by Martin Ford
Imagine a world in which there was a Star Trek transporter, and anything could appear at any place at any moment. You’d never be able to learn from that. What allows us to learn about the world is the fact that objects do move on paths that are connected in space and time, and over a billion years of evolution that might have been wired in as a way of getting you off the ground faster.
MARTIN FORD: Let’s talk about the future. What do you see as the main hurdles to getting to AGI, and can we get there with current tools?
GARY MARCUS: I see deep learning as a useful tool for doing pattern classification, which is one problem that any intelligent agent needs to do. We should either keep it around for that, or replace it with something that does similar work more efficiently, which I do think is possible.
At the same time, there are other kinds of things that intelligent agents need to do that deep learning is not currently very good at. It’s not very good at abstract inference, and it’s not a very good tool for language, except things like translation where you don’t need real comprehension, or at least not to do approximate translation. It’s also not very good at handling situations that it hasn’t seen before and where it has relatively incomplete information. We therefore need to supplement deep learning with other tools.
More generally, there’s a lot of knowledge that humans have about the world that can be codified symbolically, either through math or sentences in a language. We really want to bring that symbolic information together with the other information that’s more perceptual.
Psychologists talk about the relationship between top-down information and bottom-up information. If you look at an image, light falls on your retina and that’s bottom-up information, but you also use your knowledge of the world and your experience of how things behave to add top-down information to your interpretation of the image.
Deep learning systems currently focus on bottom-up information. They can interpret the pixels of an image, but don’t then have any knowledge of the object the image contains.
An example of this recently was the Adversarial Patch paper (https://arxiv.org/pdf/1712.09665.pdf). In the paper, they show how you can fool a deep learning system by adding a sticker to an image. They take a photo of a banana that is recognized with great confidence by a deep learning system and then add a sticker that looks like a psychedelic toaster next to the banana in the photo. Any human looking at it would say it was a banana with a funny looking sticker next to it, but the deep learning system immediately says, with great confidence, that it’s now a picture of a toaster.
The deep learning system is just trying to say what the most salient thing in the image is, and the high-contrast psychedelic toaster grabs its attention and it ignores the perfectly clear banana.
This is an example of how deep learning systems are only getting the bottom-up information, which is what your occipital cortex does. It’s not capturing at all what your frontal cortex does when it reasons about what’s really going on.
To get to AGI, we need to be able to capture both sides of that equation. Another way to put it is that humans have all kinds of common-sense reasoning, and that has to be part of the solution. It’s not well captured by deep learning. In my view, we need to bring together symbol manipulation, which has a strong history in AI, with deep learning. They have been treated separately for too long, and it’s time to bring them together.
MARTIN FORD: If you had to point to one company or project that’s going on now that is the closest to being on the path to AGI, who would you point to?
GARY MARCUS: I’m very excited about Project Mosaic at the Allen Institute for AI. They’re taking a second crack at the problem Doug Lenat was trying to solve, which was how you take human knowledge and put it in computable form. This is not about answering questions like where was Barack Obama born—computers actually represent that information pretty well and can extract it from available data.
There’s a lot of information, though, that’s not written anywhere, for example, toasters are smaller than cars. The likelihood is that no one says that on Wikipedia, but we know it to be true and that allows us to make inferences. If I said, “Gary got run over by a toaster,” you would think that’s weird because a toaster’s not that big an object, but “run over by a car” makes sense.
MARTIN FORD: So this is in the area of symbolic logic?
GARY MARCUS: Well, there are two related questions. One question is, how do you get that knowledge at all? The other is, do you want symbolic logic as a way to manipulate that?
My best guess is that symbolic logic is actually pretty useful for it, and we shouldn’t throw it out the window. I’m open to somebody finding another way to deal with it, but I don’t see any ways in which people have dealt with it well, and I don’t see how we can build systems that really understand language without having some of that common sense. This is because every time I say a sentence to you, there’s some common-sense knowledge that goes into your understanding of that sentence.
If I tell you I’m going to ride my bicycle from New York to Boston, I don’t have to tell you that I’m not going to fly through the air, go underwater, or take a detour to California. You can figure all that stuff out for yourself. It’s not in the literal sentence, but it’s in your knowledge about humans that they like to take efficient routes.
You, as a human, can make a lot of inferences. There’s no way you can understand my sentences without filling in those inferences, where you’re effectively reading between the lines. We read an enormous amount between the lines, but for that whole transaction to work there has to be shared common sense, and we don’t have machines that have that shared common sense yet.
The biggest project to do that was Doug Lenat’s Cyc, which started around 1984 and by most accounts, it didn’t work very effectively. It was developed 30 years ago in a closed form. Nowadays we know much more about machine learning, and the Allen Institute for AI is committed to doing things in open-source ways in which the community can participate. We know more about big data now than we did back in the 1980s, but it’s still a very difficult problem. The important thing is that they’re confronting it when everybody else is hiding from it.
