There are, he conceded, aspects of the world ChatGPT is describing that it does not understand. But he rejected LeCun’s belief that you have to “act on” the world physically in order to understand it, which current AI models cannot do. (“That’s awfully tough on astrophysicists. They can’t act on black holes.”) Hinton thinks such reasoning quickly leads you towards what he has described as a “pre-scientific concept”: consciousness, an idea he can do without. “Understanding isn’t some kind of magic internal essence. It’s an updating of what it knows.”
In that sense, he thinks ChatGPT understands just as humans do. It absorbs data and adjusts its impression of the world. But there is nothing else going on, in man or machine.
“I believe in [the philosopher Ludwig] Wittgenstein’s position, which is that there is no ‘inner theatre’.” If you are asked to imagine a picnic on a sunny day, Hinton suggested, you do not see the picnic in an inner theatre inside your head; it is something conjured by your perceptual system in response to a demand for data. “There is no mental stuff as opposed to physical stuff.” There are “only nerve fibres coming in”. All we do is react to sensory input.
The difference between us and current AIs, Hinton thinks, is the range of input.
Thanks that’s super helpful and a super interesting interview, missed that.
I don’t know about anyone else but I’m getting “Don’t Look Up” vibes every day an expert comes out with an alarmist take. It’s like a conspiracy that’s published from the start but just ignored/downplayed by society at large. But maybe I’m just also an alarmist lol, a “thinking machines are here” at the very least.
> The book’s core argument is that an artificial intelligence that could equal or exceed human intelligence—sometimes called artificial general intelligence (AGI)—is for mathematical reasons impossible.
Mathematically impossible? Really? I wonder how much of this book has already been disproved by the existence of GPT-4. (One of the chapter titles is "Why machines will not master human language". LOL.)
GPT-4 can answer test questions, but could it have come up with what Einstein published in 1905 if it had all the information about the world up to say, 1902?
Is “can develop theories otherwise unknown to anyone but one of the most accomplished scientists of all time” really a logical bar to have to meet for “an artificial intelligence that could equal… human intelligence”?
Totally agree, great read. AFAICT the core argument is about how “complex systems” are complex, and machines are less complex than brains because brains have quantum magic. I feel like that’s a MASSIVE oversimplification of complexity theories (based on the very light touches I got in school years ago), kinda like papers about how the presence of “imaginary” numbers proves that there’s a multiverse or other terminology misapplications.
Plus this chapter list is funny:
Part III: The limits and potential of AI
9. Why there will be no machine intelligence
9.1 Brain emulation and machine evolution
9.2 Intentions and drivenness
9.3 Consciousness
9.4 Philosophy of mind, computation and AI
9.5 Objectifying intelligence and theoretical thinking
10. Why machines will not master human language
10.1 Language as a necessary condition for AGI
10.2 Why machine language production always falls short
10.3 AI conversation emulation
10.4 Mathematical models of human conversations
10.5 Why conversation machines are doomed to fail
11. Why machines will not master social interaction
11.1 No AI emulation of social behaviour
11.2 AI and legal norms
11.3 No machine emulation of morality
Of all those, I think _one_ subsection clearly relates to math?
I hope this isn’t too snarky, thanks to original commenter for posting it - I love looking back at stuff like this either way, and this post in particular brought me much joy. Sadly not gonna pick this one up lol
One of the authors is a philosopher (mostly realist, logic and ontologist) so that's a different angle than a math based book that's for sure. And that's also what makes it interesting compared to most other books that went out recently.
GPT is not understanding language we have many examples that show that.
It has some other qualities that make it good at interacting with humans using language, but it is never "understanding".
That's one part of the points they make in the book.
One of the authors made a video recently about exactly your point https://m.youtube.com/watch?v=s6DA4EcsLm8 This may be easier for you to grasp than the book that require careful reading (and understanding) of the introduction part where they define their premises.
As far as I know you can only get one of them under a microscope, which seems to support your view, but "mind-brain identity theory", positing that mind and brain are literally the same thing, is still a thing in philosophy though not quite the hot item it was 60 years ago.
It looks like an interesting book. Published in 1995, it is mostly focused on cognitivist / functional approaches in computational linguistics rather than the current ML paradigm, though there is a section on connectionism, a decentralized bottom up approach which favored the emergence of linguistic behaviour and reasoning from artificial neural networks over cognitivist ideas of implementing symbolic reasoning directly.
The authors agree with the connectionist critique of cognitivists like Fodor (correctly, and on good Wittgensteinian grounds, IMO) but then critique the positive views of the connectionists that "connectionist modelling will pave the way for a better understanding of brain-behaviour (or, more contentiously, brain-'mind') relationships" as "incoherent". I think that latter claim of incoherence is looking a bit tattered now in view of the undoubted practical linguistic interest of LMMs.
