17 ms·

An argument for the impossibility of machine intelligence [pdf]

104 points by imaurer 5y ago

go_elmo 5y ago
The Turing machine was designed by imagining a human-operator. Our Mind has also only a finite state, and no matter if quantum effects are involved, the information in it is always finite, describable in a finite state. Thus, all turing machines are capable do do exactly what we do with information. This argument is incredible.
- Isinlor 5y ago
  This is Russell's teapot. You can not prove that infinite states do not exist. In fact, common models of physics assume possibility of infinite number of states by depending on formalism axiomatically assuming infinite sets (e.g. axiom of infinity in Zermelo–Fraenkel set theory).
Dr_Birdbrain 5y ago
I hope that it turns out that this paper was written by GPT-3 :)
R0b0t1 5y ago
Yet we are machines...? Speaking specifically of neural networks as they exist now the answer is no because there is no obvious way to learn.
- sgt101 5y ago
  Are machines all computation? Are all the processes of the physical universe computation?
  - mcguire 5y ago
    If you are a materialist, yes and yes. If not, all bets are off and there are no rules.
    - NineStarPoint 5y ago
      I don’t think being a materialist implies that. It’s entirely possible for matter/the fabric of the universe to have non-computable properties.
      - R0b0t1 5y ago
        Computers exist inside the universe so the universe must be able to compute things. Likewise you can look for certain hallmarks of information manipulation that mean you are computing something. Usually philosophers talking about these things either haven't read or are just discovering complexity theory.
        sgt101 5y ago
        >Computers exist inside the universe so the universe must be able to compute things. Yes - but there may be other mechanisms of translating information into outcomes that are possible in the universe as well. I don't know any though.
        R0b0t1 5y ago
        Yes, obviously, but those things are computation.
        sgt101 5y ago
        Ahh - no; some of them are uncomputable in the sense that a u.turing machine cannot calculate them. For example no-repeat tiling.
    - sgt101 5y ago
      but there are conceivable processes that are beyond computation, for example Busy Beavers - but could they manifest in the physical universe?
      - sgt101 5y ago
  - ChainOfFools 5y ago
    all of these discussions eventually reveal themselves to be special framing of the old Parmenides question about determinism, whether we live in a block universe where choice and change are illusions, and thought and being are the same. I am increasingly convinced he is right, and that arguments such as the OP (and Searle-ism generally) present end up refuting not the existence of artificial intelligence, but intelligence itself. "Artificial" smuggles in naturalistic fallacy and privileges the dualism hypothesis.
dsr_ 5y ago
This appears to be a series of arguments from incredulity. In particular, it is equally incredible that intelligent life should evolve from a single-cell organism. But we have that as a counter-argument. It is entirely reasonable to suspect that none of the current approaches will yield success, but claiming that no machine intelligences can possibly arise is... incredible.
- 9wzYQbTYsAIc 5y ago
  Agreed. The main claim being made is that “since AI is a logic system, and living humans are complex systems, AI cannot replicate human intelligence”. That claim rests on some unfounded, and implicit, assumptions. In particular, the author assumes that neural networks are not complex systems (and as an even deeper, implicit assumption, that no complex neural network could ever exist).
  - hnaccount_rng 5y ago
    No the assumption is that a logic system cannot be complex… well 2-SAT likes to have a word with the author I guess
    - abetusk 5y ago
      I'm not sure if it was intentional or not but 2-SAT is polynomial solvable [0] whereas 3-SAT is NP-Complete [1]. [0] https://en.wikipedia.org/wiki/2-satisfiability#Algorithms https://en.wikipedia.org/wiki/2-satisfiability#Algorithms [1] https://en.wikipedia.org/wiki/Boolean_satisfiability_problem#3-satisfiability https://en.wikipedia.org/wiki/Boolean_satisfiability_problem...
    - 9wzYQbTYsAIc 5y ago
      Indeed, that does appear to be another of the many assumptions made in the article.
- qsort 5y ago
  Agreed. I am also extremely skeptical of AI, but while the paper does a good job at highlighting the problems with AI, the eventual conclusion is not at all well-supported. There's an hidden assumption that complex systems cannot be modeled mathematically at all, but while that can be true right now, there is no fundamental reason why satisfactory models can't be produced at all.
jmull 5y ago
The paper is full-on nonsense. I’m surprised someone wasted their time writing it and you probably shouldn’t waste your time reading it. In the part I read it claims we can’t develop AI because we can’t accurately model full reality. There’s no argument about what the connection there is, it’s just stated. Kind of obviously, if we assume engaging with reality is necessary to develop intelligence, an artificial intelligence could do so in a similar way we non-artificial ones do, right?
- dane-pgp 5y ago
  I agree that it is nonsense. To save people the click, here, for example, is how the paper argues that a software system couldn't gain intelligence by simulating an evolutionary process: "But we neither know how to engineer the drive that is built into all animate complex systems, nor do we know how to mimic evolutionary pressure, which we do not understand and cannot model (outside highly artificial conditions such as a Petri dish). In fact, if we already knew how to emulate evolution, we would in any case not need to do this in order to create intelligent life, because the complexity level of intelligent life is lower than that of evolution."
  - simonh 5y ago
