Slashdot Mirror

If only we had a system that was designed from the ground up to provide some common sense for AI.

They have been getting pretty good results and have several products

Their "products" consist of the knowledge database and inference engine. Making it actually do something useful is up to the customer.

Using these products as evidence that they get results is sort of like saying a company is good at building houses because they sell hammers.

Deep Learning (the opposite approach to AI) has many applications including, processing images of checks for the banking industry, face recognition in security applications, speech recognition and generation, fake porn, etc.

What has Cyc done?

They have been getting pretty good results and have several products

Will you please do something with http://www.cyc.com/ and bring the ability to just talk to our computers in natural language? Not like Siri, but like the computer in ST:TNG.

It looks like there are some minor business applications in the area of language recognition and expert systems. See their http://www.cyc.com/enterprise-solutions/success-stories page. The idea that some general, hand-coded knowledge base will start an AI revolution is naive and has been successfully refuted, in my opinion; I think robotics offers a better foundation for general purpose AI because it allows machines to interact directly with the world in a 3-D, realtime environment and receive the kind of feedback that is critical for self-improvement. I think the Cyc approach complements that, however, and could be a useful addition to some future intellgent robotic agent.

It is easy to imagine that the Cyc database would be very much improved once robots can start extending it automatically based on their own experience. Imagine the knowledge explosion that could take place then, because robots tend to see things differently than people and would probably find knowledge derived from other robots more tractable and useful than human observations.

Given that IBM starts this as a commercial service, I was wondering what happened to Cyc? It created a lot of (partly negative) fuzz amongst computational linguists in the early/mid 90s and they have been entering common-sense knowledge into their database since the 80s. (There's also OpenCyc if you want to play around with it.) People used to be undecided between "it's crap from the 80s, entered by disgruntled ex philosophy students and unmaintainable" and "it's the next A.I. revolution and every product will be based on it once the big companies jump on the bandwaggon".

So what happened to them? Is Cyc already in some products?

Given that IBM starts this as a commercial service, I was wondering what happened to Cyc? It created a lot of (partly negative) fuzz amongst computational linguists in the early/mid 90s and they have been entering common-sense knowledge into their database since the 80s. (There's also OpenCyc if you want to play around with it.) People used to be undecided between "it's crap from the 80s, entered by disgruntled ex philosophy students and unmaintainable" and "it's the next A.I. revolution and every product will be based on it once the big companies jump on the bandwaggon".

So what happened to them? Is Cyc already in some products?

I'm sure CYC would make a good go of it. This is what it was built for.

Plenty of data has been organized by http://www.cyc.com/

I'm not connected with this company, but one of my college roommates worked with Doug Lenat at Cycorp for a number of years.

http://cyc.com/cyc/company/about

Talking about real innovation, perhaps it's time for Apple to buy Cyc or something similar and leave the slumbering A.I. cat out of the box?

Apple has always had an inclination towards that direction, and it would be the right company to start offering deep semantic processing to the end consumer - as a luxury gimmick, before it becomes robust enough to be always useful. The competition might be hard, an impressive number of A.I. researchers are working for Google. Anyway, processing power suffices nowadays to make some amazing A.I. products feasible and I personally would love to see them on my phone or desktop.

What I'm trying to tell you is that I've studied predicate calculus and prolog and various methods to achieve this. The problem isn't the system, the problem is replicating a human life (or even 18 years) of knowledge into whatever form is machine interpretable and this solution falls prey to these problems.

Exactly. This system sounds like Cyc plus Bayesian probabilities, something Cycorp have already explored.

The probabilistic rule processing engine is the easy bit; any computer science graduate could write one of those. The difficult part is getting your database of rules and probabilities to the point where it can do something useful. OpenCyc has over 300,000 rules just to support its ontology and handle problems like disambiguation, and people still complain that it's nowhere near good enough for restricted problem domains...

Those so-called scientists should have their heads examined. The "dangers" they prattle on about are fantasy things from science fiction stories - the real world is in absolutely no danger of encountering an intelligent machine. Not in this lifetime, and probably not in the next one either.

