Pathways to Artificial General Intelligence

November 19th, 2007

Barney Pell is the founder of Powerset, a stealth-stage startup developing advanced AI technologies to deliver breakthroughs in search and navigation. In his 2007 Singularity Summit presentation he offered a framework for comparing different approaches to artificial general intelligence, in which we view any intelligent behavior as a combination of architecture and development, both of which can be characterized as more or less human-brain-like.

The following transcript of Barney Pell’s 2007 Singularity Summit presentation “Pathways to Artificial General Intelligence: Architecture, Development, and Funding” has not been approved by the author.

Pathways to Artificial General Intelligence

So first, in terms of advanced general intelligence, this could mean many different things. For the purposes of my talk, I wanted to come up with one specific concrete scenario. This is really when do robots become human-competitive in the sense that they compete with us for ordinary jobs. So, the scenario is that a humanoid robot that competes with high school graduates for jobs in the broad variety of jobs that a high school graduate can apply for. And the interesting thing here is that these robots all can get the job. Not always, but often.

What does that mean? Well, they have to be general, because high school graduates are not specialized in any particular topic. They have to be trainable to learn the job, whether it’s working at Starbucks or working in a bank. They have to be robust, because this is what is required of humans who do things. They also have to be social, because they are inhabiting a world of humans, and that involves also using language. And they have to be affordable. If these things come and they cost a million dollars, no matter how competitive they might be, they’re just not worth it.

So, how is that going to come about? A few years ago, I went and asked a bunch of colleagues in the field of AI to describe that exact scenario, and I said, “When do you think this is going to happen?” There was actually very broad consensus, within a hundred years. Broadly, I would say it was between 30 and 70 years, with the outliers being at a hundred. (Bruce Klein has since published another survey on the Singularity Institute website that you could go and check out.)

Then I asked a follow-up question: What will be the major milestones in approaches along the way, given that you are so confident that this is going to happen. At that point people said, “I have no idea.” And I said, “Well, do you think that your work is going to be one of the critical milestones along the way?” And they generally said, “Probably not.” The broad consensus from people within the field was that something major was going to happen within this century, most of whom felt that their work was probably not actually going to be on the critical path when all is said and done.

So, there was not even much of a debate. Why is that? One reason is that most of the field in the history of artificial intelligence has not actually worked on the problem of artificial general intelligence. It’s a really interesting point. In general, for most of history that has been viewed as something that’s too hard. But also there has been a preference on specialist versus general systems.

The path has been: pick some domain where humans do something interesting, start working on that problem, understand the depth of it, and then try to show that we could make programs that could do that kind of problem really well. That applies to chess programs, and all medical diagnostic systems, those kind of things. Another benefit from the field of working on that kind of approach is that you can focus on incremental research. You can work to advance what was the state-of-the-art a little bit, and you can also take advantage of low-hanging fruit, how these things really make applications in the world, instead of making something that will be useful a hundred years out.

The result of this kind of thinking has been that there has been no roadmap that really emphasizes generality. The bigger part is that there has to be funding. Most bleeding-edge funds funders have been in the Department of Defense, and even the DoD wants timeframes in five years or ten years, and increasingly put pressure on researchers to get these things done. The same is true in corporations that have long-term research labs. Over the past 20 or 30 years, most of the labs have gone away, and people have been forced to focus on low-hanging fruit.

So, thinking about a framework for how might these things actually come about, there are actually two key dimensions upon which different approaches can vary. In general it is centered on whether the approach is going to be similar or different from humans in terms of architecture, that which is built into the system at the start. And/or in terms of development, which is the process by which it comes to have its ultimate behavior.

So, any behavior you look at is going to a combination of what was built into the system at the start and of the developmental process by which the behavior comes about. So, at the bottom left here is fundamentally the same human architecture and fundamentally the same developmental process by which the system grows to have its behaviors. At the origin here is human babies, which we generally know how to do, and do very successfully. Close, we share a very similar architecture and a very similar developmental process. This is a strong brain-inspired model.

