I think so. I think there is no question that consciousness is linked to general intelligence because general intelligence means that you need to able to create an abstraction of the world, which means that you need to be able to go beyond observing it, but also be able to understand it and to experience it. So, I think that is a very simple way to put it. What I’m actually wondering is whether consciousness was a consequence of biology and whether we need to replicate that in a machine, to make it intelligent like a human being is intelligent. So essentially, the way I’m thinking about this is, is there a way to build a human intelligence that would seem human? And do we want that to seem human? Because if it’s just about reproducing the way intelligence works in a machine, then we shouldn’t care if it feels human or not, we should just care about the ability for the machine to do something smart. So, I think the question of consciousness in a machine is really down to the question of whether or not we want to make it human. There are many technologies that we’ve built for which we have examples in nature, which perform the same task, but don’t work the same. Birds and planes, for example, I’m pretty sure a bird needs to have some sort of like, consciousness of itself of not getting into the wall, whereas we didn’t need to replicate all those tiny bits for the actual plane to fly. It’s just a very different way of doing things.
So do you have a theory as to how it is that we’re conscious?
Well, I think it probably comes from the fact that we had to evolve as a species with other individuals, right? How would you actually understand where to position yourself in society, and therefore, how to best build a very coherent, stable, strong community, if you don’t have consciousness of other people, of nature, of yourself? So, I think there is like, inherently, the fact that having a kind of ecosystem of human beings, and humans in nature, and humans and animals meant that you had to develop consciousness. I think it was probably part of a very positive evolutionary strategy. Whether or not that comes from your neurons or whether that comes more from a combination of different things, including your senses, I’m not sure. But I feel that the need for consciousness definitely came from the need for integrating yourself into broader structure.
And so not to put words in your mouth, but it sounds like you think, you said “we’re not close to it,” but it is possible to build an AGI, and it sounds like you think it’s possible to build, hypothetically, a conscious computer and you’re asking the question of would we want to?
Yes. The question is whether or not it would make sense for whatever we have in mind for it. I think probably we should do it. We should try to do it just for the science, I’m just not sure this is going to be the most useful thing to do, or whether we’re going to figure out an even more general general-intelligence which doesn’t have only human traits but has something even more than this, that would be a lot more powerful.
Hmmm, what would that look like?
Well, that is a good question. I have clearly no idea because otherwise—it is very hard to think about a bigger intelligence and the intelligence that we are limited to, in a sense. But it’s very possible that we might end up concluding that well you know, human intelligence is great for being a human, but maybe a machine doesn’t have to have the same constraints. Maybe a machine can have like a different type of intelligence, which would make it a lot better suited for the type of things we’re expecting the machine to do. And I don’t think we’re expecting the machines to be human. I think we’re expecting the machines to augment us, to help us, to solve problems humans cannot solve. So why limit it to a human intelligence?
So, the people I talk to say, “When will we get an AGI?” The predictions vary by two orders of magnitude—you can read everything from 5 to 500 years. Where do you come down on that? You’ve made several comments that you don’t think we’re close to it. When do you think we’ll see an AGI? Will you live to see an AGI, for instance?
This is very, very hard to tell, you know I mean there is this funny artifact that everybody makes a prediction 20 years in the future, and it’s actually because most people when they make those predictions, have about 20 years left in their careers. So, you know, nobody is able to think beyond their own lifetime, in a sense. I don’t think it’s 20 years away, at least not in the sense of real human intelligence. Are we going to be able replicate parts of AGI, such as, you know, the ability to transfer learning from one task to another? Yes, and I think this is short-term. Are we going to be able to build machines that can go one level of abstraction higher to do something? Yes, probably. But it doesn’t mean they’re going to be as versatile, as generalist, as horizontally thinking as we are as humans. I think for that, we really, really have to figure out once and for all whether a human intelligence requires a human experience of the world, which means the same senses, the same rules, the same constraints, the same energy, the same speed of thinking, or not. So, we might just bypass, as I said—human intelligence might go from like narrow AI, to a different type of intelligence, that is neither human or narrow. It’s just different.
So you mentioned transferred learning. I could show you a small statue of a falcon, and then I could show you a hundred photographs, and some of them have the falcon under water, on its side, in different light, upside down, and all these other things. Humans have no problem saying, “there it is, there it is, there it is,” you know just kind of find Waldo [but] with the falcon. So, in other words, humans can train with a sample size of one, primarily because we have a lot of experience seeing other things in lowlight and all of that. So, if that’s transferred learning it sounds like you think that we’re going to be able to do that pretty quickly, and that’s kind of big deal if we can really teach machines to generalize the way we do. Or is that kind of generalization that I just went through, that actually is part of our general intelligence at work?
