- Transcript from YouTube video of Event to Commemorate the 5th Printing of Intellectual Production of Engineers (55). Claude 3 Opus to extract and summarize the questioned parts. After looking at the output, I pointed out the parts that I thought were missing some words and asked the students to supplement them.
- Thoughts: a few of these were actually answered but dropped.
- I also got a list of Slide questions from the management and noticed when I tried to answer the unanswered ones.
Q: Do you have plans to write a book on intellectual production for engineers who use LLM? A: We are currently writing on Scrapbox. However, given the development of LLM, we believe that the book format may not be optimal; creating a system that LLM can read and answer may be more valuable for the intellectual product.
- (Supplemental note: Mr. Nishio mentions the definition of âwritingâ. If writing is considered writing, then he is currently working on writing on Scrapbox. On the other hand, he does not mention whether he plans to publish a paper book, but considering the development of LLM, he thinks that it would be more valuable to create a system that LLM can read and answer questions rather than an intellectual product in the form of a book.
Q: Is there anything you do or are aware of to make sure you are ready to output the knowledge you gain while working? A: It is important to first write down the knowledge you gain during your work in a place where only you can see it. Whether or not you make it public is another story. By writing it down and storing it as digital data, when you look up related topics later, your past findings will be searchable and reusable.
- (Supplement) Mr. Nishioâs definition of âoutputâ includes writing things down in a place where only he can see them. In other words, he believes that it is important to accumulate and organize knowledge within oneself first, rather than assuming that it will be made public from the beginning. He states that by doing so, when researching related matters later, past knowledge can be found and reused in searches.
- (PS) This area of discussion is related to tokorotenâs âIncreased abstraction capabilities will help you avoid NDAsâ.
Q: Do you use several LLM services for personal use? If so, how do you use them? A: I use several different LLM services. For example, Anthropicâs Claude is excellent at understanding long contexts, so we use it to process long documents such as article summaries. On the other hand, OpenAIâs GPT-4 is highly customizable, so when I have a clear task in mind, I use it by devising system prompts. However, which service is best suited for what is changing day by day, so you need to actually try various services and determine for yourself.
Q: I feel that if there is no opportunity to review the information compiled in Scrapbox or other systems, the functionality of the links cannot be utilized, the breadth and connections of knowledge cannot be expanded, and inputted information is forgotten. Do you have an opportunity to review the information on a regular basis? A: I think it is important to look for past records that are relevant to the topic you are currently interested in, rather than mechanically reviewing them once a month. Even if you are not using Scrapbox, it is important to make good use of the functionality of the system you are using, such as searching for related posts about what you are thinking about now.
Q: In 10 years, what do you think the art of intellectual production with rapidly evolving LLMs will look like? A: To be honest, it is difficult to predict what will happen 10 years from now, and I expect to see more areas where LLM will replace human jobs. But at the same time, we also need to focus on areas that AI cannot do, i.e., areas that LLMs cannot replace, and develop our capabilities in those areas. The important thing is to observe the changes around you and keep moving in the direction you think is best. I believe it is important to focus on the task at hand, rather than being swayed by abstract future predictions.
- (Supplemental) The message seems to be that it is important to focus on the specific task at hand and make steady progress.
These are the details of the main questions and answers from the lecture. From Mr. Nishioâs answers, I could sense that he sees LLM not as a mere tool, but as an entity that will change the very nature of intellectual production, and that he is committed to exploring new methodologies through practice. At the same time, I think he emphasized the importance of sincerely tackling what is in front of us while valuing our own sense of reality because we live in an era of rapid change.
Other questions (reprinted from answers in [/villagepump/intellectual production of engineers who use LLM](https://scrapbox.io/villagepump/intellectual production of engineers who use LLM))
Q: Tips for casual output from seniors to juniors A: I donât know the context (=context)
- You can think about what kind of context you might want to give them to elicit responses that are useful to you. Conversations to humans and prompts to LLMs are the same.
- Maybe repeating the âchallenge of getting ChatGPT to give me the output I expectâ will improve my âskill of asking questions that include appropriate contextâ and be beneficial to my conversations with humans.
Q: When I am on the front lines of education, there seems to be a polarization between those who can make good use of LLM and those who canât (just as there are those who can use Google search well and those who canât). I would like to see as many people as possible take advantage of LLM (in the same way that some people can use Google search well and others canât).
A: I had received this in advance.NISHIO Hirokazuâs Assistantă«çăăăăŠăżă
-
- The following is a summary of Scrapboxâs description of how to prevent the polarization between those who can successfully use LLMs (Large Language Models) in education and those who cannot, and to encourage more people to use LLMs effectively.
