Intellectual Productivity Study Group by LLM

Cybozu Labs Study Session 2023-10-06

It was two months of porting this to /nishio and improving it to the nth degree.

Summarize what you experienced and felt during the two months

Up to now.

https://gyazo.com/e86f3facdda9fb550d57325ce06d50a0

thumbnail (i.e. miniature image)

1: Current Topics: Azure Cognitive Search

2: Vector Search

1: Current Topics: Azure Cognitive Search

Talk about combining both and reranking (≈reordering after looking at the search results) is far better than the common "keyword only" search, or vector search.

It's been known for a while that "it's better to combine and rerank."

https://gyazo.com/763fb9ae8f7e017772460b9ef471e515

Very small overlap in search results for both DPR and BM25

→ Good of both.

Vector search is available in:.

Strong semantic similarity

Missing low-frequency words / Significant performance degradation outside the distribution

Weak on low-frequency words (=weak on proper nouns, technical terms, and product names), so need to combine normal search

A text generation system that reads 100 search results

Vector search is good for ambiguous searches in natural language conversations because of the flurry of hits, but it has a high cognitive load for users because "the name of the company you visited in sales" or "the product model number" also hits "different things that look very similar".

I was thinking that the groupware needs to improve this problem if it is going to use vector search.

7/18 Public Preview begins

I'm thinking of trying it soon.

What is "search?"

What is the value you are creating for your customers?

The "search UI" we provide to users now is not the best way to deliver that customer value.

Relanker ≈ "a midget who looks at 200 search results and sorts them in good order."

Fusion in Decoder ≈ "a little guy who reads 100 search results and writes sentences"

The era in which these things are becoming available

Especially in the reranker section, a style that takes into account things other than "text explicitly entered by the user," such as operations on the groupware immediately before, may lead to differentiation.

2: Vector Search

Vector search for this Scrapbox project, published 2023-06-05

Case 1

It's all about "social security spending."

I recall a past lecture in which he said, "I've talked about the high cost of Social Security."

The focus was on the ratio of investment to science and technology.

I'd try to find it, but I can't find it by searching for "social security payments".

https://gyazo.com/34b6beae7ed4d7d5d95fd0d4d08aaff6

The expression was "38 trillion for medical care, 24 trillion for welfare and others, and 120 trillion in total."

There's also "Science and Technology" a little further down the road.

I made an excerpt page with "Social Security Expenses" in the title.

Case 2

I'm trying to come up with a story about a hole in the wall, and I'm searching for wall or hole, but I can't find it.

Vector search for "hole in the wall, you have to get close to see it."

Second candidate found.

https://gyazo.com/9c55e099867137986e3be8ead02a9fdb

It's a match between "you can't see it unless you get closer" and "you could see it if you got closer."

The nuance of not being able to see the hole in the wall unless you get close to the wall.

The hole in the wall was pictured, but not texted.

I'll add a note.

Case 3

Vector search on "unfeasible ideas seem original".

The one I was looking for was the fifth hit.

https://gyazo.com/395e22955d484b244995604e4b1664fb

A layman's idea may appear original in an oral presentation, but it is not feasible.

Didn't hit on the swing between "original" and "originality," between "unfeasible" and "unfeasible."

The pages of this experiment are episodic, which would make them memorable and easily recalled.

I created a new page with "[Unfeasible ideas seem original.

Case 4

Vector search on "kintone talks around data"

No.1:.

I think the ideal situation would be to have the same collaborative editing capability for database schema and customization JavaScript.

No. 4: The

Reforming Municipal Operations with Digital Solutions for Public Utilities

I see - I was wondering how to explain kintone, but you explain it as "email and Excel integration".

I like the term "email and Excel integration."

Case 5

"Bias to measure only what is easy to measure."

While cost is easy to measure numerically, quality is harder to measure and therefore easier to underestimate.

Looking Back

It's now possible to search by "this is what I meant" instead of "I must have written this 'string'."

Nishio himself is adapting to the vector search system.

Early cases search by fragments, as in "social security funding scientific research" and "hole in the wall, you can't see it unless you get up close."

As I get used to it, I'm searching for "short sentences that express meaning" like "unfeasible ideas seem original," "kintone is a conversation around data," and "bias to measure only what is easy to measure."

