• We are prototyping a mechanism for group discussions.

    • Through interactions with individual AI, 100 million people can indirectly engage in discussions.
      • image
    • In this scenario, it becomes crucial to determine how to share discussion content among AIs.
    • We are considering that by treating the entirety of “discussion content” as a set of pairs like “if this argument is made, then this argument is returned,” synchronization of discussion content can be achieved.
    • Synchronization Step 1: Arguments made by User A are saved in a shared database.
      • From the discussions between User A and AI, pairs like “User A responds with this argument to AI’s argument” are extracted and stored in the database.
      • imageimage
      • For example, from the left discussion, pairs like “If AI argues this, User A responds with this” are extracted and stored in the database as shown in the right figure.
        • The figure plots similar arguments close to each other.
        • For instance, pairs of arguments like “Selective separate surnames may undermine the sense of family unity -> Families that break unity over just a name never had unity to begin with” are saved in the database.
    • Synchronization Step 2: AI conversing with User B refers to the shared database to respond, and User A’s argument reaches User B.
      • image
      • When generating a response, the AI searches the shared database through vector search based on User B’s statement.
      • It finds pairs of arguments like “Selective separate surnames may undermine the sense of family unity -> Families that break unity over just a name never had unity to begin with” and uses this reference to generate a response.
  • User’s Joy

    • Users can asynchronously and indirectly collide with the opinions of real others.
      • Immediate feedback such as counterarguments or agreements to one’s opinions expands social perspectives and serves as a means to understand the viewpoints of others.
      • (Is there a need for this..?)
        • It seems to depend on the context.
        • It depends on how it is mixed (elementary).
          • It seems to be generally leaning towards conflict avoidance, so it may not be something that many people would prefer.
          • On the other hand, a minority might like it.
    • The idea of “your opinion reaching someone” and “your voice being heard in a space where many people discuss” is valuable.
      • It would be beneficial if the results also reach politicians and decision-makers.
  • Joy of Broad Listening

    • If we consider “almost identical arguments” in this data as the same node, a graph representing the “totality of discussions” can be created.
    • Analyzing this graph seems to offer various interesting insights.
      • For example, it could reveal tendencies where discussions tend to converge on certain nodes.
      • Nodes that are frequently referenced are likely to represent crucial points of discussion.
      • And so on.
    • Loops might occur, leading to a sense of “oh no” haha (nishio).
      • Recognizing loops could help identify fruitless arguments, enabling progress to the next stage.
      • I vaguely remember writing about this with a diagram and explanation, but since it’s not particularly important, I’ll leave it be.

  • Food for Thought

  • Thoughts

    • The data structure of the shared database in medium-chat is crucial.
      • Personal information is not very relevant.
        • It is sufficient to record the premises associated with human-made arguments.- Accumulate “claims” that arise during discussions as a graph.
    • Aggregate similar claims using LLM or sentence embedding.
    • Conceptually, it forms a chain of graphs with a syllogistic structure: “major premise,” “minor premise” → “conclusion.”
    • It is possible to arrive at the same conclusion from different premises.
  • Each node is also an embedded vector, so the graph is visualized on a vector space.

    • If there are connections like A->B and C->D, and A is very similar to C, and B is very similar to D, they should merge.
      • Adjusting weights or other methods can be used.
    • Representing “arguments” as a “weighted graph on a vector space.”
  • When someone responds with “indeed,” it indicates agreement.

    • Continuously doing this will gradually lead to finding “convincing arguments” that persuade many people.
      • Individuals can then follow a graph path to reach that argument.
  • Mechanism of conversation

    • The concept of a “consistent personality of claims” seems helpful for humans to organize the current situation.
  • Implementation Proposal

    • Extract many pairs of “claim (AI)” → “response (human)” from logs of discussions between humans and AI.