Neri Neri's value is slowly decreasing.

Neri Neri's value is slowly decreasing.

I've been feeling it for a while now.

Why is that?

https://gyazo.com/98ff798e89104a87ec36e7effbf82f66

1: I put a certain query Q1

2: omni says "A or B or C is related to

I said, "I see! I hadn't thought of that idea! That's interesting!" And I'm like.

3: A lot of thoughts develop with that omni stimulus.

A new D is being created.

4: I put query Q2 in relation to that D

5: omni "relates" to my recent C1, B1, and D

I'd be like, "I know..."

People are too slow on their feet.

The speed of development of human thought is slow compared to the power of vector search leaps.

The parable of [nomadism

When I first created omni-private, I tried a little and immediately removed my own origin fragments from the search.

Why it happens is because your writing looks like your writing.

Many people interpret vector search as a search by meaning, but in reality, it is a vector that involves not only meaning but also style and other factors.

Then I started including self-derived fragments for experimentation.

New findings made in [Experiments to find relationships

When I search with this, I get almost no hits on my text.

Fragments hit from my Scrapbox are also "scraps of other people's statements."

The ratio of self and other fragments in the search results may give you an idea of the amount of grass.

However, the title has been changed and those pages have been hit in searches.

Maybe so. I haven't checked properly yet.

That's the story that bothered me when I made Scrapbox a search index in the first place.

The specification is to leave it as "useful because it provides a clue to find the renamed one, even if it's a hit."

It's not a good idea to have all the slightly different output generated by the AI as a search target.

Because the current search is structured as "N cases from the one with the highest vector similarity", when such content is hit, the search results are filled up.

As a result, similar content is generated again.

This is one of those hammy modes caused by a flaw in the specs.

Older implementations would collapse update time data on import to Scrapbox, resulting in full telomere updates.

Similarly, the updater information for each line has been filled in.

After all, was it necessary here to let the AI use users who are not me in the flesh, and to handle them differently?

The "when a ningen comments in AI-generated content, it's useful" story should also have done a better job of detecting and cutting it out

---