Fractal Summarizer Feedback Summary

2024/4/26 (Nishio)

• Used the summary of Fractal Summarizer when introducing a book at an in-house study session.
  • /nishio/Plurality Study Session 1
  • Convenient to have a summary at the desired granularity for the material.
  • Identified issues:
    • 
    • There is a section break in the middle of a chunk.
    • The content in the first half and second half differs too much, leading to the latter being ignored.
    • However, the latter part serves as a navigation point to indicate what will be discussed next, so it was not desirable to discard it.
  • Reflection:
    • Noticing this issue is a benefit of not having a one-size-fits-all summary.
    • It might be worth considering an approach to detect section breaks.
      • Proposed actions (blu3mo):
        • One rough method could be to “feed it to LLM to detect breaks” (blu3mo).
        • However, I am thinking of a faster and more user-controllable mechanism (blu3mo).
    • As someone who wants to provide easily readable summaries of books, seeing such occurrences makes me want to control the breaks and regenerate the summary (although the interface design would be complex).
      • Rather than 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” (blu3mo).

2024/4/24

(Nishio) I was doing something else and suddenly thought, if the input at the end is “diverse opinions of people,” it could be considered as a form of broad listening.

• Currently chunking based on the order of input, so if diverse opinions with a random order are input, it might result in a confusing fractal summary (blu3mo).

  • I envisioned using embedded vectors and clustering with varying k in k-means as a solution (Nishio).
    • I see (blu3mo).

• Chunking based on meaning rather than order could lead to generating useful summaries to understand diverse opinions.

• A leap of thought from there:

  • Is it possible to mimic the structure of a Neural Network’s encoder?
    • Imagining performing weighted summarization by LLM instead of linear combinations in each perceptron.
    • 
  • If edges can be appropriately weighted, it might be possible to achieve “meaning-based chunking.”
    • The extraction of features gradually occurs from input to output.
    • Visualizing this process with a UI like Fractal Reader could ensure traceability.
  • The challenge lies in 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 feasible?
  • What is being done in Simulation of Research Activities with LLM x Tree Search is somewhat similar (blu3mo).

• It seems plausible to hypothesize that with a “low-cost LLM,” one could generate high-quality summaries more affordably than with a “high-cost LLM.”

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

2024/4/23

(Nishio) People found the discussion on fractal summarization interesting within the company.

A spontaneous remark that came up: “If AI were to create meeting minutes, there would definitely be moments where you’d wonder, ‘Did we really say that?’ So, instead of a ‘black box,’ it’s more like a ‘summary with traceability’ that allows for deeper exploration as needed.”

---

Considerations based on the above content:

• Want to rethink how lines (the smallest unit of division) are separated.
  • Simply dividing by line breaks can sometimes result in undesirable splits.### Fractal Summary (Claude)
• During the session after the lunch break, the mechanism of the Fractal Summary developed by blu3mo was introduced and tested by the participants.
  • The Fractal Summary is a tool that gradually summarizes the 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 the impressions of using the page that connects all the contents of the Fractal Summary, or the reasons why motivation did not arise to use it! (blu3mo)

• I am curious about what you found valuable or not valuable (blu3mo)

Impressions of the page that connects all the contents of the Fractal Summary (healthy-sato)

• I wanted to read a Japanese translation without a summary on a screen filled with summaries (imagining it as a column between Summary 4 and the original text).
• In an ideal scenario, if there was a way to navigate to the corresponding Scrapbox page, it would be easier to write down impressions…
  • I resolved it by splitting the screen
  • 
• Mentioning the positioning of each chapter around Summary Level 1-2 might enhance readability?
  • 
  • I see (blu3mo)(blu3mo)
    • The structure of the summaries around Level 1-2 seems off, so I want to improve it (blu3mo)

I might concentrate for about 1 Pomodoro and try reading the Fractal Summary… (nishio)

• It emphasizes the intersection of individual and collective identities and suggests that digital technology development should aim to utilize diversity.

  • I was intrigued by the ”intersection of individual and collective identities” so I dug deeper and found the summary below, which made sense to me.
  • I wrote about this on the ”intersection of individual and collective identities” page
    • I thought it was good to have phrases like this that can be kept as a single concept in Scrapbox.
  • I wish there was an explanatory page and an icon for the ”Fractal Summary“.

• Oliver Wendell Holmes stated that the greatest value of civilization is to complicate the means of living and enable a richer life, while Carlo Rovelli explained that reality is not a collection of things but a network of processes, and the approach of ⿻, which emphasizes complexity, emergence, and multilayered organization, offers a new perspective on social systems.

  • 
    • The differences in content were confusing.
    • The experience of the summary here is not very good.
    • Perhaps the original text has too many leaps in the story, making it more of a list of information…

• John Dewey stated that the new public must destroy the 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.

  • 
  • Hmm, there are leaps…
  • Maybe in terms of implementation, Levels 3 and 4 are just operating in parallel with different input ranges, but as a user, you would expect some continuity between 3 and 4.
    • I see (blu3mo)(blu3mo)
    • So either 3 takes 4 as input, or 4 takes 3 as input, or maybe repeating some adjustments a few times…- J.C.R. Licklider argues that decisions regarding the development and use 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 to create Polis.”
    • So, make it possible to jump from each question to each chapter.

• 

  • There are a lot of discussions about Luna (a fictional character), huh?
    • 
      • Oh, I see.
      • I mistakenly interpreted the excessive line breaks as paragraphs.
      • Maybe include a process to combine them up to a certain length if they are too short?
      • Or should I use the data converted from the PDF into content by page that I have on hand?