Speaking to the Web
Posted by Jeriaska on September 29th, 2007
SIAI Interview Series - John Smart
The following transcript of the SIAI Interview with John Smart has been edited for clarity with footnotes by the speaker. Video and audio are available at the Singularity Institute website.
Speaking to the Web
“When will artificial intelligence out-compete humans?”
I think we have had fifty years of it, basically. We have had some successes. This thing I have got on my hand here, this digital watch with a calculator, this is a calculational singularity to you and me. Like James Burke said, there has never been so much in our technological world about which we know so little. That’s going on all around us. So, this has out-competed me at my ability to do arithmetic. This is a small piece of Earth’s intelligence that has shot past human beings in a very specific module: time-keeping. But we are getting more and more of those kinds of pieces. Some graduate student made a fifteen node neural net, something small, that actually was able to find supernovas in the night sky, because it is an easy image recognition problem. The exploding star’s image gets large and white, and then over a week or so, it slowly fades off. Easy enough that a very simple program could be made to steer the telescope. So now supernovas are not hunted for by human beings anymore.
When you and I use our credit cards, they are monitored by a neural net called Falcon, created by HNC, now Fair Isaac. This is a $200 million a year program. It’s a big project with a lot of people on it. Basically, it trains off of all your previous purchases so you go out and buy Pampers and you haven’t bought them in ten years, that flags a human being. But it is not a human that is looking at the primary data anymore.
These are all pieces of AI that have out-competed human beings in very specific areas. Each of them are narrow, but we have fifty years of these kinds of successes. So we can show, piece by piece, higher and higher order things. Now, the ability to do visual recognition of faces, we are just on the edge of it. Some of the military systems are just on the edge of this kind of an HNC system that would flag something unique, pop it to a human, and then the human could train the net, saying, “No, you’re not looking for that. You’re looking for something a little different.”
The one that I am the most excited about is what I call the conversational interface. You and I were using something called Altavista. A pretty primitive search engine in 1998. The average query that we did to Altavista was 1.3 words. How long did it take for us to double our queries to 2.6 words? Seven years. By 2005, the average query length was 2.3 words. How long is it going to take for the next doubling to come? In my essay,[1] I went on record saying it’s going to be seven years. Because what is actually happening, I think, is an exponential curve, which is basically code breaking. If you look at how code breaking works, when you are trying to figure out a foreign code, it goes in an S curve. It’s hard to get the primers at the beginning because you don’t really understand the system, and then you start getting them, and the more of them you get, there’s this positive feedback loop where you figure out more and more about this foreign code you are trying to break, and then you reach the midpoint where you’ve got pretty much all the easy stuff that you need and then there is not as much new stuff to get and so then it starts to level out.
What I suggested is humans speaking to the web using sentences, whether they are typed or spoken, is actually on a code-breaking function, which means that it is going to look exponential for the first few doublings.[2] When you and I make a query to each other, the average sentence question is something like eleven, fourteen words. So, I argue that we are going to go from 2.6 to 5.2… to 10.4, and that’s 2019. That’s getting up to human-level questions, assumedly with human-level grammars and subtlety. What I think is going to happen somewhere around 2019 is every kid in the world is going to have a cell phone, because they are going to be dirt cheap by then. Every kid in the world is going to learn as fast as their curiosity drives them, just talking to Google. That’s really powerful.
“What’s the big deal with Google these days?”
Google’s got an optical cortex: Google Maps; Google Earth, sensors and databases, etc.. Google is the fastest-learning baby on the planet right now. It’s the biggest, largest, most interesting word-based, and some images-based, web. It’s really a primitive intelligence. Because Google learned the word “near,” you and I can say “Coffee shops near Marina.” Now, that’s four words. The first link is going to be a bunch of pins that are centered around the downtown district of Marina, or San Francisco, or whatever. That’s Google Maps. So, we have now been teased into four words. That’s more than the current average, which is 3.5 words. Google is shortly going to learn time. So, you are going to be able to ask Google about a webpage that I saw a year ago, two years ago. You are going to be able to say those words and Google is going to be smart enough to go back to its database and give you only the time-based searches relative to what you are looking for. So, step by step, all these companies that are using these new natural language processing tools for search are going to tease us into talking to our computers. What’s exciting about that is, once you are talking to a computer, you are going to want that computer to have a face. Two-thirds of our information in face-to-face communication is non-verbal. You and I are nodding to each other right now, so I know I’m reaching you. If you weren’t, I would have to change what I’m saying. So, it’s incredibly efficient to have verbal and nonverbal at the same time. What I have just done is sold you a bill of goods. I’ve sold you on the idea that avatars become useful so that you can have a conversational interface. Not just useful for you talking to the machine, to your car, to your house, to your robo-kitchen, whatever. But you talking to other human beings through your avatar.
