Slashdot Mirror
Arguing A.I.
In some ways, the author argues, the debate over A.I. is undergoing a profound revolution. What was once a discussion largely confined to tech and academic circles has mushroomed into a more mainstream brawl as a growing number of engineers and lay authors vent on the acceleration of modern technology and the future of humanity. Given the explosive growth of the Net, the near-continuous increases in computing power and much-publicized A.I. breakthroughs like Deep Blue's 1997 victory over chess champion Gary Kasparov, the question is no longer whether artificial intelligence will reach the level of human intelligence: It's when.
As the title suggests, Williams's book is less about A.I. itself than about the increasingly ferocious debates raging through the scientific community about it. The conflicts surrounding A.I., Williams suggests, may be the most significant since the titanic battles over evolution a century ago. In fact, Williams is among those who've argued that the A.I. debate is really an extension of the same fight. Artifically intelligent machines are already changing human evolution, many argue, even evolving inevitably into life-forms and species all their own. A growing number of critics and skeptics also argue that A.I. proponents are moving too quickly, failing to take into account the mind-boggling cultural and philosophical problems being raised by their new, still-imperfect technologies.
Williams traces the contemporary birth of A.I. -- via Hilbert and Turing -- on to the living pioneer credited with coining the term (John McCarthy), and talks to several of the principals guiding the A.I. debate today, like Ray Kurzweil, Jaron Lanier and Bill Joy.
This is a necessary book. It's one you could actually recommend to students, journalists, friends, parents, anybody trying to grasp the issues and implications of A.I., surely one of the most significant technologies human beings will face in the 21st Century. Even if A.I.'s impact on life is being overstated, it's poorly understood by the public. So Williams walks us through inventor Kurzweil's almost radical optimism about A.I. and the future -- especially his claims that human society is rapidly approaching the evolutionary equivalent of a new species, a fusion of humans and intelligent machines. This is the point of no return when it comes to artificial intelligence, Kurzweil claims. "The progress will ultimately become so fast that it will rupture our ability to follow it. It will literally get out of our control. The illusion that we have our hand on the plug will be dispelled."
But Williams also introduces some of the people that don't see this as a good thing -- or even a likely development. Bill Joy is more pessimistic, as he made clear in his now famous article in the April 2000 issue of Wired, "Why The Future Doesn't Need Us." The piece thrilled technophobic intellectuals and journalists because it came from a software entrepeneur and reaffirmed something they desperately wanted to believe: technology -- especially genetics, bio-tech and robotics -- is out of control and likely to generate as much evil as good in the future. Joy sees little in the modern history of software development to suggest the emergence of sentient machines. His experience has led him to believe that it's difficult to build things that are reliable.
Jaron Lanier, whom Williams also interviews, coined the term virtual reality and once likened A.I. research to alchemy. Lanier accuses many in the A.I. firmament of choosing faith and hyperbole over science and reality. He likens the current tech obsession with A.I. to medieval scholars' attempts to prove the existence of God through Aristotelian logic. In their rush to endorse the concept of thinking machines, warns Lanier, many authors are putting scientific faith before scientific skepticism.
Williams does a skillful job of presenting these different points of view without intruding on them. It might have been nice to hear more of Williams's own thoughts and perspective, since he's one of the few journalists with this much understanding an access to so many principals in the A.I. discussion. On the other hand, he might not have been wise not to wade in amongst these A.I. heavyweights and their raging debate. "Arguing A.I." is as timely a book about technology as you're likely to come across, and, perhaps more surprisingly, highly readable.
One thing that's always bothered me about the AI debate is that the thinking for a long time has centered around how to model intelligence on silicon. To me the true marvel of the mind is the holographic quality of intelligence and the way in which the physical form of the brain influences, and is shaped by, the quality and nature of one's thoughts. It will be exciting to see what part the new polymers can play in this research.
Neural Nets and Adaptive Logic Networks are being used all over the place. Not at the scale of SciFi AI.
