3 ms·

This is the third response I've seen like this. Is this some sort of training material? "What to say when people doubt the machine" pamphlet? "Just squeeze the

This is the third response I've seen like this. Is this some sort of training material? "What to say when people doubt the machine" pamphlet? "Just squeeze the stone harder" is bad advice for getting blood.

omeid2 3y ago
If I understand correctly, you're not happy with the way ChatGPT can realise it's own mistake? or that it makes the mistake? I am not really sure, what is the problem exactly?
- genewitch 3y ago
  it doesn't "realize its own mistake", it requires negative feedback from the user to spur it into "apologizing" and trying to fix the "mistake(s)". It doesn't matter if the mistake is real or imagined. even the uncensored (less RLHF'd) models do this - apologize after negative feedback and attempt to fix. And "it's" only means "it is", the possessive is "its". You probably knew that, but didn't realize your own mistake without negative feedback.
  - omeid2 3y ago
    >It doesn't matter if the mistake is real or imagined. Is this based on your experience? My experience with ChatGPT is that rarely it apologises or attempts to fix something that is correct, though sometimes it does apologise and give you an alternative or the exact same working solution.
amenhotep 3y ago
That you think this is weird advice or advice that makes LLMs look good shows that your position is one of reflective opposition rather than understanding and experience with the things. If you interact with them much it's an obvious course of action. It will come up with nonsense, and if you tell it it's nonsense it will immediately apologise and explain why the nonsense was nonsense. It will come up with good information, and if you tell it the information was bad it will immediately apologise and offer nonsense explaining why. You can easily get it in a loop of it agreeing that solution A is bad and offering solution B, then agreeing that we previously excluded solution B and offering solution A instead. It is a very quick way to get an intuitive understanding of their limitations.
- krainboltgreene 3y ago
  > It will come up with nonsense, and if you tell it it's nonsense it will immediately apologise and explain why the nonsense was nonsense. This only makes sense if you believe LLMs are something more than they are. As someone who knows how LLMs work pretty well I don't. The presumption here is that the following advice is 100% better or correct and I also don't believe that.
  - omeid2 3y ago
    The assumption that anything is "100%" is absurdist. It is not a meaningful or useful discourse.

omeid2 3y ago

If I understand correctly, you're not happy with the way ChatGPT can realise it's own mistake? or that it makes the mistake? I am not really sure, what is the problem exactly?

genewitch 3y ago
it doesn't "realize its own mistake", it requires negative feedback from the user to spur it into "apologizing" and trying to fix the "mistake(s)". It doesn't matter if the mistake is real or imagined. even the uncensored (less RLHF'd) models do this - apologize after negative feedback and attempt to fix. And "it's" only means "it is", the possessive is "its". You probably knew that, but didn't realize your own mistake without negative feedback.
- omeid2 3y ago
  >It doesn't matter if the mistake is real or imagined. Is this based on your experience? My experience with ChatGPT is that rarely it apologises or attempts to fix something that is correct, though sometimes it does apologise and give you an alternative or the exact same working solution.

amenhotep 3y ago

That you think this is weird advice or advice that makes LLMs look good shows that your position is one of reflective opposition rather than understanding and experience with the things. If you interact with them much it's an obvious course of action. It will come up with nonsense, and if you tell it it's nonsense it will immediately apologise and explain why the nonsense was nonsense. It will come up with good information, and if you tell it the information was bad it will immediately apologise and offer nonsense explaining why. You can easily get it in a loop of it agreeing that solution A is bad and offering solution B, then agreeing that we previously excluded solution B and offering solution A instead. It is a very quick way to get an intuitive understanding of their limitations.

krainboltgreene 3y ago
> It will come up with nonsense, and if you tell it it's nonsense it will immediately apologise and explain why the nonsense was nonsense. This only makes sense if you believe LLMs are something more than they are. As someone who knows how LLMs work pretty well I don't. The presumption here is that the following advice is 100% better or correct and I also don't believe that.
- omeid2 3y ago
  The assumption that anything is "100%" is absurdist. It is not a meaningful or useful discourse.