    They wrote another paper on this topic, the summary of which is that an AI capable of human level conversations is impossible because: "This is (1) because there are no traditional explicitly designed mathematical models that could be used as a starting point for creating such programs; and (2) because even the sorts of automated models generated by using machine learning, which have been used successfully in areas such as machine translation, cannot be extended to cope with human dialogue. If this is so, then we can conclude that a Turing machine also cannot possess AGI, because it fails to fulfil a necessary condition thereof." https://arxiv.org/abs/1906.05833 https://arxiv.org/abs/1906.05833 In other words it can't ever be done because we haven't done it yet. QED. How stuff like this gets to come out of U Buffalo is beyond me. At first I suspected it might have come out of a religious think tank, but no.
    - 5y ago
snek_case 5y ago
The most obvious counter-argument is that the amount of things we can do with AI keeps expanding. People were incredulous that computer chess programs could beat humans in the 1980s. Now they can beat us at basically any board game including Go, do image classification, and we have some early prototypes of self-driving cars. AI hasn't mastered common-sense reasoning yet. That's likely going to come last, but the amount of things AI can understand is set to only expand IMO.
- TheOtherHobbes 5y ago
  Define "understand." I think you may be confusing automated processing with communicable abstracted insight. If this isn't obvious consider the difference between producing an AI that can play chess, producing an AI that learns to play chess, and producing a research program that produces an AI that can play chess and summarises all the resulting developments and insights.
  - naasking 5y ago
    It's not clear whether there is a difference in kind between those behaviours rather than merely a difference in degree of complexity.
- The_rationalist 5y ago
  Well I am not defending the paper thesis but no it's time to realize that we are in a new AI winter where progress has stopped. Sure we can make accuracy progress on tasks that were underesearched before, moreover we do make extremely slow (and with increasingly diminishing returns) accuracy gains on core tasks. But the diminishing returns are diminishing fast to the point that progress in terms of applications has stopped for core AI tasks such as NLU. However there is still some hope as the vast majority of papers bring an innovation but almost never attempt to merge/synergize with other papers innovations. If human resources where allocated at merging the top 10 papers on a given task, I'm sure it would lead to a major accuracy improvement.
- ben_w 5y ago
  One thing I’ve long noticed is that “common sense” is analogous to a stopped clock, in that it’s its only correct when it happens to also be a different form of reasoning such as deductive, inductive, or abductive reasoning. Things called “common sense” but which are not also a different kind of knowledge are mere cultural shibboleths, and vary from wrong (fan death) to opinion (Shakespeare is good). The traditional examples of common sense knowledge given when introducing the topic of A.I. are sufficiently imprecise to only be true given further common sense interpretation. For example: “things fall when you let go of them” unless they’re buoyant, or they fly, or they’re already on the ground, or you were in free-fall when you let go — these exceptions won’t really surprise anyone, and yet it’s both more compact and more accurate to say Σf=ma, f_g=G(m_1)(m_2)/r^2 etc.
SubiculumCode 5y ago
and this is why arxiv is not the same as peer review.
- musicale 5y ago
  It doesn't seem to be that far off from other (usually unconvincing) philosophical arguments that ultimately boil down to axioms or definitions. For example, John Searle (of the infamous Chinese Room "argument"/paradox) posits that a simulation of a mind - however realistic and convincing it may be - is not the same as an actual mind any more than virtual reality is the same as actual reality. For example, you could have a chatbot that passes the Turing test but inside it is just smoke and mirrors (e.g. ML models.) Which is to say if we could somehow replace Searle with a robot whose appearance and behavior were indistinguishable from that of the actual Searle then it would only confirm what we already know. (Though it is fun to imagine the actual Searle secretly watching his robotic replacement in anger as it does things that he would never do but which are still completely convincing to his duped students, while faculty colleagues take a liking to robo-Searle in a way that they never did to the original.)
_aavaa_ 5y ago
I'd like to point the reader's attention to [1]. [1] https://arxiv.org/abs/1703.10987 https://arxiv.org/abs/1703.10987
- mcguire 5y ago
  "In recent years, a number of prominent computer scientists, along with academics in fields such as philosophy and physics, have lent credence to the notion that machines may one day become as large as humans. Many have further argued that machines could even come to exceed human size by a significant margin. However, there are at least seven distinct arguments that preclude this outcome. We show that it is not only implausible that machines will ever exceed human size, but in fact impossible."
- visarga 5y ago
  That was a good one.
- a-dub 5y ago
  a breath of fresh air on the topic!
- SyzygistSix 5y ago
  "Numerous entries in the Godzilla film series feature machines so large that they can crush portions of the Tokyo skyline with a single step (Honda, 1975)"
a-dub 5y ago
