Slashdot Mirror
Cyc System Prepares to Take Over World
Scotch Game writes: "The LA Times is running a story about the soon-to-be-even-more-famous Cyc knowledge base that has been created by Cycorp under the leadership of Douglas B. Lenat (bio here). It's a pop piece with little technical information, but it does have some enticing bits such as the suggestion that the Cyc system is developing a sense of itself. If you're not familiar with Cycorp and its goals then take a look. Of course, you should realize that this is, in fact, the system that will one day send Arnold Schwarzenegger back in time in order to kill a young pretty lass by the name of Sarah Connor. But for now the system is pre-sentient and pretty cool ..." See also OpenCyc.
And very good ones at that, which demonstrate the underlying principles of Turing machines, and show how they cannot produce semantic understanding, merely syntactical manipulation of data.
They really only suggest that Turing machines can't produce semantic understanding. I mean, it takes more than mere arguments to be a proof, particularly in the mathematical world that surrounds Turing machines.
Bzzzt! Wrong... the Turing test says nothing about whether something is intelligent, merely whether it can fool a person. Blind adherence to rules is not intelligence.
Well, how do you define intelligence then? If you can't tell by observing behavior, how do you decide? Is something only intelligent if it operates exactly like a human brain? Why does the operation make a difference?
Now here's your category error. You are assuming that the brain is also a Turing machine and that by some miracle of "emergent behaviour" intelligence arises. But that's obviously not true, as Searle showed, because Turing machines cannot be intelligent!
You're arguing that we aren't Turing machines because we are intelligent and Turing machines can't be. But there is no actual proof of that. And it is not obvious otherwise that we aren't Turing machines.
Consider this: Imagine a computer, no different from your desktop only insanely more powerfull and with effectively unlimited memory. On this computer is running a simulation of a human brain, accurate to the limits of our knowledge of physics. Every quark (or every string, if you prefer) is perfectly simulated on this machine.
Is the machine, which is a Turing machine, intelligent?
If your answer is no, then I ask what is it that occurs in the human brain that isn't occuring in the machine?
The enemies of Democracy are
Of course, you should realize that this is, in fact, the system that will one day send Arnold Schwarzenegger back in time in order to kill a young pretty lass by the name of Sarah Connor.
;-)
Use that old fashioned off switch before it gets up to any dirty tricks. It does have an off switch, right? Even Data has an off switch...
While I agree that Cyc isn't the future of intelligent computing, I have to disagree with you on another point.
Searle has _not_ proved anything of the sort. he argues for his position fairly well, but on closer inspection thay are just arguments, not any kind of proof. For a good rebuttal, read Dennet for instance.
For those that haven't heard about it, It's the 'chinese room' thought experiment, where a room contains a person and a large set of rulebooks. A story - written in chinese - and a set of questions regarding the story is put into the room. The person then goes about transforming the chinese characters according to the rules, then outputs the resulting sequence - which turns out to be lucid answers to the questions about the story. This is supposed to prove that computers cannot think, as it is 'obvious' that humans work nothing like this. Problem is, it isn't at all obvious that we do not work like this (no, not rulebooks in the head, or even explicitly formulated rules, that's not needed for the analogy).
You want to know more, I can heartily recommend a semester of philosophy of the mind!
/Janne
Trust the Computer. The Computer is your friend.
Cyc already exhibits a level of shrewdness well beyond that of, say, your average computer running Windows.
Now if they could only come up with something more shrewd, devious, conniving, underhanded & backstabbing than the CREATORS of your average computer running Windows®
try { do() || do_not(); } catch (JediException err) { yoda(err); }
(1) CYC is one of the few survivors of the "A.I." speculative bubble of the mid-1980s. Though this bubble was not as large as the recent InterNet bubble, there was a lot of hype. The US computer industry feared it would lose the "A.I. war" against Japan's "Fifth Generation Project". This project was going to build an intelligent supercomputer using expert systems. It was almost a complete bust.
(2) A major contention behind CYC is that so-called "expert systems" will be useful once they pass a certain level of critical knowledge, particulary incorporating trivia called "common sense". Most early expert systems were very small and narrow, with just a few hundred or thousand pieces of knowledge. They frequently broke. CYC is a thousand times large than most other expert systems with a couple million chunks of knowledge.
