3 ms·
I see this parroted but no, it's not like humans at all in this regard. While some humans may speak confidently incorrectly they internally know whether they kn
by alfalfasprout 4y ago
I see this parroted but no, it's not like humans at all in this regard. While some humans may speak confidently incorrectly they internally know whether they know something or not. Large language models don't inherently provide an accurate confidence.
- SpeedilyDamage 4y agoThat’s not strictly true, an AI for sure can provide a confidence interval for its answer, but the sentiment I feel is generally correct.
- Retric 4y agoChatGPT can’t provide any score for how correct it’s answer is. Basically everything it says is a lie, some of those lies happen to randomly be true.
- SpeedilyDamage 4y agoIt absolutely can, just check out the OpenAI Azure documentation to understand more: https://learn.microsoft.com/en-us/azure/cognitive-services/openai/reference https://learn.microsoft.com/en-us/azure/cognitive-services/o... Edit: This also might be helpful: https://learn.microsoft.com/en-us/azure/cognitive-services/qnamaker/concepts/confidence-score https://learn.microsoft.com/en-us/azure/cognitive-services/q...
- Retric 4y agoI don’t know what you think that shows, but none of it corresponds to “truth.”
- SpeedilyDamage 4y agoI wasn't trying to suggest it did, just that the concept of "confidence" is a key one in ML.
- Retric 4y agoGotcha. Rereading what I said, “correct” rather than “factual” was poor word choice on my part. From it’s perspective correct is just whatever results in a high score.
- SpeedilyDamage 4y agoYeah, it's well trained by humans who do know what is "factual" (or at least better than ChatGPT does), and the "correctness" score gets used to figure out why ChatGPT gave a "wrong" answer. Over time, the model is trained this way to better align "correct" with "factual", but yeah ChatGPT needs humans to tell it what's "factually" accurate and what's not.
- Retric 4y ago> well trained by humans Yes and no, the goal wasn’t focused on just being a source of truth. They wanted it to be able to write fairy tales, poetry, and song lyrics so creative answers where rewarded even if they weren’t factually accurate. There was an attempt to reduce harmful and deceitful responses, but the team acknowledge the tendency for “plausible-sounding but incorrect or nonsensical answers” calling them hallucinations. Which becomes common when prompts are outside of it’s training set. So yes there is definitely a link between the internal “Correctness” score and “Truth” but they are measuring very different things.
- willhslade 4y agoHow confident is it in it's confidence interval?
- skizm 4y agoIn what practical or observable sense does that matter if someone never acts as if they know what they're saying is incorrect? Also, assuming the Dunning-Kruger effect is mostly accurate, there will be large portions of the population genuinely believing they are correct when they are not.
- notahacker 4y ago> In what practical or observable sense does that matter if someone never acts as if they know what they're saying is incorrect? People usually do act as if they know what they're saying is incorrect, especially if it's a case of lack of recollection rather than a conscious and strategic lie. ChatGPT defaults to offering a continuation with the correct format and plausible text strings, without regard to truth values it doesn't know because it's indifferent to them. And even if human learning or human motivation did lead a to person to internalise nothing about, say, the titles of scientific papers but the general format and authors well enough to make up similar sounding ones on demand in future, they probably wouldn't be as good as it...