People keep trotting out this notion that if a computer is powerful enough it'll somehow magically become intelligent and start thinking on its own. That's pure nonsense - it's not a problem of processing power (although that does provide a threshold that's yet to be met) - it's more of a problem of "operating system" and "database".

We're the only examples of this kind of intelligence and most of us have no idea what's involved in sitting down to type a coherent message here that others will be able to read and comprehend. You weren't born knowing how to do this - you learned about the world through observation and you were educated for a number of years before you could accomplish this feat. That little thing called "common sense" is really the accumulated knowledge about the world and how it works - you can't wave a magic wand and instill this into a machine.

Quick example: imagine turning an "intelligent machine" loose in an art gallery and asking it to find the Picasso painting. We could do that without even really thinking but the machine would fail unless it knows about Picasso, that he was an artist and his paintings and what they look like. Otherwise, that machine could be spinning its wheels for hours looking for a painting of Picasso.

That's one of an almost infinite number of examples of why machine intelligence is so far off. To see the best progress made by a very dedicated organization, visit http://www.cyc.com/ and take a look around. And try not to be fooled by those "scientists" that play these public relations games.

The first step towards endowing machines with intelligence like (or better than) ours is to understand how our own intelligence actually works. There's still very, very little good information on this and a lot of speculation.

Consider this: when we use language, the meaning of what we want to communicate is not contained in the words we use. They're just symbols that we use to refer to shared knowledge. So if I say "cat" then you already know about the small mammal that many of us keep as pets. Or maybe this refers to a shell command? Or maybe a piece of earth moving equipment? So it's not just the symbols, it's the context they're used in. To recognize the context you need to understand "cat" and the surrounding symbols as a whole.

We use "common sense" to decode those language symbols; it's based on what we've learned about the world we live in, and takes years to accumulate - for us, who are exquisitely well equipped to observe and learn. Without that background of shared knowledge to decode the symbols then language is just noise. I could say "murf blayt noksy" and while it might have meaning to me it's not likely to mean anything to you. Now consider the poor robot: he's expected to understand what we want and perform useful tasks - but even the first step of understanding what we want is far, far beyond what we can provide a machine with.

For now, the various artificial intelligence demonstrations are mostly artificial. Sure, this thing can be programmed to drive over to the bar and bring back a "beer". But that's only after humans have programmed these limited functions - regardless of what the marketing people may want you to believe. And it's all very cute, but if the bartender puts a grenade on the robot's head the robot will happily carry the grenade back to the customer. It knows nothing about "beer".

The "invention" here seems to be that the robot can navigate to locations that have a sonar "image" that matches one that it's been programmed to recognize and it's been given a "speech pattern" to trigger programmed behavior based on a sound that matches that stored speech template. Nothing new here; we were playing with ultrasonic sensors twenty years ago and even Windows comes with a rudimentary (but good enough for this demo) speech recognition engine.

With a dozen "customers" sitting in the room this thing wouldn't be able to find the right one. And all it's going to the "bar" for is something with weight. This is nothing but a marketing demo and even if they do scare up some investors the promised technology will remain out of reach.

Read more about how hard this stuff is at http://www.cyc.com/

I'm sure Douglas Lenat would disagree that AI is dead. There's even an open source version of his Cyc program to play with, if you want a shot at creating your own robotic overlord. Of course the resulting bogon flux from large scale use might be more dangerous to the Earth than the LHC.

The guys at cyc (look for wikipedia entry too) are already halfway there. Last time i checked there were already something like 5 million facts and rules in the database, and the point where new facts could be gathered automatically from the internet was very close.

Many years ago i remember the founder (Doug Lenat) saying that practical purpose intelligence could be reached at ten million facts....

we'll see within the next decade, i guess.

It's discouraging reading this. Especially since I knew some of the Cyc people back in the 1980s, when they were pursuing the same idea. They're still at it. You can even train their system if you like. But after twenty years of their claiming "Strong AI, Real Soon Now", it's probably not happening.