(3) One of the more interesting parts of CYC is its "ontology". You could think of it is a giant thesarus for computerized reasoning. What is the best way of doing this? Previous examples are the philosophers' systems of categories descended from Aristotle and the linguists' meaning dictionaries called thesarii. CYC uses neither of these because they are not useful for computerized reasoning. It developed its own exlucidating hidden human assumptions of space, and time, and object, and so-on. The CYC ontology is publically available on the net at the cyc web site . The ontology is much more sophisticated than a mere web of ideas (called semantic net in A.I. jargon). It has a web, it has declarative parts like Marvin Minky's frames. It has procedural parts, or little embedded programs for resolving holes and contradictions. Again this is on the web site.
I saw the words "Knowledge Base" and "pre-sentient" and immediately images of the future came to mind. Images of article Q219872 saying, "Life is like a box of Outlook macros, you never know what you're gonna get," and article Q207843 replying, "Those look like comfortable dongles."
--
--
"Outlook not so good." That magic 8-ball knows everything! I'll ask about Exchange Server next.
Minsky, I believe. The version I heard was "The ability for a machine to perform tasks which if carried out by a human being would be perceived as requiring intelligence."
I also like another definition of AI, as provided by that greatest of scholars, Anonymous: "Artificial Intelligence is the science of making computers that behave like the ones in the movies."
++ Say to Elrond "Hello.".
Elrond says "No.". Elrond gives you some lunch.
Here's something quite cool for you...
When my son was born he was strong enough to roll himself over, which isn't typical. When I talked, he rolled and looked at me. A baby, less than 10 minutes after being born, could recognize my voice. Not to mention those that have noticed a baby reacting to a voice while still in the womb. Very cool. He's ten months old now, and it's quite amazing how smart he grows daily.
What I've always wondered is exactly how we could recognize intelligence in a machine. I already knew that my child had the ability to be intelligent because he is human... but will it take a truly amazing act before we acknowledge intelligence in something that "shouldn't" have it?
If it looks intelligent, and acts intelligent in all conceiveable circumstances, then we'll be forced to conclude that it is intelligent, even if we know what's going on under the hood. Are you suggesting that, should we one day discover the secrets of the emergent behaviour of the human brain (reducing it, therefore, to "a simple rules system"), that we will suddenly cease to be intelligent?
Unless a machine builds its knowledge of the world through its senses, it will never have common sense. No machine will ever understand enough about nature from being fed a bunch of facts, regardless of how how many inferences it can make. The interconnectedness of intelligence is intractable to formal symbolic means. We have the common sense to hold a cup of coffee upright and level without spilling it from experience and our ability to coordinate millions of sensory nerve impulses so as to trigger the right sequence of motor neurons.
Holding a cup of coffee is just one in myriads of highly detailed knowledge that one can learn through experience. A machine cannot gain this sort of knowledge from being spoon fed facts via a keyboard. Cyc is merely a glorified database with a fancy query language, one that requires experienced human data entry slaves to maintain. Unless a machine is given the ability to learn from and interact with its environment via its sensors and effectors it's not intelligent. Sure Cyc is a cool hack but it has about as much to do with intelligence as MySQL. To associate it with intelligence is an insult to computational neuroscience researchers around the world whose goal is true human level AI. Sorry.
AI systems of this class are comparable to "how-to" books. If the author anticipated your question, they're useful, and otherwise they're not.
I agree. The symbolic crowd is holding on for dear life to an obsolete science. Their approach to AI is an insult to those of us who know that the only viable solution will be a neural one. Why? Because intelligence is so atrononomically complex as to be intractable to formal symbolic means. Intelligence must be acquired through sensory experience. The ability to interact with the environment via motor neurons is a big plus.
The symbolic clan (that includes people like Marvin Minsky, Lenat, etc...) have taken to equating AI with inference engines, expert systems and glorified databases. To defend their moribund approach to AI, they invent cute little phrases like "there is no such thing as a free lunch". It's sad. It goes to show you that delusion has its rewards.
The article is good, but this is a poor quote. As others have pointed out, what "the human brain alone can do" is a moving target. Remember when only humans could play world-class chess? prove theorems? Add two numbers together?
"That which makes us human and not just machines" is often defined simply as "the stuff that machiens can't do" ... yet.