> But we neither know how to engineer the drive that is built into all animate complex systems, nor do we know how to mimic evolutionary pressure, which we do not under- stand and cannot model (outside highly artificial conditions such as a Petri dish). In fact, if we already knew how to emulate evolution, we would in any case not need to do this in order to create intelligent life, because the complexity level of intelligent life is lower than that of evolution. This means that emulating intelligence would be much easier than emulating evolution en bloc. Chalmers is, therefore, wrong. We cannot engineer the conditions for a spontaneous evolution of intelligence. this is the thing i've always sort of loved about philosophy. they just kinda make shit up, provide their own definitions that are rooted in a bamboozling by use of flowery language, and then once they've stated all their definitions with their conclusions baked in, they hop, skip and jump down the path which now obviously leads to the conclusion they started with. it's kind of like a form of mathematics where they define their own first principles in each argument with the express purpose of trying to build the most beautiful path to their conclusions. it really is a beautiful form of art, like architecture for ideas.
- Rd6n6 5y ago
  You can’t pick your favourite bad argument and ridicule an entire field. You are incidentally using ideas from several different old, influential philosophies to even formulate Luther comment Philosophy includes questions like “how do we decide whether something is true or trustworthy,” or “what constitutes a good or a bad way to make a case for something.” If you’re going to throw philosophy out, you can’t question anything any more
  - imbnwa 5y ago
    Science used to be called 'natural philosophy'
  - threatofrain 5y ago
    Philosophy may refer to the specific branch in academia and its current practice, as opposed to any philosophical inquiry. Every field already pursues their own philosophical inquiry, and yet philosophers and mathematicians are in separate departments. Such is the current practice and organization of academics. If we were to consider mathematics and computer science as part of philosophy, then we might say that as a mode of inquiry, philosophy has had great success in achieving multidisciplinary consensus and international impact. But if we were to consider philosophy as a specific branch of academic organization, then we might be disappointed at the fruits emerging from that field.
Traubenfuchs 5y ago
Should we ever attain hardware, software and understanding of the human brain good enough to emulate a human brain, we have done it. There is absolutely no reason why this shouldn‘t be possible. Actually, we could already do it if we understood the brain enough and could model it good enough, even if the emulation might not be real time.
visarga 5y ago
What a funny a priori paper. Maybe the authors lost a bet and had to write it.
mcguire 5y ago
"Though the infinitesimal definition of utility in (1) and the penalisation of complexity in the definition of Υ provide a statistically robust measure of the kind of surrogate intelligence those working in the general artificial intelligence (AGI) field have decided to focus on, the definition is too weak to describe or specify the behaviour even of an arthropod. This is not only obvious from the issues already mentioned above, but also from the fact that algorithms which realise the reward-schemes proposed in (1) and (2) (for example, neural networks optimised with reinforcement learning) fail to display the type of generalisable adaptive behaviour to natural environments that arthropods are capable of, for example when ants or termites colonise a house." Ok, I don't like the mathematical definitions of intelligence either (although I might be convincable and they do have some advantages over other definitions I've seen), but this refutation seems to be a prime example of proof-by-assertion. "Brooks defines an AI agent, again, as an artefact that is able ‘to move around in dynamic environments, sensing the surroundings to a degree sufficient to achieve the necessary maintenance of life and reproduction'." And this definition implies many things we know to be intelligent (i.e. people) are not. So there's that. "There are three additional properties of logic systems of importance for our argument here: 1. Their phase space is fixed. 2. Their behaviour is ergodic with regard to their main functional properties. 3. Their behavior is to a large extent context-independent." Aaaaand here we go... "As we learn from we standard mathematical theory of complex systems [23], all such systems, including the systems of complex systems resulting from their interaction, 1. have a variable phase space, 2. are non-ergodic, and 3. are context-dependent." Ok, to the extent that the first statement is true about "logic systems", it is also true about any physically realizable, material system. On the other hand, the "complex system", to that same extent, is not physically realizable. (Consider "a variable phase space means that the variables which define the elements of a complex system can change over time" or "a non-ergodic system produces erratic distributions of its elements. No matter how long the system is observed, no laws can be deduced from observing its elements." and question how much information is required for this in the authors' sense.) And there we have the intrusion of the immortal soul into the argument that artificial intelligence is impossible.
erdewit 5y ago
In the same vein that heavier-than-air flying machines are impossible.
tehchromic 5y ago
It's not likely to be a popular opinion with technologists as AI's potential has lit the technopopular imagination, however this question has bothered me for a long time. I think strong emergent AI suffers philosophical problem that won't go away, and to the extent that the conversation revolves around evolution and consciousness rather than logic and intelligence, then we are having the right conversation. I'll put my argument out there and let the flames come as they will. Strong AI is about as likely to emerge from our current state of the art AI machinery as it is to emerge suddenly out of moon rocks. That's to say the fear of machines becoming self-conscious and posing an existential threat to us, especially replacing us in the evolutionarily sense, is completely unfounded. This isn't to say that building machines capable of doing exactly that isn't possible - we and all living things are proof that it's possible - it's to say that achieving this level of engineering is on par with intergalactic mass transit or Dyson spheres - way out of our league for the foreseeable. And, even if we