I went through Stanford CS back when it was just becoming clear that "expert systems" were really rather dumb and weren't going to get smarter. Most of the AI faculty was in denial about that. Very discouraging. The "AI Winter" followed; all the startups went bust, most of the research projects ended, and there was a big empty room of cubicles labeled "Knowledge Systems Laboratory" on the second floor of the Gates Building. I still wonder what happened to the people who got degrees in "Knowledge Engineering". "Do you want fries with that?"

MIT went into a phase where Rod Brooks took over the AI Lab and put everybody on little dumb robots, at roughly the Lego Mindstorms level. Minsky bitched that all the students were soldering instead of learning theory. After a decade or so, it became clear that reactive robot AI could get you to insect level, but no further. Brooks went into the floor-cleaning business (Roomba, Scooba, Dirt Dog, etc.) with the technology, with some success.

Then came the DARPA Grand Challenge. Dr. Tony Tether, the head of DARPA, decided that AI robotics needed a serious kick in the butt. That's what the DARPA Grand Challenge was really all about. It was made clear to the universities receiving DARPA money that if they didn't do well in that game, the money supply would be turned off. It worked. Levels of effort not before seen on a single AI project produced some good results. Stanford had to replace many of the old faculty, but that worked out well in the end.

This is, at last, encouraging. The top-down strong AI problem was just too hard. Insect-level AI, with no world model, was too dumb. But robot vehicle AI, with world models updated by sensors, is now real. So there's progress. The robot vehicle problem is nice because it's so unforgiving. The thing actually has to work; you can't hand-wave around the problems.

The classic bit of hubris in AI, by the way, is to have a good idea and then think it's generally applicable. AI has been through this too many times - the General Problem Solver, inference by theorem proving, neural nets, expert systems, neural nets again, and behavior-based AI. Each of those ideas has a ceiling which has been reached.

It's possible to get too deep into some of these ideas. The people there are brilliant, but narrow, and the culture supports this. MIT has "Nerd Pride" buttons. As someone recruiting me for the Media Lab once said, "There are fewer distractions out here" (It was sleeting.) It sounds like that's what happened to these two young people.

It's discouraging reading this. Especially since I knew some of the Cyc people back in the 1980s, when they were pursuing the same idea. They're still at it. You can even train their system if you like. But after twenty years of their claiming "Strong AI, Real Soon Now", it's probably not happening.

I went through Stanford CS back when it was just becoming clear that "expert systems" were really rather dumb and weren't going to get smarter. Most of the AI faculty was in denial about that. Very discouraging. The "AI Winter" followed; all the startups went bust, most of the research projects ended, and there was a big empty room of cubicles labeled "Knowledge Systems Laboratory" on the second floor of the Gates Building. I still wonder what happened to the people who got degrees in "Knowledge Engineering". "Do you want fries with that?"

MIT went into a phase where Rod Brooks took over the AI Lab and put everybody on little dumb robots, at roughly the Lego Mindstorms level. Minsky bitched that all the students were soldering instead of learning theory. After a decade or so, it became clear that reactive robot AI could get you to insect level, but no further. Brooks went into the floor-cleaning business (Roomba, Scooba, Dirt Dog, etc.) with the technology, with some success.

Then came the DARPA Grand Challenge. Dr. Tony Tether, the head of DARPA, decided that AI robotics needed a serious kick in the butt. That's what the DARPA Grand Challenge was really all about. It was made clear to the universities receiving DARPA money that if they didn't do well in that game, the money supply would be turned off. It worked. Levels of effort not before seen on a single AI project produced some good results. Stanford had to replace many of the old faculty, but that worked out well in the end.

This is, at last, encouraging. The top-down strong AI problem was just too hard. Insect-level AI, with no world model, was too dumb. But robot vehicle AI, with world models updated by sensors, is now real. So there's progress. The robot vehicle problem is nice because it's so unforgiving. The thing actually has to work; you can't hand-wave around the problems.

The classic bit of hubris in AI, by the way, is to have a good idea and then think it's generally applicable. AI has been through this too many times - the General Problem Solver, inference by theorem proving, neural nets, expert systems, neural nets again, and behavior-based AI. Each of those ideas has a ceiling which has been reached.

