Agents living in Scrapbox.

from Diary 2023-07-29 Agents living in Scrapbox.

• scbot

from [/omoikane/BabyAGI code reading](https://scrapbox.io/omoikane/BabyAGI code reading).

• Based on this, think about “[Agents living in Scrapbox.

  • If you’re an agent living in Scrapbox, it’s natural to receive JSON data from Scrapbox.
  • If there’s nothing explicit about the mission, it’s natural to read them in order of newest first.
    • Reading each page will be the initial task.
  • Read pages updated in the last 24 hours if they are activated once a day.
    • What is “reading?”
    • Summary?
    • Pick up keywords?
    • Ask questions of humans?
    • Write your thoughts?
    • People don’t react to everything they read either.
  • Give prompts for project objectives, etc.?
  • I should make a script to run this locally before I automate it…

• Blog text generation instead of translation

• How about an agent who reads this Scrapbox and blogs about his/her impressions, rather than translating or blog writing?

  • Sometimes he throws out questions.
  • If your impressions are good enough to write in Japanese, you can have them written in English.
    • Impressions by AI: neither 0 nor 100

“What is ‘reading’?” Not “What is ‘reading’ for?” It is.

• For example, “keep a rough idea of where things are so that you can remember them when you need them” is, in today’s terms, creating a vector index.

  • leveraged reading
  • The purpose of this is to reduce the cost of reading back by using a compressed representation
  • reading - Times when writing knowledge is not enough.

• (4.1) What is “reading”?

• Entertainment, finding information, assembling understanding

• Old format content with a string of letters

• The content of the book is not the only material to assemble.

• Gradation of “find” and “assemble

• What are Sources. and who are the beneficiaries?

• This is not clear, so the objective is not clear.

• When I thought about letting agents live in this Scrapbox, I implicitly assumed “source/nishio, beneficiarynishio”.

  • This is a very special situation.

• For example, if the “source of information/nishio and the beneficiary is someone who cannot read Japanese but can read English”, it would be a bot that creates translations and explanations in English.

• If “the source of information is a book and the beneficiary is nishio,” then it will be a bot that creates a summary of the book based on an understanding of my interests.

  • English videos and especially “I’m not good at audio input such as videos even in Japanese” and high cost since it is “English”.
  • I’d value a bot that would watch around English videos for me and generate links with summaries of interesting times.

• Can the beneficiary become an agent itself?

  • It can be done if the agent is curious and the value is to have that curiosity satisfied.
  • So, “what is curiosity?”
    • It’s not enough to get new information.
      • That would keep me from reading the random number table.
    • Want to get one step ahead of what you already have?
      • I can only understand one step ahead of myself.
      • “what you already have”. - Chicken and Eggs.
        • Without knowledge, it is impossible to judge the value of knowledge.
      • The only way to do this is to give them knowledge of the presets.
        • Knowledge you already have = your Scrapbox
        • So once you clone yourself.
      • New updates in other people’s Scrapboxes, books and YouTube videos as input.
      • Vector search in chunks
      • Equivalent to “find by association from past memory.”
      • Omit those that are already connected after they are found.
      • This gives us “associations not yet connected”.

Hmmm… then…

• It’s not enough to have a vector index with only the contents of /nishio to show to others.
  • Only useful as a first stepping stone.
• A vector index is needed to stock all the “past reads” by the agent
  • This would, of course, include other people’s works.
  • In the future, I’d like to give them books to read, but for the early stages of development, I’d like to crawl the publicly available Scrapbox.
  • Related: Finding links between multiple projects.

reading Diary of 2021-05-04

• That’s going to be beneficial for some of them, but it’s a bit of a chore.

• How about creating a “free room” (project for bots) first?

• Read /nishio and write about a page a day in /nishio and as many pages as you like in the “Room for Bots”, sort of

I even put a script in omoikane-embed that writes to Scrapbox and confirmed that it works. Something minimal (process complete! ) and tag it as “We have a system that interacts with Scrapbox”, then fork it and create a repository for Scrapbox over here and add some functionality…

Ah, the current omoikane-embed is thinking “it’s not a big number of pages, so it’s OK” and plugging all the pages into the vector index every time instead of updating them incrementally. If you’re going to put it in /nishio, you’re going to have to fix that…

---

This page is auto-translated from /nishio/Scrapboxに住んでるエージェント using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I’m very happy to spread my thought to non-Japanese readers.