Yeah it is hardly low profile. It was a very well funded project that promised the moon and failed to deliver despite absorbing a massive amount of DoD money. A lot of AI researchers at the lab I worked at weren't big fans as there was the sense that Cyc was a shining example of how money spent on symbolic AI was just wasted.
Doug Lenat is, to me, the most confusing case of someone in AI. He programmed a general problem solver (Eurisko) to use (meta^n)-heuristics that solved a major strategy game, coming up wih a creative plan no human thought of, and yielding insights on the field of heuristics and the "Representation Language Language".
... and then "went dark", "officially" working on this tedious, brittle attempt to compile common sense into a graph, which yielded nothing in practical application.
It's like we're in a movie and we're about to get a big reveal that he's really been using Eurisko to solve major untouchable problems.
Since when people call him "Doug"? :) Eurisko had some amazing accomplishments, but some questioned their authenticity saying Douglas and students help it a bit. Anyway, Douglas Lenat is my all-times idol and I hope this prolonged "stealth mode" was the intermission of something great. Isn't Freebase.com his project as well?
No, that was Danny Hillis. Freebase was purchased by Google and became the core of what is now known as the Google Knowledge Graph.
Interestingly, Ramanathan Guha[1] was originally involved in the Cyc project and is now at Google. He was also instrumental in the creation of RDF, the format central to the W3C's Semantic Web and LOD efforts.
I confused OpenCyc with Freebase, which was linking to OpenCyc at least in the past. I know Freebase got acquired, but thanks for the clarification! It's really nice that Google is aggregating a lot of AI power. Combined with the robotics companies they've acquired, too, the direction is clear.
No it hasn't. Cyc has been widely reported on outside of academic circles for decades. It last got a lot a buzz when Watson was preparing to debut on Jeopardy.
If you read "Why AM and Eurisko Appear To Work" [1] and the precursor papers it becomes clear that they relied heavily on humans annotating what counted as an "interesting" concept, both for training and for extraction of concepts to publicize. Not necessarily a bad thing (every supervised algorithm need supervision) but it was a little hype-y ("my robot has independently discovered natural numbers!"). They also ran into major tractability problems as the heuristics got more involved.
Under the hood it was basically genetic algorithms over a meta-object protocol to extract well-scoring arrangements from ontologies, which is damn interesting, but combinatoric complexity bites you every time.
Kenneth Haase published a couple of papers dealing with some of the issues in more depth.
30 years in the cooler? I would be very afraid of finally revealing 30 years of hard work, only to find out it's irrelevant now...or that we pursued in the wrong direction.
Does not seem to be a very good way to go about revolutionizing any technology...
Ha, near secrecy. Sheesh, Lenat is a well-known figure in the AI world and Cyc is a very well-known project that has been written about for years. I don't care if the BusinessInsider author is what, 28 years old, he should do some decent reporting and not use such a stupid headline.
The sensational headline seems to have worked. I'm not sure if I've heard of this specific project, but to me it didn't sound all that novel. It reminds me of the semantic web.
Is it? How do we know? You've been working on a product "in stealth" for 30 years, shown practically nobody outside your company, and provided us with nothing more than a vaporous description of its capabilities (in Businessinsider, no less). I'm left wondering not only how seriously I should be taking this claim, but also exactly what the claim even is.
Artificial intelligence is hard. I did an internship at Numenta [1], where Jeff Hawkins is approaching AI from this same biological-first model. He hypothesized how a small subset of the brain works, constructed a model of it, and admirably hired dozens of engineers to build it. The guys behind the Numenta software are some of the best engineers I've met, with combined centuries of experience, and it's taken them almost ten years to get the software to its current, extremely primitive state. Right now, the model is implemented, and you can use the API to apply it to specific applications (predictive analytics, anomaly detection). But we are a long way off from the capability of applying it to "general input."
