Everything hinges on his belief: "I believe that it is impossible for human beings to create a human-equivalent artificial intelligence."
I agree that it's difficult to give formal specifications for a "human-equivalent artificial intelligence," but assuming the human brain is an entirely physical object, why shouldn't we be able to at least theoretically create or simulate a computer system with as much complexity and the same functionality?
I agree with you. Also, I don't believe humans aren't limited by the same issues that machines have. Halting problem comes to mind, nothing shows that humans aren't bound by it too. I believe (another belief ;p ) that theoretically machines can be as intelligent as human considering that were both bound by the same limits. But I'm no AI expert as my post will probably prove. IANAAIE
not sure about hypercomputation but Godel did try to prove that humans had intuition something that would allow them to somehow go beyond the limitation of his incompleteness theorem. Of course he did go crazy thinking about all this and he never did finish his proof about intuition. I personally don't believe it's possible. Humans are as limited as machines.
> I don't believe humans aren't limited by the same issues that machines have. Halting problem comes to mind, nothing shows that humans aren't bound by it too.
Math has started, quite a while ago, to show that we are already probing some sort of inherent limits of knowledge for the human mind.
> theoretically machines can be as intelligent as human
Intelligence is the easy problem. It's just a function of complexity after all. It's consciousness / awareness that's harder, more important and more interesting.
Deeper than that, everything hinges on defining intelligence as what is solvable/not solvable in theory given infinite resources (time etc). So any process that can blindly enumerate all solutions to a problem (no matter how often some wrong solutions might be repeated) is just as "intelligent" as if it could solve the problem in constant or log* or whatever time. This has the interesting result of us being by definition intelligently designed, since the random mutation / natural selection process is maximally intelligent (can result in solutions (possibly through an unbounded number of intermediates such as humans) to any problem that has a solution).
Clearly, a practical definition of intelligence needs to include time. Otherwise, any problem whose solution can be described as a number and quickly checked could eventually be solved by a counter.
Even if that were true, we could still see rather drastic advances emulating intelligence far below the capability of humans.
Imagine if we could program a self replicating machine group similar to an ant colony? Self propagating, powering, healing. Any excess resources the colony has goes into building our houses, growing our food, building cars etc...
'Real AI' may require new hardware like quantum computers, and may take a very long time, but even 400 years ago, did anyone expect us to be sending small robotic vehicles around and out of the solar system?
I do believe that real AI will happen, but it may seem very alien to us.
Searle's 'experiment' is silly. Searle is trying to make readers empathize with the 'person in the box'. Let's concede, for the moment, that the person in the box doesn't understand Chinese in any meaningful way.
That says nothing about whether or not the SYSTEM on a WHOLE understands Chinese. Given the constraints of Searle's example (that every input produces a 'correct' output, but does so via this labour-intensive process), it seems there IS intelligence going on there at the SYSTEM level.
Ultimately, the point of that story is saying 'the cpu isn't smart'. Fine, that's not the argument behind AI/PI. And the whole 'there's nothing going on that isn't in the person already' also misses the point: what if, instead of a slow human translating, we had a fast human doing the lookups at the magnitude of a trillion lookup operations per second? They'd be able to do substantially more 'translating' work than the author of the software. So now, say that we program 'learning' instead of 'translation'. Sweet, now the computer can't learn any BETTER than the human, but it can learn FASTER, and there may be some substantial benefit to that.
His own argument is self-refuting: "in fact it was doing exactly what I had told it to do and nothing which I couldn't have done myself longhand given the time." Right. But you're limited by death, and a computer isn't limited in the same way. So say AI never makes it to 'smarter than human given enough time'. Fine. If it does in 30 seconds what it would take me 17,000 years to complete, I'm pretty happy.
The FAR more interesting question (to me) and probably one that can't be solved until we 'get there' is this: will computer intelligence be subject to the same foibles and hangups as a human? So, for instance, could "Google AI" respond to my search request by saying, "Wow, what a boring question, stop bothering me."
Since human motivations and preferences are caused by evolutionary forces, I don't see how an AI would acquire such things, aside from what we build into it. And if we wanted to build motivation into it, it shouldn't be difficult, once we have an artificial mind capable of understanding the necessary concepts.
1) A keyboard is not an evolutionary influence condom. Those influences affect code decisions that we make. So imagine we create AI that 'desires' the continuation of human life at some level, and assume this is a manifestation of our evolutionary drive to survive. Such a beast may answer a query such as, "How do I build an atomic bomb?" with feigned boredom (to hide an unwillingness to answer my question).
