Creating Tools of Narrow Superintelligence
“People have been saying for 50 years that artificial intelligence is just around the corner. Why the failure?”
At the start of AI in the ’60s, people got very excited. They said, now we have this new thing, computers. We could probably solve this problem very quickly. And then we discovered over time, it was harder than we thought. We thought that thinking was just being logical, being able to follow rules A, B, to C, and we discovered that it is more than that. It is analogical, it is dealing with uncertainties, that there is much more out there in the world. Understanding and dealing with it is harder than we thought.
“What is artificial intelligence?”
When I did my textbook, I had to define what AI was. One of the things I wanted to avoid was all these philosophical debates. The question of even is AI possible, what is thinking, what is intelligence, what is consciousness? In some definitions those all come into play. The way I tried to define the field was to say AI is building the best possible programs. So, you are given a task to do, you are given some robot or some computer to do it with, and there are so many possible programs. To make it simple, there are n or 2^n possible programs, and AI is finding the best one. Or at least finding a pretty good one.
“Are we heading toward a Singularity?”
I think it is interesting. I remember reading Vernor Vinge’s stories and being very interested by that. Going back and rereading I.J. Good’s ideas. It’s an interesting idea to think about, and I think that as scientists it is our responsibility to consider these kinds of ideas. We should always be saying, “What we’re working on, is it going to be a force for good or for bad? What are the effects on society going to be? And this is one of the possibilities. So, it is something we have to think about. Is it coming? If it is, what’s going to happen and how do we get ready for it?
“Is the Singularity near?”
So, I’ve looked at it and I don’t see it yet. It is like trying to get a seismograph and say, “Is there an earthquake here?” There are no signals yet, as far as I can see. On the other hand, sometimes earthquakes just come up by surprise.
“Do you know when the Singularity will occur?”
I think that is the whole problem with it. You can’t. I said, “Can you predict by the signs that it is coming?” No, the signs are not there yet. So that means I can’t predict. We just have to wait and see.
“Do you agree with Kurzweil’s forecasts regarding the Singularity?”
Ray is great and I appreciate all that he has done. I appreciate all of the advances in AI that he has done in reading and music. He is a great technologist. I think in terms of these predictions, he is doing the best he can by saying, “I want to nail down this prediction. I am going to look at the data that we have. I am going to extrapolate from that data and come up with an answer. I think the only problem is that the data is not conclusive. You look at Moore’s law and yes, you can say that computers are going at this particular rate. But I don’t think you can necessarily say that when we pass a certain number of transistors, that corresponds to a certain number of neurons, and therefore the two are equivalent. I think that’s where the analogy breaks down.
“What markers would indicate that the Singularity is upon us?”
You would look in two ways. You would look at the theoretical. Are we almost ready to build something that we think could do the whole job? I think now, we’re not. There is no consensus on how you would go about building a machine that would improve itself indefinitely. The other thing you would want to look at is the practical. What is it that machines are capable of doing now and how fast is that progressing? So, certainly we have seen a lot of advances. We have seen machines doing more and more. But it does not seem like we are at the point where it is imminent.
“Do you see technology’s exponential growth or double-exponential growth at play in Google’s operation?”
You have to look at different applications. In many places we see more like linear growth. One of the good examples is work in machine translation, translating from Chinese to English. There we have taken an approach by saying the more you know about language, the better you do. We collect more and more examples of language, feed them into our learning algorithms, and as you add more language examples, performance goes up linearly. We don’t really expect it to be exponential. If anything, we expect it to start to level off at some point, but we have not reached that point yet. We still put in more data and it performs better. In other places you see examples where you need to reach a certain threshold of data before you get results at all. Maybe something like Google as a whole is an example of that. When the web was teeny, a search engine would not be that useful. But now that there are tens of billions of pages out there, almost any query you can do, you get something back. So, it reaches this threshold of usefulness just because of the data that is there, and on top of that we are trying to improve it all the time.