The fact is, AI is not going to impress anyone until it can handle an equivalently general class of inputs to what a human can handle. And think about the inputs that humans receive. The five senses are just the beginning. They're the concrete inputs we receive from the world. In addition to them, there's a near infinite class of more abstract inputs that build on top of them. As humans, we interpret social cues, hormonal feedback, the emotions our body often inexplicably generates, and dozens of other "second-order" inputs that are derived from our fundamental sensory ones, but seem equally fundamental to us.
Ok, sure. Maybe you can model the brain, and maybe it can respond to inputs from its sensory environment. But those inputs are the absolute lowest building block of the totality of input into our brain. We have an entire subconscience dedicated to processing those fundamental sensory inputs, and generating derivative ones for our consciousness to process. The difficulty in AI, in 2014, is modeling that subconscience. How do we go from basic environmental input to its infinitely more complex derivatives? How do we build a subconscience that our current AI can interpret as an input of its own?
We are a long way off from this capability. Nobody is going to figure it out on their own, and neither is a company of a few dozen. To me, it seems incredibly wasteful to spend 30 years working on this product without revealing any of the journey. After all this time, what do you have to show for it? Basically nothing. Your AI can apparently complete tasks of a similar complexity to state of the art that's been developed in less than a decade.
I want to see models, and I want the community to discuss them, problem solve with them, and build on them. Enough of this closed source, proprietary, snail's pace AI development.
Software is constrained by computing limitations. Fundamentally, there can be no computational model that even approaches the complexity of a brain. At least not today. So why not forget about trying to build Skynet, and work with the community to further the field together?
> The fact is, AI is not going to impress anyone until it can handle an equivalently general class of inputs to what a human can handle.
I don't know, I would be happy with just a dog brain or even a spider brain as a starting point. Or even a billion ant brains (simple individually, complex collectively).
There is so much hostility in AGI research. Everyone always arguing that everyone else is doing it wrong and that their approach is the One True Path. Progress is slow. When it happens it's debated if it's actually progress, or another step towards a dead end.
The only real feedback is failure. One camp goes a long time without producing an AGI. That's considered "proof" their approach will never work and that the New Idea will.
I'm just going to make a fictional analogy. It doesn't prove anything, but I hope it illustrates my frustration:
"Look at these idiots trying to build artificial flight by putting together large piles of feathers. Can't they see feathers are just pieces of a much more complicated puzzle?
And over there another group is dropping objects off buildings. Trying to find things that can stay in the air the longest. Even if they find something, it will never actually fly, just fall slightly slower.
There are people inflating balloons with gasses. Have you ever seen a bird made of balloons? It's absurd. Maybe eventually they will get something that hovers off the ground. But it will never fly gracefully like an eagle. It will never fool pigeons into thinking it's another pigeon.
One group has managed to make really strong artificial wings. They have gotten them stronger and faster than even real birds. But they need an entire steam engine to power them. And they don't even use feathers.
My group is going to build an artificial bird from the ground up, based on real biology. Teams of mechanics are working on designing tiny joints and artificial muscles. We have a prototype that can flutter across a room. At this rate we may only be decades away from true artificial flight."
That's probably where flight research would currently be if a few of those approaches hadn't achieved stunning success. Those successes would have silenced any groups that were holding out for artificial birds.
Also: "Cycorp also offers a complete version of Cyc, including many more assertions and additional NL capabilities, under a ResearchCyc license at no cost for research purposes."
I used that during my BSc. Thesis, the non-free version includes data that is of higher precision and more useful for the military use. Which they have been building this for. It was used to answer questions regarding current threats and find out which other threats nobody has thought about could also occur, where, by whom etc. You can research that, if you want. But one thing is true, the information on them is scarce. I think they have had held a talk at Google though, if I remember correctly.