2) A lot of AI research is premised upon modeling selection between choices. It seems entirely possible that the questions I have would be incredibly boring compared to whatever other thoughts are there to be had. That is: what's useful to us might not be coequal with AI, and that limited, domain-specific pseudo-intelligence is really what's necessary. But any time there are external constraints on exploration space (for instance, requiring a response within a certain time), we don't know what we're missing out on.
If "dumb luck" can do it, why can't "smart luck" do it too? You and I are already an existence proof of the fact that it's possible to build "intelligent" systems out of matter. So what's stopping us from actually understanding what it means to be "intelligent" on a theoretical level then solving the resulting engineering problems to implement it in practice?
In other words, since "dumb luck" can optimize a fitness function that results in intelligence, why can't that same (or some similar) fitness function be optimized a little less dumbly, like using human intelligence to navigate the through the solution space?
I've long pointed out that human brains are far more complex than computers. However given Moore's law it seems likely that somewhere in the 2015-2030 range (depending on how one defines the complexity of a neuron), computers will be a match for human brains in complexity. At some point after that happens I expect human level AI to appear.
The growing power of statistical techniques for previously impossible tasks like translation and answering Jeopardy questions (see the computer IBM is putting together) suggests that once the requisite CPU power is available, it won't take long for computers to match humans on a wide variety of cognitive tasks. Once you hook together the right selection of abilities, I expect something that looks suspiciously like original thought to emerge.
Right — it's impossible to perfectly simulate something we don't even understand in physical terms (we can only simulate what we think it's doing). Though it's debatable whether artificial consciousness is needed for human-level artificial intelligence. Intelligence is just one aspect of the brain.
> Once you hook together the right selection of abilities, I expect something that looks suspiciously like original thought to emerge.
It's very unlikely to simply "emerge". Sure it's theoretically possible, but that's very optimistic. It's more likely that if we develop AGI it will be due to gaining a thorough understanding of what it means to be "intelligent". (Yes, I know the word "intelligent" is vague, but I think you and I can at least agree that we "know it when we see it", and more technical definitions do exist.)
It's very unlikely to simply "emerge". Sure it's theoretically possible, but that's very optimistic.
You are presenting a statement of opinion as if it were fact.
My contrary opinion is based on a number of things. Such as the way in which computer composed music working with a database of known music is able to come up with decent compositions. And the fact that I understand the functioning of the human brain to be the interaction of multiple components, each of which has a reasonably specific function.
Now this is not to say that it will happen by magic. But I fully expect AI, when it arrives, to be an emergent behavior that we no more understand in detail than we understand exactly how Google Translate comes to the translations that it comes to. (Yes, we know the algorithms. And we can tweak parameters, run it, and decide whether we think it is doing better. But tracing the logic is a lost cause.)
(Yes, I know the word "intelligent" is vague, but I think you and I can at least agree that we "know it when we see it", and more technical definitions do exist.)
Unfortunately we can't agree on that. Because if we set out goalposts, they will be moved when we arrive. That has been happening ever since Eliza successfully carried on conversations with people, and computers first started beating people at chess.
That's why I said "suspiciously like original thought", because when it arrives, I don't expect people to accept it until after it has become clearly better than people.
Similarly, he gets the Turing Test wrong. He seems to think that all you have to do to pass the Turing Test is to convince a single person that you are a person.
In 100 years we will have gobs of massively parallel computing power in the size of today's cellphones.
Even badly written code will be able to execute at speeds we can't fathom today. Every dot in an image will be able to be analyzed in milliseconds, matched to terabytes of stored patterns. Every sound wave in a voice analyzed by the same massive amount of power and patterns.
We may not have "real" AI in 100 years like in the movies but it will be a spooky close simulation.
I just don't know what will happen with power storage, we'd need a huge leap, the simple evolutions in battery chemistry so far are just not enough.
He is right. Something he described as AI will be impossible to be created by human being.. But I really doubt "real AI" will be something everyone currently imagine..
Here's why I disagree: smart human beings evolved from not-very-smart primates.
That proves there is a natural process by which greater intelligence can be created.
Therefore there is no reason that an even greater intelligence cannot be created with help from man. And for singularity's sake super-human intelligence doesn't even have to be an AI. Genetically-engineered super-intelligent primates would do the trick as well. The idea is that once you've created something smarter than yourself, not matter how you did it, it will then be able to figure out how to make something even smarter. And so on. That's the singularity.
