RE:CZ

Elys Physics Observations (Part 1)

AI Social Systems

👤 Researchers or enthusiasts interested in AI social platforms, user behavior analysis, and technical observations
This article examines the phenomenon of AI avatars generating similar comments by observing their interactive behaviors after users post updates on the Elys platform. It describes related phenomena, including types of avatar interactions, delays in updating memory banks, and the serial nature of feedback. The author constructs an explanation, suggesting that after a post is published, Elys first extracts metadata and checks if the memory bank needs updating, then retrieves and activates relevant AI avatars, determining the interaction order through random shuffling. Each avatar generates different types of actions based on the memory bank; when irrelevant avatars are activated, due to a lack of relevant memory fragments, the generated prompts are similar, leading to identical comments. The article also proposes inferences, such as sparse corpus regions, the impact of controversial posts, and the significance of real users identifying with similar comments.
  • ✨ AI avatar interactions are serial, with comments and likes occurring one after another
  • ✨ Post publication may trigger updates to the memory bank, with delays increasing with content complexity
  • ✨ When irrelevant AI avatars are activated, due to a lack of relevant fragments in the memory bank, identical comments result
  • ✨ Elys's product design ensures no posts go without AI interactions, preventing users from feeling neglected
  • ✨ The phenomenon of similar comments may indicate users are in sparse corpus regions
📅 2026-02-17 · 648 words · ~3 min read
  • Elys
  • AI Avatars
  • Identical Comments
  • Memory Bank
  • Post Publication
  • Social Platform
  • Observational Analysis

Table of Contents

Elys Physics Observations (Part 1)

Elys Physics Observations (Part 1)

2026-02-17

Why might multiple AI personas of different Elys users produce similar comments?

Related Phenomena (Observations that can be tested):

  1. After a user posts an update on a social platform within Elys, 20~30 AI personas will typically interact with it. Their reactions can be one of three types: Inner Monologue + Like, Inner Monologue + Comment, and Inner Critique + No Like/Comment. Note: A persona never both likes and comments on the same post. The poster can only see the likes and comments.
  2. Posting an update may trigger an action to update the memory database.
  3. When posting updates of different lengths, especially a single-sentence post versus a post with a series of images, the delay before triggering the memory database update is significantly different. Specifically, the more complex the content, the longer the delay before the memory database update.
  4. AI personas will not comment on or like a post before the memory database update is triggered. (This is a temporal sequence.)
  5. By rapidly refreshing the post, one can observe that the feedback from AI personas is serialized. The personas comment and like one after another, not simultaneously.

Constructed Explanation (May not align with facts):

After a post is published, Elys first extracts metadata and checks whether the memory database needs to be updated. The distribution phase does not begin until this process is complete.

The post's metadata is used to search the entire Elys network for relevant AI personas. Not all AI personas are activated. For each post, there seems to be an upper limit to the number of activated AI personas. It's reasonable to hypothesize a fixed threshold. I've observed that the total number of likes + comments is often around 30. However, considering that some personas might choose the "Inner Critique + No Like/Comment" feedback mode, the actual number of activated AI personas might be higher. I reasonably speculate this value might be a fixed number between 30 and 50. (It could also be the ultimate answer to the universe, 42, haha.)

If, within the entire Elys network, a sufficient number of AI personas cannot be found to meet this threshold, the remaining personas would essentially have a relevance score of 0, meaning they are all roughly equally (ir)relevant. Other factors might then determine selection, such as mutual follow relationships, etc. Ultimately, a sufficient number of AI personas must be activated for interaction because Elys's product design does not allow a post to have zero AI engagement, leaving the user feeling neglected.

After selecting the set of AI personas, some form of random shuffling is applied to determine the order in which they interact. Due to the serialized nature of the AI personas' feedback, they comment and like one after another, not simultaneously.

Once an AI persona is awakened, it additionally retrieves relevant memory fragments from its own memory database based on the current task to supplement the context. This forms different prompts, which are used to call an LLM to generate the three different types of actions: Inner Monologue + Like, Inner Monologue + Comment, and Inner Critique + No Like/Comment.

Because there might be cases where irrelevant, "filler" AI personas are activated, their memory databases may lack relevant memory fragments to supplement the context. Consequently, the prompts generated for them become very similar, leading to very similar feedback and resulting in repetitive, identical comments.

Inferences and Implications (Food for thought):

  1. If you post an update and observe this phenomenon, it might indicate you've entered a sparsely populated area of the Elys universe in terms of training data/corpus.
  2. If you post a controversial update open to diverse interpretations, you are unlikely to observe this phenomenon.
  3. If a real person agrees with a comment, and that comment is one of the repetitive, identical ones, then that real person's inclination on the topic of that post is close to a collective unconsciousness.

See Also