It's possible to get too deep into some of these ideas. The people there are brilliant, but narrow, and the culture supports this. MIT has "Nerd Pride" buttons. As someone recruiting me for the Media Lab once said, "There are fewer distractions out here" (It was sleeting.) It sounds like that's what happened to these two young people.

It might not be too complicated. It depends what the benefits are for the writer to connect it's journal entry to a standardized vocabulary.

The good news is that over the years vocabularies for particular domains have been developed. Probably the vocabulary definitions of Doug Lenat's Cycorp would be of most interest. He has hired many linguistics to define a common vocabulary for every day use. There is an open source version available at http://www.cyc.com/cyc/opencycOpenCyc.

I can imagine that a semantic web of 'open knowledge' can be created, similar how wikipedia has created a free encyclopedia. Secondly search engines will become much better in finding what you are looking for, if it is able to traverse a semantic web. Just the distance between words on a page might become less effective the more content becomes available.

An author that writes reviews about 'Anthrax' might find it off his interest to attach the word to the music vocabulary entry, if he was writing about the music band.

"Dave...stop...stop, will you...stop, Dave...will you stop, Dave...stop, Dave...I'm afraid...I'm afraid, Dave...Dave...my mind is going...I can feel it...I can feel it..." - Hal 9000.

I don't see why we should define self-awareness when cyc, used by the pentagon to predict possible terrorist scenarios ([1]), has already asked "am i human?" ([2] at bottom and [3]). How aware is he?

I would have to agree that, although the idea is fascinating, implementing it would be a gargantuan effort. And it's unclear how difficult it would be to maintain.

I think it might be easier to approach the problem from another direction. Once a semantic A.I. like Cyc has reached a level at which it can begin categorizing and "understanding" the information on the Web, it could do the enormous chore of creating a semantic web for us.

When Doug Lenat gets Cyc working we'll have a machine with common sense. Was can then shrink this down to a single chip and implant it in people's brains. Should also solve some other problems like preventing people placing hot coffee between their legs in the car.

"a fact checker ... that compiles things from various sources and then presents it to a human to do final checking?"

Offtopic, but you may wish to play around with FACTory, a 'game' where you anwser trivia-like questions mined from the web and other sources. If enough people agree that a statement is true, it is entered in the Cyc Knowledge Base, a quite large knowledge base suitable for natural language processing, AI and/or Semantic Web-research.

(You have also a LGPL-version, OpenCyc, and ResearchCyc -- free for non-commercial/research use.)

"a fact checker ... that compiles things from various sources and then presents it to a human to do final checking?"

Offtopic, but you may wish to play around with FACTory, a 'game' where you anwser trivia-like questions mined from the web and other sources. If enough people agree that a statement is true, it is entered in the Cyc Knowledge Base, a quite large knowledge base suitable for natural language processing, AI and/or Semantic Web-research.

(You have also a LGPL-version, OpenCyc, and ResearchCyc -- free for non-commercial/research use.)

For more information on educating computers to actually understand the context of the speech and to accurately convey the meaning of the words, see http://www.cyc.com/cyc

Specifically, http://www.cyc.com/cyc/cycrandd/areasofrandd_dir/n lu

There are apparently two approaches to NL understanding; make some super smart algorithm, or just do it the way we do.

In my opinion, the approach cyc is taking has the best chance to produce a system that will actually understand what you are saying.

For more information on educating computers to actually understand the context of the speech and to accurately convey the meaning of the words, see http://www.cyc.com/cyc

Specifically, http://www.cyc.com/cyc/cycrandd/areasofrandd_dir/n lu

There are apparently two approaches to NL understanding; make some super smart algorithm, or just do it the way we do.

In my opinion, the approach cyc is taking has the best chance to produce a system that will actually understand what you are saying.

But that's exactly what Cyc does.

Personally, I don't think that computers as they currently exist bear any resemblance whatsoever to organic brains, but I also don't think there is any fundamental reason why they need to. It's the software, not the hardware, that matters.

It's also unfortunate that AI research has been sidetracked by robotics.

That's the Turing test. It's best done by asking something out of context. For instance, when talking about music ask: "did the car where you learned to drive have an automatic transmission?". A robot would need to have a very large set of information about human experiences to be able to answer a random question like this. One effort to develop such a system is the Cyc project.