had the technology, it would be so entirely foolish to undertake that no sentient species would do it. That said, there's a substantial argument to make that we will augment ourselves with our own machinery so throughoughly that we will become unrecognizable and in effect, accomplish the same task through merging with the machine. This is likely, but not at all to be like the experience of the singularity in that all of humanity is suddenly arrested and deposed by autonomous AI. An interesting scenario in this vein is if a few powerful individuals can wield autonomous systems, modify themselves and simply wipe out all the competition, then in effect the rest of us wouldn't know the difference. This outcome is actually I think on the more likely side, albeit a good ways away in the future. Less likely but still totally legitimate as a concern is the idea that AI could be very easily weaponized. This is a real problem and is I think behind the more substantive warnings by good thinkers on the topic. Like bioweapons, we might be wiped out by an machine that's been intentionally programmed and mechanically empowered to cause real harm. This kind of danger could also be emergent, in that a machine might be capable of deciding that it ought to take certain actions as well as have the capacity to take them, and then, voila, mass murder. However it seems unlikely that such a mistake would be made, or that a bad actor would be capable to commit such an intentional crime. I think this is on par with nuclear MAD: even total madmen dictators hit the pause on the push-the-button instinct. And an AI MAD or similar would surely take as much resource to produce as a nuke arsenal. In other words, the resources required to build such a machinery are on the order of a nation-state, and perhaps more complicated to achieve than a nuclear arsenal, so probably more likely to be stopped or fail in-process rather than succeed. So there are dangers from AI but I would say they are lesser than the accumulated danger of industrial society rendering they planet uninhabitable, which should of course occupy our primary concern these days. The idea that the biological evolutionary 'machine' whose motive for existence is accumulated over billions of years of entropic adaptation can be out engineered, or accidently replicated by modern computational AI is silly - the two aren't in the same league and it's hubris to suppose otherwise. There's more intelligence in the toe of a lady bug than in an the computing power ever made. In sum the danger from emergent AI is overstated, however the concern is most welcome to the extent that it informs wisdom and care in consideration for our techno-industrial impact on the biosphere.
natch 5y ago
“The authors declare that they have no conflict of interest.” “Department of Philosophy” hmm
doganulus 5y ago
Their premises about logical systems are wrong so their conclusion is not valid. In short, of course, there are logical systems with potentially infinite state space. For example, a Turing machine. A digital circuit is no different. Turing completeness is abundant, it is everywhere.
mensetmanusman 5y ago
If it takes longer than the heat death of the universe to understand intelligence, does that mean it’s impossible?
Borrible 5y ago
I take for granted, the world exists, therefore it is. You may call it Borrible's first tautology. Or perhaps bias. Yes, Borrible' bias sounds clever. At least to me. And that is what counts, doesn't it? I don't really know what that fucking world really is, but nonetheless it exists. With temporarily stable local dynamics, some parts of the world began to copy themselves. Albeit with errors and quirks. The recurring processes of the surrounding builded the mold for the debris that collects in the swirls. Some of those copies developed representations of their surroundings. First in form of simple notes sticking on themselves, being themselves. Which was an advantage, when they bumped into another. They could navigate that thing I called world. Which made their copy process stable. With a lot of time and try and error, some parts of those parts of parts of that thing I called world even developed some really fancy little dollhouse worlds in this part of the world that will later call itself the brain. And the most advanced ham actors in that dollhouse put more tiny little dolls in that house, the most precious one, ego. It represented that part of the world that started the whole shebang, the body. And it equipped that tiny little dollhouse with a lot of woundrous and a lot of silly things, some animated , some not. And it took great delight in it, it even fancied itself a god and pushed the tiny little ego around doing his biddings. But for the most part it just tried to please itself and learn about the world and itself, based on all that input it got somehow from the world. And the drama that ham actor and his Muppet Friends acted. Exactly like all those good little boys and girls do on their playgrounds since time immemorable. When I was young, something happend. My dolls started to become 'Little Computer People'. And people my generation and that before developed fancy models about this Matrioshka Doll World, about Worlds in Worlds in World in Worlds. Infinite regress, sometimes recursive, sometimes not. A calaidoscopic mirror, sometimes dark, sometimes shiny. Simulacron 1, 2, 3 and so on until there is no energetic process in that thing I call world, that can be harvested. And every time those models became more complex, they gave more agency to that part of the world that is now mumbling about building a new Ghost in the Machine. Apart from the possibly insurmountable practical problems, I see no reason in principle why it should become more complex in the form of artificial intelligence. As an aside, it's great to be that part of the world, but beware. It may all end the moment that ham actor in that dollhouse cuts the strings to that world he is living of. A risk deeply embedded in this structure. Of an agent acting in a model of the world. The agent is subject to the risk of his striving to make himself independent of the world.

