"Autocompletion" isn't an adequate explanation when neither the prompt nor the output has ever existed before. Something else is going on... something that obviously has a lot in common with how our own minds work.
That's not a lot of different sentences when you consider the combinatorical explosion.
I could also understand an argument saying that it's impossible for a computer to think, but I don't think I'd agree with it.
I have seen many claims that transformer models can't "think" but I have not seen any purposed tangible tests to verify or refute the claim.
Simpler models look very good on the surface, but there are some examples where this starts to break own. Example:
"Is it legal for a man to marry his widow's sister?"
The catch here is the man is dead, which is the reason he can't marry. GPT-4 gets this right. Simpler models however can be prompted to perform step-by-step reasoning where they figure out that dead men can't marry, and that a "man's widow" means he's dead, but the model still can't be coaxed into figuring out that he can't marry in the scenario above.
I'm guessing GPT-4 still has similar hangups where it gets stuck and can't get out no matter the coaxing, although I don't have any examples handy. Anyone?
This doesn't mean GPT-4 is a stochastic parrot, just that there are still flaws in its reasoning abilities.
Turns out posthumous marriages are legal in some places so a French man can legally marry his widow’s sister assuming it doesn’t run afoul of other laws like (which I didn’t check the legality of) polygamy.
LLMs are great at being interfaces to the data they were trained on. Ask them anything beyond that, and it becomes clear they're not capable of thinking like humans do, i.e. produce novel ideas.
It's kind of absurd we're having this discussion, TBH. The only reason it's difficult to distinguish mimicry from thinking is because of the sheer volume of data we're talking about here, and some clever algorithms that find correlations in it, and can synthesize a human-like response.
>the sheer volume of data we're talking about here, and some clever algorithms that find correlations in it, and can synthesize a human-like response.
If you swap the volume of data from "terabytes of UTF-8" to "decades of continuous input from the 5 senses," this doesn't sound very different than what you or I do already. The fact that it's difficult to differentiate mimicking from true thinking strongly suggests that there really isn't a difference, or at minimum not one that has any real world practical significance. If there is a scenario where the difference does have practical significance, what scenario is that? Because I sure can't think of one.