Engineering Computers to be Computer Engineers
Posted by Jeriaska on September 27th, 2007
SIAI Interview Series – Peter Voss
The following transcript of the SIAI Interview with Peter Voss has not been approved by the author. Video and audio are available at the Singularity Institute website.
“If we have computers that are significantly smarter than we are and can work 24-7, imagine a medical researcher in an AGI. That intelligence would be much smarter than a human researcher, plus you can copy it a hundred thousand times and have a hundred thousand of these PhD researchers chipping away at disease and aging problems, pollution and whatever.”
Engineering Computers to be Computer Engineers
“What is artificial intelligence versus artificial general intelligence?”
The area that I am working on we call AGI: artificial general intelligence. That is to contrast it from the field of AI because the field of AI today almost exclusively deals with narrow AI. It is solving specific problems. You build a computer system to solve a particular problem. That is pretty much what AI does. So AGI are systems that can potentially learn a very broad spectrum of different things.
If you have a system that can learn like a human, that can acquire knowledge like a human, you can think about a system acquiring the ability to become a good computer scientist. If it becomes a good computer scientist, it could improve its own design. Then you have recursive self-improvement. So the computer becomes smart enough to understand its own design and to improve on it, which will give it even more ability to improve its design and become incredibly smart very quickly.
“What future technology interests you?”
One of the things that is very dear to my heart is longevity: life extension. I would like to have the opportunity, and I would like other people to have the opportunity, to live really, really long lives. Let’s call it an indefinite lifespan. So that we do not have a hundred thousand people a day dying involuntary death. That to me would be a really good thing. Now, we are not smart enough to fix that problem. We are getting there very, very slowly by improving medical science, nanotechnology, and so on. But that’s really, really hard. We are not smart enough to do that. We do not have enough smart people to solve these technical problems. If we have computers that are significantly smarter than we are and can work 24-7, imagine a medical researcher in an AGI. That intelligence would be much smarter than a human researcher, plus you can copy it a hundred thousand times and have a hundred thousand of these PhD researchers chipping away at disease and aging problems, pollution and whatever. They are going to be able to solve those problems a lot quicker. I think that’s a big deal.
“What is ‘greater than human’ intelligence?”
We don’t know. We’re not smart enough. That is kind of the flip answer, but it is really true. What I am certain of is that it will be a lot smarter than humans. First of all, it can reach very difficult conclusions, problems that I might struggle over for weeks or months before solving. It would compress that duration and be able to come up with conclusions within seconds or minutes. It would be able to solve problems much faster. They will also be able to solve more difficult problems, plus you can pool the knowledge of many of these smart AI’s that have been working in different fields. They can almost instantaneously pool their knowledge, which for humans is very, very difficult, to transfer our knowledge, our insights, our experience to other people. It’s a very tedious, slow process. In computers that can happen instantaneously. That will be a big deal as well. So, machines working 24-7 at a much higher speed, being able to exchange information, being able to clone it and copy it. Once you have a person who has a certain level of expertise, I wish I could clone myself with the years I have spent painfully acquiring knowledge I need for AGI design. If there were a thousand people around with my knowledge and background, we would be able to make progress much faster. So, it would represent a fundamental shift.
“What work are you doing in the field of artificial intelligence?”
What I am doing is I run a research and development company. Currently we have fifteen people in the company. We have been going for six years, and our role is specifically to develop an AGI engine. To develop an actual system that can learn and think ultimately like humans can. So, we are actually physically working on developing the technology.
“People have been saying for 50 years that artificial intelligence is just around the corner. What is different now?”
Well, first of all I do say that we are getting close, and that is based on the work we are doing ourselves. Plus, some other things. There are other companies also working on stuff. What is different? I have also been around long enough to do a comparison. When I first started with computers compared to what I am using now, to give you an example, some of the experiments we do on our system right now, we might give it some kind of a problem and we would expect for it to give us a reply within, say, a second. If you are talking to a person, you want to have a conversation, you would reasonably expect to have a response within a second, if it’s something not that difficult. Now, that same algorithm that we are running now, running that on a machine of twenty years ago would have taken one week to give us an answer. Now that is a big difference. You cannot do any experiments if you type in a question and you have to wait a week to get a response to see if your algorithm is actually working. That is very different from getting a response within a second.