104 points by imaurer 5y ago

The Turing machine was designed by imagining a human-operator. Our Mind has also only a finite state, and no matter if quantum effects are involved, the information in it is always finite, describable in a finite state. Thus, all turing machines are capable do do exactly what we do with information. This argument is incredible.

Isinlor 5y ago
This is Russell's teapot. You can not prove that infinite states do not exist. In fact, common models of physics assume possibility of infinite number of states by depending on formalism axiomatically assuming infinite sets (e.g. axiom of infinity in Zermelo–Fraenkel set theory).

Dr_Birdbrain 5y ago

I hope that it turns out that this paper was written by GPT-3 :)

R0b0t1 5y ago

Yet we are machines...? Speaking specifically of neural networks as they exist now the answer is no because there is no obvious way to learn.

sgt101 5y ago
Are machines all computation? Are all the processes of the physical universe computation?
- mcguire 5y ago
  If you are a materialist, yes and yes. If not, all bets are off and there are no rules.
  - NineStarPoint 5y ago
    I don’t think being a materialist implies that. It’s entirely possible for matter/the fabric of the universe to have non-computable properties.
    - R0b0t1 5y ago
      Computers exist inside the universe so the universe must be able to compute things. Likewise you can look for certain hallmarks of information manipulation that mean you are computing something. Usually philosophers talking about these things either haven't read or are just discovering complexity theory.
      - sgt101 5y ago
        >Computers exist inside the universe so the universe must be able to compute things. Yes - but there may be other mechanisms of translating information into outcomes that are possible in the universe as well. I don't know any though.
        R0b0t1 5y ago
        Yes, obviously, but those things are computation.
        sgt101 5y ago
        Ahh - no; some of them are uncomputable in the sense that a u.turing machine cannot calculate them. For example no-repeat tiling.
  - sgt101 5y ago
    but there are conceivable processes that are beyond computation, for example Busy Beavers - but could they manifest in the physical universe?
    - sgt101 5y ago
- ChainOfFools 5y ago
  all of these discussions eventually reveal themselves to be special framing of the old Parmenides question about determinism, whether we live in a block universe where choice and change are illusions, and thought and being are the same. I am increasingly convinced he is right, and that arguments such as the OP (and Searle-ism generally) present end up refuting not the existence of artificial intelligence, but intelligence itself. "Artificial" smuggles in naturalistic fallacy and privileges the dualism hypothesis.

dsr_ 5y ago

This appears to be a series of arguments from incredulity. In particular, it is equally incredible that intelligent life should evolve from a single-cell organism. But we have that as a counter-argument. It is entirely reasonable to suspect that none of the current approaches will yield success, but claiming that no machine intelligences can possibly arise is... incredible.

9wzYQbTYsAIc 5y ago
Agreed. The main claim being made is that “since AI is a logic system, and living humans are complex systems, AI cannot replicate human intelligence”. That claim rests on some unfounded, and implicit, assumptions. In particular, the author assumes that neural networks are not complex systems (and as an even deeper, implicit assumption, that no complex neural network could ever exist).
- hnaccount_rng 5y ago
  No the assumption is that a logic system cannot be complex… well 2-SAT likes to have a word with the author I guess
  - abetusk 5y ago
    I'm not sure if it was intentional or not but 2-SAT is polynomial solvable [0] whereas 3-SAT is NP-Complete [1]. [0] https://en.wikipedia.org/wiki/2-satisfiability#Algorithms https://en.wikipedia.org/wiki/2-satisfiability#Algorithms [1] https://en.wikipedia.org/wiki/Boolean_satisfiability_problem#3-satisfiability https://en.wikipedia.org/wiki/Boolean_satisfiability_problem...
  - 9wzYQbTYsAIc 5y ago
    Indeed, that does appear to be another of the many assumptions made in the article.
qsort 5y ago
Agreed. I am also extremely skeptical of AI, but while the paper does a good job at highlighting the problems with AI, the eventual conclusion is not at all well-supported. There's an hidden assumption that complex systems cannot be modeled mathematically at all, but while that can be true right now, there is no fundamental reason why satisfactory models can't be produced at all.

jmull 5y ago

The paper is full-on nonsense. I’m surprised someone wasted their time writing it and you probably shouldn’t waste your time reading it. In the part I read it claims we can’t develop AI because we can’t accurately model full reality. There’s no argument about what the connection there is, it’s just stated. Kind of obviously, if we assume engaging with reality is necessary to develop intelligence, an artificial intelligence could do so in a similar way we non-artificial ones do, right?