“What is accelerating change?”
I remember as a kid seeing Sagan’s Cosmic Calendar. You put all the important emergent events on a calendar year and all the interesting stuff happens after June. In the early months of the ‘calendar,’ everything is so slow: you’re creating galaxies, it takes forever. You first get the earthlike planets right in the middle of the calendar year, and then what does it look like? The whole second half of the calendar year looks like a J curve.
Why is that? Why does all the acceleration occur toward the very end? There is something about the universe that we live in that biases it towards acceleration of information. And you know what? Our physics textbooks don’t talk about that. They tell us about the second law of thermodynamics, that everything is running down, entropy is going up. But we know in our bones that life does the exact opposite. Life is ordering itself. Life is accelerating itself. There is a piece of physics, there is a piece of philosophy, there is a piece even of some spirituality that’s missing here. I would say that most spiritual systems got that way before the scientists did. And now we are trying to figure out a way to reconverge science and spirituality into this realization that there is something about this universe that is taking us to higher, more sublime levels, and there are good paths that we can make towards those things.
“Why does change accelerate?”
People think it comes from the special structure of our universe, the way it is organized. Every complex thing has to throw its trash away, and we throw all of our trash away, which is the heat, out into the vacuum of space, which is the most efficient way to throw it away. If we did not have huge vacuums of nothingness between all the interesting things in our universe, they would not be very efficient at creating local order. So, here is my $64 answer to you of where intelligence goes. I don’t think it goes to outer space as it gets smarter. I think it goes to inner space. I think people looking at intelligence going into outer space actually have it 180 degrees out of face. We have to leave earth. That’s obvious. We already know we have maybe five billion years more and then the earth is going to get heated up by an expanding red giant sun. Intelligence has got to leave earth. Most futurists currently think it is going into outer space. I think that’s 180 degrees wrong. I think we are going into inner space. And the things we are finding out about the inner space today are just mind-blowing.
For example, programmable matter: you’ve heard of graphene transistors, single-electron transistors? Pretty soon, we are going to be able to compute with one electron at a time. Right now, our computers send huge clouds of electrons through every gate. Hugely inefficient, you know how hot your laptop gets. We are on the verge right now of computing with one electron at a time. Fat-fingered 21st century human beings computing with one electron at a time, creating seven-qubit quantum computers. Entangling photos so that when we unentangle them, seven miles apart down a fiber optic cable, you get spin down on one, almost instantaneously, the other one is spin up. What does that tell you? It tells you that things at the quantum scale don’t respect space and time. That is so weird, so interesting, and I think there is lots and lots of evidence, things like the quantum space, things like black holes, things like these theories of the multiverse, string theories, M-theories, that there is so much unbounded complexity there and potential, that that’s where we go.
“Why support the Singularity Institute?”
The Singularity Institute is opening the dialog. We have to give ourselves permission to discuss these things. If someone is figuring out a way to get press, to get us to discuss these things seriously, that is absolutely what we need to be doing. Everything that I have just said is just my single opinion, having studied these subjects. I would love to see what James Surowiecki calls the wisdom of crowds in this space, because once you get the nice bell curve of all the possible thought, I think somewhere in that envelope is going to be the reality. As he says, when you take a guess as to how many jellybeans are in that jar, your guess is going to be plus or minus 20%. You get ten of us who are cognitively diverse in a room and make that guess, it’s going to be plus or minus 5%. That is what the internet is going to give us. That is what publicizing these issues through groups like the Singularity Institute is going to give us. It going to give us much better insight on the pieces of the AI puzzle that are the low hanging fruit that are going to take us to the next step and surprise us, because we want to open that newspaper everyday and be surprised.
“Will the future always surprise us?”
I think so. I think that’s the nature of the universe, what Kurt Gödel called “computationally incomplete.” No matter how complex your formal logical system, there are always questions you can ask that cannot be proven or disproven from within that system. So, you’ve got to design a new system.
