Creating a reader AI 2023-11-13

from Create a reader AI Creating a reader AI 2023-11-13 from Diary 2023-11-13 Create a reader AI

pEnglish Publication of English Articles

• 1: First of all, /nishio is fine.
  • Small Start
  • There may be synergy with the automatic translation to /nishio-en by DeepL now
  • Small step up from "just machine translation of DeepL"
  • 1-1: GPT3.5 makes you "write" instead of "translate
    • Eventually we will probably stop translating the current DeepL translation and replace it
      • I wonder if it will be placed directly in /nishio-en.
      • Optimal placement is a consideration.
    • Are we really going to do it all?" "Wouldn't it be better to improve it through trial and error?"
      • Yes, we should not do everything from the beginning.
      • Small experiments should be repeated and improved.
      • I would suggest prompting them at first, but eventually fine-tuning them.
      • Readers don't assume they are native English speakers.
        • Target many people around the world who have learned English as a second language.
        • I want to output in a language that is easy to understand and machine translate.
    • This will be written by a machine and then reviewed and reworked by me.
      • So, in effect, it could be interpreted as "I wrote it with the assistance of a machine.
      • Then you can en.
  • 1-2: Building a readership - It's easier to write if you don't expect others to respond.
    • Japanese is I am the reader, so I can write without expecting others' reactions, and that's why it continues.
    • However, when the article is translated into English, the effect is that "if there is also a Japanese article, I will read the Japanese one" and readers will be lost.
    • As a result, we unconsciously seek "others' reactions to English articles".
      • This is not good.
    • You can create a being that is not a "stranger," that reads and responds back.
      • Reader AI
      • logs
        • I did this all together as a messy experiment, but I don't think we should split up the steps and not have them share memories.
          • 1: Japanese to English (human translation, translation AI, or writing AI)
          • 2: English to English (English reader AI, write observations in English)
          • 3: English to Japanese (Translation AI)
        • If it were separated like this, it would not be possible to translate "so was so wa" in the third step
          • I can look at it and realize, "I see, 'so-and-so' is an unknown concept.
          • How do we resolve this?
            • The 1 AI should be able to handle that neatly, making "sowasowa" the link and also the "sowasowa" correspondence.
            • This is not just about this case, the various "links" on Scrapbox are conceptual handles and should be properly connected
• 2: Talk about putting it on Github
  • Vague.
  • Github Pages? wiki? self-hosted in Vercel? Use (mem.nhiro.org)?
  • Certainly, some of the mechanical outputs do not feel appropriate to be placed in the Scrapbox
    • As long as it's saved in a form that can be checked later, it's a good thing.
  • Needs are not clear.
    • Do I read it, does someone else read it, or does an AI read it?

Next Action

• 1-2 1-2, 2, 3 to work on the command line.
• Move it with Github Actions and put it in a new private project

This page is auto-translated from /nishio/読者AIを作る2023-11-13 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.