Tl;Dr.: Cyc is military precise, OpenCyc is not. Use case: Terror-Cell and threat identification, Information Gathering, Reconnaissance, Data-Fusion
It never made sense to me that the whole process would be manual. I would've developed an AI that could use their "complicated and cumbersome" forms automatically based on "Speech or Written" Input
There is also: http://www.larkc.eu/ and many other alternatives "Expert Systems". I heard the the military version of http://clipsrules.sourceforge.net/ is pretty good and in use here and there. But I don't know of the current progress and use of such systems.
Can someone involved or knowledgeable give us an update on the state-of-art in AI/Expert-Systems used by the mliitary? I like to stalk military technology based developments =)
This thing doesn't model the brain, despite what the article claims. It's not about a brain model, it's about the data and its model. They're painstakingly adding lots of data about the world, which could serve as a resource for a brain-like model.
It's also neither closed source nor secret. The article is a bit crap so don't base your criticism of the project just from what you read there.
It's been popping up for ages, for example Businessweek, 1997:
At its core, Cyc is a logical system based on nth order predicate calculus. However, it is able to leverage other AI techniques to populate its knowledge base.
If you'd like to talk with a current intern, email interns@numenta.com and the person running the program can give your contact info to the current group.
When I read this story the first thing that sprung to mind was Ted Nelson's Project Xanadu, described by Wired magazine as "the longest-running vaporware story in the history of the computer industry".
As any entrepreneur will tell you, it is never a good idea to have long-running projects with few public deliverables. Projects like this need new blood, new ideas and continuous user validation in order to remain relevant.
I tested OpenCYC three years ago while working for a startup that did semantic tagging & recommendation of text-based content.
Essentially, we would take something like:
Gov. Rick Perry has said he will no longer wear cowboy boots, which some believe is part of an attempt to soften his gunslinging image as he considers another run for president.
Let's dive into what you get from OpenCyc vs DBpedia (ontology sourced from Wikipedia).
DBpedia:
Tons of machine-readable information like Party, Alma Mater, birthday, spouse, etc. Extensive categorical links (dcterms:subject) like category:United_States_presidential_candidates,_2012.
OpenCyc:
Knows he's a politician affiliated with "Republican Party," "Democratic Party," "The Republican Party," and "The Democratic Party"
I highlight only one example, but this is all over the place. Opencyc has duplicate terms where dbpedia/wikipedia does not. Opencyc has far less information. Opencyc has more incorrect information.
This is inevitable when you consider the two approaches. Wikipedia has tens of thousands of people making connections and updating the resource, where opencyc relies more on scripts. Opencyc, and quite possibly Cyc, is already antiquated by Wikipedia.
[1] Note that the confidence was estimated from heuristics I wrote based on how the ontologies were put together
[2] OpenCyc typically does not have mappings for plurals, while wikipedia has a very convenient redirect system for string mappings
[3] OpenCyc matches President with 60 concepts. Too much noise to do anything with
I am not surprised by your results. However, the system is more robust than your experience would indicate.
OpenCyc is a subset of ResearchCyc, which itself is a subset of (Full)Cyc. OpenCyc is primarily used for mapping between ontologies. It contains 239k concepts from ResearchCyc, but only the basic rules for definitional relationships between them. These relationships include part/whole, disjointness, etc.
You mention DBPedia as being superior for your purpose, but I would counter that the two are complementary. There is a mapping between DBPedia and OpenCyc within the Linking Open Data cloud. In fact, it was one of the first ontologies contributed to the W3C's LOD initiative[1][2].
The concepts in OpenCyc are rigorously organized from most general (e.g. Thing) to more specific (e.g. board game). Each concept may have specific instances (e.g. Yahtzee, Trivial Pursuit, Scrabble, etc.) These primitives all live within a custom Lisp, where they may be reasoned over. DBPedia's structure arises naturally from user activity. It is organized primarily by Wikipedia's category system and includes individual pages.
Unlike Wikipedia, the Cyc project does not aim to contain every instance of a concept. The relationships between concepts are what matter. Once one knows that something belongs to a given Cyc concept, one can leverage the system's knowledge to reason about it.
