09: Gary Marcus — Making AI More Human

Listen to the episode on iTunes here

Gary Marcus, founder and CEO of Geometric Intelligence

AMLG: Gary I’m super excited to have you today, thanks for coming on the show. We first met a few years ago in New York when I was running a tech meetup, the Singularity society, and you kindly came and spoke. You’ve been a professor of psychology at NYU for many years where your work has focused on language, biology, and the human mind. You’re the author of four books and you’re now a popular contributing author to the New Yorker, Wired and the New York Times.

You’ve spent decades studying how children learn, and then in 2015 you founded this startup called Geometric Intelligence, focused on mining cognitive psychology for insights into building better machine learning techniques. Just this past December you were acquired by Uber to run their newly founded AI labs — congratulations on that exit.

An example of a classification problem solved by a convolutional neural net — a series of labelled inputs go into the system, then layers of nodes modeled on neurons detect low level characteristics and move up a hierarchy toward higher level characteristics, then system generates an output

So your algorithms offer an alternative approach to what is now a very popular branch of machine learning, called deep learning. Let’s talk about deep learning — it’s a sexy buzzword which is thrown into about every startup pitch I see these days, and many corporate presentations, so I’m sure listeners have heard the term. What it really is is a rebranding of an old technique of using neural nets, which dates back to the 50s. Neural nets basically mimic the human neocortex, and by feeding in massive amounts, gigabytes of data and using tons of computational power, the algorithms are able to recognize patterns. Part of the reason why this technique is back in vogue is the combination of increasingly powerful computers combined with the massive training datasets that companies are building up. So there’s been a flurry of activity, and the Googles and Facebooks of the world are throwing resources at the technique. As just one example, Facebook, using the over 400 billion photos people have uploaded, has built something called DeepFace, an image recognition tool that’s now better than humans at recognizing whether two different images are of the same person.

Gary Marcus at EmTech Digital Conference, May 2016. Watch the full presentation.

Gary you are well known as a critic of this technique, you’ve said that it’s over-hyped. That there’s some low hanging fruit that deep learning’s good at — specific narrow tasks like perception and categorization, and maybe beating humans at chess, but you felt that this deep learning mania was taking the field of AI in the wrong direction, that we’re not making progress on cognition and strong AI. Or as you’ve put it, “we wanted Rosie the robot, and instead we got the roomba.” So you’ve advocated for bringing psychology back into the mix, because there’s a lot of things that humans do better, and that we should be studying humans to understand why they do things better. Is this still how you feel about the field?

GM: Pretty much. There was probably a little more low hanging fruit than I anticipated. I saw somebody else say it more concisely, which is simply, deep learning does not equal AGI (AGI is “artificial general intelligence.”) There’s all the stuff you can do with deep learning, like it makes your speech recognition better. It makes your object recognition better. But that doesn’t mean it’s intelligence. Intelligence is a multi-dimensional variable. There are lots of things that go into it.

Perception is just one slice of what makes up “intelligence”

In a talk I gave at TEDx CERN recently, I made this kind of pie chart and I said look, here’s perception that’s a tiny slice of the pie. It’s an important slice of the pie, but there’s lots of other things that go into human intelligence, like our ability to attend to the right things at the same time, to reason about them to build models of what’s going on in order to anticipate what might happen next and so forth. And perception is just a piece of it. And deep learning is really just helping with that piece.

In a New Yorker article that I wrote in 2012, I said look, this is great, but it’s not really helping us solve causal understanding. It’s not really helping with language. Just because you’ve built a better ladder doesn’t mean you’ve gotten to the moon. I still feel that way. I still feel like we’re actually no closer to the moon, where the moonshot is intelligence that’s really as flexible as human beings. We’re no closer to that moonshot than we were four years ago. There’s all this excitement about AI and it’s well deserved. AI is a practical tool for the first time and that’s great. There’s good reason for companies to put in all of this money. But just look for example at a driverless car, that’s a form of intelligence, modest intelligence, the average 16-year-old can do it as long as they’re sober, with a couple of months of training. Yet Google has worked on it for seven years and their car still can only drive — as far as I can tell since they don’t publish the data — like on sunny days, without too much traffic…

AMLG: And isn’t there the whole black box problem that you don’t know what’s going on. We don’t know the inner workings of deep learning, it’s kind of inscrutable. Isn’t that a massive problem for things like driverless cars?

