app.hyper-limit.bookmark

{
  "url": "https://l4.pm/wiki/Personal%20Wiki/bluesky/the%20Donut%20algorithm.html",
  "$type": "app.hyper-limit.bookmark",
  "title": "the Donut algorithm",
  "pinned": false,
  "summary": "The article titled \"the Donut algorithm\" by the author Luna discusses developments in a project called Nagare, which is a general-purpose feed side project for Bluesky. The primary focus is on addressing the \"low following\" problem within user feeds, particularly how users with varying follower counts experience different issues. \n\nKey points include:\n\n1. **User Feed Challenges**: The article highlights the problems faced by users with low follower counts (10-100) who experience stale feeds due to insufficient activity. In contrast, users with high follower counts (1000+) face issues related to overwhelming activity, making it difficult to manage and store data.\n\n2. **Solution for High Following**: A solution was implemented for users with high follow counts by creating a ranking function that sorts the accounts they follow based on usefulness. This involved scoring posts based on the authors, likers, and the content of the posts, allowing users to receive a curated feed of their top followed accounts.\n\n3. **Strategies for Low Following**: The author discusses various ideas to tackle the low following problem, such as virtual follows, miniclusters, subcommunity detection, and spiderclusters. Each of these concepts aimed to enhance feed personalization but faced challenges in implementation.\n\n4. **Development of the Donut Algorithm**: The most progress was made with the spidercluster approach. After collaboration with another tester named Kara, the concept of the Donut Algorithm was developed. This algorithm creates a feed by sampling interactions from similar users, creating a “donut” structure where the outer ring consists of users who like similar content. \n\n5. **Algorithm Implementation**: The Donut Algorithm involves scraping a seed user’s recent interactions and identifying co-interactors to fetch their interactions as well. From this data, the algorithm computes candidate authors and likers, helping to curate a personalized feed. \n\n6. **Real-time Adaptation and Storage Challenges**: A significant challenge mentioned is the need for the algorithm to be real-time aware, adapting to changing user preferences without overwhelming the storage capabilities of the VPS hosting the project. The author proposes using reference counting to efficiently manage and evict outdated data from the system.\n\nIn conclusion, the article describes an innovative approach to enhancing user social feeds in the Bluesky platform through the Donut Algorithm, which seeks to balance engagement for users with both low and high follower counts. The ongoing challenges of real-time data management and storage limitations are recognized as critical aspects of further development.",
  "archived": false,
  "imageUrl": "https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:ghmhveudel6es5chzycsi2hi/bafkreidvf26so7pxmpfavp5bbeqb63ohukqpaxivuolppxsjvaosftwr2q@jpeg",
  "createdAt": "2025-09-02T05:01:14.685Z",
  "encrypted": false,
  "updatedAt": "2025-09-02T05:01:28.576Z",
  "description": "![running1.png]  #author_luna #bluesky #atproto #recommendation_algorithms   %at=2025-09-01T23:05:27.151Z  this is an article related to nagare, my general-purpose bluesky feed side-project. previousl..."
}

did:plc:5j2yklrr4pozy7yhrdq5xfn7 | at://did:plc:5j2yklrr4pozy7yhrdq5xfn7/app.hyper-limit.bookmark/3lxtdkg5exo2u