“Could emerging technologies actually cause us to disengage with reality?”
My optimism would be that there are these hidden factors operating against disengagement. You can set up your wearable computer to filter whatever kind of reality you want to you. You could be in a complete echo chamber cocoon, hearing only your own ‘voices’ and interests, when you get that technology, and completely zoom off from reality. But most people, I don’t think are going to do that. They are going to do it enough to keep their culture, but they are going to be fully conscious of how those technologies interact with them. Look at the Amish. I love the Amish. One of my favorite cultures. They debate whether they want to use rubber bands. Think about the power of that culture. They can decide, is a rubber band taking away from my community? Of course, they use cell phones on their own terms, they use rubber bands on their own terms. That’s a really wonderful culture. I want that kind of culture to exist right past the Singularity, like Spider Robinson says, in these little Earthpark bubbles, where there is going to be deep nanotech surveillance on everybody to make sure they are not making nukes in their basement, but if they want to live that way, more power to them, man. I love it. Because then there’s more diversity, more ways that I can think and be, but you think of all the cultures that have successfully done that in the long-term and they snap back to reality to some degree.
There always seems to me to be real strong forces that pull people back, even though they like to go off into those separate worlds. I have these laws of technology. My third law of technology is that first-generation technologies are usually dehumanizing, because you don’t get the interface right and they are primitive. Second-generation technologies are indifferent to humanity, so they are a wash, a net neutral effect, really. Some benefits and some negatives. And with luck, third generation technologies are net humanizing. I think that’s the way with videogames toda. First generation, kids go off into their own space, they forget a lot of their social skills. We are just entering the second generation now where kids can use team speak and can talk to each other using voice, collaborating as a group to solve a complex problem. So there are some impressive aspects. Plus, there are some serious games where kids can actually play chemistry simulations and maybe learn chemistry twice as fast as if they had had to play with the nasty chemicals, and the simulations are realistic enough that they snap back to skills they can transfer to the real world. But then there are all these negative effects, still. So we are not even yet at the full second generation where you could say it’s neutral, but we’re getting close.
Think of a third generation where you are wearing the game with you everywhere. A bird flies by and your knowledge management system tells you, “You want to know the calculus behind that? I know you have an AP test in three weeks. It’s only going to take two minutes. I’ll give it to you. And I promise you, you’ll have it nailed.” In that kind of a world, you can imagine a kid who is so tuned in that you turn off the whole net and they could be Survivor Man, get food out of the ground for two weeks, because they uploaded that little module. They can make change, which kids can’t do today. They’ve lost that ability, because we are still in that first generation with regard to calculator interactions. In that kind of a world, the kids could actually be smarter than their parents were (never smarter than the machines, that’s not possible) but smarter than their parents were at all the skills that matter to them even when they turn the machines off. And that’s a truly empowering world, because then you can deal with the machines on your own terms.
I do believe that in the long run, we won’t be using words like “transhumanism” and “separation of identity.” We are just going to start to feel that these things, third generation technologies of all types, are extensions of us. When your mom dies in 2050, you are not going to go to a tombstone to grieve. You’re going to fire up Digital Mom. You’re going to get all of her stories, all of her attitudes, you’re going to be in some city in France where your grandmother was 100 years ago, and she’s going to whisper a little story to you through some digital scrapbook. It’s not going to be your true grandma, but it’s going to be close enough that you are going to feel as if you are in all these places.
Then, at some point, DigiMom gets good enough, and you get good enough that there really isn’t much of a difference between my digital and my electronic self. I think that successive approximation type of AI is the one I predict we are going to see. Now, there are other people at this summit who think we are going to see a very significant what they call “hard take-off.” More power to them. I’m glad there are people out there doing that kind of work. I personally don’t think we are going to see hard take-offs, except in these narrow areas that we have seen in the past fifty years. I could be wrong.
Footnotes.
1. John Smart. “The Conversational Interface: Our Next Great Leap Forward,” 2005. http://www.accelerationwatch.com/lui.html
2. (Ed Note: One idea in Smart’s essay is that the ‘phase space’ of interesting questions and simple answers that humans speak to each other in conversation is a finite and modelable system. While that ‘space’ is growing, the web’s ability to archive all past and current question and answer pairs, and extrapolate future ones, is likely growing at a much faster rate).