What if natural evolution has become quick enough now (and it does seem to accelerate all by itself, if you put the evolution of species on a time scale) so that the "next step" (the superhuman beings) will emerge without any conscious and/or voluntary input from us?
I'm not saying this is certainly what's going to happen, it's just a (literal) what-if question.
The author's argument hinges on humans creating more intelligent AI, not humans genetically engineering beings capable of evolving greater intelligence than humans. I would also argue that humans engineering new species isn't exactly "natural" in an evolutionary sense. The author even points out that he isn't considering natural evolution in his argument.
What a ridiculous argument (something of a given intelligence can not possibly create something more intelligent)! It's kind of like saying no one who can only run at 15 miles per hour could ever create something that can go faster than 15 miles per hour.
The given enough time part is the logical flaw i think.
It's kinda like saying that a man who can only run 15 miles per hour will eventually be able to run 30 miles and thus even if he creates a machine to do it faster, he was still capable of doing it.
The fact that any single human will not be able to understand how such AI works to the minimal detail, means nothing about the way such AI can be created.
For example: Evolutionary programming, Iterative design, Gödel Machines. All of them possible approaches.
I've never understood how one could be so sure about this. It seems to me all we need to program a human in a box is:
1. An atomic-level scan of a single human being.
2. A computer with program powerful enough to simulate all the atoms in a single human being and their interactions.
If you can get this, you have a computer as smart as a human (modulo quantum effects, if those do indeed play any part in cognition.
This would be a horribly inefficient way of doing it but I'd be very surprised if it were not possible to get a computer this fast in real life, or to conduct an approximated atomic level scan of a human being. It may not happen soon, but it would be truly shocking to me if it doesn't happen eventually, provided we do not destroy ourselves before reaching that technological level.
What irks me most about the AI debates is that a machine that thinks like a human would actually be artificially delusional, not artificially intelligent. For true, strong AI all the SETI argumants apply: would such an intelligence resemble our own, or could it be so different that we wouldn't even recognize it?
"Clearly intelligence is already an infinite-dimensional beast"
Pure fluff. This author seems to have no more than a layman's appreciation of AI. He's also either prone to hyperbole or unaware of the illogic/fluffiness of such statements, or both.
This essay makes asserts that, for example, a human could never write a chess program that is better at playing chess than the human is, because the human could always just run the chess program by hand, given time.
This is like saying that no computer is more powerful than any other computer, given sufficient storage and time. But limits on storage and time are THE limits. This is a very poor way to define "power".
1. Figure out what the brain does chemically to create thought.
2. Replicate that function electronically.
Responses to his objections about complexity greater than ourselves:
If we can't create something smarter than ourselves, then how did amoeba do it? If you can figure out step #1 then you can also diff a smart brain vs a dumb one. Take whatever makes a brain smarter and double it in step #2.
IMHO hardware's not the problem. It's reverse engineering the software our brains use. I expect when we figure this out, we'll be suprised how simple we really are.
For practical purposes what does it matter? He is basically saying if humans cannot comprehend something like the fifth dimension then we couldn't build an AI that could either. So what.
I do believe humans could build a fusion reactor so a computer with expert level knowledge in all areas related to fusion reactors could build one thousands of times faster that it would take a human to do. So anything you think a human could do a high-level AI could do dramatically faster. So even if the AI could never construct something smarter than itself, it could make something faster, and in turn that AI could make one even faster. That is the singularity.
the blogger closed discussion before I could reply so I'll reply on here for what good it does me.
Whether a technological singularity is desirable is debatable. Whether development of human-equivalent AI would lead to technological singularity is debatable. On the other hand, human-equivalent AI seems inevitable to me. Author asserts some things without much in the way of proof:
"You could take a precise reading of the structure of the human's brain and simulate that brain inside a computer. But taking this initial reading is impossible in practice right now, and may remain so indefinitely, and computers need to be, conservatively, ten orders of magnitude more powerful before the simulation step becomes possible."
Why would this remain impossible in practice indefinitely? It's only a matter of time and effort until the human brain's goings-on can be reproduced outside a human brain, no matter how many orders of magnitude "more powerful" thee equipment needs to be. Unless there is truly some human-inaccessible spiritual process that goes on alongside physical processes in the human brain to produce human intelligence, there's no reason purely physical processes can't be reproduced outside a human brain.
