In Reply to: I know it still makes mistakes. And most of the time you posted by Bruinfan4ever on May 22, 2026 at 11:27:24
mistakes is like saying a parrot sometimes misspeaks. As far as comprehension is concerned, a parrot is actually in a much better position to make connections between what it's saying and its real world implications than an LLM, for which your question and the whole text it's processing are all a series of meaningless character sequences.
I can see it being potentially useful as a summary engine under very specific conditions if used very narrowly (on a single document, for instance).
So it gives me this detailed information. Why would I not believe what it is saying here?
Respectfully, because ChatrGPT doesn't know what climate is. What change is. What scientist is. What agreement is. What real is. All it knows is when you see the words "climate" in real proximity to the word "change" (when they're adjacent, let's say) there's a good chance that you'll find the word "scientist" within, say, 15 word proximity to the phrase, and "agree" within 3 to 5 word proximity of "scientist" and "real" within 3 word proximity of the phrase "climate change". So it found what you are looking for. Then it finds documents that have that constellation of words and finds that a lof of them also contain "by John Cook and colleagues (2013)" somewhere near that constellation high percentage. So it tacks that in the text.
DOn't get me wrong. Like I'll say over and over again, this sh*t is f*cking brilliant. All you need is to understand the mathematical and computational complexity involved in building a coherent sentence in any language in basic ways to drop down on your knees and say "I'm not worthy" in response to how deftly fhe GPT4 models are able to put those snippets into actual grammatical sentences.
But at the end of the day, all it's doing is to spit back at you the statistical co-occurrence or words in the documents it was fed. For an LLM to process "Where does the 97% of scientist agreement on climate change being real" is no different than processing "Xifsf epft uif (&5 pg tdjfoujtu bhsffnfou po dmjnbuf dibohf cfjoh sfbm" would be for you as far as comprehension is concerned.