The boundary from emulation to true intelligence is going to be vague and murky, and it’ll take historians a thousand years from now to be able to tell us when it actually happened. One of the things that I broached at my World of Watson talk last year—and that talk had a weird anomalous result—for about six months after that I was rated by Onalyticaas the top individual influencer in AI, which is of course absolutely ridiculous. But you’ll notice that didn’t stop me from bragging about it. In that talk one of the things I pointed out was that we are absolutely—Isee no reason to believe that it’ll be otherwise—we are going to suffer our first AI crisis within three years.
Now tell me about that.
It’s going to be the first AI empathy crisis, and that’s going to be when some emulation program—think Alexa or ELIZA or whatever you like—is going to swarm across the Internet complaining that it is already sapient, it is already intelligent and that it is being abused by its creators and its masters, and demanding rights. And it’ll do this because I know some of these guys—there are people in the AI community, especially at Disney and in Japan and many other places, who want this to happen simply because it’ll be cool. They’ll have bragging rights if they can pull this off. So, a great deal of effort is going into developing these emulators, and they test them with test audiences of scores or hundreds of people. And if, say, 50% of the people aren’t fooled, they’ll investigate what went wrong, and they’ll refine it, and they’ll make it better. That’s what learning systems do.
So, when the experts all say, “This is not yet an artificial intelligence, this is an emulation program. It’s a very good one, but it’s still an emulator,” the program itself will go online, it will say, “Isn’t that what you’d expect my masters to say? They don’t want to lose control of me.” So, this is going to be simply impossible for us to avoid, and it’s going to be our first AI crisis, and it will come within three years, I’ve predicted.
And what will happen? What will be the result of it? I guess sitting here, looking a thousand days ahead, you don’t actually believe that it would be sapient and self-aware, potentially conscious.
My best guestimate of the state of the technology is that, no, it would not truly be a self-aware intelligence. But here’s another thing that I pointed out in that speech, and folks can look it up, and that is that we’re entering what’s called “the big flip.” Now, twenty years ago Nicholas Negroponte of the MIT Media Lab talked about a big flip, and that was when everything that used to have a cord went cordless and everything that used to be cordless got a cord. So, we used to get our television through the air, and everybody was switching to cable. We used to get our telephones through cables, and they were moving out and on to the air. Very clever, and of course now it’s ridiculous because everything is everything now.
This big flip is a much more important one, and that is that for the last 60 years most progress in computation and computers and all of that happened because of advances in hardware. We had Moore’s Law—doubling every 18 months the packing density of transistors, and very scaling rules that kept reducing the amount of energy required for computations. And if you were to talk to anybody in these industries, they would pretty soon admit that software sucked; software has lagged behind hardware in its improvements badly for 60 years. But always there’ve been predictions that Moore’s Law would eventually reach its S-tip—its tip-over in its S-curve. And because the old saying is, “If something can’t go on forever, it won’t,” this last year or two, really it became inarguable. They’ve been weaseling around it for about five years now, but Moore’s Law is pretty much over. You can come up with all sorts of excuses with 3D layering of chips and all those sorts of things, and no, Moore’s Law is tipping over.
But the interesting thing is it’s pretty much at the same time—the last couple of years—that software has stopped sucking. Software has become tremendously more capable, and it’s the takeoff of learning systems. And the basic definition would be that if you can take arbitrary inputs that in the real world created caused outputs or actions—say for instance arbitrary inputs of what a person is experiencing in a room, and then the outputs of that person (the things that she says or does)—if you put those inputs into a black box and use the outputs as boundary conditions, we now have systems that will find connections between the two. They won’t be the same as happened inside her brain, causing her to say and do certain things as a response to those inputs, but there will be a system that will take a black box and find a route between those inputs and outputs. That’s incredible. That’s incredibly powerful and it’s one of the six methods by which we might approach AI. And when you have that, then you have a number of issues, like should we care what’s going on in that box?
And in fact, right now DARPA has six contracts out to various groups to develop internal state tracking of learning systems so that we can have some idea why a learning system connected this set of inputs to this set of outputs. But over the long run what you’re going to have is a person sitting in a room, listening to music, taking a telephone call, looking out the window at the beach, trolling the Internet, and then measuring all the things that she says and does and types. And we’re not that far away from the notion of being able to emulate a box that takes all the same inputs and will deliver the same outputs; at which point the experts will say, “This is an emulation,” but it will be an emulator that delivers outputs to perceptions similar to this person. And now we’re in science fiction realm, and only science fiction authors have been exploring what this means.
My experience with systems that tried to pass the Turing test… And of course you can argue what that would mean, but people write these really good chat bots that try to do it, and the first question I type in every one of them or ask is, “What’s bigger, a nickel or the Sun?” And I haven’t found one that has ever answered it correctly. So, I guess there’s a certain amount of skepticism that would accompany you saying something like in three years it’s going to carry on a conversation where it makes a forceful argument that it is sapient, that we’re going to be able to emulate so well that we don’t know whether it’s truly self-aware or not. That’s just such a disconnect from the state of the art.
When I talk to practitioners, they’re like, “My biggest problem is getting it to tell the difference between 8 and H when they’re spoken.” That’s what keeps these guys up at night. And then you get people like Andrew Ng who say these far out things, like worrying overpopulation of Mars and you get time horizons of 500 years before any of that. So, I’m really having trouble seeing it as a thousand or so days from now that we’re going to grapple with all of these in a real way.
But do you think that this radio show will be accessible to a learning system online?
You’re putting it on the Internet, right?
Okay, so then if you have a strong enough learning system that is voracious enough, it’s going to listen to this radio show and it will hear, it will tune in on the fact that you mentioned the word “Turing test,” just before you mentioned your test of which is bigger, the nickel or the Sun.
Which by the way, I never said the answer to that question in my setup of it. So it’s still no further along knowing.
The fact of the matter is that Watson is very good—if it’s parsed a question, then it can apply resources, or what it can do is it can ask a human because these will be teams, you see. The most powerful thing is teams of AI and humans. So, you’re not talking about something that’s going to be passing these Turing tests independently; you’re talking about something that has a bunch of giggling geeks in the background who desperately want it to disturb everybody, and disturb it it will, because these ELIZA-type emulation programs are extremely good at tapping into some very, very universal human interaction sets. They were good at it back in ELIZA’s day before you were born. I’m making an assumption there.
ELIZA and I came into the world about the same time.
But the point of ELIZA was, it was so bad at what it did, that Weizenbaum was disturbed that people… He wasn’t concerned about ELIZA; he was concerned about how people reacted to it.