- tuning and customization to niche operations
- LLM is not a âkiller appâ for everyone at this time, but it can be tuned and customized to niche operations to great effect. It is important to apply it to highly specialized areas and focus on use cases that will make operations overwhelmingly more efficient.
- systematization and sharing of methodologies
- By systematizing the methodology of intellectual production using LLM and sharing the knowledge gained through practice and case collection, many people will be able to use LLM. It is necessary to present not only specific methodologies, but also theoretical considerations of how intellectual production changes through the use of LLM.
- providing education and training
- Develop the ability to use LLM by offering a variety of educational programs and training; more opportunities to learn specific uses and applications of LLM through hands-on training and workshops are effective.
- integration into actual operations
- By incorporating LLM into actual work and showing concrete examples of how it can be used, many people will become familiar with LLM. In particular, it is effective to show specific examples of initiatives, such as the review of insurance documents and their application to financial contracts.
- knowledge expansion, reproduction, and utilization
- Using a strategy of using knowledge to gain time and money and then investing those resources in further knowledge acquisition will allow for sustainable growth. Through this strategy, the ability to leverage LLM can be enhanced.
- Through these methods, we can promote the effective use of LLMs and prevent polarization in the educational field.
- tuning and customization to niche operations
- This LLM output is a âlow-cost beatâ; there is no need for the LLM output to become an unprocessed final product.
- In the actual response, the LLM output was used to discover âsectionâ as a reference opinion from a different point of view.
- Tuning and customization for niche businessâ, I realized that the question I want more people to use LLMâ How can people get along with LLMâ is only vaguely aware of the use cases in the first place.
- He said that it is useful to first identify use cases that are âuseful for thisâ in specific âniche operationsâ and then share the knowledge gained through practice and case collectionâ.
Q: How to deal with AI agents from a developerâs point of view A: Of course I see it as âwhat I make of it.â
Q: Wouldnât the need for memory be diminished if structured information like the knowledge graph is left to machines? A:
- Yes, if you can leave it to me.
- This question does not clarify âhow to build a knowledge graphâ or âwhat to leave to the machineâ
- If you can concretize it at a workable level and you think it would be beneficial to do it, then practice it and verify it.
Q: If you read a book that is not a PDF, what do you do to get that information into scrapbox? I can generate my own summary/get an AI to summarize it for me, but as long as itâs not the full text, the information will be lost. âŠ? A:
- It is inevitable that information will be lost. Any summary will be less informative than the original.
- Summarizing means discarding information in the original text according to subjective values.
- For the first half, see p. 133, footnote 26.
-
I am cutting and scanning more and more paper books, digitizing them, OCRing them, and making them cross-searchable. It is more efficient to be searchable than to remember.
-
Q: I often want diagrams for understanding, if not images, would you recommend DOT language or something else? A:
- Personally, I think DOT is less time efficient for output than handwriting.
- Itâs a decision of which to choose based on the merits of DOTâs âtext format data so that the current AI can read itâ and the âlead time until the AI can read handwritten images without problemsâ.
Q: I can see the benefit of keeping thoughts and ideas out there and available, but what do you do with ideas that you want to keep secret? A:
- I think itâs a confusion of âout of the brainâ and âout into the world.â
- The secret stuff is âout of the brainâ and not âout in the worldâ, specifically in the PRIVATE Scrapbox project.
- Iâd like to do this, let the AI read this⊠but the implementation isnât ready, thereâs nothing technically wrong with it.
- Private development is lagging behind because of the priority given to social value production that occurs when things that can be made available to the public are done first and made public.
Q: Are you using a local LLM? A:
- I didnât use it.
- I think it would be beneficial to use it and accumulate knowledge.
- There are a lot of âthings that would be beneficial to do,â so Iâm prioritizing my efforts, and as a result, local LLMs arenât high on my priority list, so theyâre on the back burner.
From Twitter
u_g_cccc I think that literature (in a broad sense) will be written on the assumption that it will be summarized anyway, and the volume of text will become shorter. If this happens, the significance and meaning of âsystematizing knowledge and informationâ will be fundamentally changed.
yappy0625 Iâm afraid it might only pick up convenient information or accelerate echo chambers. You could say that itâs customized for the individual, but Iâm not sure if itâs just growing into a convenient AI that will do the discovery for us.
This page is auto-translated from [/nishio/Forkwell Library#55èłȘç](https://scrapbox.io/nishio/Forkwell Library#55èłȘç) 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. Iâm very happy to spread my thought to non-Japanese readers.