“Google and AI are so often mentioned in the same sentence. What’s Google doing that other search engines are not doing?”
We have the majority of users compared to the other engines. Two, I think we have more of a technological focus. Yahoo seems like they are focused on a full solution to the user. They are a content company, an entertainment company, as well as a technology company. We seem much more focused on technology. I think people look at that and say, “If something’s going to happen, maybe it’s going to come from these guys who are so focused on the technology.”
“What is going on at Google that reflects advancement in AI?”
Ai is prevalent throughout Google in the sense that we have data everywhere. There is far too much data for a human programmer to make sense of it. So, we have been forced to go to machine-learning algorithms throughout the company for almost everything we do. So, in that sense, it is one of the most AI-driven companies that I know of. Now, we have not done anything at the general intelligence level. What we have not tried to say is that we are here to solve all your answers for you. You don’t even need to do the Google queries anymore, we’ll find out what you need and drive to the store and buy it for you rather than give you information on it. We are not attempting to do that. We are just trying to do a better job of what we can do, which is connect you with the information.
“Can we build greater-than-human intelligence?”
Sure. We already have that in so many ways. My tiny little desktop calculator has greater-than-human intelligence. It does square roots much faster than I can. Google has greater intelligence in that it knows about more pages and can find them faster than any human could. Now, it doesn’t understand them the way a human does, so it’s capabilities are very limited. It has superhuman intelligence in some directions and subhuman intelligence in other directions. Maybe someday we will have a more general superhuman intelligence. Until that day, what is interesting is putting it together. What can we take from people to connect them with other people and use these narrow superintelligent tools to make those connections better?
“What are your thoughts on SIAI’s Singularity Summit?”
I think it is a fascinating area just thinking about what the possibilities are for the future. I think it is our responsibility to think about it. If there is going to be some possible negative consequences, we need to prepare for that. If we can have positive consequences quicker, it is our duty to try to do that to make the world a better place. Regardless of how high a probability or what timeframe in the future you put the Singularity, you have got to start thinking about it now. And the other reason I’m here is that a bunch of my friends are here that I wanted to talk to. A bunch of good thinkers that I have not met yet are here that I wanted to meet. So, it just seemed like a great place to be.
“Does Google anticipate future technologies?”
Certainly you want to plan for technological progress. You want to plan even for the unknowns. You want to say, I want to be in a position at some time in the future where I can take advantage of these opportunities. I don’t think that you can say the answers are always going to come out right. If I say that in six years I am going to know, maybe in six years maybe you will know, maybe you won’t. Maybe in six years you will find out there is no answer. That’s it’s impossible. There are limits on what you can do, and some of the advances in computer science have been finding those limits. In basic mathematics, going back to Gödel, showing that there are mathematical truths that cannot be proved, there are limits to what mathematics can do. In computer science, showing that there are problems that are exponentially difficult, that even with this exponential growth in computers we are not going to be able to solve. Now, the question is, where is general intelligence? Is that something that is below this line of a solvable problem, or is it above the line? I don’t think we know yet.
“Why support the Singularity Institute?”
Because it is a fascinating problem and it affects everybody in the world. This could change our lives. I think we want to know about that. We want to be able to participate and have the best chance we can for a good outcome.
“Is Friendly AI possible?”
I think the idea of Friendly AI is very interesting. At Google, we are not doing anything specifically on that, but in my earlier work at NASA. We did work on program verification and properties of programs. At NASA it is very important that software does not crash because if the software crashes, the spaceship crashes, there are people on board, they die. Bad things happen. One of the things you want to be able to do is verify that software is correct. Some of that is done through traditional testing, a quality assurance type of approach, and some of it is done through mathematical proofs of the program. Proofs of correctness or proofs of properties. You could prove a property like, this program will not get into a deadlock state. That’s a good thing to be able to know. Now, if you could define what “friendly” means, then there is a chance of using that same technology to say, “I can prove that this program will never be unfriendly.” The problem is that it is much easier to define the state of deadlock than it is to define the state of Friendliness.
