So I'm curious — we're all here because we at least hate the current state of AI with hallucinating facts, being used to undress women and children, and all the fuckery that goes along with it.
I grew up watching Star Trek: The Next Generation, which takes place on a ship with a perfect AI that does everything right and basically does nothing wrong. It never hallucinates information; it's always right. It has never been used to undress people against their will; however, the Holodeck is kind of an extension of that and was used for that on Deep Space Nine, when operated by a Ferengi (capitalist alien race in a world where humans are communist). But the Enterprise holodeck would never do that. The shipwide AI also does not traditionally carry on conversations. The one time it does, the human was hallucinating — sort of. The doctor was in a pocket universe, people were disappearing, and at one point the AI told her she was the only crew person on the Enterprise, and no, that did not make sense, but that that was still how it was. Because, in her pocket universe, it was true.
So the question is... would you want a perfect AI that was incapable of lying or harbouring anything untrue? Basically you could ask it anything and it would give you the correct answer.
The one fault I can find with that fictional AI is when Data (the android), dressed like Sherlock Holmes, asked the computer to "create an enemy which rivals my intelligence." He meant to say Sherlock Holmes's intelligence, who he was cosplaying, but the computer made a self-aware malicious AI that got out of the Holodeck and tried to destroy the ship... because it was told to do so. Other than that, though.
...I'm not trying to mislead anyone, so I will drop the other shoe, answer the begged question now. I've always felt that to get to that level of AI, we need to wade through the shit we're in now. So yeah, before you ask, that's kind of the point of the thought exercise. However, I will also say that I do not think we will get to Star Trek AI, I think we will get to Terminator AI, destroying the world rather than lifting people up. I think maybe in the Star Trek universe, AI didn't really take off until people realised that war wasn't the answer, after WW3/the Eugenics Wars, and so they were making AI to make things better, not worse. We are not in that timeline. I look at what is happening now, IRL, and the timeline in the Terminator franchise, and it's clear to me that that one is more realistic.
That said, I still wonder if anyone would want AI if it did not have any of the problems.
The thing with comparing sci-fi AI to modern LLMs is that virtually no science fiction AI that I know of actually acts the way LLMs do. They tend to be good at things modern AI is bad at, like logical reasoning or advanced math, but bad at things that AI can already do, like generate images that at least look like they could be human art, or write text that appears emotionally charged. They also tend to be directly programmed in ways that a singular (usually genius, but still) individual can pick through and understand, rather than being trained in a black box sort of manner that is very difficult for a human to reverse-engineer.
That isnt surprising, sci-fi writers arent oracles after all and just having AI of some kind probably makes a story more realistic than just assuming the technology never gets invented even far into the future, but in my view these kinds of sci-fi AI are basically a different, hypothetical technology going for the same end result. As such, I dont really expect even a very advanced iteration on what we have will look like star trek AI, any more than modern cars tend to fly or run off miniature nuclear reactors the way sci-fi of decades ago saw cars of the future. I dont think it will look like skynet either. I do think we might get some interesting science fiction in the coming decades exploring what a very advanced version of the technology we do have might end up like though. It probably wont be terribly accurate either, but Id bet it will be closer than works where the basis for extrapolating AI tech is "what if the calculator could talk and think".