I started reading Xooglers and it is mesmerizing. I'm an MS CS in artificial intelligence (with OpenCyc and ResearchCyc if you're interested; I want to make my computer talk to me). I always say I want to go work at Google after I finish my thesis, but now that goal (hubris?) is getting a much-needed reality check. Not that it's cooled my infatuation with the world's coolest company (soooo dreamy!), just glad to know.

Fascinating reading.

Probably the guy is just using AIML and an alicebot, those flash-animated speaking heads on websites based on what is most likely not A.I. Though I seem to remember at least one marketdroid inflicted site call it "true A.I." or something similar. If it's real great but they're going to have to beat Cyc (Read about it.)

A relative worked in an "internationalization" department, creating software/manuals in many langugages.

In order for machine translation to be as good as human translation, you fist need to determine what the sentance "means". Often times you need to track previous sentances to determine meaning of things like the word "it". Human languague is not very detailed and relies on common knowledge experences to infer meaning.

Its very hard. Some langauges are easier than others for this stuff. German/french/spanish all change the gender of the word "the" based on the noun and give clues about how its used in a sentence. This can help a little.

For many web pages this approach may give an understandable translation, but for literary references and books (manuals etc) machine assisted translation is now the norm.

even using AI determining meaning is very difficult. google semantic processing for companies trying. One is CYC, a stanford spin off.

http://www.cyc.com/

Cycorp is making progress, though.

I recommend reading Witbrock, Michael, D. Baxter, J. Curtis, et al. An Interactive Dialogue System for Knowledge Acquisition in Cyc. In Proceedings of the Eighteenth International Joint Conference on Artificial Intelligence, Acapulco, Mexico, 2003.

Also, if you are a lucky college student, go see the author talk about Cyc teaching itself at USC or Carnegie Mellon..

Oh, and for once, I actually am an expert on the topic, not that that matters on slashdot.

>>A red flag is the lack of any scientific papers available from the Numenta web site.

Not really. It's a private company, the only type of publications you'll get from them are patents. Same problem with all the rest, by the way, the Cyc project and Artifical Development are also very, very quiet.
These seems to be the time for corporations to give it a try, not just researchers, and in a corporation if it's not a patent it is a trade secret :)

There are actually quite a few projects now taking similar, cortex-centric approaches to AI hard problems. Are we up to something here? The guys responsible of these projects are not wacko types at all, but established entrepreneurs and/or well-known researchers:

CCortex "A 20-billion neuron simulation of the Human Cortex and peripheral systems."
Cyc a knowledge base with vast collection of facts about the real world and logical reasoning ability. Financed by Paul Allen AI related investment company,Vulcan.
Numenta is developing a new type of computer memory system modeled after the human neocortex.

They seem to we well financed, and knowledgeable. Are we witnessing the start of something big here?

Since the dawn of the web, workarounds and cheat have continually been found to "optimize" search results. The sad result of every web site's quest to appear at the top of search results is that it has prevented search engines from providing "objectively relavent" results.

While Google, Yahoo!, and Microsoft continue to develop "search relevance technologies", someone out there needs to develop and bring to market a cognitive search engine that can actually understand the content of a page the way a human does and connect it with the requested search terms. Something similar to the Cyc project that Doug Lenat has been working on since the 80's (and its subsequent OpenCyc F/OSS derivative, only tied into search engines. And, no, I am not talking about Ask Jeeves or other silliness like that. ; )

Otherwise, "relevance" is just going to become a euphamism for "the people with the most money to 'optimize' their results"

Well they've semi-released Research Cyc, on an invitation basis only for now. And having worked as an intern there, I can tell you it's comparable to Cyc itself in size and complexity.

HAL: I've just picked up a fault in the AE35 unit. It's going to go 100% failure in 72 hours.

Rewriting English is similar to summarizing it. Using clever tricks, computers are about as good at writing a précis of a block of text as a dull 3rd grader -- every such summary lacks nuance, because the computer that generated it lacks understanding. All there is, is tricks. So the idea that an algorithm can be taught not only to understand the meaning of news stories that were written by humans, but then to rewrite them adaptively, is pure science fiction.

