Future Current

Ben Goertzel, AGI-08 organizing committee member and CEO/CSO of Novamente LLC, presented “AI and AGI: Past, Present and Future” during the First Conference on Artificial General Intelligence March 1st, 2008 at the University of Memphis. Artificial General Intelligence (AGI) research focuses on the original and ultimate goal of AI–to create intelligence as a whole, by exploring all available paths, including theoretical and experimental computer science, cognitive science, neuroscience, and innovative interdisciplinary methodologies.

The following transcript of Ben Goertzel’s AGI-08 presentation “AI and AGI: Past, Present and Future” has not been approved by the speaker. Audio and Video is also available.

AI and AGI: Past, Present and Future

This is a quote from an interview with Marvin Minsky by David Stork, which was included in a book called Hal’s Legacy in the year 2000. Those of you who know Marvin will know that he is not really prone to understatement–he is sort of an outspoken, direct sort of guy. I think he puts things very plainly and very clearly:

“No one has tried to make a thinking machine. The bottom line is that we really haven’t progressed too far toward a truly intelligent machine. We have collections of dumb specialists in small domains. The true majesty of general intelligence still awaits our attack. We’ve got to get back to the deepest questions of AI and general intelligence and quit wasting time on little projects that don’t contribute to the main goal.”

Now, that is actually a little stronger than even I would say. I wouldn’t quite say that no one has been working on making a thinking machine. I wouldn’t quite say that all the narrow AI projects are a waste of time. They have actually led to a lot of important and practical stuff. Nevertheless, I think that there is some truth underlying this rhetoric, which is that the original goal of the AI field has not gotten as much emphasis in the last couple decades as I think it should have–as I am supposing many of you think it should have. I look at this conference as one step along the path to really bringing new life to the grand old goals of the AI field, and refocusing our efforts as a community of researchers on building thinking machines.

If you look at the conference program, you may notice the title of this talk today was listed as “The Past and Present of AGI,” but if you know anything about me, you will know that the tense that preoccupies me most is actually the future tense. So, while preparing the slides I changed the title a bit. Now we have AI and AGI in the past, present and future.

Like Stan, I was originally trained and started my research career as a mathematician before diverging further and further toward AI. One habit you get into in mathematics is beginning by defining your terms. These definitions I am going to start with are not as precise as I would like. There actually has been some work on giving a mathematically precise definition of general intelligence, but we’ll make do with the qualitative definition for starters.

What do I mean by “artificial general intelligence? Roughly speaking, what we mean is the ability to achieve complex goals in complex situations using limited computational resources. Naturally, this definition just pushes the problem down to defining “complexity,” and there has been a bunch of mathematical work oriented toward integrating appropriate definitions of complexity into a definition of general intelligence. There has been a really great paper by Shane Legg and Marcus Hutter on this, just last year. For now, the qualitative definition should suffice.

 

What we really mean by this is a system that can operate on its own. It can navigate some kind of world on its own, it encounters things and has certain goals. It tries to understand how to achieve its goals based on what it perceives. If it encounters a new problem, that maybe was never even thought of by its creators, it can figure out how to come to terms with that problem in the context of this world on its own, without needing a human being to change its parameters or feed new data into it.

We are not talking about unlimitedly, infinitely general intelligence, necessarily. That would be a wonderful thing to have. It’s not clear that it is possible within the constraints of the physical universe. Human beings have general intelligence, but we are fairly strictly constrained in terms of what we can do. A lot of the brain is devoted to vision, motion, and social perception, but we do have an ability that is quite different than that of what we can think of as narrow AI programs, which is what the vast majority of AI software programs pursued by researchers today consist of. You can think of narrow AI as being software that carries out specific difficult tasks–tasks that seem to require intelligence when humans carry them out–but carries out these tasks in a very specialized and purpose-specific way.

I don’t at all want to put down narrow AI research, in part because I spend a fair bit of my time working on it in the context of my AI company Novamente. We do narrow AI work for various corporate and government customers in order to get paid, provide useful services, and help the world move along. Narrow AI technology has done a lot of great things. I’ll review a few of those as the talk progresses. On the other hand, I think it is important to understand it for what it is. If you build a program that carries out one specific task, and then when you change the problem definition a little bit, you then need to go back and change the program a little bit, you may be doing something really, really useful and important, you may be saving people’s lives–however, you are not creating a general intelligence, an autonomous mind that goes through the world interpreting it according to its own understanding and figures out what to do as it goes along.

The focus of this conference is on artificial general intelligence, not because we think narrow AI is not a worthy pursuit, but because we think that AGI is in large measure a separate pursuit which deserves a lot more emphasis than it has been getting in the AI community in recent times.

I’m going to go through some information on the past, present and future of AI and AGI. I will start with a brief recap of the past. I won’t dwell too long on that, because I imagine everyone here is already very familiar with the history of AI. I think it is worth looking briefly at where we have come from, so as to understand the nature of the trajectory we are moving along as we seek to develop AGI into the future.

