AI writes daily research notes: how to select chunks

from AI writes research notes daily

AI writes daily research notes: how to select chunks 2023-08-11 There are a number of options for how to select chunks

• A: Random
• B: Recently updated chunks
• C: Chunk of recently created new pages
• D: The older the update date/time, the lower the probability
• E: Vector search on previous note I think A or D would be appropriate under the condition that it be done automatically once a day.
• In B and C, we see things from the same perspective as humans.
  • The need for “AI to see and develop what humans are seeing now.”
  • If that’s what you’re looking for, don’t wait for a once-a-day opportunity to make an explicit request to the AI.
• E is “AI reads and thinks about Scrapbox in its own interest, independent of humans”.
  • This is not a bad streak.
  • The problem is that naive implementation is where you read the same thing over and over again.
  • Need an implementation that remembers “what you’ve read in the past” and leaves it out.
  • On the other hand, “Do I really have to read something once and never again?” I mean, well, that’s not true.
    • Well, maybe the last 100 or so.
  • Added on 8/14/2023
    • You could decide based on the amount of notes you had last time.
    • 0 means random
    • If there are a lot of them, it’s unlikely that they are derived from a single existing chunk, so using the results of a vector search won’t result in “all the same ones”.
    • If the previous note has N tokens, there will be zero additional chunks to begin with.
    • I’d love to put one in at random.
  • Added on 2023-08-18

    • If there are a lot of them, it’s unlikely that they are derived from a single existing chunk, so using the results of a vector search won’t result in “all the same ones”.

    • It never happened.
      • I put a lot of different things in 🤖2023-08-18 02:22 but
      • 🤖2023-08-18 02:27 showed a commitment to a specific chunk
        • I could abstract the word here and get out of it.
        • 🤖2023-08-18 02:37
• A is the easiest to implement, so I experimented with this implementation.
  • As a result, Hatena Diary became a hit, which worked rather well for me!
  • It’s imported because it’s worth zero if it’s not a search hit.
  • but there are no Scrapbox-like links, so there are no Scrapbox-like suggestions.
  • This has resulted in a situation in which the product is contained but underutilized.
  • I don’t want to parse Hatena notation and change it to Scrapbox notation because it’s too much trouble.
  • AI will interpret this in natural language and make summary notes, a good “dig back”.

---

This page is auto-translated from /nishio/AIが毎日研究ノートを書く:チャンクの選択方法 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.