My Karma: ran over your Dogma
StrawberryFrog
I don't know about you guys but I am really scared here. This sort of thing makes us have to ask ourselves fundemental questions about what is right and wrong. Hollywood actors (that aren't chicks getting naked) should not have personal websites. Do we really want our children accidently browsing to Arnold's sight?
Someone you trust is one of us.
Kind of like when Jurrasic Park was released, they revived Tony Bennit's career (brought a dinasour back to life).
Someone you trust is one of us.
The only reason this story is getting printed is because Steven Spielberg's AI movie is coming out soon, and his studio is trying to drum up interest in the subject.
The coincidence is neat, but this story is important in itself, at least for a significant proportion of AI researchers.
Doug B. Lenat is one of the guys who gave me the AI "vvirus". I remember reading an old book about the "first generation" of AI, and of all the things I saw in it none impressed me nearly as much as Lenat's Eurisko. It was a kind of modern fairy tale for the little boy that I was at the time.
Cyc was mentioned in that book as a "long-term project". I remember visiting their website once, and thinking how all this definitely looked like the ultimate vaporare story.
In itself, Cyc is simply a continuation of Lenat's previous work, that is, a monumental, "new generation" expert system. It is to traditional expert systems what the internet is to telegraph : it does basically the same things, but the technical difference lead to a qualitative leap. It is neither intelligent (it was not designed to pass the standard Turing test) nor "conscious" (it knows about itself, but just as much as a Java class that can do introspection). But when it comes to practical applications about analyzing abstract data and drawing abstract conclusions, it can crush the competition any time.
Bloody hell, they've finally done it. Yes, this is important. Don't let the journalists' hype fool you: this guy is worth your attention, and you might pretty well hear about him again over the next few years.
Thomas Miconi
But have they told Cyc not to use humans as batteries?
Good ontologies are a big part of this -- identifying and distinguishing different contexts, associated with their likely possible properties.
The work CYC have done in finding good ways to represent such ontologies is important, but only goes so far -- in particular it seems to be essentially static. What impresses me more is some of the work that has been done elsewhere to automate the process of the discovery and maintenance of ontology -- extracting it dynamically from the associations revealed in a large pile of documents.
One example of a site which is an end user of such technology is the well known news portal moreover.com, powered mostly (I believe) by Autonomy
Surely one of them would have had to be a woman : )
Um, am I the only one creeped out by this? And presumably they've told it all sorts of other moral stuff too, but who gets to decide it's morals? It's all kinda subjective. And how do we know that they've not said anything like "It's worse to let any single one of us die than it is to let any number of other people die" or something (I doubt very much that they'd do that, but I'm just trying to come up with an example and it's Friday afternoon and I'm off for the weekend in 2 hours)?
Ultimately, Cyc isn't actually making decisions, but re-gurgitating what it's been told previously - the people programming it make the decisions. I've formed opinions about a great many things, and some of those opinions contradict what a lot (or all) of my friends and family think, but I reached them myself - Cyc needs to be able to do this before it will be sentient - right now it's just a big, sophisticated database (only a way further along the line than Jeeves).
When it can make worthwhile posts to Slashdot I might look at it again : )
Cyc is the definitive last gasp of expert systems. The basic idea is to take statements about the world, encode them in a formal language that is somewhere between predicate calculus and SQL, put all those statements in a database, and crunch on them. Predicates like IS-A and PART-OF are used heavily. The database contains thousands of statements like (IS-A COW MAMMAL) The result is a kind of encyclopedia. It doesn't learn, it just gets updated manually. There's internal consistency checking, though; it's not just a collection of unprocessed data. If "A implies B", "B implies C", and "A implies not C" get put into the database, it detects the contradiction.
The Cyc project has generated hype for many years. Lenat used to issue press releases announcing Strong AI Real Soon Now, but after the first decade that got to be embarassing.
For a while, there was a natural language front end to Cyc on the web, out of the MIT AI lab, but I can't find it right now. It was supposed to be able to do induction, and it was supposed to have MIT-related location information. So I tried "Is MIT in Cambridge", and it replied Yes. "Is Cambridge in Massachusetts", and it said yes. "Is Massachusetts in United States" returned yes. But "Is MIT in United States" returned "I don't know". That was disappointing. I'd expected it to be able to at least do simple inference. My overall impression was that it was about equal to Ask Jeeves in smarts.