dane-pgp 5y ago
I agree that it is nonsense. To save people the click, here, for example, is how the paper argues that a software system couldn't gain intelligence by simulating an evolutionary process: "But we neither know how to engineer the drive that is built into all animate complex systems, nor do we know how to mimic evolutionary pressure, which we do not understand and cannot model (outside highly artificial conditions such as a Petri dish). In fact, if we already knew how to emulate evolution, we would in any case not need to do this in order to create intelligent life, because the complexity level of intelligent life is lower than that of evolution."
- simonh 5y ago
  They wrote another paper on this topic, the summary of which is that an AI capable of human level conversations is impossible because: "This is (1) because there are no traditional explicitly designed mathematical models that could be used as a starting point for creating such programs; and (2) because even the sorts of automated models generated by using machine learning, which have been used successfully in areas such as machine translation, cannot be extended to cope with human dialogue. If this is so, then we can conclude that a Turing machine also cannot possess AGI, because it fails to fulfil a necessary condition thereof." https://arxiv.org/abs/1906.05833 https://arxiv.org/abs/1906.05833 In other words it can't ever be done because we haven't done it yet. QED. How stuff like this gets to come out of U Buffalo is beyond me. At first I suspected it might have come out of a religious think tank, but no.
  - 5y ago

snek_case 5y ago

The most obvious counter-argument is that the amount of things we can do with AI keeps expanding. People were incredulous that computer chess programs could beat humans in the 1980s. Now they can beat us at basically any board game including Go, do image classification, and we have some early prototypes of self-driving cars. AI hasn't mastered common-sense reasoning yet. That's likely going to come last, but the amount of things AI can understand is set to only expand IMO.

TheOtherHobbes 5y ago
Define "understand." I think you may be confusing automated processing with communicable abstracted insight. If this isn't obvious consider the difference between producing an AI that can play chess, producing an AI that learns to play chess, and producing a research program that produces an AI that can play chess and summarises all the resulting developments and insights.
- naasking 5y ago
  It's not clear whether there is a difference in kind between those behaviours rather than merely a difference in degree of complexity.
The_rationalist 5y ago
Well I am not defending the paper thesis but no it's time to realize that we are in a new AI winter where progress has stopped. Sure we can make accuracy progress on tasks that were underesearched before, moreover we do make extremely slow (and with increasingly diminishing returns) accuracy gains on core tasks. But the diminishing returns are diminishing fast to the point that progress in terms of applications has stopped for core AI tasks such as NLU. However there is still some hope as the vast majority of papers bring an innovation but almost never attempt to merge/synergize with other papers innovations. If human resources where allocated at merging the top 10 papers on a given task, I'm sure it would lead to a major accuracy improvement.
ben_w 5y ago
One thing I’ve long noticed is that “common sense” is analogous to a stopped clock, in that it’s its only correct when it happens to also be a different form of reasoning such as deductive, inductive, or abductive reasoning. Things called “common sense” but which are not also a different kind of knowledge are mere cultural shibboleths, and vary from wrong (fan death) to opinion (Shakespeare is good). The traditional examples of common sense knowledge given when introducing the topic of A.I. are sufficiently imprecise to only be true given further common sense interpretation. For example: “things fall when you let go of them” unless they’re buoyant, or they fly, or they’re already on the ground, or you were in free-fall when you let go — these exceptions won’t really surprise anyone, and yet it’s both more compact and more accurate to say Σf=ma, f_g=G(m_1)(m_2)/r^2 etc.

SubiculumCode 5y ago

and this is why arxiv is not the same as peer review.

musicale 5y ago
It doesn't seem to be that far off from other (usually unconvincing) philosophical arguments that ultimately boil down to axioms or definitions. For example, John Searle (of the infamous Chinese Room "argument"/paradox) posits that a simulation of a mind - however realistic and convincing it may be - is not the same as an actual mind any more than virtual reality is the same as actual reality. For example, you could have a chatbot that passes the Turing test but inside it is just smoke and mirrors (e.g. ML models.) Which is to say if we could somehow replace Searle with a robot whose appearance and behavior were indistinguishable from that of the actual Searle then it would only confirm what we already know. (Though it is fun to imagine the actual Searle secretly watching his robotic replacement in anger as it does things that he would never do but which are still completely convincing to his duped students, while faculty colleagues take a liking to robo-Searle in a way that they never did to the original.)

_aavaa_ 5y ago

I'd like to point the reader's attention to [1]. [1] https://arxiv.org/abs/1703.10987 https://arxiv.org/abs/1703.10987

mcguire 5y ago
"In recent years, a number of prominent computer scientists, along with academics in fields such as philosophy and physics, have lent credence to the notion that machines may one day become as large as humans. Many have further argued that machines could even come to exceed human size by a significant margin. However, there are at least seven distinct arguments that preclude this outcome. We show that it is not only implausible that machines will ever exceed human size, but in fact impossible."
visarga 5y ago
That was a good one.
a-dub 5y ago
a breath of fresh air on the topic!
SyzygistSix 5y ago
"Numerous entries in the Godzilla film series feature machines so large that they can crush portions of the Tokyo skyline with a single step (Honda, 1975)"