Ray Kurzweil has written about this in his book and people have talked about it earlier today. We will basically instantiate that on a simulation, where the simulation represents a neural scan. And voilà, we will have a human simulated brain. Crank it up a few more iterations of Moore’s law and it will get a thousand or a million times faster, as Eliezer was saying. You could then go to looking at a different kind of architecture but the same kind of developmental process. You might make an artificial system based on top-down engineering approaches, but it starts as a human child and then it develops in human schools, socializes with humans, learns language in all the same ways.

You could imagine scenarios where we take fundamentally human approaches but we actually start training systems in completely different ways. This could actually be intelligence amplification, where you have computers augmenting people, but it could also be viewed as a totally different way of developing capabilities for these systems that are human-like but are trained with images and massive speeds, and all these different things so that they are much more powerful than brains normally trained. A couple other points, the chess program is an example of a very different architecture, developed in a different way.

Great chess programs like Deep Blue were not based on how humans actually approach the game of chess. And the architecture was not based on pattern recognition, but in taking advantage of all the power of computing. There is a fair amount of work in the field of AI that starts with these more top-down freedom of engineering approaches, and maybe we will get there by building this kind of system. Maybe it will have nothing at all to do with what we look like as humans.

We might find radically different development and radically different architecture. One example of this would be simulated evolution. So, we don’t even worry about how humans work at all, or worry about architecture. We create an artificial environment where we start out with something, simulate a lot, find out what sort of environmental pressures led to the evolution of intelligence, maybe it just happens. But through a massive number of simulations over time, actual artificial beings start evolving. They probably start generating their own language capabilities and social structures. So, that’s a very different sort of architecture, and a very different kind of development. As Rod Brooks said in his talk, for all we know, it’s possible that’s already happened. It’s possible they don’t even notice us and we don’t even notice them.

Of course you have strengths and weaknesses for each of these approaches. The brain simulation approach takes advantage of Moore’s Law, both for the imaging and for the simulation. It may be possible to create AI without ever understanding how brains really work, and then we leave the hard work of understanding the brain to those newly created beings. It’s also closest to the existence proof we have of a general intelligence, which are humans. There is nothing even close to it.

So those are strengths. The limitations of the approach are, there may be so much more about how brains actually work that when you go down that path you get almost everything and you are missing something and you get nothing that is anywhere near an intelligent system. Without the understanding we just have no idea. It’s also not clear what it means to have software improvements if you are taking this hardcore approach of just brain simulation. It’s also subject dependent and though it is subject to Moore’s law and can go faster and faster, it’s really not clear what fundamental improvements can be made to the systems beyond just normal speed-up.

I think that what is actually going to happen is an intermediate position. The top-down engineering is going to keep on going, and a lot of things on the market will be increasingly driven by and demonstrating general intelligence. And in parallel, the human-inspired brain-centered approach is going to be gaining in capability, and we will start seeing some fusion of these two. There is this central virtuous cycle here. At some point, better virtual artificial intelligence actually becomes central to the creation of better products. To the extent that AI has been valuable at all, it has been valuable in the context of very specialized systems.

And something new is happening. We now have some level of autonomous vehicles. Electronics is in cars. Within perhaps ten years, we will actually have cars driving us automatically and being taxis for us in societies, which could lead to a major transformation in the design of urban architecture. I encourage you to go and read about Sebastian Thrun’s work at Stanford. Well, once that starts happening, we have strong drives for more and more general AI.

Similarly, in the DoD, there are now autonomous planes, and all kinds of autonomous vehicles and robots on the battlefield. Whereas there used to be an emphasis on special purpose systems, there is an increasing emphasis now on general purpose systems, because the nature of warfare and the nature of combat has changed from top-down centralized systems to asymmetric warfare. So there has been a huge shift in the DoD in terms of how we organize their use of robotics and the emphasis on generality and learning their systems, because the rules of combat change all the time.

Then there’s elderly care robotics. It’s big. Imagine once you start having good elderly care robotic systems and there’s multiple companies competing in these massive markets, the ones that then have more ability to communicate with you and empathize with you. There are a couple other domains: virtual worlds and video games. Video games have now passed the point where the realism of the graphics is not as important anymore to the gameplay as the AI’s. That means that an industry that is now bigger than Hollywood is going to be driven by AI’s in making these virtual worlds more realistic.