AI systems of this class are comparable to "how-to" books. If the author anticipated your question, they're useful, and otherwise they're not.
I'm not quite sure I follow you, especially in light of this:
If you believe a brain and its reactions can be simulated by a computer, why is that not sufficient for intelligence?
Is this belief associated, in any way, to theological beliefs?
Please explain your position, as I am genuinely interested in understanding it.
---
(Not that the company isn't real or working hard on the area, but just take this with a grain of salt...)
GollyGee Blocks -- 3D creativity software for kids.
Do you know. I have to work with people that talk like this all the time. They like to think they are intelligent too.
I was thinking myself it would be nice to use Cyc to train neural networks. That way you might be able to 'grow' (the beginning of) a real AI. Does this sound feasable?
Monkey sense
The Mongrel Dogs Who Teach
Penrose argued (in a nutshell) that (1) intelligence may be driven by quantum events
(2) quantum events are nondeterministic
(3) computers are deterministic
Ergo, computers can never be intelligent, QED.
Ok, where should we start to demolish this argument? Heck, you can probably do it yourself now.
I wonder if in Cyc's early years, instead of being shrewd enough to ensure it knows what you're talking about, it kept asking "Why?, Why?, Why?" to everything you explained to it.
Having trouble making friends online? Can't even find one person who will put you on their buddy list? For an additional $9.95 per month on your AOL account, you can have an artificial Buddy to chat with who's online 24 hours a day. His/Her screen name is Cyc342
-
A robot may not injure a human being, or, through inaction, allow a human being to come to harm.
- A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.
- A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.
I'm not sure someone with $50M invested is going to put 2. and 3. in that order, though[100% ISO 646 Compliant]
SVM, ERGO MONSTRO.
I'll file a bug report with the Vatican.
Sorry folks, but carbon-based biology has nothing to do with the future of intelligence. It might be nice to believe that a group of cells randomly firing electrical signals at one another can create a sentient being, but such thoughts are naive. Sure, each individual cell might be alive, but it doesn't mean that a group of dumb cells working together would make an intelligent being. Its like believing that since one rock is dumb, a mountain would be bright.
Unfortunately, some people look at the behavior of h. sapiens and shout "intelligence!" Sure, it may appear that h. sapiens is intelligent, but only with a short examination. A human may pass a Turing test, but even though the human proved that he or she is indistinguisable from an intelligent entity, it doesn't mean anything, because I feel that I can make up any arbritrary decision I like so I can declare that a being that is indistinguishable from a sentient entity is still not sentient. :)
Seriously though, the so called "intelligent" h. sapiens owes its "intelligence" to a group of electrical impulses and a few simple chemical reactions among the many millions of cells that makes up the creatures "brain". With a powerful computer, we could simulate the reaction of chemical/electrical impulses of h. sapiens, but no one 'cept an undergraduate would be foolish enough to call such a simulation "intelligent". It can be argued that h. sapiens runs mainly on instinct and conditioned responses, its very clear that humans seem uncapable of long-term thinking, a sign of intelligence, and are thus doomed to ruin their habitat through environmental neglect and ever more damaging wars.
So remember, humans aren't intelligent, they only think they are.
Actually, I think HAL killing the crew of the space ship was a lack of morals, instead of a set of misplaced morals. Assume that HAL was incapable of lying, through either not being programmed with the capability, or else having implicit instructions not to lie. Also assume that HAL was never programmed with the 1st law, or was programmed with a flawed instance of the first law. Both assumptions are reasonable. I don't see why someone would program in a set of lying subroutines into a program designed to run a ship, since we want the astronauts to be told the truth about the ship's sensor information, how much fuel is in the tank, etc. HAL couldn't have a strict first law, due to the fact that it might have to sacrifice one member of the crew to save the rest.
