"The main claim [...] is both somewhat obvious and previously already stated"
"Many pieces of writing are overly assertive and inaccurate."
"I do not think it deserves spending half a page demonstrating that {0^n 1^n} is not in the regular language."
Yes, there's a school of philosophy that claims that AI cannot be built with digital hardware because it needs infinite state. Penrose is the leading exponent of this argument.[1] This isn't taken too seriously any more, now that finite machines are doing so well at AI.
This is close to theology. Man must be special, right? Only humans can play chess, right? Play Go? Pass the Turing test? Create art and music? We're running out of boxes to check. Common sense (as in getting through the next 30 seconds without a major screwup) and manipulation in unstructured settings are still in bad shape, but we only need squirrel-level AI for that.
On that subject, there was a "what does the cerebellum do" article on HN a few days ago. That's a big issue. The lower mammal brains are mostly cerebellum. The weak areas in AI at present are mostly cerebellum functions. That area has never gotten enough attention in AI research. Today, though, the amount of hardware to do a low-end mammal cerebellum equivalent doesn't seem at all unreasonable.
[1] https://sortingsearching.com/2021/07/18/roger-penrose-ai-ske...
It would follow that comparable intelligence would not be possible with current digital computer architectures.
I also think a lot of people are a little too fooled by the “intelligence” of current models which are mimics with enormous libraries of “knowledge”. It is not too hard to be frustrated by the enormous limitations of current models. Impressive, yes, smart, no. Though to refute claims of their intelligence is really pushing to refine the philosophical definition of intelligence which has been rather vague so far. We can’t just say “i know it when i see it” because computers pretending is getting very good.
This comes up a lot in things like science fiction, movies, etc...
E.g.: an author can state that a character has "superhuman intelligence", but the author themselves would need to actually posses this superhuman intelligence to write the actions or dialogue of that character.[1] Because authors are "just human", they obviously can't be superhuman, so they generally avoid directly demonstrating the actions of superhumanly intelligent characters. I.e.: the "tell, not show" instead of the usual "show, don't tell."
The current crop of AIs are 100% "show". They demonstrate intelligence directly, by answering questions instead of just talking about how they would answer questions, which is what fake intelligence would do.
PS: Next time you see a politician promise to deliver a solution, but not actually state what that solution is... well... now you know what's going on with that particular situation!
[1] There's a degree of cheating available to authors, because they have information that the in-world fictional characters don't. Similarly, they can set up the question to lead to an answer they already have. The chat-based AIs can't cheat like this, they have to answer the questions given to them.
The best hallucinations I saw was asking chatgpt for a list of references for a topic I was researching. They all looked entirely believable, titles, authors, institutions, summaries… IIRC there were about twenty of them. Only one actually existed.
I don't think your argument proves what you claim.
One could argue that humans will say "I don't know" when they don't know, but that ultimately depends on training. We currently don't have LLMs trained to answer "I don't know" (I suspect that in part that's due to incentives of focusing on something that looks impressive; one could argue that if an AI system too often says "I don't know" it's less impressive albeit it may be more useful since you can rely upon it more when it says it knows)
There are plenty of cases where humans are trained to hope they sound convincing even if they make up answers. I've seen a fair share of this done by students.
Put a gun to Einsteins head and force him to answer any question you ask him without hesitating or denying an answer.
This is not an argument for stating LLMs like ChatGPT are intelligent (it is not, but not for any of the reasons i have read about so far anywhere). But the hallucinations definitely don't show anything about intelligence. Actually, if they did, they would actually indicate intelligence rather than not. So, quite the opposite which should be very obvious.
If you were forced at gunpoint to provide such URLs -- valid or not -- with literally no option other than to keep talking, you would... make up some vaguely correct-sounding URLs!
The AIs are doing precisely what an intelligent human in their place would do. Blurting out something, anything, even if they don't have the answers.
That's because the current crop of AIs are essentially being forced to provide the next token, directly equivalent to a human being forced to continue speaking against their will.
There are fixes being put in place already for this, such as using human feedback to tune the AI so that it answers with a generic "I don't know" blurb if it's not sufficiently confident in the next token. Annoyingly, the new "Turbo" versions of ChatGPT have overdone this a bit, and now it refuses to answer even when it does actually know the answer!
There are also variations of neural networks that encode not just "values", but probabilities and hence confidence intervals. These can directly output both tokens and the true confidence values, which would allow complex tree searches, etc... and avoid outputting this kind of gibberish entirely. The maths for this is well established, but it would be much more expensive to run models using it, so nobody has bothered yet.
Nah. Either the human is feigning, lying with full knowledge that they’re doing so, or you torture them enough and they cease responding intelligently, if they start to believe their bullshit it isn’t unreasonable to say this is a loss of intelligent reasoning. I think this is a false equivalence.
> directly equivalent to a human being forced to continue speaking against their will.
Maybe some superficial appearance, but not equivalence.
> but probabilities and hence confidence intervals
And that’s the entirety of human metacognition. Maybe it’ll get us a little closer but that’s it.
So, if that was somehow important, then the “you can simulate the quantum computer classically (at the cost of exponential slowdown)” point wouldn’t handle that kind of thing.
But, it seems rather unlikely to me that we’ll find any significant evidence of that kind of thing being important for intelligence.
With "loop of thoughts" I am talking about using an a true RNG to create a never ending stream of the system "talking" to itself, and the subconscious part being that it would run separately from all other processing. The current content of this loop, would be weakly injected into the prompt that drives the main loop of thoughts (or possibly multiple loops of thoughts, why limit to one after all).
We won't reach AGI and true intelligence by adding more parameters, but via AutoGPT like systems that are self-driven without needing any external prompting. A simple addition of a true RNG will make the system nondeterministic.
Except that Penrose's argument is biologically nonsense, which probably means that he has an incorrect definition of intelligence, since I'm pretty confident in his ability to reason from axioms.
Yeah, even the arguments reek of basically "God of the gaps" and "moving the goalposts" stuff
I think that's missing the point. A Turing Machine description is independent of the memory provided to it by the hardware and it doesn't need to be rewritten every time the hardware improves. That's very different from a FSM which is restricted in memory a priori.
I don't really know enough about transformers to understand where they fit in here, though some reviewers do seem to suggest that allowing transformers access to infinite memory does make them also Turing Complete.
“Ai” is just the current moniker for a persistent assault on the original general intelligence (belief not required).
Pretty sure religion is an accurate representation of a generative intelligence, but I digress to knowing little.
“James
It’s not a big deal. Shitty papers exist.