I think transferred learning is necessary to build AGI, but it’s not enough, because at the end of the day, just because a machine can learn to play a game and then you know have a starting point to play another game, doesn’t mean that it will make the choice to learn this other game. It will still be you telling it, “Okay, here is a task I need you to do, use your existing learning to perform it.” It’s still pretty much task-driven, and this is a fundamental difference. It is extremely impressive and to be honest I think it’s absolutely necessary because right now when you look at what you do with machine learning, you need to collect a bunch of different examples, and you’re feeding that to the machine, and the machine is learning from those examples to reproduce that behavior, right? When you do transferred learning, you’re still teaching a lot of things to the machine, but you’re teaching it to reuse other things so that it doesn’t need as much data. So, I think inherently the biggest benefit of transferred learning will be that we won’t need to collect as much data to make the computers do something new. It solves, essentially, the biggest friction point we have today, which is how do you access enough data to make the machine learn the behavior? In some cases, the data does not exist. And so I think transferred learning is a very elegant and very good solution to that problem.
So last question I want to ask you about AGI and then we can turn the clock back and talk to issues closer at hand is as follows: It sounds like you’re saying an AGI is more than 20 years off, if I just inferred that from what you just said. And I am curious because the human genome is 2 billion base pairs, it’s something like 700 MB of information, most of which we share with plants, bananas, and what-not. And if you look at our intelligence versus a chimp, or something, we only have a fraction of 1% of the DNA that is different. What that seems to suggest to me at least is that if the genome is 700 MB, and the 1% difference gives us an AGI, then the code to create an AGI could be a small as 7 MB.
Pedro Domingos wrote a book called The Master Algorithm, where he says that there probably is an algorithm, that can solve a whole world of problems, and get us really close to AGI. Then other people on another end of the spectrum, like Marvin Minsky or somebody, don’t even know that we have an AGI, that we’re like just 200 different hacks—kind of 200 narrow intelligences that just kind of pull off this trick of seeming like a general intelligence. I’m wondering if you think that an AGI could be relatively simple—that it’s not a matter of more data or more processing, but just a better algorithm?
So just to be clear, I don’t consider a machine who can perform 200 different tasks to be an AGI. It’s just like an ensemble of, you know, narrow AIs.
Right, and that school of thought says that therefore we are not an AGI. We only have this really limited set of things we can do that we like to pass off as “ah, we can do anything,” but we really can’t. We’re 200 narrow AIs, and the minute you ask us to do things outside of that, they’re off our radar entirely.
For me, the simplest definition of how to differentiate between a narrow AI and an AGI is, an AGI is capable of kind of zooming out of what it knows—so to have basically like a second-degree view of the facts that it learned, and then reuse that to do something completely different. And I think this capacity we have as humans. We did not have to learn every possible permutation; we did not have to learn every single zooming out of every fact in the world, to be able to do new things. So, I think I definitely agree that as a human, we are AGI. I just don’t think that having a computer who can learn to do two hundred different things would do that. You would still need to figure out this ability to zoom out, this ability to create abstraction of what you’ve been learning and to reapply it somewhere else. I think this is really the definition of horizontal thinking, right? You can only think horizontally if you’re looking up, rather than staying in a silo. So, to your question, yea. I mean, why not? Maybe the algorithm for AGI is simple. I mean think about it. Deep learning, machine learning in general, these are deceptively easy in terms of mathematics. We don’t really understand how it works yet, but the mathematics behind it is very, very, easy. So, we did not have to come up with this like crazy solution. We just came up with an algorithm that turned out to be simple, and that worked really well when given a ton of information. So, I’m pretty sure that AGI doesn’t have to be that much more complicated, right? It might be one of those E = mc2sort of plugins I think that we’re going to figure out.
That was certainly the hope, way back, because physics itself obeys such simple laws that were hidden from us, and then once elucidated seemed, any 11th gradehigh-school student could learn, maybe so. So, pulling back more toward the here and now—in ’97, Deep Blue beat Kasparov, then after that we had Ken Jennings lose in Jeopardy, then you had AlphaGo beat Lee Sedol, then you had some top-ranked poker players beaten, and then you just had another AlphaGo victory. So, AI does really well at games presumably because they have a very defined, narrow rule set, and a constrained environment. What do you think is going to be, kind of, the next thing like that? It hits the papers and everybody’s like, “Wow, that’s a big milestone! That’s really cool. Didn’t see that coming so soon!” What do you think will be the next sort of things we’ll see?