So, the gov't comes along and tells HAL to lie, although probably not in those words, since HAL doesn't know what a lie is. Maybe it was worded that HAL couldn't let the astronauts find out the information. So, HAL being a learning entity, starts to worry about the humans asking it for the information, and "knows" that he can't tell them when they ask. So, it looks at the possible solutions, say 1) Shut down, 2) Kill the crew, etc... If it shut's down, there is nobody to control the ship, and thus the mission is in danger. If the crew is dead, then they can't ask for the information, and HAL probably is allowed actions that result in the death of one or more crew members, to "save" the mission. Its just that nobody ever told HAL to keep at least one crew member alive. So, part of the mission is the "don't lie" command given with a high priority (something along the lines of "must be obeyed to complete the mission), and the astronaut's lives have a slightly lower permission. In short, HAL was buggy.
Searle proved no such thing as your assertion; he merely provided a series of thought experiments which force us to think about what intelligence might actually consist of.
And very good ones at that, which demonstrate the underlying principles of Turing machines, and show how they cannot produce semantic understanding, merely syntactical manipulation of data.
If it looks intelligent, and acts intelligent in all conceiveable circumstances, then we'll be forced to conclude that it is intelligent, even if we know what's going on under the hood.
Bzzzt! Wrong... the Turing test says nothing about whether something is intelligent, merely whether it can fool a person. There are already some pretty good pieces of software out there about this, and they'll get better in the next few years. But they won't be intelligent. Blind adherence to rules is not intelligence.
Are you suggesting that, should we one day discover the secrets of the emergent behaviour of the human brain (reducing it, therefore, to "a simple rules system"), that we will suddenly cease to be intelligent?
Now here's your category error. You are assuming that the brain is also a Turing machine and that by some miracle of "emergent behaviour" intelligence arises. But that's obviously not true, as Searle showed, because Turing machines cannot be intelligent!
I first read about Cyc in Discover Magazine back when I was a Junior in HS. I thought it was the coolest thing since frozen bread. Then I read up on the topic of AI.
I have no doubt that one day AI will come to pass. I mean that in the strongest possible terms--a piece of software will pass the rigorous Turing Test and will be agreed by all to be intelligent in exactly the same sense humans are.
I *DO* have doubts that Cyc will be at all related to this outcome. Think about it: When I say "Joe is intelligent" do I mean "Joe knows a lot of facts?" No. Do I mean "Joe is good at symbolic logic?" No. I mean "Joe pursues goals in a flexible, efficient and sophisticated manner. He has a toolbox of methods that is continually growing and recursive." Does this description apply to Cyc?
No. Lenat and friends created a bunch of "knowledge slots" that they have preceded to fill in with pre-digested facts. What do I mean by "pre-digested"? For instance, Cyc might come up with an analogy about atoms being like a little solar system with electron planets "orbiting" the nucleus sun. Great, but that analogy came about because of how the knowledge was entered. Put in "Planets Orbit Sun" and "Orbit mean revolve around" and "Electron revolves around Nucleus" and then ask "What is the relationship of Electron to Sun?"--the analogy just falls out with some symbolic manipulation. It would be a lot more impressive if Cyc made analogies based on data acquired like a human: full of noise, irrelevance and error based on self-generated observations.
Cyc is a highly connected and chock-full database with a flexible select language. As a product that's awesome. As a claim to AI it's pretty weak.
--
324006
The fact is that no modern computer, no matter how powerful it gets, will ever be capable of creating true AI.
What do you mean by "true AI"? Artificial Intelligence is defined differently by different people but one widely accepted definition is "The ability for a machine to perform tasks which are normally thought to require intelligence". Intelligence can also have a number of definitions but keys factors are generally the ability to acquire and use knowledge and the ability to reason.
Cyc is doing things that previously machines have not been able to do so it has a lot to do with the future of AI.
You are right to mention that rules based systems will not bring us Strong AI but you make the mistake of thinking that Strong AI == AI. Strong AI is not the only goal of AI research. Many AI researchers are, like the developers of Cyc, trying to create machines that can do things that have previously been the preserve of the human brain. Their work is just as valid as those striving for String AI and at the moment is having more impact on the world.
Sorry, but Cyc is just a nice toy and of no use in serious AI research.
I doubt the defense department would be so interested in Cyc if it were "just a nice toy". :)
"The fact is that no modern computer, no matter how powerful it gets, will ever be capable of creating true AI."
I'b be carefull with reasoning like that. If Moore's law will keep on going we might very well have powerfull enough CPU time in 100 years. That's not tomorrow, but certainly not never