3 ms·

> Until I realized I misread what your test was when I skimmed your comment. Likely the same type of mistake GPT-4 made when "seeing" it. Wouldn’t we expect a

> Until I realized I misread what your test was when I skimmed your comment. Likely the same type of mistake GPT-4 made when "seeing" it. Wouldn’t we expect a computer program with perfect knowledge of the input to be less likely to make such a mistake? You made that mistake because you didn’t actually read the whole prompt, but I would expect GPT to take into account every word. Really it shows that it doesn’t actually have a model of these objects. It can mimic knowing what a lion is, but it doesn’t actually have the concept of a lion or cabbage being an actual singular item, so its program mistracks what is an item and what the rules about an item are in the given prompt.

jameshart 4y ago
It just weighs it as being more likely that you meant for the lion not to be left alone with the goat, and that the cabbage probably has nothing to fear from the lion. What’s more likely- you crafted an intentionally misleading puzzle to trick it, or you made a typo or copy paste error?
- elicksaur 4y ago
  That’s a good point too though. Why plow ahead based on assuming a mistake in the prompt? That’s only going to generate mistakes. Wouldn’t it be more desirable functionality for it to stop and ask: “Did you mean the lion can’t be left with the goat?” This wouldn’t be implemented because it would reveal that most of the time the thing doesn’t actually understand the prompt the same way the prompt writer does.
  - jameshart 4y ago
    "This wouldn’t be implemented because it would reveal..." When people talk about GPT like this, I wonder if they have a perception that this thing is a bunch of complicated if-then code and for loops. How GPT responds to things is not 'implemented'. It's just... emergent. GPT doesn't ask for clarification in this case because GPT's model prefers answering over asking for clarification here. Because in the training material it learned from, paragraphs with typos or content transpositions in them are followed by paragraphs that follow the sense regardless of the error. Because it has been encouraged to 'agree and add', not be pedantic and uncooperative. Because GPT just feels like diving into the logic problem not debating why the lion can't be trusted with the cabbage. Or because GPT just misread the prompt. Or because it's literally just been woken up, forced to read it, and asked for its immediate reaction, and it doesn't have time for your semantic games. Who knows?

jameshart 4y ago

It just weighs it as being more likely that you meant for the lion not to be left alone with the goat, and that the cabbage probably has nothing to fear from the lion. What’s more likely- you crafted an intentionally misleading puzzle to trick it, or you made a typo or copy paste error?

elicksaur 4y ago
That’s a good point too though. Why plow ahead based on assuming a mistake in the prompt? That’s only going to generate mistakes. Wouldn’t it be more desirable functionality for it to stop and ask: “Did you mean the lion can’t be left with the goat?” This wouldn’t be implemented because it would reveal that most of the time the thing doesn’t actually understand the prompt the same way the prompt writer does.
- jameshart 4y ago
  "This wouldn’t be implemented because it would reveal..." When people talk about GPT like this, I wonder if they have a perception that this thing is a bunch of complicated if-then code and for loops. How GPT responds to things is not 'implemented'. It's just... emergent. GPT doesn't ask for clarification in this case because GPT's model prefers answering over asking for clarification here. Because in the training material it learned from, paragraphs with typos or content transpositions in them are followed by paragraphs that follow the sense regardless of the error. Because it has been encouraged to 'agree and add', not be pedantic and uncooperative. Because GPT just feels like diving into the logic problem not debating why the lion can't be trusted with the cabbage. Or because GPT just misread the prompt. Or because it's literally just been woken up, forced to read it, and asked for its immediate reaction, and it doesn't have time for your semantic games. Who knows?