X平台在GitHub开源了其“For You”信息流的最新推荐算法。该算法融合来自已关注账号的内容和通过相似性搜索发现的网络外内容，并利用基于Grok架构的Phoenix Transformer模型进行统一评分。模型通过分析用户近期互动，同时预测点赞、回复、转发等多种互动概率，并综合成最终相关性分数。排序过程包括候选内容收集、上下文信息补充、AI模型评分、多样性调整及垃圾信息过滤等步骤。此次开源提供了完整的可运行管道、预训练模型、内容理解服务及内置广告混合模块，核心计算完全由Transformer模型驱动。
NEWS: 𝕏 has published the latest recommendation algorithm on GitHub.
Grok Summary of the new algorithm:
The system that decides exactly which posts appear in your For You feed. It blends two types of content and ranks everything using a powerful AI model (a version of the same transformer architecture behind Grok).
How it works
• Two sources of posts:
• In-network (Thunder): Posts from accounts you already follow.
• Out-of-network (Phoenix Retrieval): Fresh posts the system discovers for you using smart similarity search across millions of posts.
• The AI brain (Phoenix): A Grok-based transformer model looks at your recent activity (what you liked, replied to, reposted, watched, etc.) and predicts how likely you are to engage with each post. It doesn’t use old-school manual rules — it learns directly from real user behavior.
• Multi-action prediction: Instead of just guessing “you’ll like this,” it predicts probabilities for many actions at once: like, reply, repost, click, watch video, expand photo, follow the author, mute, block, etc. Then it combines them into one final relevance score.
• Smart ranking steps:
- Gather candidates from both sources.
- Enrich them with extra context (author info, media type, engagement numbers, etc.).
- Run the AI model to score every post.
- Apply diversity (so you don’t see 10 posts from the same person in a row).
- Filter out spam, blocked/muted accounts, old content, etc.
- Pick the top posts for your feed.
What’s new in today’s open-source release
- Full end-to-end runnable pipeline — anyone can now test retrieval + ranking locally.
- Pre-trained mini Phoenix model included (ready to run out of the box).
- New Grox content-understanding service (spam detection, post categorization, policy checks).
- Built-in ads blending module with brand-safety controls.
- Much richer user & candidate data (topics, mutual follows, impression history, media detection, etc.).
- Completely removed almost all hand-engineered features — the transformer does the heavy lifting.
- It’s fully open source (Apache 2.0 license).
来源：DogeDesigner · x.com