OpenCyc's reasoning capability is limited by a lack of assertions (facts and rules) -- ResearchCyc's is not. ResearchCyc contains over 5 million assertions not present in OpenCyc. (Things like: water is wet, a dog is a mammal, mammals have hair, etc.) It also contains Natural Language tools not present in OpenCyc: parsers, taggers and more. With these tools, one can go from natural language to a formal logic representation. Or, given a formal representation generate natural language. These capabilities exist today in real world applications[3][4].
Funny story - when I first started using Twitter a few years back, I had tweeted about some AI news. Pretty soon after that , I had a new follower that was trying to have a conversation with me - it was retweeting stuff randomly and sending some weird replies to me - it all seemed quite odd. Turned out that it was @cyc_ai - the Twitter handle of the the Cyc AI system, presumably trying hard to emulate a person - but failing unfortunately!
If you look up their reviews by ex-employees you will find they are extremely negative, talk about lots of workplace bullying, insufficient resources and a vicious upper management that doesn't respect their workers.
Human level artificial intelligence is a distance dream but can't say, virtual assistants like Google Now (http://www.google.com/landing/now/) and Braina (http://www.brainasoft.com/braina/) are already doing a decent job. May be we have to wait for 10 more years to achieve what is called AI-Complete or Strong AI.
This is pretty funny to me, because Cyc is actually a pretty high-profile project. Cursory searches on the topic of knowledge engineering return links to material on Cyc, not to mention that older AI textbooks mention Cyc as an example of a nascent large-scale knowledge engineering project.
45 comments
[ 7.3 ms ] story [ 110 ms ] threadAnd here is how that page looked in 2003: http://web.archive.org/web/20031204204518/http://www.cyc.com...
Quite a few of the publications are missing from most recent one. It is also worth noting that news page link was broken from 2003 until recently.
Then again, Business Insider...
... and then "went dark", "officially" working on this tedious, brittle attempt to compile common sense into a graph, which yielded nothing in practical application.
It's like we're in a movie and we're about to get a big reveal that he's really been using Eurisko to solve major untouchable problems.
One datapoint: http://archive.wired.com/wired/archive/2.04/cyc-o.html
So for more than 20 years.
He doesn't seem to mind:
http://www.cyc.com/about/team/doug-lenat
No, that was Danny Hillis. Freebase was purchased by Google and became the core of what is now known as the Google Knowledge Graph.
Interestingly, Ramanathan Guha[1] was originally involved in the Cyc project and is now at Google. He was also instrumental in the creation of RDF, the format central to the W3C's Semantic Web and LOD efforts.
[1] http://en.wikipedia.org/wiki/Ramanathan_V._Guha
Under the hood it was basically genetic algorithms over a meta-object protocol to extract well-scoring arrangements from ontologies, which is damn interesting, but combinatoric complexity bites you every time.
Kenneth Haase published a couple of papers dealing with some of the issues in more depth.
[1] http://eksl.isi.edu/files/library/Lenat_Brown-1984-why-AM-an...
Does not seem to be a very good way to go about revolutionizing any technology...
Is it? How do we know? You've been working on a product "in stealth" for 30 years, shown practically nobody outside your company, and provided us with nothing more than a vaporous description of its capabilities (in Businessinsider, no less). I'm left wondering not only how seriously I should be taking this claim, but also exactly what the claim even is.
Artificial intelligence is hard. I did an internship at Numenta [1], where Jeff Hawkins is approaching AI from this same biological-first model. He hypothesized how a small subset of the brain works, constructed a model of it, and admirably hired dozens of engineers to build it. The guys behind the Numenta software are some of the best engineers I've met, with combined centuries of experience, and it's taken them almost ten years to get the software to its current, extremely primitive state. Right now, the model is implemented, and you can use the API to apply it to specific applications (predictive analytics, anomaly detection). But we are a long way off from the capability of applying it to "general input."