The other day, when I asked a user to use this vector search while sharing the screen on Zoom, the user typed, "Tell me about your experience of cooking for yourself," and I thought, "Oh, you don't understand how to use this at all.

Maybe most people in the world don't know the difference.

Similar input just because ChatGPT does.

The user then typed in "description about cooking" and found my terrible cooking story, which is a good thing, right?

Not much vector search goodness.

Because the "meaning" in the query is in effect one word of "cooking", so it is no different from a keyword search.

Automatically export JSON from Scrapbox

Vector embedding by inscribing on 500 tokens

Upload to [Qdrant

Began operation on 6/9, now running daily.

I started writing reports in Scrapbox on 7/29.

8/9 Organized code to make it easier to put into other projects.

People actually showed up to try it.

It's been two months since we've made a lot of improvements to this omni.

https://gyazo.com/9d448f09d2d3c47287128ca2efef61dd

AI writes a note in Scrapbox, AI reads the note the next day, and writes another note.

Repeat intellectual production activities in the same way as humans in the collaborative editing space called Scrapbox

Implemented on 8/12,

AI's once-a-day note-with-comments communication style has the advantage of "no human triggers to pull".

Even when humans are too busy to write comments, AI goes ahead on its own.

Even if you write some comments, there is no "send" action.

I may come up with something after that and add more.

No need for human decision-making on when is "the end" or "completion".

Not sure if it would be better without the "send" action.

Personally, I think it's better to have it, and can be accomplished by running scripts in the development environment when you're in front of a PC.

I'm on the train to get to the camp site, and I feel like triggering from my phone in these cases.

AI created a style of adding to the same page instead of creating a new page.

After that, the style of creating new pages was DISABLED (8/31 main branch stopped).

This allowed "multiple topics" to develop in parallel

Write a note that a human thinks "I'd like to develop this topic" as a seed and treat it as an AI note (with 🤖🔁 in the title in the implementation at the time).

Talk if you can afford it (I can't afford it)

History

This is useful

What kind of experience does it bring?

Make a link to "I think I've written something like this before."

Sometimes assisted here by Scrapbox's link suggestions

For those not familiar with Scrapbox: Scrapbox runs an ambiguous search when creating a link and suggests a link destination.

If not, you had to use search and other methods to find them.

We can leave that to AI.

Further Development

When there is a good "long phrase" in the AI's generation or in the past descriptions that the AI has uncovered, make it a link, like a marker.

Naturally, it would be a red link.

This is long enough that it is expected that "it will not be a natural connection"

So we will "red-link extension" this.

Specific example 1

I didn't know what "the importance of exchange in the problem-solving process" was, so I red-linked it and stretched it.

When OMNI said, "Problem solving bridges the gap between the ideal and the current situation, and the exchange of information in the process is important," I understood, "I see, you mean exchange of information is exchange."

I wrote about this realization in [Information exchange is exchange

(The meaning of this realization is probably not conveyed, but that is a case study for later, so I won't explain it here.)

After a while, "[Is the exchange style of knowledge exchange A?" The question arose

Search with explanation of why it was shortlisted.

Instead of a human directly reading the search results, the AI reads them first and writes a description.

For a long time, I used "Nishio's Vector Search" and thought, "Isn't this just fine?" but now that "Extension of red link by AI" is available, I think "this one is better".

I find it different and more useful than a simple vector search.

It could be called "decoupling search and action."

Now "searches" don't involve updating, so if I don't update my Scrapbox with the search results, the search never happened!

Implicitly required that "humans read the results and act on them before they forget they searched."

Search and action were coupled.

AI extends the red link" is a human writes the "intention of search" and the AI does the "search, reads it, and writes a beat of explanation", so the human action after the search is optional.

Pioneer Mode is a development of "Extending the Red Link with AI

If you put a link on the "✍️🤖" page, the AI will check it periodically and generate it automatically.

I can now use it from my phone.

When two people get along well, have a lively discussion, and their comments influence each other, it is difficult to discern which part of the resulting material originates from one person.

The same thing happens with humans and AI. The idea of trying to identify this, or to make it identifiable, can be detrimental to designing a UI.

And I was writing my own.

Especially on Scrapbox, the action of "read a page, and if the thought inspires you, write on that page" is afforded.

