Bluesky's Attie App Uses AI for Feed Control

🔑 Enhanced Key Takeaways

• •Attie utilizes a 'Semantic Filtering Layer' that sits between the AT Protocol firehose and the user interface, allowing Claude to perform real-time sentiment and topic-based pruning of posts before they render.
• •The transition of Jay Graber from Bluesky CEO to lead the Attie project is part of a broader 'decentralized ecosystem strategy' where Bluesky PBC encourages core team members to spin off specialized, interoperable clients.
• •Unlike standard Bluesky feeds, Attie stores user-defined 'preference vectors' locally on the device, ensuring that the AI-driven feed curation remains private and is not stored on the AT Protocol PDS (Personal Data Server).

🛠️ Technical Deep Dive

• •Integration via AT Protocol 'AppView' API, which allows Attie to ingest the global firehose stream.
• •Uses Anthropic's Claude 3.5 Sonnet API for low-latency inference on post metadata and content classification.
• •Implements a vector-based preference engine that maps user interaction history to latent space embeddings for personalized feed ranking.
• •Supports 'Client-Side Logic' execution, meaning the AI processing happens after data retrieval but before DOM injection, minimizing server-side compute costs.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline