RFC (repost): Priority of algorithm usage across different (Twitter) features

[twister21]

tl;dr: fewer algorithms for home feed generation, more algorithms for content & author discovery across the core features Topics, Trends, Lists and Search

Introduction

Twitter, a text-focused service, shouldn't serve a “For You” feed by default that includes any embedding space (e.g., recommended communities, subscribable creators, tweets and videos “based on your likes”) or a large amount of social graph content, apart from likes and retweets if this hasn't been explicitly disabled, like other popular platforms.

Instead, it's more helpful to introduce new sorting and filter options for (now multiple) following-only home feeds (see Feed restructuring) and provide separate methods to discover follow-worthy accounts and notable tweets through better use of

• a user's topic following list, which should be fully viewable via /topics, and individually via /t (human-readable identifiers are more useful than generic (snowflake) topic ids for direct URL access), topic feeds should have Standard sorting & filter options
• Twitter's home-built "SimClusters" system, which should be better integrated with topics (e.g., show all major topics accounts tweet about, and potentially also allow for (account, topics) following pairs)

as well as

• highlighting platform-wide popular data records & extended recommendations
• improving search through semantic understanding.

It's very unlikely that an algorithm can outperform a human being in picking whose tweets best fit in one's home feed and content moderation happens most efficiently by following the right and unfollowing the wrong accounts + topics, but the current feed system doesn't allow that and needs improvement.

Content discovery

Presently, the Explore page only shows trends and tweets from (some of?) the topics a user follows. (Why do the recommendations take precedence over the topics you actually follow after the first five list entries?)

Instead, the following automatically curated content should be presented:

• popular data records (affiliates of organizations should be followable via auto-generated lists, separately for personal and other organizational accounts)
• grouped (temporary) trends and related (permanent) topics, which should be clickable
• recommended data records based on social graph ("followed by one/multiple account(s) you follow") and like history

Besides Standard sorting & filter options, these categories should additionally be filterable by data record's type (tweet|Space|topic|account).

This way, users can remove the recommended content from their own timeline and not miss out on highly relevant content without losing the ability to discover new things. In practice, this could be previewed on the home feed's "who to follow" sidebar card, which would then be renamed to "See more of Twitter" or "More activity".

Feed restructuring

For you feeds are generally not well suited for users who follow a wide thematic variety of accounts and topics. Improving content curation options and platform interoperability (shared following lists, superfeeds and semantic supersearch) is more reasonable than cloning YouTube, Substack/Medium, LinkedIn or Discord (https://github.com/orgs/community/discussions/60763).

• "For you" should be replaceable by a "Most interacted" tab, serving (unseen and/or Standard filter options (datetime range) respecting) top tweets from the following list
• Accordingly, "Following" would then be "Latest"
• Once all content of the home feed has been seen, the (enhanced, see above) explore feed should appear
• Also, the user's custom topics feed should be moved from the explore page to a separate topics page, and topic-social-proof's automated keyword labeling as well as trend presentation (more effective grouping, free country selection, preventing hashtag abuse) need improvement

And for the configuration of the reply section:

• providing Standard sorting & filter options is the minimum
• additionally, it should be possible to (sort)/filter/group by
  • content density score (factual statements, questions, statistics, external links, graphics)
  • spam score/marked as spam
  • reactions-only (emojis, punctuation marks, and potentially meme replies)
  • sentiment ([positive/neutral]|negative)
  • context/recurring keywords (replies that refer to the same thing)

This would greatly improve the experience, and the original "Community Notes" are replies, curated in a meaningful way.

Standard sorting & filter options

sortable by:

• datetime of creation (latest|oldest)
• popularity (top|least relevant): number of
  • tweet, Space: replies + likes + retweets (= engagement) or views (most seen), also engagement/views ratio (best)
  • topic, account: followers (+ account subscribers)

filterable by:

• datetime range (last 24h|72h|7d|30d|6m|1y|10y; custom timespan; all-time)
• tweet: isThreaded (true|false)
• account: personal|organizational, subscribable (true|false), affiliated (true|false)