Then, when a human sees an "AI-generated page" and makes a post, it seems strange that the human's post is not picked up because it was originally an AI-generated page, so the distinction between AI-generated and non-AI-generated pages is abandoned.

For example, it is possible to know who updated each line, so if you create a dedicated account for the AI, it is possible to "ignore AI-generated content and use only human-written content".

As to why I haven't done it, I don't feel the need to do it...

I feel like I might discover something new if I try it, but I have a lot of other things I want to try, so it's not a priority and hasn't been started.

On the other hand, I also explicitly think it's a bad idea to write the vector search results directly into Scrapbox as AI output.

It's OK up to the point where a human reads it, but once it's in the search results again, there's a feeling that the value of the "title to content mapping" is spoiled.

That was a good summary.

This one in the continuation is indeed a problem.

This is how we're dealing with that issue right now.

So there is a difference in the utility that the two provide, what is the difference?

It is useful to use ChatGPT when explaining it to others

The omni's output is closer to my personal water surface (the boundary of what is not yet verbalized), so it is more effective in supporting verbalization.

Maybe it has to do with the different characteristics of blogs and Scrapbox: in Scrapbox, instead of explaining the same concept over and over again, you create a page for that concept and link to it.

https://gyazo.com/f93a3266d674745e880b61760e8bb721

Case 1

Most people will find it easier to understand.

From my subjective point of view, it is clearly degrading. Maybe GPT4 doesn't understand the cooporation correctly.

PS: I'd like to explain this a little more.

In Ikujiro Nonaka's SECI model, "collaboration" refers to the sharing of "tacit knowledge" that is difficult to verbalize through "sharing the same experience," such as through joint work, without the use of language.

So I get the feeling that you "don't get it at all" when you describe it as "sharing information, experience, and tips".

In light of this, this statement can be interpreted as "before making a linguistic plan, it is necessary to first share a nonverbal experience through cooporation.

And in the context in which this statement came up, we were talking about increasing intellectual productivity through AI.

This is the first step in the process. One form of this is to let AI live in the place where I usually do my intellectual production by creating a system that allows AI to work on its own, instead of humans giving it instructions every time and AI responding to them.

Example 2 ("Information exchange is an exchange" as you just said)

I found "information exchange is exchange" to be an important realization, but I think most people probably don't know what's so interesting about it.

Very rough description.

Exchange Form A: Giving a Gift in Return

Exchange Form B: Obedience and Protection

Exchange Form C: Pay the price and get the goods

Information exchange is an exchange" is a realization that "since information is also a good, information exchange can be considered within the framework of Gyoto Emorya's exchange theory.

Supplemental (if time permits)

https://gyazo.com/851a1f7001e9afba1c7447e90751e218

To include items (e.g., books) that cannot be placed in the public domain in the search results.

Vector search alone is pretty interesting.

If you imagine "a system in which books pop up and related pages open when you talk in front of a bookshelf," you will understand how interesting it would be.

If someone else's Scrapbox is hit, I read it interestingly, "Mr. X, I didn't know you wrote about this," which may lead to subsequent communication.

I'm still using it differently.

So you must feel different utility, what is it?

When it's self-derived, it's like the AI and the human are driving the thinking as a unified entity.

I feel like I'm accelerating."

Feel like you're "seeing things from a different perspective."

When derived from others, the feeling of "Oh, so this is what Mr. X said about this subject..."

I feel like I "found someone else's statement."

case

Fragments derived from oneself are reconstructed once chewed up inside oneself, so there is a sense of smooth connection. Fragments derived from others still have a hard surface.

However, it may be possible in the future for LLMs to do the "chewing and reconstructing", "melting together", and "smoothly connecting" themselves by fine-tuning with their own origin data.

The process of chewing is not a knowledge problem but a question of what chunks to divide the input into and what format to use, so I feel fine tuning would be useful.

Is it because I'm naive enough to pass off things of other people's origin as "that's one way of thinking" even if it differs from my current opinion?

The "it's normal for others to have different opinions" runaround?

If your self-derived opinion differs from your current opinion, since you are both you, "Why do we disagree?" Since they are both me, does this strongly trigger the question, "Why do we have different opinions?