The fact is, AI is not going to impress anyone until it can handle an equivalently general class of inputs to what a human can handle. And think about the inputs that humans receive. The five senses are just the beginning. They're the concrete inputs we receive from the world. In addition to them, there's a near infinite class of more abstract inputs that build on top of them. As humans, we interpret social cues, hormonal feedback, the emotions our body often inexplicably generates, and dozens of other "second-order" inputs that are derived from our fundamental sensory ones, but seem equally fundamental to us.
Ok, sure. Maybe you can model the brain, and maybe it can respond to inputs from its sensory environment. But those inputs are the absolute lowest building block of the totality of input into our brain. We have an entire subconscience dedicated to processing those fundamental sensory inputs, and generating derivative ones for our consciousness to process. The difficulty in AI, in 2014, is modeling that subconscience. How do we go from basic environmental input to its infinitely more complex derivatives? How do we build a subconscience that our current AI can interpret as an input of its own?
We are a long way off from this capability. Nobody is going to figure it out on their own, and neither is a company of a few dozen. To me, it seems incredibly wasteful to spend 30 years working on this product without revealing any of the journey. After all this time, what do you have to show for it? Basically nothing. Your AI can apparently complete tasks of a similar complexity to state of the art that's been developed in less than a decade.
I want to see models, and I want the community to discuss them, problem solve with them, and build on them. Enough of this closed source, proprietary, snail's pace AI development.
Software is constrained by computing limitations. Fundamentally, there can be no computational model that even approaches the complexity of a brain. At least not today. So why not forget about trying to build Skynet, and work with the community to further the field together?
[1] www.numenta.org
I don't know, I would be happy with just a dog brain or even a spider brain as a starting point. Or even a billion ant brains (simple individually, complex collectively).
The only real feedback is failure. One camp goes a long time without producing an AGI. That's considered "proof" their approach will never work and that the New Idea will.
I'm just going to make a fictional analogy. It doesn't prove anything, but I hope it illustrates my frustration:
"Look at these idiots trying to build artificial flight by putting together large piles of feathers. Can't they see feathers are just pieces of a much more complicated puzzle?
And over there another group is dropping objects off buildings. Trying to find things that can stay in the air the longest. Even if they find something, it will never actually fly, just fall slightly slower.
There are people inflating balloons with gasses. Have you ever seen a bird made of balloons? It's absurd. Maybe eventually they will get something that hovers off the ground. But it will never fly gracefully like an eagle. It will never fool pigeons into thinking it's another pigeon.
One group has managed to make really strong artificial wings. They have gotten them stronger and faster than even real birds. But they need an entire steam engine to power them. And they don't even use feathers.
My group is going to build an artificial bird from the ground up, based on real biology. Teams of mechanics are working on designing tiny joints and artificial muscles. We have a prototype that can flutter across a room. At this rate we may only be decades away from true artificial flight."
Eh?
http://www.cyc.com/platform/opencyc
Also: "Cycorp also offers a complete version of Cyc, including many more assertions and additional NL capabilities, under a ResearchCyc license at no cost for research purposes."
Tl;Dr.: Cyc is military precise, OpenCyc is not. Use case: Terror-Cell and threat identification, Information Gathering, Reconnaissance, Data-Fusion
It never made sense to me that the whole process would be manual. I would've developed an AI that could use their "complicated and cumbersome" forms automatically based on "Speech or Written" Input
There is also: http://www.larkc.eu/ and many other alternatives "Expert Systems". I heard the the military version of http://clipsrules.sourceforge.net/ is pretty good and in use here and there. But I don't know of the current progress and use of such systems. Can someone involved or knowledgeable give us an update on the state-of-art in AI/Expert-Systems used by the mliitary? I like to stalk military technology based developments =)
[1] http://videolectures.net/coinactivess2010_witbrock_lkc/
It's also neither closed source nor secret. The article is a bit crap so don't base your criticism of the project just from what you read there.
