prev Weekly journal 2023-08-21~2023-08-28

2023-08-28 10:00

  • Although the period is a little short and it is a Monday, I decided that it is appropriate to make a break here since it is the time to return to normal operations after three consecutive camps and a rest after returning from those camps.

What we did last week

  • I created a version of the reading notes while reading a locally placed file.

    • Current reading notes: Gist
      • Version in which the unit of information input is rows instead of chunks
      • Of the 956 lines in total, the most recent reads lines 367 to 405.
    • At this point, I have not yet done a vector search against my Scrapbox.
      • Questions about workflow have arisen.
        • Shouldnā€™t the book content be included in the vector index first?
        • Iā€™ve decided to implement the reading order from the head by analogy with the way humans read, but is that really appropriate?
          • Iā€™m beginning to think itā€™s not.
          • Reading from the beginning is not a good way to read, even for humans.
          • I think it would be better to do an analogy of Whole Mind System.
    • Pending once I try to do another experiment that I can do with my phone as I get home from the camp.
  • AI notes on Scrapbox can be progressed in multiple pages in parallel (multi-head).

    • Tried ā€œ22 parallel, 1 hour updateā€ as an experiment on the train home from the camp.
    • This did not produce the expected results due to a bug.
    • Moving during waking hours doesnā€™t sound good.
      • I get a release notification from Github at 16:07, but then GPT4 runs 20 after this, huh? So the overwrite will run in about 30 minutes? Thatā€™s hard to writeā€¦ I should still move when Iā€™m sleeping --- 64e5b04eaff09e00004978a3

    • Regarding frequency and quantity
      • Especially in the early stages, it was overcapacity for me as well, due to the fact that people did not know what to do and mass-produced pages with miscellaneous keywords to observe what would happen, and by the fact that it was 22 pages in parallel!
        • It could have something to do with fatigue from the camp.
      • Right now, as a quick implementation, Iā€™m randomly selecting 10 items from the target page.
      • In earnest, I think Spaced Repetition update is a good idea.

There is one more concept that does not have a clear name, but is important.

  • AI named: page memory.
    • page memory is difficult to see the value in the short term, but clearly beneficial in the long term

8/21 multi-head - page memory 8/23 Ignore used pages from all history 8/25 Update page limit 8/26 Pinning effect on the topic

This page is auto-translated from /nishio/週čؘ2023-08-28~2023-09-02 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. Iā€™m very happy to spread my thought to non-Japanese readers.