43 comments
[ 3.5 ms ] story [ 92.6 ms ] threadI agree that it's difficult to give formal specifications for a "human-equivalent artificial intelligence," but assuming the human brain is an entirely physical object, why shouldn't we be able to at least theoretically create or simulate a computer system with as much complexity and the same functionality?
Math has started, quite a while ago, to show that we are already probing some sort of inherent limits of knowledge for the human mind.
> theoretically machines can be as intelligent as human
Intelligence is the easy problem. It's just a function of complexity after all. It's consciousness / awareness that's harder, more important and more interesting.
Imagine if we could program a self replicating machine group similar to an ant colony? Self propagating, powering, healing. Any excess resources the colony has goes into building our houses, growing our food, building cars etc...
'Real AI' may require new hardware like quantum computers, and may take a very long time, but even 400 years ago, did anyone expect us to be sending small robotic vehicles around and out of the solar system?
I do believe that real AI will happen, but it may seem very alien to us.
That says nothing about whether or not the SYSTEM on a WHOLE understands Chinese. Given the constraints of Searle's example (that every input produces a 'correct' output, but does so via this labour-intensive process), it seems there IS intelligence going on there at the SYSTEM level.
Ultimately, the point of that story is saying 'the cpu isn't smart'. Fine, that's not the argument behind AI/PI. And the whole 'there's nothing going on that isn't in the person already' also misses the point: what if, instead of a slow human translating, we had a fast human doing the lookups at the magnitude of a trillion lookup operations per second? They'd be able to do substantially more 'translating' work than the author of the software. So now, say that we program 'learning' instead of 'translation'. Sweet, now the computer can't learn any BETTER than the human, but it can learn FASTER, and there may be some substantial benefit to that.
His own argument is self-refuting: "in fact it was doing exactly what I had told it to do and nothing which I couldn't have done myself longhand given the time." Right. But you're limited by death, and a computer isn't limited in the same way. So say AI never makes it to 'smarter than human given enough time'. Fine. If it does in 30 seconds what it would take me 17,000 years to complete, I'm pretty happy.
The FAR more interesting question (to me) and probably one that can't be solved until we 'get there' is this: will computer intelligence be subject to the same foibles and hangups as a human? So, for instance, could "Google AI" respond to my search request by saying, "Wow, what a boring question, stop bothering me."
2) A lot of AI research is premised upon modeling selection between choices. It seems entirely possible that the questions I have would be incredibly boring compared to whatever other thoughts are there to be had. That is: what's useful to us might not be coequal with AI, and that limited, domain-specific pseudo-intelligence is really what's necessary. But any time there are external constraints on exploration space (for instance, requiring a response within a certain time), we don't know what we're missing out on.
In other words, since "dumb luck" can optimize a fitness function that results in intelligence, why can't that same (or some similar) fitness function be optimized a little less dumbly, like using human intelligence to navigate the through the solution space?
I've long pointed out that human brains are far more complex than computers. However given Moore's law it seems likely that somewhere in the 2015-2030 range (depending on how one defines the complexity of a neuron), computers will be a match for human brains in complexity. At some point after that happens I expect human level AI to appear.
The growing power of statistical techniques for previously impossible tasks like translation and answering Jeopardy questions (see the computer IBM is putting together) suggests that once the requisite CPU power is available, it won't take long for computers to match humans on a wide variety of cognitive tasks. Once you hook together the right selection of abilities, I expect something that looks suspiciously like original thought to emerge.
It's consciousness that's harder to contemplate.
It's very unlikely to simply "emerge". Sure it's theoretically possible, but that's very optimistic. It's more likely that if we develop AGI it will be due to gaining a thorough understanding of what it means to be "intelligent". (Yes, I know the word "intelligent" is vague, but I think you and I can at least agree that we "know it when we see it", and more technical definitions do exist.)
You are presenting a statement of opinion as if it were fact.
My contrary opinion is based on a number of things. Such as the way in which computer composed music working with a database of known music is able to come up with decent compositions. And the fact that I understand the functioning of the human brain to be the interaction of multiple components, each of which has a reasonably specific function.
Now this is not to say that it will happen by magic. But I fully expect AI, when it arrives, to be an emergent behavior that we no more understand in detail than we understand exactly how Google Translate comes to the translations that it comes to. (Yes, we know the algorithms. And we can tweak parameters, run it, and decide whether we think it is doing better. But tracing the logic is a lost cause.)