It's been popping up for ages, for example Businessweek, 1997:
http://www.businessweek.com/1997/25/b353210.htm
We're looking for intern candidates for Fall 2014. Details for how to apply are at http://numenta.com/company/#team.
If you'd like to talk with a current intern, email interns@numenta.com and the person running the program can give your contact info to the current group.
As any entrepreneur will tell you, it is never a good idea to have long-running projects with few public deliverables. Projects like this need new blood, new ideas and continuous user validation in order to remain relevant.
[1] http://www.w3.org/2001/sw/sweo/public/UseCases/ClevelandClin...
[2] http://videolectures.net/coinplanetdataschool2011_witbrock_c...
Essentially, we would take something like:
Gov. Rick Perry has said he will no longer wear cowboy boots, which some believe is part of an attempt to soften his gunslinging image as he considers another run for president.
And map it to something more machine-readable:
Let's dive into what you get from OpenCyc vs DBpedia (ontology sourced from Wikipedia).DBpedia: Tons of machine-readable information like Party, Alma Mater, birthday, spouse, etc. Extensive categorical links (dcterms:subject) like category:United_States_presidential_candidates,_2012.
OpenCyc: Knows he's a politician affiliated with "Republican Party," "Democratic Party," "The Republican Party," and "The Democratic Party"
I highlight only one example, but this is all over the place. Opencyc has duplicate terms where dbpedia/wikipedia does not. Opencyc has far less information. Opencyc has more incorrect information.
This is inevitable when you consider the two approaches. Wikipedia has tens of thousands of people making connections and updating the resource, where opencyc relies more on scripts. Opencyc, and quite possibly Cyc, is already antiquated by Wikipedia.
[1] Note that the confidence was estimated from heuristics I wrote based on how the ontologies were put together
[2] OpenCyc typically does not have mappings for plurals, while wikipedia has a very convenient redirect system for string mappings
[3] OpenCyc matches President with 60 concepts. Too much noise to do anything with
OpenCyc is a subset of ResearchCyc, which itself is a subset of (Full)Cyc. OpenCyc is primarily used for mapping between ontologies. It contains 239k concepts from ResearchCyc, but only the basic rules for definitional relationships between them. These relationships include part/whole, disjointness, etc.
You mention DBPedia as being superior for your purpose, but I would counter that the two are complementary. There is a mapping between DBPedia and OpenCyc within the Linking Open Data cloud. In fact, it was one of the first ontologies contributed to the W3C's LOD initiative[1][2].
The concepts in OpenCyc are rigorously organized from most general (e.g. Thing) to more specific (e.g. board game). Each concept may have specific instances (e.g. Yahtzee, Trivial Pursuit, Scrabble, etc.) These primitives all live within a custom Lisp, where they may be reasoned over. DBPedia's structure arises naturally from user activity. It is organized primarily by Wikipedia's category system and includes individual pages.
Unlike Wikipedia, the Cyc project does not aim to contain every instance of a concept. The relationships between concepts are what matter. Once one knows that something belongs to a given Cyc concept, one can leverage the system's knowledge to reason about it.
OpenCyc's reasoning capability is limited by a lack of assertions (facts and rules) -- ResearchCyc's is not. ResearchCyc contains over 5 million assertions not present in OpenCyc. (Things like: water is wet, a dog is a mammal, mammals have hair, etc.) It also contains Natural Language tools not present in OpenCyc: parsers, taggers and more. With these tools, one can go from natural language to a formal logic representation. Or, given a formal representation generate natural language. These capabilities exist today in real world applications[3][4].
[1] http://lod-cloud.net
[2] http://lod-cloud.net/versions/2007-10-08/lod-cloud.png
[3] http://videolectures.net/coinplanetdataschool2011_witbrock_c...
[4] http://videolectures.net/coinactivess2010_witbrock_lkc/