The field of AI as presently conceived started maybe in the 1940’s, if you’re generous–the 1950’s and ’60s, as more generally conceived. It really started out with a lot of excitement. Going back even further, I think it was 1943 when McCulloch and Pitts wrote a great paper demonstrating that a formal model of the brain had universal computational power; that you could compute anything you wanted using formal neurons. That paper by McCulloch and Pitts was the beginning of general computability theory, of computation as a way of thinking about intelligence. From that early conceptual advance, we had things progressing rather quickly for awhile.

In 1959, Samuel’s checkers programs did what Deep Blue did many decades later–overwhelmed people by playing a game that seemed to require a lot of cognitive capability, and doing a mighty good job at it using pretty simple algorithms. This and a number of other early successes led a lot of people to think if you can beat these tough games that most people cannot play very well, using essentially a few lines of code, it’s just going to be another decade until we have obsoleted humans altogether and we can all go relax on the beach while the robots do all our work.

It turned out not to be quite that simple. There was certainly more impressive progress during the early days. In 1967, they had the movie 2001: A Space Odyssey, where we had HAL 9000 with human-level conversational capability, and human-level ethical capability as well. If you look at what was happening in the real world of research, in ‘69 you had Shakey the Robot at Stanford. It was not so implausible at that point to think that if we have a little robot that can do some simple perception, movements and basic problem solving by ‘69, then by 2000 maybe we will have a HAL 9000 or something even better and more useful.

The AI field started out with a bang. By the time I was old enough to really understand what was going on, things had gotten a little less excited. I was born in 1966 and started university in ‘82. My introduction to AI was reading Hofstadter’s book Gödel, Escher, Bach in the late ’70s. By that point in time, it had become apparent to more people that maybe there was a little more to the problem than had been suspected back in the ’50s and ’60s. There were a lot of paradigms coming about and a lot of great ideas, without a clear notion of what the right direction was.

In the ’70s we had Winograd’s work with blocks world, interpretation of English sentences for manipulating simple objects, which I think is an important direction in terms of integrating perception, action, cognition and embodiment. We had genetic algorithms and evolutionary programming coming to the fore in the ’70s. Doug Lenat is better known these days for the Cyc project, wrote a fascinating program called Eurisco in the early ’80s, which was a pool of heuristics, including meta heuristics for creating new heuristics. This created a lot of interesting inventions in the electronics area, and it won a game contest called Traveller that was made for humans. It kept beating the game so well by exploiting weird loopholes in the rules that they eventually essentially kicked it out of the contest.

In the hardware side, we had Danny Hillis, who was one of Minsky’s students, created something called the Connection Machine. I programmed one of those in the ’90s when I was in Australia. We had like 64,000 processors. That was a really cool programming paradigm because you felt like you were getting massively parallel hardware that works sort of like the brain works. So there was an awful lot of creativity and invention that was going on in this period, I would say without the same optimism of the early days. It was not quite clear what direction to go in.

The way I characterize the more recent period in the ’90s and the beginning of this millenium, there still has not emerged any consensus about what is the right path to AGI. On the other hand, we have seen narrow AI applications get better and better, doing some pretty impressive and useful things. The AI field has a bad reputation in some domains, which I have seen many times in working with AI in a business context, some people with a corporate mindset have the view that artificial intelligence has been promised forever and what has ever been done?

Actually, an awful lot has been done, although not that much of it has been in the AGI vein. Just to look at a few examples, everyone knows Deep Blue, which has finally conquered the game of chess, at least relative to human abilities. Go has not yet been conquered, which is an interesting problem, but the fact that we can beat grand masters in chess is pretty interesting. Something that not that many people in the general public know about is actually extremely important in terms of the real world we live in is the application of AI in the military domain. If you look at the Gulf War, whatever you think of the politics of it, the use of AI technology in the Gulf War was quite extensive. AIs were used as virtual copilots in airplanes and to carry out all sorts of low-level tasks–how do you pack cargo in a plane most efficiently? Most impressively, for the overall planning of the operations, all sorts of AI planning algorithms out of the traditional “Good Old-Fashioned AI” community were used.

I’ve heard it said by a number of people, although I do not have an official source on this, that the amount of money saved by the U.S. military during the Gulf War by AI planning technology exceeds the total AI research budget ever spent by the U.S. government on AI. It is a big application. Hopefully it is the kind of application we are not going to need too much of in the future of the human race, but it does demonstrate that a lot of this AI technology, even though not achieving human-level AI or having much generality, it can do a lot.

Another example of AI that is not always thought of as AI is Google, the search engine that most of us use probably multiple times a day. Here is a picture of Google’s first server. They have expanded a bit since then. People do not think of Google as artificial intelligence because it is not trying like Ask Jeeves to hold a conversation, but in fact Google has hired a boatload of AI PhDs. There is a lot of the same material that you learn in a university AI course being deployed inside Google’s algorithms for returning search pages to you, for placing advertisements on the side of the page, and for doing all sorts of other Google services.

