372
ChatGPT 5's PHd level intelligence is breathtaking \s
(lemmy.world)
I saw this on reddit, thinking it’s a joke, but it’s not. GPT 5 mini (and maybe sometimes 5) gives this answer. You can check it yourself, or see somebody else’s similar conversation here: https://chatgpt.com/share/689a4f6f-18a4-8013-bc2e-84d11c763a99
It seems as though you read the first sentence I wrote and not any of the sentences afterward.
Indeed I did. Especially the parts where you made excuses for it, saying..
Nobody claimed it was magic. They gave it a very reasonable prompt that a grade 1 child could answer, and it failed. And this..
Again, you're claiming the prompt is misuse, "tHEyRe uSInG iT wRonG". Going on to say it's the 'AI companies fault really' for pushing it to everyone instead of just domain experts is again not getting the point. The AI should never respond with a confident answer to a prompt it has no idea about. That's nothing to do with the user or the targeted audience, that's just shit programming.
Agreed. But the technology isn't there yet. It's not shit programming, because the theory of how to solve this problem doesn't even exist yet. I mean, there are some attempts, but nobody has a good solution yet. It's like you're complaining that cars can't go at 500 miles per hour since the technology limits them to 200 mph or so, and blaming this on bad car design when it's actually the user's expectation that's the problem. The user has been mislead by the way things are presented by AI companies, so ultimately it's the AI company's fault for overmarketing their product.
(Fuck cars btw).
LLMs don't work like grade 1 children. The real problem is that AIs are being marketed in such a way that people are expecting them to be able to be at least as good as anything a grade 1 child can do. But AIs are not humans. They are able to do some things better than any human yet on other tasks they can be outperformed by a kindergartner. This is just how the technology is.
Blame expectations, blame marketing, fuck AI in general, but you've been totally misled if you're expecting it to be able to, say, count the number of letters in a word or break a kanji into components when all it sees are tokens; not letters, not characters.