MARTIN FORD: What do you think the timeframe is for AGI?
GARY MARCUS: I don’t know. I know most of the reasons why it’s not here now and the things that need to be solved, but I don’t think you can put a single date on that. What I think is that you need a confidence interval—as a statistician would describe it—around it.
I might tell you that I think it’ll come between 2030 if we’re phenomenally lucky and more likely 2050, or in the worst case 2130. The point is that it’s very hard to give an exact date. There are lots of things we just don’t know. I always think about how Bill Gates wrote the book The Road Ahead in 1994, and even he didn’t really realize that the internet was going to change things as it did. My point is that there could be all kinds of things that we’re just not anticipating.
Right now, machines are weak at intelligence, but we don’t know what people are going to invent next. There’s a lot of money going into the field, which could move things along, or alternatively it could be much harder than we think that it is. We just really don’t know.
MARTIN FORD: That’s still a fairly aggressive time frame. You’re suggesting as soon as 12 years away or as far away as 112 years.
GARY MARCUS: And those figures could of course be wrong. Another way to look at it is while we’ve made a lot of progress on narrow intelligence, we haven’t made nearly as much progress so far on general intelligence, AGI.
Apple’s Siri, which began life in 2010, doesn’t work that much differently from ELIZA, an early natural language computer program created in 1964, which matched templates in order to give an illusion of understanding language that it didn’t really. A lot of my optimism comes from how many people are working at the problem and how much money businesses are investing to try and solve it.
MARTIN FORD:
It definitely is a massive change in terms of AI not being something just done as a research project at a university. Now AI is central to the business models of big companies like Google and Facebook.
GARY MARCUS: The amount of money being spent on AI far eclipses anything before, although there certainly was a lot of money spent in the 1960s and early 1970s before the first so-called AI Winter. It’s also important to acknowledge that money is not a guaranteed solution to the problems of AI, but is very likely a prerequisite.
MARTIN FORD: Let’s focus on a prediction for a much narrower technology: the self-driving car.
When are we going to be able to call for something like an Uber that’s driven by nothing but an AI, and that can pick you up at a random location and then take you to a destination that you specify?
GARY MARCUS: It’s at least a decade away, and probably more.
Martin Ford: You’re almost getting into the same territory as your AGI prediction.
GARY MARCUS: That’s right, and for the principal reason that if you’re talking about driving in a very heavy metropolitan location like Manhattan or Mumbai, then the AI will face a lot of unpredictability. It’s one thing to have a driverless car in Phoenix, where the weather is good and the population is a lot less densely packed. The problem in Manhattan is that anything goes at any moment, nobody is particularly well-behaved and everybody is aggressive, the chance of having unpredictable things occur is much higher.
Even simple road elements like barricades to protect people can cause issues for an AI. These are complex situations that humans deal with by using reasoning. Right now, driverless cars navigate by having highly detailed maps and things like LIDAR, but no real understanding of the motives and behavior of other drivers. Humans have an OK visual system, but a good understanding of what’s out there and what they’re doing when they’re driving. Machines are trying to fake their way around that with big data, and it’s not clear to me that you get to the accuracy levels you need for driving in Manhattan simply by adding more data to these big data-driven systems. You might get to 99.99% accuracy, but if you do the numbers on that, that’s much worse than humans, and it’s far too dangerous to have that scale on the road, especially on busy streets like in Manhattan.
MARTIN FORD: Maybe then there’s a nearer term solution, where rather than a location of your choice, it takes you to a predefined location?
GARY MARCUS: There’s a possibility that very soon we may well have that in Phoenix, or another limited location. If you can find a route where you never need to take a left turn, where humans are unlikely to be in the way, and the traffic is civilized, you might be able to do that. We already have monorails at airports that work in a similar way following a predefined path.
There’s a continuum from super-controlled circumstances like the monorail at the airport where nobody should be on that track, to the Manhattan streets where anybody and anything could be there at any time. We also have other factors like the fact that weather is much more complicated than in Phoenix. We get everything. Sleet, slush, hail, leaves, things that fall off trucks; everything.
The more you go into an unbounded open-ended system, the more challenge there is and the more you need to be able to reason the way an AGI system does. It’s still not as open-ended as AGI per se, but it starts to approach that, and that’s why my numbers are not totally different.
MARTIN FORD: Let’s talk about an area that I’ve focused on a lot: the economic and job market impact of AI.
Many people believe we’re on the leading edge of a new Industrial Revolution; something that’s going to completely change the way the labor market looks. Would you agree with that?
