Comments

• ctietze

  August 10

  The <something> pseudo XML markers helped earlier generations (GPT 2 I think?) to structure input and responses at all because the models had such a hard time compared to today. It's optional nowadays but can still help save words (tokens) expressing that something is e.g. an instruction, something else is 'meta'. It would often be harder to convey with prose alone, so not obsolete at all from the practictioners in the field I know.

  Author at Zettelkasten.de • https://christiantietze.de/
• tomp

  August 10

  @harr said:

  @tomp said:
  The reason for using simple markup is that chatbots are very very good at recognizing patterns. This way the instructions don't use many tokens and don't get mixed up with the actual requests you want to make.

  I played with the blocks and asked AI and Google for advice. Apparently you don't need to markup every single line anymore. Instead Markdown was recommended.

  Uhmm, Markdown is a (simplified) markup language. "Markup" doesn't just mean XML.

  That prompt looks really good.
• tomp

  August 10

  @harr said:
  Another experiment is in the works. This time a default prompt to start chats about books. Nothing to see yet, but I wanted to share the idea.

  The basic idea is not to ask the AI what it thinks about the book, but ask the AI what everybody else is thinking about the book. How was the book received by the audience? What ideas are commonly talked about? What ideas are controversial? What ideas are generally accepted? What ideas have been influential?

  I'm curious how it will affect my reading notes and actual reading.

  The act of writing such a prompt has an interesting effect. I'm looking at my own reading practice and reading goals from a different perspective. What is it, that I typically want to get out of a book? Why do I care, what others say? Do I want to fact-check the book? Or is the book a source for fact-checking others?

  Has anybody else tried out a default prompt for books?

  This thread's initial question remains interesting!

  I ask questions like that. I have so far not used a chatbot with respect to reading a book. But for questions about something I'm interested in, I will ask e.g. what the usual or well-accepted writings say. I will ask if there is research that is regarded as reliable. I would ask for the most common dissents but I haven't had the occasion yet.

  My thinking is that the chatbot is best at constructing replies using the most likely token probabilities so such replies should be more reliable.
• andang76

  August 17

  @harr said:
  Another experiment is in the works. This time a default prompt to start chats about books. Nothing to see yet, but I wanted to share the idea.

  The basic idea is not to ask the AI what it thinks about the book, but ask the AI what everybody else is thinking about the book. How was the book received by the audience? What ideas are commonly talked about? What ideas are controversial? What ideas are generally accepted? What ideas have been influential?

  I'm curious how it will affect my reading notes and actual reading.

  The act of writing such a prompt has an interesting effect. I'm looking at my own reading practice and reading goals from a different perspective. What is it, that I typically want to get out of a book? Why do I care, what others say? Do I want to fact-check the book? Or is the book a source for fact-checking others?

  Has anybody else tried out a default prompt for books?

  This thread's initial question remains interesting!

  Sometimes I ask for information about a book to see if it might be useful to me (or ask for a recommendation based on a specific need), but I really dislike having the ideas I’ll encounter revealed to me beforehand. That is an analysis I want to conduct myself—and, if anything, I’d rather evaluate their quality only after I've read them.