4 ms·

I'd like to provide you with some constructive feedback on your comment: 1. You should include what you think is interesting about the large language model's a

I'd like to provide you with some constructive feedback on your comment: 1. You should include what you think is interesting about the large language model's ability to solve the word puzzle. It's not obvious to me as a reader. 2. You shouldn't post the step-by-step solution to a puzzle, regardless of how it was derived. It's unlikely that anyone will read through it, and if they do, they're less likely to try the puzzle themselves.

rokkamokka 2y ago
I mean, I thought it was an interesting read... I'm sorry you didn't
- yesfitz 2y ago
  What did you find interesting?
jsnell 2y ago
It's not actually solving the puzzle, for one. Like half the steps are obviously nonsense.
- jw1224 2y ago
  Are you sure about that? It correctly solves the puzzle. Each step (though a little confusing to parse) does make sense in context.
  - hulium 2y ago
    The first line is already nonsense. The answer is obviously not "room". Getting the correct final answer tells you nothing about the reasoning. The LLM will solve the puzzle even if you only pass it the sentence "the [] de Milo is discovered by a Greek []".
  - jsnell 2y ago
    > Each step (though a little confusing to parse) does make sense in context. The proposed intermediate answers don't make sense, and they're the actual puzzle. You must answer all the intermediates correctly, you can't skip levels. The following ones clearly aren't correct (I think, I hadn't actually done that puzzle): ["the of nowhere"] -> middle; the model gets that right, but garbles the answer due to not correctly extracting the clue and having trailing garbage [your [middle] one can get you grounded] -> finger; "room" as suggested by model makes no sense [two-dimensional [finger] puppet projected on a wall] -> shadow; the model gets this right, but only by ignoring the clue completely, which means the justification is nonsense [dial ([shadow]-based clock)] -> sun; the model says sundial, which then forces it to make an inane "a sundial represents the sun" argument at the next level. [[red] spandex halloween costume, maybe] -> devil; the model doesn't give an answer word at all for this, but just says that "a red spandex costume is not a pitchfork"
SamBam 2y ago
I was actually quite interested, even though it's more of a comment about AI than about the game. I would note that I think the LLM had trouble because copying and pasting from the puzzle omits all the blank word underlines ("____"). I've dealt with this when copying into my own notepad for trying to work it out myself.
- yesfitz 2y ago
  What did you find interesting about it?

rokkamokka 2y ago

I mean, I thought it was an interesting read... I'm sorry you didn't

yesfitz 2y ago
What did you find interesting?

jsnell 2y ago

It's not actually solving the puzzle, for one. Like half the steps are obviously nonsense.

jw1224 2y ago
Are you sure about that? It correctly solves the puzzle. Each step (though a little confusing to parse) does make sense in context.
- hulium 2y ago
  The first line is already nonsense. The answer is obviously not "room". Getting the correct final answer tells you nothing about the reasoning. The LLM will solve the puzzle even if you only pass it the sentence "the [] de Milo is discovered by a Greek []".
- jsnell 2y ago
  > Each step (though a little confusing to parse) does make sense in context. The proposed intermediate answers don't make sense, and they're the actual puzzle. You must answer all the intermediates correctly, you can't skip levels. The following ones clearly aren't correct (I think, I hadn't actually done that puzzle): ["the of nowhere"] -> middle; the model gets that right, but garbles the answer due to not correctly extracting the clue and having trailing garbage [your [middle] one can get you grounded] -> finger; "room" as suggested by model makes no sense [two-dimensional [finger] puppet projected on a wall] -> shadow; the model gets this right, but only by ignoring the clue completely, which means the justification is nonsense [dial ([shadow]-based clock)] -> sun; the model says sundial, which then forces it to make an inane "a sundial represents the sun" argument at the next level. [[red] spandex halloween costume, maybe] -> devil; the model doesn't give an answer word at all for this, but just says that "a red spandex costume is not a pitchfork"

SamBam 2y ago

I was actually quite interested, even though it's more of a comment about AI than about the game. I would note that I think the LLM had trouble because copying and pasting from the puzzle omits all the blank word underlines ("____"). I've dealt with this when copying into my own notepad for trying to work it out myself.

yesfitz 2y ago
What did you find interesting about it?