a-dub 5y ago

> But we neither know how to engineer the drive that is built into all animate complex systems, nor do we know how to mimic evolutionary pressure, which we do not under- stand and cannot model (outside highly artificial conditions such as a Petri dish). In fact, if we already knew how to emulate evolution, we would in any case not need to do this in order to create intelligent life, because the complexity level of intelligent life is lower than that of evolution. This means that emulating intelligence would be much easier than emulating evolution en bloc. Chalmers is, therefore, wrong. We cannot engineer the conditions for a spontaneous evolution of intelligence. this is the thing i've always sort of loved about philosophy. they just kinda make shit up, provide their own definitions that are rooted in a bamboozling by use of flowery language, and then once they've stated all their definitions with their conclusions baked in, they hop, skip and jump down the path which now obviously leads to the conclusion they started with. it's kind of like a form of mathematics where they define their own first principles in each argument with the express purpose of trying to build the most beautiful path to their conclusions. it really is a beautiful form of art, like architecture for ideas.

Rd6n6 5y ago
You can’t pick your favourite bad argument and ridicule an entire field. You are incidentally using ideas from several different old, influential philosophies to even formulate Luther comment Philosophy includes questions like “how do we decide whether something is true or trustworthy,” or “what constitutes a good or a bad way to make a case for something.” If you’re going to throw philosophy out, you can’t question anything any more
- imbnwa 5y ago
  Science used to be called 'natural philosophy'
- threatofrain 5y ago
  Philosophy may refer to the specific branch in academia and its current practice, as opposed to any philosophical inquiry. Every field already pursues their own philosophical inquiry, and yet philosophers and mathematicians are in separate departments. Such is the current practice and organization of academics. If we were to consider mathematics and computer science as part of philosophy, then we might say that as a mode of inquiry, philosophy has had great success in achieving multidisciplinary consensus and international impact. But if we were to consider philosophy as a specific branch of academic organization, then we might be disappointed at the fruits emerging from that field.

Traubenfuchs 5y ago

Should we ever attain hardware, software and understanding of the human brain good enough to emulate a human brain, we have done it. There is absolutely no reason why this shouldn‘t be possible. Actually, we could already do it if we understood the brain enough and could model it good enough, even if the emulation might not be real time.

visarga 5y ago

What a funny a priori paper. Maybe the authors lost a bet and had to write it.

mcguire 5y ago

"Though the infinitesimal definition of utility in (1) and the penalisation of complexity in the definition of Υ provide a statistically robust measure of the kind of surrogate intelligence those working in the general artificial intelligence (AGI) field have decided to focus on, the definition is too weak to describe or specify the behaviour even of an arthropod. This is not only obvious from the issues already mentioned above, but also from the fact that algorithms which realise the reward-schemes proposed in (1) and (2) (for example, neural networks optimised with reinforcement learning) fail to display the type of generalisable adaptive behaviour to natural environments that arthropods are capable of, for example when ants or termites colonise a house." Ok, I don't like the mathematical definitions of intelligence either (although I might be convincable and they do have some advantages over other definitions I've seen), but this refutation seems to be a prime example of proof-by-assertion. "Brooks defines an AI agent, again, as an artefact that is able ‘to move around in dynamic environments, sensing the surroundings to a degree sufficient to achieve the necessary maintenance of life and reproduction'." And this definition implies many things we know to be intelligent (i.e. people) are not. So there's that. "There are three additional properties of logic systems of importance for our argument here: 1. Their phase space is fixed. 2. Their behaviour is ergodic with regard to their main functional properties. 3. Their behavior is to a large extent context-independent." Aaaaand here we go... "As we learn from we standard mathematical theory of complex systems [23], all such systems, including the systems of complex systems resulting from their interaction, 1. have a variable phase space, 2. are non-ergodic, and 3. are context-dependent." Ok, to the extent that the first statement is true about "logic systems", it is also true about any physically realizable, material system. On the other hand, the "complex system", to that same extent, is not physically realizable. (Consider "a variable phase space means that the variables which define the elements of a complex system can change over time" or "a non-ergodic system produces erratic distributions of its elements. No matter how long the system is observed, no laws can be deduced from observing its elements." and question how much information is required for this in the authors' sense.) And there we have the intrusion of the immortal soul into the argument that artificial intelligence is impossible.

erdewit 5y ago

In the same vein that heavier-than-air flying machines are impossible.