After locating the book at one of the popular public academic material utilities, my first instinct was to look for my my favorite anti-cognitivist anti-nativist quote from Philosophical Investigations about the operation of meaning (or intentionality, to use the technical term):
> "There must surely be a further, different connexion between my talk and N, for otherwise I should still not have meant HIM."
> Certainly such a connexion exists. Only not as you imagine it: namely by means of a mental mechanism.
(PI, §689)
This isn't directly quoted in the book, but they do say:
> one of the points that Wittgenstein is making with respect to rules and language is that there are no mechanisms of grammar, no calculus of determining rules, behind reasoning, behind language. Rules can be a guide to action, but they do not constrain it in the way in which the viewpoint nourished by computational linguists proposes.
17 comments
[ 3.1 ms ] story [ 54.1 ms ] threadhttps://www.newstatesman.com/long-reads/2023/06/men-made-fut...
There are, he conceded, aspects of the world ChatGPT is describing that it does not understand. But he rejected LeCun’s belief that you have to “act on” the world physically in order to understand it, which current AI models cannot do. (“That’s awfully tough on astrophysicists. They can’t act on black holes.”) Hinton thinks such reasoning quickly leads you towards what he has described as a “pre-scientific concept”: consciousness, an idea he can do without. “Understanding isn’t some kind of magic internal essence. It’s an updating of what it knows.”
In that sense, he thinks ChatGPT understands just as humans do. It absorbs data and adjusts its impression of the world. But there is nothing else going on, in man or machine.
“I believe in [the philosopher Ludwig] Wittgenstein’s position, which is that there is no ‘inner theatre’.” If you are asked to imagine a picnic on a sunny day, Hinton suggested, you do not see the picnic in an inner theatre inside your head; it is something conjured by your perceptual system in response to a demand for data. “There is no mental stuff as opposed to physical stuff.” There are “only nerve fibres coming in”. All we do is react to sensory input.
The difference between us and current AIs, Hinton thinks, is the range of input.
Further discussion here: https://marginalrevolution.com/marginalrevolution/2023/06/wh...
I don’t know about anyone else but I’m getting “Don’t Look Up” vibes every day an expert comes out with an alarmist take. It’s like a conspiracy that’s published from the start but just ignored/downplayed by society at large. But maybe I’m just also an alarmist lol, a “thinking machines are here” at the very least.
It is dense and opinionated but even if you may disagree at the end, it is absolutely worth a read.
Mathematically impossible? Really? I wonder how much of this book has already been disproved by the existence of GPT-4. (One of the chapter titles is "Why machines will not master human language". LOL.)
It’s also unknowable because of course we can’t collate that input data.
Plus this chapter list is funny:
Of all those, I think _one_ subsection clearly relates to math?I hope this isn’t too snarky, thanks to original commenter for posting it - I love looking back at stuff like this either way, and this post in particular brought me much joy. Sadly not gonna pick this one up lol
It has some other qualities that make it good at interacting with humans using language, but it is never "understanding".
That's one part of the points they make in the book.
One of the authors made a video recently about exactly your point https://m.youtube.com/watch?v=s6DA4EcsLm8 This may be easier for you to grasp than the book that require careful reading (and understanding) of the introduction part where they define their premises.
As far as I know you can only get one of them under a microscope, which seems to support your view, but "mind-brain identity theory", positing that mind and brain are literally the same thing, is still a thing in philosophy though not quite the hot item it was 60 years ago.
If you're curious, see for example: https://en.wikipedia.org/wiki/Type_physicalism or https://en.wikipedia.org/wiki/Anomalous_monism
The authors agree with the connectionist critique of cognitivists like Fodor (correctly, and on good Wittgensteinian grounds, IMO) but then critique the positive views of the connectionists that "connectionist modelling will pave the way for a better understanding of brain-behaviour (or, more contentiously, brain-'mind') relationships" as "incoherent". I think that latter claim of incoherence is looking a bit tattered now in view of the undoubted practical linguistic interest of LMMs.
After locating the book at one of the popular public academic material utilities, my first instinct was to look for my my favorite anti-cognitivist anti-nativist quote from Philosophical Investigations about the operation of meaning (or intentionality, to use the technical term):
> "There must surely be a further, different connexion between my talk and N, for otherwise I should still not have meant HIM."
> Certainly such a connexion exists. Only not as you imagine it: namely by means of a mental mechanism.
(PI, §689)
This isn't directly quoted in the book, but they do say:
> one of the points that Wittgenstein is making with respect to rules and language is that there are no mechanisms of grammar, no calculus of determining rules, behind reasoning, behind language. Rules can be a guide to action, but they do not constrain it in the way in which the viewpoint nourished by computational linguists proposes.