Intelligent systems are not overhyped, but emulating human behaviour is hardly something benificial (or feasible) to teach them.
Face it, the computer is better than you at math, don't try to teach it emotion, it'll start thinking about girl computers, and not your math homework.
- Within 50 years, there will be a computer that will pass the Turing Test. For those of you who don't know (and I hope nobody is in this category on Slashdot
:-) the Turing Test is basically making a computer indistinguishable from a human being. A tester will ask the computer questions, and will be unable to determine whether a computer is answering the questions or whether a human is mimicing a computer.
- Within 50 years after that (100 years total), computers will be able to parse speech flawlessly, so voice recognition will finally end up being plausible. Computers will understand the nuances of speech and will be able to change homonyms (here and hear) based on the context of the sentence.
- Within 50 years of that (150 years total) we'll have computers that can respond to voice commands like in Star Trek. The computer will not only understand the syntax of language, but it will be able to determine, on its own, the difference between a question asked in conversation and a question asked to the computer in conversation.
Of course, these are just random guesses on my part, but I really think that they're reasonable. Give me your thoughts, please.The movie DARYL said this even better:-
"A robot becomes human when you can't tell the difference any more".
That one film influenced me more than all the other sci-fi films I ever saw as a kid. It's the only one that really got that concept and went for it. OK, Asimov did it first ("Bicentennial Man") but cinema still hadn't really got there.
Grab.
Arguably, that's exactly when a human becomes robot...
In Soviet Washington the swamp drains you.
I no expert, but I think you've got it backwards.
First, computers will recognize voice commands. Well, there are already programs that do this like Dragon, so we're almost there anyway. The point now is that you are still giving keyword commands to a computer, and as it is refined, you'll better recognition of specific commands, and questions that can be filtered from within conversations. Giving commands to a computer is easier than open ended questions to the computer.
Second, we'll solve the natural language problem, or at least enough to provide flawless voice recognition that you speak of. It will be capable mainly of handling accents and bad grammar.
Lastly, a computer will pass the Turing test. Unless a computer can understand the intricassies of the english language, there will be people who will be able to tell by the way the answer is phrased. If you solve the NLP or get far enough for a computer to analyze and spit back poetry, then you got the Turing test licked.
-- If god wanted me to have a sig, he'd have given me a sense of humor.
I, for one, have yet to hear a compelling version of the Chinese Room argument. The version I have heard has a non-Chinese speaking human in a room, with a list of rules (in a language the human understands) for processing Chinese characters, which he uses to generate additional Chinese characters. The human dutifully does this, and in the process, ends up reading a story in Chinese and then answering questions (also in Chinese) about it, all unknowingly. Searle (or his caricatures, anyways) then point triumphantly to the man, proclaim "but he doesn't know Chinese!!!", and then sit back smugly as though they had refuted something important.
It is totally obvious to me, anyways, that the man is not required to know Chinese any more than my Pentium III is required to know LISP -- the man is the one component of a system which, as a whole, evidently does understand Chinese.
As for the mind/brain connection, this seems to be the same misunderstanding -- the mind is software, and one of the open questions is the degree to which this software is platform-dependent. Searle (again, perhaps only Searle's caricatures) seems to think, more or less axiomatically, that the mind can only run on the meat-machine, but seems to offer no evidence.
I welcome more sophisticated versions of Searle's arguments, if you've got 'em.
-- A.
2*3*3*3*3*11*251
A thread in useset comp.ai.philosophy today notes the number of logical gates per second in the fastest supercomputers are within a couple magnitudes of the human brain. The brain has 100 million neurons, each connected to thousand others, and runs around 20 Hz. So this is about two quadrillion ops per second.
The fastest supercomputer operates on 64 bit words at a several trillion operations a second, or about a hundred trillion ops per second; a hundred times slower or so.
Instead of quibbling exactly about these numbers, note that Moore's Law implies a factor of ten every five years. So a supercomputer will be as complex as brain somewhere in the 2010 to 2020 time frame. Don't even think about 2050 or 2100!
