马斯克开源推荐算法，完全基于AI大模型

Core Insights - Elon Musk announced that the core algorithm for the social platform's "For You feed" has been officially open-sourced [12] - The algorithm retrieves, ranks, and filters posts from two main sources: Thunder (posts from accounts users follow) and Phoenix Retrieval (posts discovered from a global corpus) [12] - The algorithm utilizes a Grok-based Transformer model to merge and rank content, predicting the adoption probability of each post [12] System Architecture - The system relies entirely on a Grok-based Transformer to learn relevance from user interaction sequences, eliminating the need for manual feature engineering [6] - Candidate models are isolated during the Transformer inference process, ensuring that post scores do not depend on the content of other posts in the batch, which guarantees score consistency and cacheability [7] - Multiple hash functions are used for embedding lookups in both retrieval and ranking processes [8] Prediction and Scoring - The model predicts multiple behavior probabilities rather than a single "relevance" score, enhancing the understanding of user engagement [9] - The final score for each post is a weighted combination of predicted adoption probabilities, with higher scores indicating a stronger recommendation for users [12] Pipeline Architecture - The candidate-pipeline framework provides a flexible structure for building recommendation pipelines, separating pipeline execution and monitoring from business logic [10] - It allows for parallel execution of independent stages and graceful error handling, facilitating the addition of new sources, hydration, filters, and scoring components [10]