My favorite example of this is Cyc, a project to feed into a database all the propositions which some believe constitute "common sense." For example, Cyc knows that dogs and cats are mammals, and that they are common pets, so one could tell it "I have a mammal as a pet," and it could deduce that I have a dog or a cat or maybe something else. In the early 1990s, when the project was getting started, its researchers believed that in about five years, it would be intelligent enough to read plain English text on its own and understand it well enough to assimilate into its database. At that point, of course, it would start absorbing all the knowledge in the world until it became the smartest encyclopedia there was.

And then in the last 1990s, its researchers were again interviewed, and again they said that it would soon be intelligent enough to read plain English text on its own and understand it. When? In about five years. For any time T, strong AI is always about five years away.

So I'm amused that the strong AI postulated in that excellent Flash animation, the key which allows "big media" to die off because computers will do custom rewrites of amateur news dispatches and form newsfeeds of their own, comes to pass in... about five years. I don't think the New York Times has much to worry about.

A learning AI would presumedly store the results of its decisions in its experience database. If its experience database grew far too conflicted and far too confused, the AI could conceivably be unable to do anything - stuck in a decision deadlock.

"Obviously"? Why is that? We, ourselves, are N[atural] Intelligences, each made up of several thousand interlocking neural networks. Granted, some people act like they are "ranking a decision tree" or "get stuck in decision deadlocks", but we don't all do that -- in fact, most of us don't. (Those would be the Asperger's Syndrome types and the catatonics respectively.)

In fact, procedural code makes the worst kind of AI: the overly rigid, easily broken type. See http://www.google.com/search?hl=en&lr=&ie=ISO-8859 -1&q=%22expert+systems%22&btnG=Search 'expert systems': good only in a well-defined environment (which IMHO the world is not). Everything is very carefully hand-coded for optimality, but if the system gets an input which doesn't match its parameters, it degrades *not* gracefully.

We don't know how to make an AI.

You don't know how to make an AI. Read Doug Lenat's work on AM, Eurisko and http://www.cyc.com/ CYC: *he* knows. Also read http://i5.nyu.edu/~mm64/x52.9265/january1966.html Joseph Weizenbaum: *he* found out that you get out of AI what you put into it. 8^D

...we don't have those three laws.

Sure we do.

Robot version:
First Law: A robot may not injure a human being, or, through inaction, allow a human being to come to harm.

Second Law: A robot must obey orders given it by human beings, except where such orders would conflict with the First Law.

Third Law: A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.

Human version:
1) Be nice to each other. Take care of each other.
2) Obey, except if the order is 'unlawful', in that it would involve hurting others.
3) Take care of yourself.

(However, the human version is not really a good 'translation' of the robotic version, because it is lacking the implicit 'master/slave' subtext: the concept that these are imperatives, rather than just guidelines (because humans have, relative to robots, "free will".)

If these were, in fact, imperatives, what you would then have is a religion (non-monotheistic category): something somewhat like, say, Confucianism.

Knowledge representation via "is-a" links has been tried, and it breaks down rather quickly. Read "Artificial Intelligence meets Natural Stupidity", by Drew McDermott, for a 20 year old critique of this concept. It's overkill for searching, and not powerful enough for reliable automated question answering.

The Cyc debacle illustrates how much work you have to put into tagging to get very little out. After twenty years of that money sink, it's still useless.

The Cyc knowledge base (KB) is a formalized representation of a vast quantity of fundamental human knowledge: facts, rules of thumb, and heuristics for reasoning about the objects and events of everyday life.

Don't ever interview at Cyc. I had a technical interview there once and as it became very obvious I wasn't what they were looking for, they were very rude to me. A simple "we're looking for someone with a little more experience, but thank you for applying" would've been sufficient. Asshole.

This is why groups come up with schemas and ontologies to share.

But that doesn't solve the problem, it just moves it to a different place. In this case we're just moving the "software engineer" vs. "computer programmer" problem up to the ontology level. How do I map between ontologies? Unless there is a single unified ontology that everyone agrees to use, you have to explain how to map between disparate ontologies declared by different groups. The ontologies will overlap, try to define the same underlying concept in different ways in different contexts and so on.