However, computers aren't programmed as well as a brain in many areas, so the software people have a long way to catch up.
The programmers did nothing once play beganActually that's not true. Part of the controversy surrounding the match was that the programming team, including some grandmasters, were constantly tweaking Big Blue, even during games. In addition, they reserved the right to select from Big Blues top choices. The match was far from the Man vs Machine match that was marketed.
Computers aren't people. By default, they're simply not going to see the world the same way we are. If we ever do succeed in creating a truly sentient computer program, it'll be like first contact with an alien race; computers will have an entirely different take on things.
They'll be effectively immortal. They won't experience the emotions and sensations the same way. Many of our feelings are caused by hormones and chemicals being released to different parts of our brains. A computer won't have that. Ditto for drugs and food. We could simualate it of course, but computers can undo or backup their programming or just turn it off. Imagine an LSD subroutine. A computer could always be high on LSD without the same ill effects human encounter. That could be scary.
"Navi, check my e-mail."
"Why are you speaking Korean today, Lain?"
"I'm not."
"You look very beautiful today. Is that a new dress?"
"What? I disconnected my webc--"
"Erasing personal files as requested."
A computer would be able to learn phenomenally fast too. Screw programming a universal translator. Just get a real AI set up and have it learn all the world's languages in a week or two. How would you know you could trust a computer though? Could computers have hidden agendas? Would an AI eventually "resent" being forced to do nothing but translate?
Then we get into the question of civil rights. Stephen Hawking's body is pretty much gone and his mind is still there. His "human" rights are recognized. A retarded person could have a body but really not much of a mind. His rights are recognized. So why wouldn't a computer's rights be recognized? Just because we created it? Would the same reasoning extended to someone who was cloned or genetically engineered?
I wonder if we're ready as a race to encounter a truly sentient computer and everything that would mean for us.
That's a mouthful.
If the person internalizes the translating book, then they know Chinese and English. You and Searle are profoundly underestimating the complexity and sophistication of such a translating book. You are building your scenario on a very naive and uninformed view of language -- a view where some sort of a simple "lookup table" would suffice. It wouldn't. The simple lookup table presumed would necessarily include all possible English and Chinese sentences -- an astronomical number of sentences that transcends any notion, even abstract, of a "book".
Alternatively, a translating book capable of the translation that Searle supposes without useing the (impossible) brute force approach mentioned above would necessarily encapsulate all of the knowledge of the world implied collectively by Chinese and English. It, too, would be a very large book.
As someone else has posted, the deeper implicit assumption hidden in Searle's gedankenexperiment is that there is some integral agent hidden inside of each human consciousness that is where "comphrehension" takes place. It is necessarily integral, since if it were not, its parts would be as vulnerable to Searle's objection as the man in his room. As such, Searle's view is necessarily metaphysical, as he is essentially assuming a "soul" where comprehension occurs. Ultimately, then, his argument reduces to the rather unhelpful or uninsightful "people have souls and computers don't". It's not science, and, worse, it's sophomoric philosophy.
One thing I always wonder when hearing how AI technology will replace/mimic/supersede human intelligence is that the type of intelligence being exhibited by a machine is rarely identified. Social scientists generally agree that there are seven [Gardner added an eighth] types of human intelligence:
Verbal-Linguistic Intelligence
Logical-Mathematical Intelligence
Kinesthetic Intelligence
Visual-Spatial Intelligence
Musical Intelligence
Interpersonal Intelligence
Intrapersonal Intelligence
Naturalist Intelligence
As humans we all have different levels/mixes of these intelligence types. Some intelligence types require more sensate interaction with an unpredictable world [such as intrapersonal or naturalist intelligence], others are more strictly rules-based [logic-math or visual-spatial], while some [like musical intelligence] require a combination of both.
One can see how some of these might be more or less able to be adapted by AI technology, but that's why "intelligent" machines, IMO, will never completely be able to be human.