“What does a post-Singularity society look like to you?”
I guess by definition you can’t know what it’s like beyond the Singularity, but if you expand the definition of Singularity a little bit to say “advanced general AI,” I like Barney Pell’s description. It’s when a computer can get most jobs that a high school graduate gets today. How does that change the world? So, one thing is, it makes the world a much richer place. It changes what each of us is able to do, because we are now freed up from doing a lot of labor that we were forced to do before. I remember reading Freeman Dyson’s biography, and one of the things he said was very interesting. The middle class women of his age were much more feminist than the middle class women of today, and the reason was that they all had servants who stayed home and did the cooking, cleaned the house and took care of the kids. So, they were free to go out and do their suffragette movements or whatever it was they were working on. Whereas today, the middle class, we don’t have servant, you often have two parents working and so there is much less free time. Now, after the Singularity in Barney’s sense, we are going to be back to that kind of a lifestyle where not just the middle class but everybody essentially has these servants that can take care of business. So, suddenly, a lot of your lifetime is freed up. Now it’s a question of what do you want to do with it?
“Will the Singularity be democratizing?”
I don’t think it will be more democratizing or more equalizing. Certainly it will raise all boats. Everyone will be at a higher level of living, have more free time, and so on. But I think that the people at the higher end of the economic scale will get an even increasing share, and so this gap between the haves and the have-nots will grow, even as the have-nots have much more.
“Do you have concerns regarding future technologies?”
I certainly do. I think there are a lot of things to worry about, and I don’t think it’s new with AI or nanotechnology or biotechnology. I think we have been dealing with it for many years. I grew up worrying about nuclear war, trying to protest and help through that a little bit. It seems like we made it through that crisis, though there is still a chance of some problems there. There have been all sorts of potential threats. So, for fifty years it has been within the bounds of a high school chemistry students to make up a batch of botulism and drop it in the water supply and kill thousands or maybe hundreds of thousands of people, but no one has decided to do that yet. I’m not sure why. I’m not sure if there really are constraints that are going to continue to hold. I know there is going to be continued threats that smaller and smaller groups of people can do more and more damage. I don’t know how much these ethical constraints or other kinds of constraints are going to stop that from happening.
“Assuming Strong AI is likely, do you anticipate a hard take-off or a soft take-off?”
I think I agree with the soft take-off for the reasons you mentioned. Also, just looking at Ray’s curve, exponential growth means the same percentage change every year. So, I think we will keep on looking at that. Now, at Google, you are very happy with your results and we are happy that we can deliver them to you. From the inside I think it’s kind of like making sausages. We worry more about the results that aren’t good than the results that are good. You do a search and get six out of ten good results and you’re really happy. We say, damn, there were four bad ones in there. We’ve really got to fix that. So, we will keep on doing that. We will keep on coming up with improvements. We will give you better results. We will give you better ways to interact with us. We know there are lots of things now that you want to be able to do but can’t do. For example, if you don’t know any of the words, what do you do? Well, we put all the onus on you. So, you have to fall back and say, I’m looking up some medical problem, I don’t know the specific medical terms, but you know enough to get to a document that will have the terms, and then you read that, and then you can go back and make the actual query. We want to make that step faster. We want to get you directly to the right answer, even if you didn’t know the right words to begin with. Throughout, we want to make the interactions faster, we want to correct you with the right documents or connect you with the right person. We think we will add continual improvements to that and you will be getting happier and happier. I don’t see a hard take-off, but I do see you being able to do more over time.
“What are your thoughts on molecular manufacturing and bio-engineering?”
It’s fascinating to me. I’m trying to learn something about it and educating myself. I have a long way to go. I think it does tie-in, because it is this mix between the physical world and the world of information. I think it is likely that in the coming century that a large part of manufacturing with cells or with atoms. I kind of think it will be more done with cells than with atoms, just because they are bigger, they are easier to manipulate. They have already got this three gigabyte computer on board, we just have to figure out how to program it.