#Bug
14 messages · Page 1 of 1 (latest)
One thing I think is helpful when talking to ChatGPT is not to think of it as a fact bot, but only a chat bot. In other words, while it's broadly intelligent, it's not designed to be correct really. It's just designed to respond in natural language. It just so happens to be so good at this that it is often correct, but not always. Hallucinations like this are not uncommon!
yep, not really a bug, it's a known limitation of these ai models, especially the older gpt models like gpt-3.5
if I ask him for a piece of information. give an answer to it, then correct it in the next one. then I ask the same question again, and he corrects it again.
nothing new, well known and documented limitation of these models.
it also cant really learn from you, if you start a new chat it wont know anything about previous converstations
but he has the information about it, he just can't decide which is the right answer.
I don't think he's got the information. From my experience you need to "preload" everything you need to talk about otherwise it won't pull it from memory. You have to bring all the needed pieces together for it. It might have very limited retrieval ability, or maybe the thinking process is capped and can't really do a thorough search through its memory.
it was a simple question. I asked which food contains the most protein. it said tuna (25-30g). then I asked if it wasn't the chicken breast (20-25g). he replied that yes it is. and this 5 times back and forth. he can't decide which one is more, I think because the answer is not a specific number. but there is more to tuna. he gives the answer, but he can't decide.
you're using gpt3.5 it's extra bad at those kinds of things, again, this is a known limitation
it's also quite bad at simple math problems
and the answer will change based on how you ask it to approach the question
this is not a bug, it's just a property of these AI models
Ok. Thank you for your help