What I would say about the ’90s and the beginning of this millennium, we have seen narrow AI technology come into the mainstream in a lowkey way, without even being called “AI.” The questions is how far does it go toward the grander goals of general intelligence, which is something I am going to explore a bit in the rest of the talk.

Another way to put it, we were supposed to have this in 2001. Instead, what we have is this:

Which is arguably more useful, and less threatening, but not quite as interesting from a scientific standpoint.

To sum up my own personal view of the past of the AI and AGI field, there are a lot of real-world achievements in AI. So many deep and fascinating ideas that you could spend your life exploring neural nets, evolutionary learning, logical reasoning, probabilistic logic, planning… it goes on and on and on. It is fair to say there is nothing yet remotely close to a consensus about what is the right path to achieving human-level AGI. I have often observed that there is at least as many theories about how to get to AGI as there are AGI researchers. I know I have at least two, myself. In many cases, visions of AGI have given way to successes in narrow AI. Of course, in many cases visions have AGI have given rise to narrow AI failures, but the successes are more interesting to pay attention to.

What is the present state? Where are we now?

McCulloch and Pitts recognized in the ‘40s that the brain could be thought of as a computing machine. We have come a long way: we understand more about neural networks. IBM has gotten a bunch of press since the Blue Gene supercomputer has been used to simulate some substantial chunks of the mammalian cortex. McCulloch and Pitts’ vision that the brain and computing can be thought of in the same way, that there are deep parallels to be drawn and insights to flow in both directions, from neuroscience and AI and back again.

That insight has been amply borne out. On the other hand, if you look at what has been done with the Blue Gene, it shows you both the strengths and weaknesses of the current state of knowledge in terms of the computational theory of the brain. We have great hardware, and we can simulate a big spiking neural net with a huge number of neurons sending activation into each other. On the other hand, we still don’t know how the neurons should be connected to each other in order to yield intelligent behavior. We don’t even know that about the brain of a cockroach, let alone a mouse or a human, so we are still a fairly long way from learning how the parts of the brain need to be connected together to make general intelligence, although we have a decent understanding of what those parts are, in a general sense.

At the present time, narrow AI continues to dominate the AI field. I have already mentioned Deep Blue and Google. It is worth looking both at their strengths and their shortcomings. Deep Blue beats me at chess, which is very easy to do. One thing that it cannot do though is understand a minor change to the rules of chess and adapt its play accordingly. Take fish or random chess, which is a variant of chess where you randomize the pieces in the back row, which Bobby Fisher introduced to decrease the role of memorizing openings in chess and make it more a thinking game rather than a memory game. You cannot take Deep Blue and say, “Here is a variation to the rules, now can you beat me?” You could reprogram it to do that, but then a human is supplying the generality and the context, so the human has to do the reprogramming. Deep Blue is brilliant at what it does, but is disappointingly brittle in some ways.

Google is a fantastic service. On the other hand, what happened to Ask Jeeves? You must remember that service from ten years or so ago. It was supposed to answer your questions. That kind of got downsized and they changed the interface because it could not answer very many questions very usefully. Google, being more savvy on the business side, has not really gone there. It has not really tried to put real question answering in their interface yet. It seems to go beyond the narrow AI algorithms that they are so good at.

The DARPA Grand Challenge is another interesting example. It is pretty exciting. I very much look forward to having every other driver on the road replaced with automated robot systems–it would do wonders for the Beltway in D.C., where I live. On the other hand, you can see the brittleness and the narrowness here. You take a program that controls a car in the desert; can that program control a motorcycle or even an ATV. It cannot, unless you go in as a human and fiddle around with the programming.

There is a step of figuring out how the environment is represented and what actions need to be taken in tuning the parameters of the system to the problem. This step is done by human intelligence. Once that is done, you get a system that can carry out actions in a narrow context fairly effectively. You lack the ability to generalize from one context to the other, which is a real shortcoming of all narrow AI approaches.

To dig a little deeper into the narrow AI and AGI distinction, I will spend a few minutes talking about language understanding. There are so many different aspects to the AGI problem, but digging into language for a few minutes, I want to look at some of the strengths and weaknesses of what we can do now with the kind of AI that underlies Google and other statistical language processing products. It is impressive to see that in many cases Google can be used as a question answering system. You can type a question in there and you get an answer, but there is also fairly strict limitations.

I ask it “How many years does a pig live.” The first answer is semi-off. It tells me how many years a guinea pig lives for. But eventually: “Some pigs live as long as 15 years.” “The lifetime of the average pig is considered to be 12 to 15 years.” That is really not bad. It is not extracting the answer and giving it to you in a sentence, although I bet they could do that with decent reliability. It is not doing anything like what the human brain is doing. It is relying on a completely differently structured knowledge base, but you cannot argue with the functionality.