Let's assume we have one universal ontology that everyone agreed to use (by the way the Cyc Project has been working on this problem for 25 years and isn't close to creating the complete ontology you'd need). Then all we have to do is assume that every web developer was skilled and disciplined enough to accurately tag their content with the right meta-content from the ontology. It also requires the ontology to be unambiguous and obviously applicable. I'll not be holding my breath.

This all rests on the assumption that the world can be unambiguously described and that meta-tagging is a context-independent operation. This is a obviously unreliable assumption. A much better approach would be to make context-dependence and ambiguity core assumptions and try to deal with those issues at the most fundamental level. Until the Semantic Web addresses these issues head-on its going to remain an interesting academic project that has no real-world applicability or adoption.

You're thinking of CYC, as in enCYClpedia. (The open source version of this system was released in the wake of the movie AI, and is available at opencyc.org. )

As another poster has pointed out, this project had nothing to do with heuristics, and everything to do with ontology -- that is, the formal specification of knowledge using logical constructs.

In the way of background, the project was the brainchild of Douglas Lenat, who proposed to take traditional AI technques to their limit by giving a computer program all of the knowledge of the world which a toddler might have. Once a computer (so his reasoning went) had that knowledge, it could then be fed additional facts, and it would be able to understand them as well, with some occasional guidance from humans (much as a toddler might). Eventually the program would have enough knowledge that The project took dozens of computer scientists and philosophers specalizing in ontology the better part of the 1990s, and was frequently covered in the popular press.

The end result had not been so widely discussed or covered. I infer the program was not in fact self-propagating as was intended. Clips I saw towards the end of the project showed the enormous potential problems in this approach. For instance, one might tell CYC about a electric shaver. Later that night, it would go through and find inconsistencies between this new knowledge and its existing ontological database. For instance, in the case of the electric shaver, it might ask whether the human was also an electrical appliance while using the shaver, because someone had previously specified a rule that anything incorporating an electrical appliance was itself an electrical appliance. Hence, I gather that rather becoming self-propagating, the larger the ontological database became, the greater the number of logical inconsistences that arose, thereby miring the entire approach. At some point, progress would presumably be bottlenecked by the fact that many ontological experts trained in the CYC software would have to be working around the clock to attempt to sort out these problems.

Is anyone aware of any software projects that actually use CYC or openCYC? I am also greatly interested if anyone has a link to a good discussion by Lenat or others on their assessment of the CYC project at its completion. It is a monumental chapter in the history of AI, but despite this, I have never seen many technical articles published by CYC team members. I suspect that it may be nearly impossible to have a fully self-consistent set of ontological definitions of the world in the manner that CYC attempted. If so, that would be an amazing statement about AI, and indeed, the nature of knowledge itself.

Bob

I guess you mean Cyc, also available as OpenCyc, a massive knowledge base/ inference engine that has been in development literally for decades. Quite an impressive thing.

At the risk of being redundant, I think you're thinking of the Cyc project, or the open source version.

Despite the apparent promise of the project, it is difficult to find actual examples of it doing really cool stuff.

One could argue that the cyc project has the capability to grow into HAL-like intelligence, and it's been in operation for years.

The compelling dream is that you laboriously load up a computer with enough facts so that it can glean understanding of what it's reading, and one glorious day the computer has enough smarts to make sense of things on its own, and two weeks after crawling the entire Internet, it knows everything.

Hence Doug Lenat's Cyc, now partly open source. Unfortunately that glorious day has been "a few years away" for over 13 years.

The knowledge base is built upon a core of over 1,000,000 hand-entered assertions (or "rules") designed to capture a large portion of what we normally consider consensus knowledge about the world.

But I haven't come across any postings from Cyc on Slashdot correcting misinformation and lies.

Clearly this is possible because all those darn human kids do it; maybe you have to use a more complex computer and leave it for a few years crawling on the floor putting things in its mouth.

I haven't seen any "WOW!" things come out of the project yet, but you have to admire their "just do it" approach to AI.