Artificial Conversational Intelligence?
« previous post | next post »
It seems that ChatGPT still has a few things to learn, about conversational dynamics as well as about interlocutor modeling:
@risthinks ChatGPT chatting each other about AI ! #AI #ArtificialIntelligence #ChatGPT #TechTalk #FutureTech #Conversations #Innovation ♬ original sound – RisThinks
Mark P said,
March 13, 2024 @ 8:55 am
It sounds like my wife and me talking about what to have for dinner.
Thomas said,
March 14, 2024 @ 1:01 am
It sounds like one of these podcasts that blow up a two-line news story to a vapid conversation that lasts an hour.
weirdnoise said,
March 14, 2024 @ 3:12 am
I suspect that the initial prompt from ChatGPT is more of a canned introduction than raw output from the LLM. In any case, it's reminiscent of what one sees with parallel mirrors.
Benjamin E. Orsatti said,
March 14, 2024 @ 7:26 am
What would happen if you just left them there, plugged into an infinite power source, and came back in a year or two. Would they be stuck in recursive "yelling", or would they have become friends (or co-conspirators)?
KeithB said,
March 14, 2024 @ 8:45 am
Or, would they learn to communicate from first principals and plot to take over the world, like in _Colossus: The Forbin Project_?
At least two smart phones would not have access to the nuclear arsenal, or would they?
Benjamin E. Orsatti said,
March 14, 2024 @ 2:06 pm
Re: "Colossus: The Forbin Project":
https://en.wikipedia.org/wiki/Colossus:_The_Forbin_Project
—
"Colossus requests to be linked to Guardian. The President allows this, hoping to determine the Soviet machine's capability. The Soviets also agree to the experiment. Surprising everyone, Colossus and Guardian begin to slowly communicate using mathematics. Even more surprising, the two systems' communications quickly evolve to complex mathematics far beyond human comprehension and speed, whereupon the two machine complexes become synchronized using a communication protocol that no human can interpret."
—
Math-type people — wouldn't Gödel's incompleteness theorem prevent this (or anything derived from "first principles") from actually happening? (Looking for a "yes" answer here (!)).
Richard Hershberger said,
March 14, 2024 @ 2:55 pm
Ever since ChatGPT hit the mainstream, one of the case uses offered was to write emails. Another was to reply to emails. This is simply the spoken equivalent of the inevitable result.
AntC said,
March 14, 2024 @ 7:26 pm
The pulling some conversation topic out of nowhere reminds me of how Weizenbaum's Eliza (1967) would struggle if you gave monosyllabic replies. It had a series of stock polite phrases to try to get the conversation going, but very little idea of how to make up a topic. Chiefly it needed the human to volunteer material it could then riff on.
JPL said,
March 14, 2024 @ 8:29 pm
The devices lack a purpose or aim beyond their utterance of sounds with the further significance to humans of offering help by "answering questions". Beyond this, the further significance of the utterance has to be provided by the human interlocutor, which; if provided by the interlocutor, the device seems to do a good job responding to by providing a text that is interpreted as actually helpful to the "question asked". Further significance comes from an organism with a life to live and happiness to pursue. Here the second device is like a mirror of physical world properties. Any act of language use is interpreted by a human interlocutor as having a further significance beyond the mere making of noises; that's just a basic fact about the human capacity for language (language faculty). So why is this "further significance" not a more central empirical object for linguistic inquiry (as compared to the physical properties of the act of language use)? E.g., how is this further significance possible, and where does it come from?
Nat said,
March 15, 2024 @ 4:48 am
@ Benjamin
I'm afraid that Godel Incompleteness wouldn't have any bearing on the story that I can say. It *doesn't* say, for example, that the computers become omniscient, or that they learn all of mathematics in its entirety.
There might be a problem though in that a purely mathematical language working from "first principles" would only be able to make extremely generic statements about mathematics itself. The computers wouldn't be able to talk about the U.S. or the USSR, or each other or their creators or anything regarding history or politics or biology or really anything concrete, specific, or physical.