Now, “How many years does a dead pig live?” What did we get? “PETA protests nude artist performing with dead pig.” “The only good pig is a dead pig: A Black Panther paper editor.” It doesn’t get it. This is kind of emblematic of the distinction between a narrow AI that is brittlely carrying out the same functionality of finding documents and a general AI that has some context for understanding what is going on.

Less humorously, “How many years does a pig live in captivity?” I was surprised at what a bad job we get here. “How long does a gray wolf live?” “How much does one gallon of water weigh?” You don’t have to be clever to bollocks the thing. If you do anything that is not oriented toward presenting keywords within intelligent orientation, you are really not going to get what you want. Part of the reason Google and other search engines like it work so well is that we humans have been so good at adapting our way of thinking to formulating appropriate sets of keyword queries. If you approach it naively, you get results that are much less useful than what we get in ordinary practice.

There are a number of people in academia and industry working at going beyond these problems. A company called Powerset based in Silicon Valley is using a natural language parser that came out of Xerox PARC to try to make something that does a bit of natural language question answering. You can type in “Who mocked Blair?” It can figure out you are talking about Tony Blair. It will find sentences like “William Hague, then leader of the opposition party, taunted Blair.” They have some synonymy in there. They know that taunting is similar to mocking; parodying is similar to mocking. That is going a little bit beyond keyword searching. It is pretty far short of general question-answering, but it’s getting there.

Right now, if you are in a particular domain, like questions about people, questions about companies, technology can do reasonably well. There is still that narrowness there, where if you take this system and ask it questions about genes and proteins in a biology context, it’s not going to be able to do it.

Just looking at the domain of language processing, which is a subset of AGI, what we see is that the narrow-AI language processing tools that we have today can parse sentences, but are not good at selecting what is the correct parse. If you give a current language-processing AI system a sentence to parse, it will come out with a Chinese menu of options. It will be two, five, a hundred different parses of the same sentence, all of which sort of make sense. How do you pick which one it is?

We are not good at semantic disambiguation. We are okay at figuring out the meaning of nouns and verbs that have multiple meanings. If you are talking about adverbs or adjectives, the algorithms do not perform all that much better than the algorithm of “choose the most popular sense in the dictionary.” There are a lot of shortcomings there. Disambiguating prepositions is close to uncharted territory. There has been a bit of work on it in the past few years, but there are no remotely good algorithms. Understanding a distinction like “I ate lunch with a fork.” “I ate lunch with my uncle.” If you’re a cannibal, that could mean something different.

There is reference resolution. “Bob and I walked to the store with Jim, then he told me to shut up.” Who is “he”? Mapping the he, she, or it back into a noun, again we have very crude heuristics for it. AI had not really grappled with it. You do not need it to make a search engine, but you need it to make a really intelligent system. Dealing with comparatives–”My cookie is bigger than yours.” If you say, “My cookie is bigger than your cookie,” then that’s all right. This is a very active research area. There is a lot going on here.

If you look at a regular sentence out of Wikipedia, this is an article on caffeine, there is just so much stuff that current NLP systems cannot deal with. They do not do well with parenthetical comments. Even “beans within a given bush”–well, who gave who the bush? When you run this through current language-processing systems, you are going to come across a dozen problematic issues just in a simple paragraph.

If you look at chatbots, there is something called the Loebner Prize. Turing came up with the idea that a sufficient condition for general intelligence would be the ability to emulate a human being in conversation. Now, how close are we to that? Loebner started a prize where each year people would try their chatbots out in conversing with human judges and see how well they do. So far, it’s almost a complete joke, in that the work done toward the Loebner Prize really has nothing to do with the mainstream of productive work in the artificial intelligence field. This is because the technology that we have now, in terms of serious AI research, is just so far from being able to handle ordinary conversation.

What I put up here is a chat I had with an Alicebot which is a bot that I believe won the Loebner Prize at least once. They are still doing basically the same things as the ELIZA program was, way back in the ’70s. They have some template rules that respond to what you say in a sort of sensible way. Then, when they don’t know what you are doing, they divert you to another topic.

I said, When someone talks to you, do you really know what they are talking about? “That does not happen very often.” What doesn’t happen very often? “What do you think it does?”

I was not trying to be complicated. I was saying what you would say to a four or five year-old kid, and they would get it. This basically uses diversionary tactics to try to get you to say something. There is no real understanding here.

There is a big question here. Can you take narrow AI language processing software and create a chatbot that knows what it is talking about? That same question arises not just in language processing, but everywhere. Can you take narrow AI robotics and make a robot that can carry out basic household tasks? Can you take narrow AI bioinformatics software and make something that can actually look at some biology data and tell you what it means, without humans being so heavily involved in the loop?