The last category of application that I think is going to be a big deal, and this strikes very close to home because it is what we are working on at Powerset, is natural language search and conversational interfaces. We have a huge revolution that is going to be happening over the next five to twenty years, which is the introduction of linguistically aware and capable systems into our routine experiences.

These trends are already happening. We’re already seeing the case of 411, local directories, voice-based interaction is now becoming a norm. I believe it’s already reached the point where more than 50% of all directory assistance calls are now being handled automatically. And that’s just going to increase. We now have microphones embedded in all of our laptops that we are using, and using electronically mediated conversations already. The preconditions are here for where voice-based interaction is going to become more and more expected. And when you have voice interaction, then you want linguistic interaction.

The progress that is being made with very limited research and funding is quite astounding, actually. Now you are going to have, over the next ten years, astounding increases in both the resources that are available and in the funding and in the emphasis on generality. So, things are going to change dramatically in the research toward AI.

Now, with the few minutes I have left, I’m going to talk about the specific thing we are working on here at Powerset. Powerset is building a conversational interface and a natural language search system. We have licensed the technology that has been developed over the last 30 years from Xerox PARC. Ron Kaplan, who led the team at PARC, is here in the audience. He is now the chief technology officer at Powerset. It’s a really interesting case because the group, when they first formed at PARC, recognized two fundamental approaches to user interface. One was a graphical user interface and the other was a conversational user interface. The graphical user interface, as we all know, took about ten years and had a huge impact on the world. And they knew that the conversational interface was going to be a very long-term project. And every five years, Ron’s group came back with fundamental improvements, but when they talked to the managers about applying it, they said they were not ready, there were still more fundamental problems to work on for the long-term.

If you look at all the other long-term research labs in the world focused on short-term apps, this one group was actually given the freedom to keep working on the hard problem and stay together as a group with sufficient funding. This was very hard and took a long time, and only within the past five years that group came back to management and said, ‘We think the fundamental problems have been solved.’ There is still a lot more work to do. This is going to be very hard, but now it’s about deploying it.

One key thing about the architecture developed at Xerox PARC, at the center of it is a language-independent core. That means that the system has a model of how language actually works and is language-independent. Then you build languages on top of it. There has already been fifteen years of work within the community testing the thesis that a language-independent system can actually support all of those languages, and essentially it has been proven that that is possible. Our deployment of this is to read every single sentence on the web, one at a time, extract out meaning from the sentence, the concepts and the relationships, index that all onto a very large database and then use that database in all kinds of amazing ways, one of which is using natural language to express what you want, and have a much better and transformative search experience. And there are many other applications.

This is a simple example of the parse of the sentence we parse on the web. This is just showing a syntactic parse, but there is a lot of richness that’s here. It’s the kind of stuff that people almost gave up on, thinking that it was impossible over the last 30 years, and now we are routinely deploying this across all of Wikipedia. We can process the whole thing at a very, very deep level within a matter of days.

There is a long way to go, but this is a very exciting time in the history of natural language and its deployment. We are going after very large markets which can support this kind of research. And Powerset is not alone. We think everyone is going to be doing this, Google and Yahoo. What this means is that these incredible, massive companies with massive markets are on the verge of where natural language kind of capabilities will become a core differentiator. These things can go very, very quickly, through all these virtuous cycles.

Now, I want to close on this thought. All this work that we are doing is from an engineering approach. I don’t see ourselves suddenly plugging it into a robot and having high school equivalent types of things. I think there is going to be a key milestone, which is going to be integration with these kinds of top-down engineering approaches in natural language with brain-style architectural components.

So, maybe a milestone to watch out for, which will indicate that we are truly on the path to achieving these kinds of Singularity milestones is when you have a top-down system like these kinds of natural language systems at Powerset, combined with associative predictive memories of the kind that Jeff Hawkins is working on, or other kinds of grounded language systems that they are working on at MIT and other places. When you actually start to see that, then we’ll start combining them into robustness and the learning abilities that human minds have, with top-down engineering approaches that are economically viable, and that’s going to be a very interesting future.

In conclusion, AGI has received only limited computational resources to date. These conditions are changing very quickly. Multiple pathways exist to get there. A combination of approaches is likely to be what works. And, while the problems are very hard, these trends are very indicative of the thought that the Singularity might actually be nearer than people expect.