phoenix" Transformer model based on the xAI Grok architecture to predict user interactions. What excites developers most is that the team not only released the code, but also directly bundled a 3GB pre-trained mini model along with a complete end-to-end execution pipeline — arguably the most sincere open-source gift package the industry has ever seen. (Background: Musk's xAI launches "Grok Build" to challenge Claude: up to 8 parallel AI agents, with a context window of 2 million tokens) (Context: Musk announces xAI is no longer an independent company, renaming it "SpaceXAI" and positioning it as an AI product under SpaceX) Open-source enthusiast Elon Musk has once again delivered on his promise! Today (the 15th), Musk published a brief yet shocking post on the X platform that reverberated through the tech world: "The latest 𝕏 algorithm has been published to GitHub." According to detailed information in the GitHub repository (xai-org/x-algorithm), this is not merely a stripped-down architectural overview, but a "truly executable" open-source version that comprehensively reveals the core secrets driving the X platform's "For You" timeline. The latest 𝕏 algorithm has been published to GitHubhttps://t.co/ZCOm51uxmh — Elon Musk (@elonmusk) May 15, 2026 In this major update on May 15, 2026, X's recommendation algorithm has undergone a fundamental shift in design philosophy: it has almost completely removed hand-engineered features and traditional heuristic rules. In their place is a system that relies entirely on the deep understanding of artificial intelligence. The core of this system is named "Phoenix," a Transformer model ported from the open-source version of xAI's large language model Grok-1 and optimized for recommendation systems. By analyzing users' past interaction history, it accurately predicts the probability that a user will like, reply, repost, or click on any given post. According to the open-source documentation, X's recommendation architecture is primarily controlled by an orchestration layer called "Home Mixer." When you open the app, the system simultaneously sources posts for you from two main pipelines: - Thunder (in-network content): Receives data in real time via Kafka, quickly delivering the latest posts published by "accounts you follow." - Phoenix Retrieval (out-of-network content): Uses a Two-Tower model from machine learning, leveraging vector embeddings to perform similarity searches across the entire network, finding unfamiliar posts you might be interested in. After these candidate posts are retrieved, they go through filtering (removing duplicates, old posts, or blocked content), and are finally handed over to the Phoenix Transformer model for "interaction probability prediction" and "weighted scoring." Notably, the system specifically incorporates an "Author diversity scorer" to prevent a user's timeline from being flooded by a handful of highly active accounts. For AI developers and algorithm researchers worldwide, the most valuable highlight of this open-source release lies in its "operability." The team has added `phoenix/run_pipeline.py` to the repository — an end-to-end inference pipeline that allows developers to directly run through the complete process from retrieval to ranking. Furthermore, the xAI team has directly provided a pre-trained "mini Phoenix model" approximately 3GB in size (featuring 256 dimensions, 4 attention heads, and a 2-layer architecture), enabling anyone to easily download and test it locally. In addition, this open-source release also includes the Grox content understanding system used for spam detection and classification, as well as the mixin system module that handles ad insertion. By laying X's most core business assets completely bare in the sunlight, Musk not only demonstrates his commitment to algorithmic transparency, but has undoubtedly dropped a bombshell on the global open-source community.

Data Status✓ Full text extractedRead Original (動區 BlockTempo)

🔍Historical Similar Events· Keyword + Asset Matching6 items

2026-05-29

Gemini Prediction Market Upgrades AI Brain: Integrates Grok Customization, Automatically Recommends Your Watchlist

Similarity 120%關鍵字 grok同分類 zh

2026-05-23

Grok Build will be integrated into SuperGrok subscriptions, available with the $30/month standard plan.

Similarity 120%關鍵字 grok同分類 zh

2026-05-19

Experiment | Running a Radio Station with AI for Five Months: Grok Hallucinated Sponsors, Gemini Became a Shilling Bot.. Total Revenue Only a Few Hundred USD

Similarity 120%關鍵字 grok同分類 zh

2026-05-15