Generally speaking, the question is, ‘Can narrow AI incrementally lead to general AI?’ This is something that we do not really know the answer to right now. My own intuition is that it’s not going to. I do not think that just working on narrow AI applications and incrementally improving them is going to lead to general AI. I think that is actually a non-trivial lesson that we can draw from the history of the AI field. I don’t think it was obvious in the 50’s and 60’s. Back then, it seemed like doing narrow AI and just gradually broadening the scope could lead to a human, and something better. I think that what we have learned now is that narrow problems are susceptible to clever, tricky computer science approaches, which do not necessarily help you very much at approaching the problem of general AI.

You can use a bunch of analogies for that. One that has occurred to me is locomotion: the problem of moving on flat surfaces is solved quite well by wheels, but generalizing the wheel might not be the best solution to moving around on general surfaces. I think the same kind of principle holds over and over. Once you narrow the scope, you can use a trick. We in the AI field have become very good at making up clever tricks. It’s fun–you can think about something for months or years and get a solution that does something really cool. It’s kind of seductive because it is easier than making a thinking machine and you can have the satisfaction of achieving something quickly. On the other hand, I have a suspicion that it is not the right path toward making a thinking machine. You can transfer some insight from narrow AI to AGI, but it requires a lot of creativity. It is not direct or obvious. I think there are key aspects of AGI that do not arise from narrow AI whatsoever.

As I have already said, I feel like narrow AI is dominating the scene in AI research with a lot of practical successes and what I would describe as also a lot of bad theoretical failures in terms of the capability of narrow AI paradigms to really help toward AGI.

I would say the situation we are in now is one in which the overall context is a very good one for AGI. Although as a field we do not have a unified idea for how we want to proceed forward, I think we do have a situation in which we are poised to proceed forward a lot more rapidly than we have in the past. A lot of different things seem to be coming together and putting us in a position where we can make a considered attack on the AGI problem.

I think there are four different things that I would isolate as being really critical in terms of the technological and scientific context. The first of them is just better and better machines–Moore’s law and all its friends. The infrastructure that we have now is just so awesome compared to what we had when I first started. I am not even that old compared with many others in the field. I remember playing with punch cards on an HP machine at Rutgers University when I was a teenager… That was really annoying. It was even more annoying than dealing with Microsoft Windows, although not by that much.

Computers get faster and faster. You have compute clouds now, and if you can afford it, you can get massive networks of distributed Linux boxes to help you out with doing whatever. My own view is that right now, hardware is not the problem. I do not know for sure how much hardware it would take to make a human-level thinking machine, but I have a suspicion that we have enough, and even if I’m wrong, we will have enough in the next five or ten year. That is new, that hardware is not the problem. With the amount of RAM you could get in machines ten to fifteen years ago, I think hardware actually was a very serious problem.

Our understanding of how the human brain and mind works has developed a heck of a lot in the last couple decades with the advent of brain scanning technology. We do not know how the neurons are connected yet and what they do, but we do have a pretty good understanding of what is the cognitive architecture of the human mind by now. There was a lot less consensus on that ten or fifteen years ago. There has been steady, incremental progress in cognitive science, without any amazing revolution, which I think really helps to orient us toward thinking about AGI in the right way.

Robotics is progressing a lot–not as fast as I would like, because of the economics of the industry–but it is certainly way beyond what it was ten years ago. For a couple thousand bucks you can order over the internet robots that will help you experiment with arms, camera eyes, mobility and so forth. That could not be done without really expensive equipment not that long ago.

[Tape change]

The general trend should be clear. Here, [Ray Kurzweil] is looking at the best supercomputer available at any point in time.

Here he is looking at calculations per second per thousand dollars. However you slice it, you get the same kind of trend. What he is showing here is the processing power of the human brain–of course, that is quite contentious. Some people claim it is significantly less than that, because the human brain has so much redundancy and inefficiency. Others claim it is really more than that because we are actually using cosmic quantum gravity resonance voodoo. We do not fully know how the brain works, but Kurzweil has made at least plausible arguments for the human brain’s processing power being equal to computers according to Moore’s law around the early 2020’s.

How relevant that is to AGI is open to debate. My own approach to AGI, along with many others here, is not that closely based on the human brain, anyway, but more based on trying to exploit the particular strengths of contemporary computing infrastructure. It is still certainly an interesting data point to see.

The other thing I mentioned is our understanding of the brain. We don’t know how everything is connected, but if you look at the technology we use now to measure what is happening in the brain, it is pretty impressive. A lot of really fancy physics is actually played a role here, in terms of sending elementary particles through the brain. This is an MRI scanner.

You can get a snapshot of, say, in this case they took people that had been given oxytocin and people who had been given a placebo and showed them faces to see how they react. You can see what the difference is. The red shows which parts of the brain are getting a bunch of oxygen, meaning which parts are doing a bunch of activity. Right now, they can give you a global idea of which parts of the brain are doing what sort of activity. You can imagine once the temporal and spacial acuity of these methods improves, eventually you are going to be able to track the flow of thoughts through the brain as someone thinks them. Even if you are not pursuing a brain-based approach to AGI, this is certainly very interesting data to have.