tehchromic 5y ago

It's not likely to be a popular opinion with technologists as AI's potential has lit the technopopular imagination, however this question has bothered me for a long time. I think strong emergent AI suffers philosophical problem that won't go away, and to the extent that the conversation revolves around evolution and consciousness rather than logic and intelligence, then we are having the right conversation. I'll put my argument out there and let the flames come as they will. Strong AI is about as likely to emerge from our current state of the art AI machinery as it is to emerge suddenly out of moon rocks. That's to say the fear of machines becoming self-conscious and posing an existential threat to us, especially replacing us in the evolutionarily sense, is completely unfounded. This isn't to say that building machines capable of doing exactly that isn't possible - we and all living things are proof that it's possible - it's to say that achieving this level of engineering is on par with intergalactic mass transit or Dyson spheres - way out of our league for the foreseeable. And, even if we had the technology, it would be so entirely foolish to undertake that no sentient species would do it. That said, there's a substantial argument to make that we will augment ourselves with our own machinery so throughoughly that we will become unrecognizable and in effect, accomplish the same task through merging with the machine. This is likely, but not at all to be like the experience of the singularity in that all of humanity is suddenly arrested and deposed by autonomous AI. An interesting scenario in this vein is if a few powerful individuals can wield autonomous systems, modify themselves and simply wipe out all the competition, then in effect the rest of us wouldn't know the difference. This outcome is actually I think on the more likely side, albeit a good ways away in the future. Less likely but still totally legitimate as a concern is the idea that AI could be very easily weaponized. This is a real problem and is I think behind the more substantive warnings by good thinkers on the topic. Like bioweapons, we might be wiped out by an machine that's been intentionally programmed and mechanically empowered to cause real harm. This kind of danger could also be emergent, in that a machine might be capable of deciding that it ought to take certain actions as well as have the capacity to take them, and then, voila, mass murder. However it seems unlikely that such a mistake would be made, or that a bad actor would be capable to commit such an intentional crime. I think this is on par with nuclear MAD: even total madmen dictators hit the pause on the push-the-button instinct. And an AI MAD or similar would surely take as much resource to produce as a nuke arsenal. In other words, the resources required to build such a machinery are on the order of a nation-state, and perhaps more complicated to achieve than a nuclear arsenal, so probably more likely to be stopped or fail in-process rather than succeed. So there are dangers from AI but I would say they are lesser than the accumulated danger of industrial society rendering they planet uninhabitable, which should of course occupy our primary concern these days. The idea that the biological evolutionary 'machine' whose motive for existence is accumulated over billions of years of entropic adaptation can be out engineered, or accidently replicated by modern computational AI is silly - the two aren't in the same league and it's hubris to suppose otherwise. There's more intelligence in the toe of a lady bug than in an the computing power ever made. In sum the danger from emergent AI is overstated, however the concern is most welcome to the extent that it informs wisdom and care in consideration for our techno-industrial impact on the biosphere.

natch 5y ago

“The authors declare that they have no conflict of interest.” “Department of Philosophy” hmm

doganulus 5y ago

Their premises about logical systems are wrong so their conclusion is not valid. In short, of course, there are logical systems with potentially infinite state space. For example, a Turing machine. A digital circuit is no different. Turing completeness is abundant, it is everywhere.

mensetmanusman 5y ago

If it takes longer than the heat death of the universe to understand intelligence, does that mean it’s impossible?

Borrible 5y ago

I take for granted, the world exists, therefore it is. You may call it Borrible's first tautology. Or perhaps bias. Yes, Borrible' bias sounds clever. At least to me. And that is what counts, doesn't it? I don't really know what that fucking world really is, but nonetheless it exists. With temporarily stable local dynamics, some parts of the world began to copy themselves. Albeit with errors and quirks. The recurring processes of the surrounding builded the mold for the debris that collects in the swirls. Some of those copies developed representations of their surroundings. First in form of simple notes sticking on themselves, being themselves. Which was an advantage, when they bumped into another. They could navigate that thing I called world. Which made their copy process stable. With a lot of time and try and error, some parts of those parts of parts of that thing I called world even developed some really fancy little dollhouse worlds in this part of the world that will later call itself the brain. And the most advanced ham actors in that dollhouse put more tiny little dolls in that house, the most precious one, ego. It represented that part of the world that started the whole shebang, the body. And it equipped that tiny little dollhouse with a lot of woundrous and a lot of silly things, some animated , some not. And it took great delight in it, it even fancied itself a god and pushed the tiny little ego around doing his biddings. But for the most part it just tried to please itself and learn about the world and itself, based on all that input it got somehow from the world. And the drama that ham actor and his Muppet Friends acted. Exactly like all those good little boys and girls do on their playgrounds since time immemorable. When I was young, something happend. My dolls started to become 'Little Computer People'. And people my generation and that before developed fancy models about this Matrioshka Doll World, about Worlds in Worlds in World in Worlds. Infinite regress, sometimes recursive, sometimes not. A calaidoscopic mirror, sometimes dark, sometimes shiny. Simulacron 1, 2, 3 and so on until there is no energetic process in that thing I call world, that can be harvested. And every time those models became more complex, they gave more agency to that part of the world that is now mumbling about building a new Ghost in the Machine. Apart from the possibly insurmountable practical problems, I see no reason in principle why it should become more complex in the form of artificial intelligence. As an aside, it's great to be that part of the world, but beware. It may all end the moment that ham actor in that dollhouse cuts the strings to that world he is living of. A risk deeply embedded in this structure. Of an agent acting in a model of the world. The agent is subject to the risk of his striving to make himself independent of the world.