Comment by wvenable

3 days ago

> If you had the LLM write the comment, then it wasn't your thoughts.

But what if I provided the LLM my thoughts? That's actually how I use LLMs in my life -- I provide it with my thoughts and it generates things from those thoughts.

Now if I'm just giving it your comment and asking it to reply, then yes, those aren't my thoughts. Why would I do that? I think the answer goes back to my original point.

If I'm telling you my thoughts and then you go and tell a friend those thoughts, would you say those are still my thoughts even though I wasn't the one expressing them directly to your friend?

4 comments

wvenable

meatmanek 3 days ago

I like to think about it in terms of output-to-prompt ratio. For HN comments, I think an output ratio of 1 or less is _probably_ fine. Examples:

    - translating (relatively) literally from one language to another would be ~1:1.
    - automatic spelling/grammar correction is ~1:1
    - Using an LLM to help you find a concise way of expressing what you mean, i.e. giving it extra content to help it suggest a way of phrasing something that has the connotation you want, would be <1:1

Expansion (output > prompt) is where it gets problematic, at least for HN comments: if you give it an 8 word prompt and it expands it to 50, you've just wasted the reader's time -- they could've read the prompt and gotten the same information.

(expansion is perfectly fine in a coding context -- it often takes way fewer words to express what you want the program to do than the generated code will contain.)

wvenable 3 days ago
I think all your examples are all perfectly fine.
As for expansion, that might just be the risk we take. I been downvoted on reddit for being "too verbose" in my replies and I'm a human. And perhaps just reading the prompt in that case wouldn't give you more information; the LLM might actually have some insight that is relevant to the conversation. What's the difference between that and googling for something and pasting it in?
- ffsm8 2 days ago
  
  The linked rule does not make such a distinction, and I don't see how this rule could be enforced with such a caveat, either.
  Hence no, none of these examples should be okay. Even if pure translation and grammar check is gonna be effectively impossible to detect too, so likely pointless to talk about
  And the last one is often detectable and very clearly against it - I'm not sure how you can come to any other conclusion
  
  1 reply →