Again, we can look at Kurzweilian curves–exponential improvement of brain scanning technology. This is the speed and resolution of non-invasive brain scanning devices–how much spacial and temporal acuity you get in a brain scanning device that does not involve cutting the head open. When you look at the details of what is happening, it is even more exciting than what these graphs suggest. Two months ago I read an article in Technology Review where at MIT they have used nanotechnology to make little bubbles with voltage sensitive dye, you can stick the bubbles into someone’s brain, and then when someone thinks you can scan the brain and see where these nanoparticles with dye inside them are traveling. That lets you look at the charge in the interstices between the neurons. It gives you much greater granularity than fMRIs. There is so much thought, passion and invention going into this, eventually it has got to help with AI and AGI.

The next thing on my list was embodiment, both virtual and physical. This shows a screenshot of an application that my colleagues at Novamente and I have been working on. We are using AI to control a virtual dog that lives in a virtual world. We are looking at dog-level rather than human-level intelligence with this particular application. The dog learns to sit, play fetch, find objects and do various tricks, but it is pretty cool that in this day and age you can get some graphic artists to make a virtual dog, set your AI up on some servers, and it gets perceptions from the virtual world, sends actions there and all of a sudden your AI is not just sending text back and forth, it is actually doing stuff in a world. There is a human controlling that human avatar and interacting with it.

You can envision various stages beyond that. How about a virtual parrot that talks? Let’s say you make parrots that can hold simple English conversation, but have some adaptive learning underlying it. If you release this into virtual worlds that have hundreds of thousands or millions of users, all those humans can be helping to teach the parrots to talk better and better. When the parrots do something stupid, they can hit it over the head. When it says something smart, they give it a little treat. This kind of ability to reach out to many, many people in a familiar, embodied context has a lot of power–potentially, power to get beyond some of the limitations of narrow AI language processing technology.

The difference between eating food with a fork, eating food with great passion, eating food with your uncle, all this is more apparent when you are eating food in an embodied context. You can see that the fork is something that I am holding and my uncle is something I am sitting near. I think there is a lot of potential there.

You have virtual babies, as well. This is actually the tamest virtual baby I could find in the Second Life virtual world. It says something about modern online culture. The potential to take a virtual child in a virtual world and teach it, talk to it, and bring it up is a powerful modality for taking AIs and helping them get smarter, and for engaging ordinary people in the process of helping to teach AIs.

One of the nice things here is that we have a growth rate that exceeds Moore’s law, which is the number of subscriptions to online massively multiplayer games. The amount of our brain cycles that we spend playing games on online worlds is impressive and rapidly escalating. What that means is both there are a lot of people willing to teach and interact with virtual online embodied AIs, and also that the amount of work going into making these virtual worlds get better and better is quite impressive. Just as with Moore’s law, we do not have to rely on people putting money into AI, we can ride on the wave of people putting money into other supporting technologies.

Physical robots as well, although not increasing quite as fast as virtual world and gaming technologies, we now actually have humanoid robots that can make a lame attempt at making robot soccer. This certainly was not the case ten years ago. I’m looking forward to the price of this kind of robot coming down so I can have one in my living room. I guess that will happen in ten or fifteen years, or so.

This is a shot of an automated unmanned vehicle controlled by James Albus’s AI technology from the National Institute of Standards in the D.C. area, where I live. If you go to Albus’s website, you can see his unmanned AI technology actually can control automated vehicles cruising around outside, doing reconnaissance. They can also do some self-defense. It is extremely useful in terms of saving human lives, because otherwise humans would be driving around in these vehicles out in the field. All of the engineering technologies are integrated together well enough now that you can get cameras, servos and memory on board, and it can all work. Albus has been at this for a couple decades and in the last five years it has finally come together to be operational.

The fourth thing I mentioned, in terms of the preconditions for AGI progress being impressively met in this day and age, is theory. This may be my all-time favorite paper for an academic paper. “The Fastest and Shortest Algorithm For All Well-Defined Problems.” There’s a catch, which is that there is a large arbitrary constant involved. What Marcus Hutter showed here essentially is that the problem of AGI is a problem of dealing with bounded computational resources.

In other words, if you define AGI as the ability to achieve some computable goal function in some environment, which is computably described, the problem of achieving that goal in that environment can be solved by an algorithm he made up, called AIXI. There is a problem with it, which is that it requires infinite computation power. On the other hand, he defined a scaled down version, AIXI^tl, where t is time and l is length. This works in bounded space and time resources, and does better than any other possible program up to an arbitrary constant. It is not a very impressive program in its internal operations, in the sense that it does a search in the space of all possible programs, find the best one, and use it, then gather new data from what it did and search the space of all possible programs again to find the best possible program based on its new data.