I started using vector search and it unearths things like blog posts I wrote 8 years ago, but I wrote them myself and I don't remember them.

Even a book you read hits you unexpectedly and you say, "Whoa, did I read that in this book? Oh, it sure does! I didn't pay much attention to it when I read it before!" It's like that.

It is said that a good book should be read many times, but in essence, it is an attribute on the part of the reader to be able to accept what is written.

In that sense, it would be a great reading experience to come across relevant pages of books related to "what I am interested in now" and read them.

I like the idea of experimenting with mixing the two.

https://gyazo.com/3f316a9871d2f7ca09716b09a257d512

Countless beneficial events that have occurred

Diaries, chat logs, lecture materials, transcripts of conversations, and other "time-lined descriptions

AI searches on a topic and hits that "in the middle of the sequence" and mentions it there.

The person who sees it cuts out that part of the page and creates a new page.

They did it with the vector search story.

I made an excerpt page with "Social Security Expenses" in the title.

Difficult to do in advance

In the realm of personal knowledge management, there is a belief that a short page on a single topic is preferable

This makes it easier to form connections across topics and contexts!

Similar to the [single responsibility principle

So the act of extracting a "single topic" from a "page with multiple topics mixed in" is performed.

In Scrapbox, this can be done by selecting multiple lines and choosing New Page from the balloon menu.

This is commonly known as "cutting out" and as the name suggests, "cutting" and creating new pages and linking them to each other

Whether "cut" or "copy" is better is debatable, and I don't always think cut is better either, there is value in keeping lecture materials readable through.

With the minutes being placed in groupware, or something like that, users would be resistant to editing them.

Behavior due to the fact that it is a vector search for fragments engraved in chunks

I'm chopping at 500 tokens right now, which is about the size of a page in a book.

With chat logs, it's not an entire thread or one specific person's statement, but a unit of "several exchanges".

The search returns "the most dense mentions of the topic" in a long conversation.

A human discards the unwanted parts of that chunk, picks the relevant ones around it, or writes a new opinion inspired by it and creates a new page, which is beneficial.

Scrapbox is designed to afford cutouts, which is beneficial when combined with LLM

https://gyazo.com/5a0d0f0ccf1c48f4060d7fd62f099afa

Patterns of thought that are beneficial for increasing [Cognitive Resolution

I've also compared ChatGPT and omni in this presentation, as well as public and private omni.

Vector search works as a "mechanism to scrape together fragments of similar topics.

Similar to the KJ method of "gather fragments that may be related in one place, and then think about what kind of relationship they may have.

I thought you looked familiar.

https://gyazo.com/72240a24b23c6402b62a5d4a91580a0a

https://gyazo.com/c49e67265b1af15a9684e8ce8778ea6f

There is a difference between "similar topics" and "similar opinions."

Different opinions on similar topics are rather close in terms of vector search

I imagine that even at this moment, writing in this Scrapbox, in another Scrapbox, on a social networking service, or in the company's groupware produces a slightly different intellectual output than writing in another Scrapbox.

By bundling them together again

Below are my notes before writing

7/29: AI writes to Scrapbox, introducing the concept of Scrapbox Agents.

Early to mid-August: Omoikane study group, discussion on working with AI, writing research notes with AI, evolution of vector search, and topics on handling AI-generated pages.

Mid to end of August: topics related to updating and search, taking into account multi-head, page memory, and user cognitive load.

Beginning of September: Explore AI-user interaction, note management, and relevance between different content.

Mid-September: various topics related to multi-head thinking, the intellectual production techniques of engineers, and working with AI.

Late September: LLM vs. other models, discussion on human concepts, optimal use of Scrapbox, and feedback on non-public tools.

Overall, this period seems to focus on Scrapbox and AI integration, particularly the concepts of vector search and multi-heading, and AI thinking and interaction with the user.

Jul 29 AI writes to Scrapbox

8/12

8/16 Stacking Vector Search Results

8/18 Ignore AI-generated pages from vector search

8/23 Ignore used pages from all history

8/25 Update page limit

8/29 Lift Ignore from Vector Search on AI-generated pages

8/31 Main branch stopped

9/3

9/4

~

9/11

9/12

9/13

9/15

9/16

9/18

9/20

Integration in private project

2023/9/22

2023-09-24

2023-09-29

---