2023-04-06
gfx Youâd be surprised how often we get extremely accurate answers when we ask GPT-4 about TypeScript. I think this is probably because there are so many good TypeScript texts in the world.
gfx However, I have no idea about the new features recently and canât get attached to them. So no satisfies will be used in the code that GPT-4 will write. This gap is not so easy⊠for those who know what theyâre doing, but it might be hard to be practical.
omochimetaru than creating a new, highly functional language, It is more efficient to use old languages with AI support for development, I donât like the idea that this would put a brake on the evolution of the languageâŠ
gfx That said, I think that concern is unfounded, as we have already achieved engines like bing that learn in real time. Rather, I am more concerned that even if a new language is created, the sample size will be too small for AI to write well and it will not spreadâŠ
omochimetaru I see. I wonder if the AI is smart enough, even if the text is only about the official specs, if it can help us understand and use new concepts from it.
nishio The technology is already in place and will soon come down to the average consumer âAbility to give LLMs a knowledge package of your area of interestâ will come down to the general public soon. When that happens, authors of new languages will create knowledge packages for LLMs, and people who want to use the language will be able to load them into LLMs and use them.
- I wrote this and implemented it in ChatGPT Plugins on 4/7, and it was provided under the name Assistants API at OpenAI DevDay on 11/7 (related: Retrieval-augmented Language Model).
nishio Creating knowledge packages is actually easier than creating documents for human readers. Because humans have limited short-term memory, it is necessary to provide knowledge in a âgood orderâ, and this is very hard for the author. - Writing is one-dimensional
nishio Rather, it is better to identify âwhat information is missingâ based on usersâ questions and add it, and it is important to create a system to do so. It is important to build a system to do that.
2023-11-10
tokoroten Iâve been able to make a custom model by feeding PDF into ChatGPT, âA book is an additional data set to the LLMâ feels like a definite future.
E-book services that can work with LLMs are going to be the next dominant e-book service. The age when a stack of books will be eaten up by LLMs, which will then be useful.
unnonouno Conversely, the writing process is going to be data creation that feeds the LLM (it no longer even needs to be in Japanese). Itâs going to be a custom LLM creation service, not publishing.
nishio Books are inefficiently efficient input to AI because information is forced to be one-dimensional for human readability. I have a feeling that the next generation will be strong in those who can create information in a data structure that is convenient for AI.
(From Nishioâs conversation log from April or so) Easier to write for LLM than to write for humans.
nishio Iâll write the date and related links later w
nishio Just yesterday we were talking about connecting FAQs to LLMs, and these FAQs are not books, âa format for transferring knowledgeâ and will be used more widely as the search gets better. Iâm sure they will be used more widely as search becomes better and more precise in answering questions.
nishio When it comes to moving from books to FAQs, âWhat should I know first?â What will I get from reading this book? etc., and for those who want to read the book in a one-dimensional way, you can give them one ânext question candidateâ for each answer.
tokoroten There are cases where what we call FAQs are really FAQs and cases where the line is drawn by FAQs, so the former will be LLM It will be absorbed, but I donât think all of it will be incorporated because of the latter feature. https://tokoroten.medium.com/æ ć ±â °Books-Uncollected Manuscript Offerings-The%EF%BC%92-c5c48fd04a69
nishio Maybe the latter will display a source label such as âofficial informationâ or âcommunityâ and the recipient will be able to make a decision based on that?
nishio Also, there was a discussion about what to generate in the âGâ part of RAG, and many people try to create âanswersâ that propagate âtruthâ but I think that is a plot evil! I prefer to generate âopinionsâ like âI thought this article might be useful in resolving your questions in terms of
âŠâ
This page is auto-translated from /nishio/äșșéăźăăă«æžăăăLLMăźăăă«æžăæčăæ„œ 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.