(Yes, I know the word "intelligent" is vague, but I think you and I can at least agree that we "know it when we see it", and more technical definitions do exist.)
Unfortunately we can't agree on that. Because if we set out goalposts, they will be moved when we arrive. That has been happening ever since Eliza successfully carried on conversations with people, and computers first started beating people at chess.
That's why I said "suspiciously like original thought", because when it arrives, I don't expect people to accept it until after it has become clearly better than people.
Yes.
Even badly written code will be able to execute at speeds we can't fathom today. Every dot in an image will be able to be analyzed in milliseconds, matched to terabytes of stored patterns. Every sound wave in a voice analyzed by the same massive amount of power and patterns.
We may not have "real" AI in 100 years like in the movies but it will be a spooky close simulation.
I just don't know what will happen with power storage, we'd need a huge leap, the simple evolutions in battery chemistry so far are just not enough.
Exponential complexity is still going to swamp available computing power for large enough N. N will be larger, though, and that will open up new uses.
That proves there is a natural process by which greater intelligence can be created.
Therefore there is no reason that an even greater intelligence cannot be created with help from man. And for singularity's sake super-human intelligence doesn't even have to be an AI. Genetically-engineered super-intelligent primates would do the trick as well. The idea is that once you've created something smarter than yourself, not matter how you did it, it will then be able to figure out how to make something even smarter. And so on. That's the singularity.
I'm not saying this is certainly what's going to happen, it's just a (literal) what-if question.
It's kinda like saying that a man who can only run 15 miles per hour will eventually be able to run 30 miles and thus even if he creates a machine to do it faster, he was still capable of doing it.
For example: Evolutionary programming, Iterative design, Gödel Machines. All of them possible approaches.
1. An atomic-level scan of a single human being.
2. A computer with program powerful enough to simulate all the atoms in a single human being and their interactions.
If you can get this, you have a computer as smart as a human (modulo quantum effects, if those do indeed play any part in cognition.
This would be a horribly inefficient way of doing it but I'd be very surprised if it were not possible to get a computer this fast in real life, or to conduct an approximated atomic level scan of a human being. It may not happen soon, but it would be truly shocking to me if it doesn't happen eventually, provided we do not destroy ourselves before reaching that technological level.
"Clearly intelligence is already an infinite-dimensional beast"
Pure fluff. This author seems to have no more than a layman's appreciation of AI. He's also either prone to hyperbole or unaware of the illogic/fluffiness of such statements, or both.
This is like saying that no computer is more powerful than any other computer, given sufficient storage and time. But limits on storage and time are THE limits. This is a very poor way to define "power".
1. Figure out what the brain does chemically to create thought.
2. Replicate that function electronically.
Responses to his objections about complexity greater than ourselves: If we can't create something smarter than ourselves, then how did amoeba do it? If you can figure out step #1 then you can also diff a smart brain vs a dumb one. Take whatever makes a brain smarter and double it in step #2.
IMHO hardware's not the problem. It's reverse engineering the software our brains use. I expect when we figure this out, we'll be suprised how simple we really are.
I do believe humans could build a fusion reactor so a computer with expert level knowledge in all areas related to fusion reactors could build one thousands of times faster that it would take a human to do. So anything you think a human could do a high-level AI could do dramatically faster. So even if the AI could never construct something smarter than itself, it could make something faster, and in turn that AI could make one even faster. That is the singularity.
1. it's not a well defined quantity.
2. under any definition, it doesn't seem to be anything useful in itself.
can some entity create an entity better at a task that the creator? obviously yes. c.f., compilers, factories, cranes, &c.
Whether a technological singularity is desirable is debatable. Whether development of human-equivalent AI would lead to technological singularity is debatable. On the other hand, human-equivalent AI seems inevitable to me. Author asserts some things without much in the way of proof:
"You could take a precise reading of the structure of the human's brain and simulate that brain inside a computer. But taking this initial reading is impossible in practice right now, and may remain so indefinitely, and computers need to be, conservatively, ten orders of magnitude more powerful before the simulation step becomes possible."
Why would this remain impossible in practice indefinitely? It's only a matter of time and effort until the human brain's goings-on can be reproduced outside a human brain, no matter how many orders of magnitude "more powerful" thee equipment needs to be. Unless there is truly some human-inaccessible spiritual process that goes on alongside physical processes in the human brain to produce human intelligence, there's no reason purely physical processes can't be reproduced outside a human brain.