2024/4/26 (Nishio)

  • Used the summary of Fractal Summarizer to introduce a book at an internal study meeting.
    • /Nishio/Plurality Study Meeting 1
    • Convenient to have a summary at the desired level for the material.
    • Identified issues:
      • image
      • Section breaks in the middle of a chunk.
      • The content in the first and second halves differs significantly, causing the latter part to be overlooked.
      • However, the latter part serves the role of guiding what the discussion will be about, being the “first paragraph of a new section,” so it was not desirable to discard it.
    • Reflections:
      • Noticing this issue itself is a benefit of not having a generic summary.
      • It might be worth exploring an approach to detect section breaks.
        • Proposed methods:
          • A rough method could involve “feeding it into LLM to detect breaks.”
          • However, considering a faster and more user-friendly mechanism.
      • As someone who wants to provide easily readable summaries of books, observing such issues may lead to a desire to control the breaks and regenerate them (though the interface design would be complex).
        • Instead of a unidirectional mechanism of “submitting once and reading the generated content,” it seems feasible to create a bidirectional mechanism of “reading the generated content while providing feedback as needed.”

2024/4/24

(Nishio) I was thinking about something else and suddenly thought, if the input at the end is “a variety of people’s opinions,” it could be considered a form of broad listening.

  • Currently chunking based on the order of input, so inputting a variety of opinions with a mixed order might result in a confusing fractal summary.

    • I thought of embedding vectors and clustering with varying k in k-means (Nishio).
      • I see (blu3mo).
  • Chunking based on meaning rather than order could generate useful summaries to understand a variety of opinions.

  • Leap of thought:

    • Can we mimic the structure of a Neural Network encoder?
      • Doing weight summaries with LLM instead of linear combinations in each perceptron.
    • If we can weight the edges appropriately, we can achieve something close to “meaning-based chunking.”
      • Extracting features gradually from input to output.
      • Visualizing this process with a Fractal Reader-like UI could ensure traceability.
    • The challenge is how to evaluate the state and weight the edges.
    • Since backpropagation is not possible, the extent to which mimicking NN is meaningful is unclear.
      • Well, LLM itself is a type of NN, so maybe backpropagation is possible?
    • Simulation of Research Activities with LLM x Tree Search is doing something similar.
    • Addition: Dynamic LLM-Agent Network: An LLM-agent Collaboration Framework with Agent Team Optimization was working on something similar.
  • It seems plausible to hypothesize that with a “low-cost LLM,” generating high-quality summaries could be cheaper than with a “high-cost LLM.”

    • With traceability and the ability to use intermediate layer data for other purposes.

2024/4/23

(Nishio) The concept of fractal summaries was well received internally.

At that time, an impromptu quote came up: “If AI were to create meeting minutes, you would definitely say, ‘Did we really say that?’ So, instead of a ‘black box,’ it’s a ‘summary with traceability’ that allows for in-depth exploration as needed!”


Considerations based on the above content:

  • Want to think about how to divide lines (the smallest unit of segmentation).

Fractal Summary

(Claude)

  • In the session after the lunch break, a mechanism for fractal summarization developed by blu3mo was introduced and tested by the participants.
    • Fractal summarization is a tool that gradually summarizes the main text, allowing readers to delve deeper into the parts they find interesting.
    • Participants provided feedback on the quality and clarity of the summaries.
    • The need for additional information to complement the content was also discussed.

I would like to know your thoughts on using the Fractal Summary Table of Contents page, or reasons why you didn’t feel motivated to use it! (blu3mo)

  • I’m interested in knowing what you found valuable or not valuable. (blu3mo)

Thoughts on the Fractal Summary Table of Contents (healthy-sato)

  • I wanted to read a Japanese translation without a summary on a screen where summaries are lined up (imagining a column between Summary 4 and the original text).
  • If there was some way to navigate to the corresponding Scrapbox page or make it easier to write feedback after reading, it would be more user-friendly…
    • I resolved this by splitting the screen.
    • image
  • Mentioning the positioning of each chapter at Summary Level 1-2 might enhance readability.
    • image
    • I see (blu3mo)(blu3mo)
      • The structure of the Summary around Level 1-2 seems off, so I’d like to improve that (blu3mo)

I might spend around 1 Pomodoro focusing on reading the Fractal Summary… (nishio)

  • ⿻ emphasizes the intersection of individual and group identities and suggests that digital technologies should aim to utilize diversity for development.

    • I was intrigued by the “intersection of individual and group identities” and after delving into it, I looked at the summary below and understood.
    • I wrote about this on the page for ”individual and group identities“.
      • I thought it was nice to have phrases like this that can be kept as a single concept in Scrapbox.
    • I wish there was an explanatory page and icon for ”Fractal Summary“.
  • Oliver Wendell Holmes states that the greatest value of civilization is to complicate the means of living and enable a richer life, while Carlo Rovelli explains that reality is not a collection of things but a network of processes, and ⿻‘s approach, which emphasizes complexity, emergence, and multi-layered organizations rooted in science, offers a new perspective on social systems.

    • image
      • There’s too much variation in the content, which is confusing.
      • The summary here doesn’t feel very good experientially.
      • Perhaps the problem lies in the original text having too many leaps in the story, making it more of a list of information…
  • John Dewey states that the new public must destroy old political forms, and pioneers like Henry George, Georg Simmel, and Norbert Wiener presented a vision of a ”connected society” that utilizes the potential for cooperation beyond diversity.

    • image
    • Hmm, the leaps are…
    • Maybe in terms of implementation, Levels 3 and 4 are just operating in parallel with different input ranges, but as users, we expect some continuity between 3 and 4.
      • I see (blu3mo)(blu3mo)
      • So, should 3 take 4 as input, or should 4 take 3 as input, or maybe iterate a few times to align them…- J.C.R. Licklider argues that decisions regarding the advancement and utilization of computer technology should not only serve the public interest but also provide a means for the public to participate in the decision-making process that shapes the future.
    • Suddenly thought, “It would be interesting to gather these kinds of arguments and create a Polis.”
      • So, make it possible to jump from each question to each chapter.
  • image

    • There seem to be a lot of discussions about Luna (a fictional character), right?
      • image
        • Oh, I see.
        • I might be mistaking line breaks for paragraphs.
        • Maybe include a process to combine them up to a certain length if they are too short?
        • Or I have data converted from PDF into content per page that I can use?