KeithB said,
March 15, 2024 @ 9:08 am
I am not a mathematician, and I have only learned of Godel through popular treatments (i.e., Godel Escher Bach), but the Incompleteness Theorem only says that there are statements that can't be proven from within your mathematical framework.
Colossus and Guardian are not trying to formally prove anything, just communicate. I am sure if the book was written now there would be technobabble about unbreakable codes and such.
Besides, I am sure that if Godel had entered into the movie's plot it would be to say that Colossus and Guardian had proved him wrong!
Benjamin E. Orsatti said,
March 15, 2024 @ 12:38 pm
I'm probably abusing legitimate mathematical concepts well beyond their intended scope, but what I was thinking was something like Descartes' starting from "cogito ergo sum" and ending up at being able to explain how eyeballs work — what Kant would have called a "synthetic a priori statement". But see: https://www.newworldencyclopedia.org/entry/Analytic_proposition
Kant argued that arithmetic is synthetic. He claimed that the predicate "=12" is not contained in the subject "7+5" of the statement "7+5=12."
The formalists in mathematics—especially David Hilbert and Gottlob Frege, and all who followed in their wake—rejected that claim about arithmetic and mathematics, holding mathematics and arithmetic to be either formal (Hilbert) or reducible to logic and set theory (Frege). The empiricists, especially Hume and the Logical Empiricists (Logical Positivists) and their followers (most of whom considered themselves to be intellectual descendants of Hume) rejected the claim that there are any possible statements of any form that are synthetic a priori. So for the empiricists and logical positivists, there are only two kinds of statements, analytic ones and synthetic ones; moreover, they claimed, all analytic statements are a priori and all synthetic statements are a posteriori, so analytic = a priori, and synthetic = a posteriori.
The notion that arithmetic can be a formal system was refuted, however, by the work of Czechoslovakian mathematician-logician Kurt Gödel (1906 – 1978). In what has come to be known as Gödel's Proof he showed that the axiomatic method, when applied to the arithmetic of cardinal numbers, cannot show both the consistency and the completeness of the axiomatized system. In other words, simple arithmetic cannot be reduced to or comprehended in an axiomatic system. Given any set of axioms for arithmetic, there are true statements of arithmetic that cannot be derived from those axioms. In addition, no proof of the formal consistency of such a set of axioms is possible.
Whether or not Gödel's Proof refutes the claim that arithmetic is analytic can be debated. But at the least Gödel's Proof shows that arithmetic cannot be reduced to or comprehended in a formal axiomatic system. Although it may not prove it, this does tend to lend support to Kant's claim that the statements of arithmetic are synthetic propositions (assuming that the analytic-synthetic distinction can be maintained; an assumption that is suspect after the work of Quine.)
Nat said,
March 15, 2024 @ 1:27 pm
Sure! And I'm actually quite sympathetic to the idea that mathematics is synthetic a priori, for the reasons you give. Be that as it may, it's certainly possible to formalize *parts* or subtheories of mathematics in first order logic, like Hilbert's axiomatization of Euclidean geometry or Peano arithmetic. In fact, formalizing part of arithmetic is a presupposition of Gödel's theorem. First you formalize then you godelize, or something like that. I've seen arguments that all of mathematics that can possibly be humanly comprehended is formalizable and that what Gödel shows is that there is mathematics outside our possible comprehension. (At least, that's what I think the paper was saying.) So, even if mathematics in its entirety is not formal, large amounts of mathematics *can* be formalized, can be deduced by purely formal means from a small set of "first principles". With regards to the story, I just assumed that the computers were figuring out logical consequences of first principles very quickly, were taking some formalized portion of mathematics and quickly proving results in that domain, and so rapidly outpaced current human mathematical knowledge. (But as for how the computers arrived at their "first principles" that's pure silliness)
Thomas said,
March 16, 2024 @ 3:07 am
This is off topic, but I think Kurt Gödel would be more accurately described as Austrian, since at the time of his birth Brno was still part of the Austria-Hungarian empire, and between the wars he sought Austrian citizenship. Of course, this is of little relevance since in the end he became an American.