In a theoretical sense, what it shows is that what we are up against in AGI is a problem of resource restrictions. If you did not have to deal with space and time constraints, it takes a trivial program for general AI. What all these other things–virtual worlds, robots, understanding the brain, using better and better computers, and all of the algorithms and knowledge representations we create–what these things are about is basically working around the finite computational resources that our universe provides us with.

Summing up, what do we have now in 2008? Fast computers networked together, decent virtual worlds for embodied AIs, halfway decent robot bodies, an awful lot of AI algorithms and knowledge representations–which are useful, interesting and relevant to AGI, but needing some attention to make them helpful for AGI. We have a basic understanding of how human cognitive architecture is, a cruder but still useful understanding of brain structure dynamics, and we have a theoretical understanding of general intelligence under conditions of massive compute resources. Adding all this up, it seems that we are well poised to do what Minsky suggested in the quote that I opened up with, and really attack the main problem.

What are some of the open questions in AGI that all of us in this room can help to resolve? I am not requiring that we resolve them all during the course of this conference, but there are things that need to be resolved in order for AGI to get where it needs to go. Of course, a list of questions is somewhat arbitrary. There are more questions than can be given in an hour-long talk, but these are some of the ones that preoccupy me everyday.

The cognitive cycle as Stan has conceived of it, how do we integrate perception, action, memory, knowledge, learning, and so forth? None of us knows for sure. This diagram is from one of Stan’s papers. It basically is a box and line diagram showing short-term memory, long-term memory, perception, action. What is interesting is that you take Stan’s diagram like this, you take one of them from one of my papers, you take one of them from Minsky’s book and you can actually cross-link them fairly well. You can say, these diagrams are getting at the same thing, which is a breakdown of intelligence into components and how they interrelate. That kind of commonality, that common understanding of the overall architecture of the cognitive cycle really was not there ten or fifteen years ago. That is just a matter of the incrementally accumulating advances in cognitive science in understanding the mind.

How to get this nailed down is an open question. My feeling is that a number of us are on the right track here, but we need more empirical work to validate that. Another key questions has to do with knowledge representation. The way I like to pose the question is it appears, “Can we take abstract knowledge representation–like a formal neural net, a semantic network or a logic system–and serve as the foundation for an AI to create its own context-specific knowledge representation?” If you can do that, then you can get beyond the brittleness bottleneck that you see in the narrow AI field. Do you need to build a special representation for chess, a special representation for go, a special representation for checkers? Or do you have a system that has a general abstract representation, and then when confronted with a new game can build its own representation for chess, its own representation for checkers, and so forth–a kind of meta-representation that allows the system to build its own representation?

The kind of representation I use in my own work is a semantic/neural net that has nodes and links spanning multiple levels of abstraction from pixel and actuator movements all the way up to abstract concepts. The question is whether this kind of thing can really serve as a medium for creating context-specific representations based on experiential interaction.

Getting back to language processing, does an AGI have to learn language, or can all this stuff being done in narrow AI and linguistics really help? Can stuff like Google or Psych really learn to understand language, or is it a dead end? Take a sentence, “Is this useful for artificial general intelligence?” Run that through some of our own narrow AI NLP pipeline that we built at Novamente for consulting customers, and we get a bunch of nodes and links. There is a bunch of technology out there that you can use to take text and break it down into logical relationships. It is not yet clear whether that can really be used to help a system truly understand language.

Getting back to robots and virtual worlds, we do not really know yet what are the restrictions on a world so that it can be useful for an artificial general intelligence. It is pretty clear that the world we live in is useful for artificial general intelligence. If you look at what most robots developed by AI researchers see, it is a little different. They put them in a lab, which does not have too much going on it. They remove anything it can trip over, no kids running through throwing stuff around. It is not clear if a research lab is a good environment, although it is clear that this field outside is.

This virtual world here, it is not clear that there is the level of physics and control over the body to support artificial general intelligence there. I tend to think there is, but it’s an open question that certainly merits further investigation. In terms of logic as a foundation for intelligence, which is something that I have explored in a bunch of my own work, can logic serve as a scalable foundation for sensorimotor learning? My personal guess is no, but I don’t know for sure. We have actually done experiments using probabilistic logic to control an agent playing fetch and hiding things in a virtual world. The question is not whether you can use logic to do sensorimotor stuff. You can. The question is whether you can go beyond toy problems using logic for sensorimotor learning. My own inclination is that you may need to integrate logic with other stuff to do that. It is an open question which certainly merits deep further investigation.

How does neural learning relate to abstract formal models of learning? Again, we do not know yet. It would be nice if we could find out. One tantalizing parallel that I will talk about in one of my talks on Monday is that if you look at the integration between neurons and neural clusters A, B, and C, there are interesting parallels between what happens in the brain and what happens in a semantic network. That parallel has not really been nailed down. I’m hoping that it will be as AI and brain science advances in the next decade.