GARY MARCUS: I do agree with that, though on a slightly slower time frame. Driverless cars are harder than we thought, so paid drivers are safe for a while, but fast-food workers and cashiers are in deep trouble, and there’s a lot of them in the workplace. I do think these fundamental changes are going to happen. Some of them will be slow, but in the scale of, say, 100 years, if something takes an extra 20 years, it’s nothing.
There is going to be a problem with AI robots and employment sometime in this century, whether it’s 2030 or 2070. At some point we need to change how we structure our societies because we are going to get to a point where there’s less employment available but still a working-age population.
There are counterarguments, like when most agricultural jobs disappeared they were just replaced by industrial jobs, but I don’t find them to be compelling. The main issue we will face is scale and the way that once you have a solution, you can use it everywhere relatively cheaply.
Getting the first driverless car algorithm/database system that works might be 50 years of work and cost billions of dollars in research, but once we have them, people are going to roll them out at scale. As soon as we reach that point, millions of truck drivers are going to be put in a position where they could lose their jobs within a matter of years.
It’s not clear that we are going to have new jobs appear that can replace existing jobs at the scale of the truck-driving industry. A lot of the new jobs that have arisen need fewer people. For example, a YouTube entrepreneur is a great job. You can make millions of dollars staying at home making videos. That’s terrific, but maybe 1,000 people do that, not a million people and not enough to replace all the potentially lost truck driving jobs.
It’s easy to come up with jobs that we will have that we didn’t have before, but it’s hard to come up with new industries that will employ large numbers of people in an era where you can build something like Instagram with 18 people.
MARTIN FORD: Probably something like half the current workforce is engaged in fundamentally predictable activities. What they’re doing is encapsulated in the data and ultimately is going to be susceptible to machine learning over some time frame.
GARY MARCUS: True, but things like that that are pretty hard right now. AI systems just don’t really understand the data as a natural language like humans do. For example, extricating information from medical records is something that’s very hard for machines to do right now. It’s predictable, and it’s not that hard work, but doing it well with a machine is a while away. In the long run, though, I agree with you. Natural language understanding will get better and eventually those predictable jobs will go away.
MARTIN FORD: Given that you believe job loss to AI will happen at some point, would you support a basic income as a potential solution to that?
GARY MARCUS: I see no real alternative. We will get there, but it’s a question of whether we get there peacefully through a universal agreement or whether there are riots on the street and people getting killed. I don’t know the method, but I don’t see any other ending.
MARTIN FORD: You could argue that technology is already having an impact of that sort. We do have an opioid epidemic in the US at the moment, and automation technology in factories has likely played a role in that in terms of middle-class job opportunities disappearing. Perhaps opioid use is tied to a perceived loss of dignity or even despair among some people, especially working-class men?
GARY MARCUS: I would be careful about making that assumption. It may be true, but I don’t think the links are ironclad. A better analogy, in my opinion, is how a lot of people use their phones as an opioid, and that smartphones are the new opium of the people. We may be moving toward a world where a lot of people just hang out in virtual reality, and if the economics work they may be reasonably happy. I’m not sure where that’s all going to go.
MARTIN FORD: There’s a range of risks associated with AI that have been raised. People like Elon Musk have been especially vocal about existential threats. What do you think we should be worried about in terms of the impacts and risks of AI?
GARY MARCUS: We should be worrying about people using AI in malevolent ways. The real problem is what people might do with the power that AI holds as it becomes more embedded in the grid and more hackable. I’m not that worried about AI systems ind
ependently wanting to eat us for breakfast or turn us into paper clips. It’s not completely impossible, but there’s no real evidence that we’re moving in that direction. There is evidence, though, that we’re giving more and more power to those machines, and that we have no idea how to solve the cybersecurity threats in the near term.
MARTIN FORD: What about long-term threats, though? Elon Musk and Nick Bostrom are very concerned about the control problem with AI; the idea that there could be a recursive self-improvement cycle that could lead to an intelligence explosion. You can’t completely discount that, right?
GARY MARCUS: I don’t completely discount it, I’m not going to say the probability is zero but the probability of it happening anytime soon is pretty low. There was recently a video circulated of robots opening doorknobs, and that’s about where they are in development.
We don’t have AI systems that can robustly navigate our world at all, nor do we have robotic systems that know how to improve themselves, except in constrained ways like tailoring their motor control system to a particular function. This is not a current problem. I think it’s fine to invest some money in the field and have some people think about those problems. My issue is that, as we saw with the 2016 US election, there are more pressing problems like using AI to generate and target fake news. That’s a problem today.
MARTIN FORD: Earlier you said AGI is conceivable as early as 2030. If a system is genuinely intelligent, potentially superintelligent, do we need to make sure that its goals are aligned with what we want it to do?