It's a bit like talking to T-800 was in Terminator 2. Good enough (I guess) when you talk about factual information uploaded into it, but when you start to go to emotional stuff (i. e. stories), it goes either frustrating or creepy really fast.
Basically it's Google on steroids plus a storytelling Markov chain.
When the hype™ promises you that GPT-4 will have 25000 inputs compared to GPT-3.5's 3000... This does not necessarily mean it's gonna be 8 times smarter (or 8 times more right). Because it's 8 times the signal but also 8 times the noise.
More likely it'll be about as wrong (maybe somewhat smarter), but just... 8 times more convincing.