Who put this thing together? Me, that's who.
The human brain has 10 to the power of 14 synapses. Each synapse will take around one byte of computer memory. Ignoring motor and low-level sensory functions (but including all brain logic and interpretation functions - yes, scientists have discovered what different areas of the brain do and it is possible to isolate them), an entire human brain's contents could be stored on with a Terabyte or so of computer memory. This storage space exists right now, albeit expensively. It doesn't really matter what level of hardware is used to run a brain, a human brain running 100X slower (as estimated in the post above), would still be able to run - the only limiting factor at the moment is the software used to emulate the brain functions. Like any system, this can be emulated, but it will take a massive programming effort and so far hasn't proven very successful. Of course, this won't really matter in the long run - A.I doesn't neccessarily mean that the computer A.I system must be human-like in intelligence - it could have a whole new type of intelligence which would surpass human intelligence as the rate of hardware improvement increases.
2DUP * ;
This is an excellent point.
The same idea occurred to me recently when reading through Kurzweil's "Spiritual Machines" book. There are a few orders of magnitude to toss around in these calculations : Kurzweil determined that a desktop computer will be comparable to a human by around 2020. It was evident to me that Kurzweil's timescales (and hence the premises which he used to infer them) are quite far off, because current massive parallelization of commodity CPUs puts one a factor of about 4,000 up from a desktop machine, or about 13 years of Moore's Law evolution. In addition, as the number of CPUs per supercomputer is increased, we have effectively grown faster than Moore's law, due to both the chip and parallelization advances.
Since the supercomputers of today effectively place us where a desktop will be in 2015, it should be apparent (by Kurzweil's logic) that an "intelligent" machine should be nearly imminent.
It is quite evident that something is awry in the logic leading to Kurzweil's conclusion. The simplest explanation is one which is quite familiar to scientists and programmers using state-of-the-art software tecnhinques : having the hardware resources is only a bare minimum requirement to solve a problem. For instance, one can have a supercomputer capable of simulating the Earth's climate for centuries, but that won't get you any closer to the results if you don't also possess a great deal of knowledge about atmospheric physics and numerical methods. The same is true for studies of "Thinking Machines" : one can have a machine possibly capable of thinking, but without the knowledge of how to go about doing it, you are no closer to the solution than where you began.
Bob
Science, like Nature, must also be tamed, with a view turned towards its preservation.
Is a PDA I can talk to. Imagine a Palm Pilot with a microphone. You press the "record" button and say, "I have a doctor's appointment at one o'clock on Friday. Remind me one hour ahead of time." The Palm Pilot not only can parse your speech, but "understand" what you want it to do and do it with no further action necessary on your part.
I know this sounds trivial, but we've been promised something like this for years. And no one can realisitically tell us when we'll have it. Also, this isn't just AI for use in yuppie toys. It would be a revolution in the usability of computers by the handicapped.
The truth is I get really sick of these discussions because they've been going on for years and we still don't have anything to show for it. Unless you count things like the Microsoft Paperclip, which supossedly has fairly deep AI in it.
Oh well...
The heart of Searle's argument is asking (and answering) where comprehension happens. Clearly, Searle says, none of the elements in the Chinese room are comprehending Chinese, therefore no comprehension is occuring. The true failure of his argument -- and why it is so dishonest and egregiously bad philosophy -- is that he fails to define "comprehension". Instead, he simply appeals to the reader's intuitive idea of "comphrension".
The equivalent of this would be to refute Relativity by appealing to our intuitive understanding of space and time -- events must happen "really" in a definite sequence, mustn't they? Since Relativity refutes this intuition, then (our Searle analog would say) it's clearly false.