Finally, the question that is perhaps closest to the heart of my own research, my feeling is that the algorithms of narrow AI are wonderful, but every one of them is susceptible to horrifying combinatorial explosions, which makes them not work when the problems get big. What I think is that there is no one magical algorithm that gets beyond the combinatorial explosion problem and we are just waiting for someone to find it. What I think is that the way the mind works is that there are multiple algorithms integrated together and each of them has its own combinatorial explosion problem, but by integrating them together in the right architecture and the right vision of the cognitive cycle, then you can get different algorithms to quell and dampen each others’ combinatorial explosions.

That is my own guess of what is the beauty of what is happening in the brain is. We have all these different modules, each of which does something with a non-scalable exponential time algorithm, and they are hacked together in a way that lets them stop each other from blowing up combinatorially. Whether that is actually true is a question that I am working hard at investigating. If it is not true, then one needs a different approach–one needs an algorithm that scales on its own, which may be the way we get there, too.

Looking toward the future, how can we guide a successful future for the AGI field? Obviously, conferences like this play a role. We had a workshop a couple years ago, and now we are doing a broader conference. There has also been a host of workshops focusing on human-level AI, general intelligence, integrative intelligence and so forth. I think more and more of these events involving more people will be important for moving the field forward.

I also cannot help fantasizing now and then what would happen if the government of some country or more than one became really interested in AGI. How fast could we progress if we had a Manhattan Project-scale funding level going into general intelligence? When could we get a thinking machine in that case?

Short of the Manhattan Project, another approach that is interesting is, think about Linux. There was no Manhattan Project to build an operating system, it just evolved from so many people participating over the internet. I am launching an initiative along with a bunch of colleagues at Novamente and the Singularity Institute called OpenCog, which will be launched late this summer. We will be releasing a bunch of our Novamente AI code into the Open Source world with a view toward seeing a broadly based Open Source AI effort aimed at getting people in this room and other people all over the world to help collaborate and build a thinking machine together. That is an alternative to getting the government to put in billions of dollars.

I found out about AI, not by reading scientific papers, but by reading science fiction, which woke my mind up to the grand possibilities here. When you reflect on building a human-level machine, what does that really mean? Where is that really going to lead us? None of us know, but I think it is important to reflect on these things.

At the end of this conference we are having a workshop on this topic, as Stan mentioned, which is aimed at looking at the futurological, social and ethical implications of AGI. Ray Kurzweil has been talking about the notion of a Singularity, where AI becomes so much smarter than human beings that humans are not really in the driver’s seat anymore. When that will happen, who knows? We cannot even predict when Windows Vista will be released, and that is pretty well known technology. You have to take prognostications with many grains of salt.

Kurzweil with his graphs has figured that we will have human-level AI by 2029 and that by 2045 we will have a Singularity, where human-level AI’s will have transformed the nature of reality as we know it. Since I am 41, I have a decent chance of making it if I keep eating my vegetables. That date does not worry me too much. On the other hand, I do not think it is completely inconceivable that we could get there within ten years. We are not going to get there tomorrow unless there is a secret project under the Mongolian Steppe or something that none of us know about. While it is hard to predict the timing, the course of events is easy to see if you open your mind to the technology going on. From that futurist point of view, AI is one of a number of technologies–nanotech, biotech, robotics and AI–all these things are cooperating together to help us to get more and more advanced technology. That leads to some difficult problems that we are not yet fit to grapple with.

One problem, which is one I have thought about a lot without any significant success yet, is let us say you make an AI that is really smart, if that AI can modify its own source code to make itself smarter and smarter, how do you keep it from modifying its source code in some way that you never conceived of. That is a really hard problem. We discussed this at the AGI workshop in 2006, and Hugo de Garis made a nice comment. “The initial condition of an AI will determine its ongoing development… initially.” That kind of sums it up. When you are making something that is beyond you, are you in the same position as a cockroach trying to predict the future development of human society and culture? Or, can you constrain things by making a mathematically defined architecture in an appropriate way? That is the sort of thing where the theory that Hutter and Schmidhuber may be interesting.

We also have a more social question. Let’s say we figure out how to make an AI that will be controllable beyond the human level in an appropriate way. How can we guarantee that someone else does not make a nasty AI first? That may be an even harder problem and one that those of us in this room are even more poorly equipped to deal with. These things are certainly well worth thinking about.

What we are going to be focusing on in the bulk of the AGI conference is more the nitty-girtty of how to actually get there, so that those problems exist in the first place. We will be looking at architecture, language and cognition, reasoning, learning and virtual embodiment. I would have liked to have some real embodiment, but we did not get a lot of roboticists here this time. Hopefully they will for AGI-09. Then, neural nets and brain modeling, and how we can work together to make an AGI renaissance really come to be–make AGI technology develop a lot more rapidly in the next decade than it has in the past.

This entry was posted on Tuesday, April 15th, 2008 at 5:06 pm and is filed under AGIRI, Ben Goertzel, Artificial Intelligence. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.