We are running tests on our system, for example, where hundreds of thousands of tests are being run, one after the other, which we can run in a few hours. It would have taken millennia to run that twenty years ago on a machine that we could afford. The development tools are much better. My productivity in developing new programs is significantly better. It’s not a million times better, as the speed has maybe improved a million-fold, but it is still significantly better. I can do things sometimes in hours that would have taken me weeks, months, or years to do.
“Do you agree with Kurzweil’s forecasts regarding the creation of Strong AI?”
It sounds right, except I think it’s too conservative, for a number of reasons. First of all, I believe his estimate for having human-level AI is way too far in the future. I think it will happen much sooner. Then, we will have a self-improving cycle. Therefore, I think we will have a much faster growth rate than what he predicts. But his approach, I think, is fundamentally looking at the wrong solution. He is talking about reverse engineering the human brain. Now, to me, that is a really, really hard way to solve the problem. It is using tools that are really not well suited for building a human brain or reverse engineering a human brain. I favor an engineering approach, where you say, “What is the problem we are trying to solve and how can we best solve it using the technology and the tools we have available?” That is the approach we are using. Reverse engineering the brain, I think is really, really hard. It could take many decades.
“What methodology are you using to create Strong AI?”
We are basically starting off by saying, What is important in human intelligence? The ability to learn, conceptualize, to think conceptually, to think contextually, to use context in your decision-making. We start by understanding the essence of intelligence and then say, “That is what we need to engineer.”
“What is Friendly AGI?”
I think it is an important question. Obviously, we are at the Singularity Summit, so that question is being asked. I have given it a lot of thought, and I continue to give it a lot of thought. The short answer is, we really don’t know the outcome of any dramatic technological change. My own thinking on the subject differs a lot from what Eliezer believes. Eliezer believes that almost any design will be unfriendly and will be a disaster, unless you specifically design Friendliness into it in some way. Now, first of all, I believe it is impossible to design Friendliness verifiably, for a number of technical reasons. I believe it is inherently impossible to achieve what he wants to do. Secondly, I also disagree with his assessment that the majority of cases would be some kind of a disaster. I think it is quite the opposite. There are many reasons why I believe an intelligent system will inherently improve people’s morality. There is an inherent safeguard built into it. That does not mean to say I am completely at ease with the situation or what the risks are. But it is going to happen. We are going to have AGI. We need to understand the process and dynamics of AGI and how they will interact with humans, as we develop the technology and make decisions as we go along. I do not believe that anybody can really anticipate it. I do have a much more positive view of the outcome.
“Assuming Strong AI is likely, do you anticipate a hard or soft take-off.”
Firm take-off. I believe essentially we have the hardware available. I don’t believe that that is a limitation. I believe the technology, the pieces of the puzzle, essentially are there. I do not think there is any fundamental new technology that has to be developed for us to achieve AGI. However, there are millions of pieces to the puzzle that need to be identified and put together in the right way. I believe the bottleneck is really having the right theory and the right focus. I mean, how many people are actually working on this problem in the world? Very, very few. So, I think if more people were working on it, we would get to it sooner. I do not think we need to wait another ten years or twenty years. It is really just that there are not enough people who believe that it is possible. In fact, on my website I have a long list of why so few people are working on it.
A lot of people teaching computer science themselves had the dream 30, 40 years ago. They tried and failed, so it is off the table. They made a promise, and could not deliver, so AI has become a swear word. A lot of people do not have the background in psychology and epistemology and theory of knowledge to really know how to go about engineering intelligence. They do not understand intelligence enough. These reasons sort of multiple out. There are just few people who are candidates to work on it. Then, of course, you cannot get funding for it because it is not popular. Another reason is that general intelligence has fallen from favor. It is basically not politically correct that there is such a thing as G or general intelligence.
I think the government gets advised by the same people in academia who essentially say it is not possible to do it now. “We tried and we couldn’t do it. We should just focus on narrow AI and all the particular applications and eventually we will get it right.” Having said that, though, there is a move towards cognitive computing at DARPA. Ron Brachman certainly put a major emphasis on that, but the results seem to be mixed for various reasons.