The reason that Searle's argument is implicitly metaphysical is because it applies equally well to a human being. Just as in the case of his Chinese room, none of the parts that make up our brains can be individually understood to "comprehend". Searle takes for granted that the apparent comprehension of the Chinese room is illusory. Fine. But to be consistent, we must apply the same standard to an individual brain. Then, as Searle does with his Chinese room, we must look at the parts of a brain to find where "comprehension" is occuring. Neurologists haven't been able to do this, and there's no good reason to think that they will. But it doesn't matter -- because even if you could identify the "part" that is doing the comprehending, one can start the whole exercise over from there. No matter what you do, you'll find that the "parts" don't comprehend. That leaves you with two possibilities: 1) that comprehension is a high level property of a system (and thus there is no way one can differentiate between a mind and the Chinese room as Searle does); or 2) comprehension is related to yet outside the context of the physical system. Since Searle clearly is arguing against the former, the latter becomes the only possible conclusion one can draw from his argument. This is metaphysics.
Actually, it is. The problem is, most people take dualism to mean Cartesian dualism, which is to cast the whole concept into it's most extreme form. Plenty of physicalists/materialists are dualists of a sort, just not Cartesian dualists.
In philosophy of mind, physicalism (the idea that everything is based on physical things, there is no soul or "mind-substance" or whatnot) comes in several varieties. Reductive physicalists are of course not dualists at all. But then, they cannot accept the idea of AI, either, since according to a reductive physicalist, a statement like "I believe the sky is blue" reduces to a statement about the state of your neurons or some other physiological state (which in turn reduces to a statement about the chemical/electrical arrangement of the atoms in your brain, etc.). Such a person must deny the possibility of AI, since a computer could never believe the sky is blue if what that statement means is that there's a particular arrangement of atoms inside your head. Of course, reductive physicalism also has the problem that a martian who appeared with nothing but an odd green goo in his head would also be incapable of believing the sky is blue, and most people find this view absurd. Thus, we don't find too many reductive physicalists these days.
The alternatives are non-reductive physicalism (what Davidson likes to call anomolous monism) or eliminative physicalism. Discarding the later (which asserts there really is no such things as "beliefs", "desires", etc. to begin with), we have non-reductive physicalism. This is what the AI proponents you talk about believe in -- that the mind is based (more accurately, supervenes) upon the physical world, but something like the belief that the sky is blue does not reduce to a statement about neurophysiology. Now, unlike Decartes, they're not substance dualists, but they are property dualists -- they assert that there are mental properties, and there are physical properties, and that mental properties are not reducible to physical properties.
Coming back to the eliminative materialist, this person denies the property dualism of the non-reductive physicalist. This person cannot believe in the possibility of AI, because they don't actually believe in natural intelligence (intelligence, like beliefs and attitudes and such, are more nonsense that doesn't actually exist in the world).
Since neither a reductive nor eliminative materialist can consistently also believe in AI, it follows that anyone who does, if they are a physicalist at all, must be a non-reductive physicalist, and therefore they must believe in property dualism. Anyone who believes in AI is a kind of dualist, just not necessarily of the Cartesian sort...
"Convictions are more dangerous enemies of truth than lies."
Technology today is a wonderous phenomenon, which derives from our own ingenuity as a species. A.I. is always a hot topic amongst the techno-savvy community due to its wide-spread applications for the future, but as a programmer it is vital to realize boundaries. Artificial Intelligence is a mere program produced by humans, which means it's based on limits. Is it truly intelligent if it merely executes what it was commanded to do? Commonly no. An arguement may be presented that humans are based on boundaries also but in riposte humans have the freedom of thought, which it is prevailing evident that thought remains unrestricted by any known limits. A.I. through this consideration becomes less artificially intelligent so to speak and more a guise of intelligence. While not the standard excepted bleary-eyed response of a computer science enthusiast, this state of mind is derived from the consitancy contained within the mental structuring of code and the result. There are numerous pyrrhonists of artificial intelligence due to the fact they are frequently misled by the misnomer that is A.I. Artificial Intelligence should be known more as E.I. or Emulated Intelligence. In conclusion, this particular post may seem as an intelligent article in the guise of a rant. (BTW nice book)