GM: It is a problem. Whether it’s an insuperable problem is an open empirical question. So it is a fact at least for now that we can’t well interpret what deep learning is doing. So the way to think about it is you have millions of parameters and millions of data points. That means that if I as an engineer look at this thing I have to contend with these millions or billions of numbers that have been set based on all of that data and maybe there is a kind of rhyme or reason to it but it’s not obvious and there’s some good theoretical arguments to think sometimes you’re never really going to find an interpretable answer there.

There’s an argument now in the literature which goes back to some work that I was doing in the 90s about whether deep learning is just memorization. So this was the paper that came out that said it is and another says no it isn’t. Well it isn’t literally exactly memorization but it’s a little bit like that. If you memorize all these examples, there may not be some abstract rule that characterizes all of what’s going on but it might be hard to say what’s there. So if you build your system entirely with deep learning, which is something that Nvidia has played around with, and something goes wrong, it’s hard to know what’s going on and that makes it hard to debug.

AMLG: Which is a problem if your car just runs into a lamppost and you can’t debug why that happened.

GM: You’re lucky if it’s only a lamppost and not too many people are injured. There are serious risks here. Somebody did die, though I think it wasn’t a deep learning system in the Tesla crash, it was a different kind of system. We actually have problems on engineering on both ends. So I don’t want to say that classical AI has fully licked these problems, it hasn’t. I think it’s been abandoned prematurely and people should come back to it. But the fact is we don’t have good ways of engineering really complex systems. And minds are really complex systems.

AMLG: Why do you think all these big platforms are reorganizing around AI and specifically deep learning. Is it just that they’ve got data moats, so you might as well train on all of that data if you’ve got it?

GM: Well there’s an interesting thing about Google which is they have enormous amounts of data. So of course they want to leverage it. Google has the power to build new resources that they give away free and they build the resources that are particular to their problem. So Google because they have this massive amount of data has oriented their AI around, how can I leverage that data? Which makes sense from their commercial interests. But it doesn’t necessarily mean, say from a society’s perspective. does society need AI? What does it need it for? Would be the best way to build it?

“CERN is a vast interdisciplinary, multi-country consortium to solve particular scientific problems, maybe we need the same thing for AI. Most of the efforts in AI right now are individual companies or small labs working on small problems like how to sell more advertising… what if we brought people together to try this moonshot of doing better science, and what if we brought not just machine learning experts, and engineers who can make faster hardware, but researchers who look at cognitive development. I think we could make some progress” — Gary Marcus

I think if you asked those questions you would say, well what society most needs is automated scientific discovery that can help us actually understand the brain to cure neural disorders, to actually understand cancer to cure cancer, and so forth. If that were the thing we were most trying to solve in AI, I think we would say, let’s not leave it all in the hands of these companies. Let’s have an international consortium kind of like we had for CERN, the large hadron collider. That’s seven billion dollars. What if you had $7 billion dollars that was carefully orchestrated towards a common goal. You could imagine society taking that approach. It’s not going to happen right now given the current political climate.

AMLG: Well they are at least coming together on AI ethics. So that’s a start.

GM: It is good that people are talking about the ethical issues and there are serious issues that deserve consideration. The only thing I would say there is, some people are hysterical about it, thinking that real AI is around the corner and it probably isn’t. I think it’s still OK that we start thinking about these things now, even if real AI is further away than people think it is. If that’s what moves people into action and it takes 20 years, but the action itself takes 20 years, then it’s the right timing to start thinking about it now.

Geometric’s Xprop algorithm systematically beating convolutional nets

AMLG: I want to get back to your alternative approach to solving AI, and why it’s so important. So you’ve come up with what you believe is a better paradigm, taking inspiration from cognitive psychology. The idea is that your algorithms are a much quicker study, that they’re more efficient and less data hungry, less brittle and that they can have broader applicability. And in a brief amount of time you’ve had impressive early results. You’ve run a bunch of image recognition tests comparing the techniques and have shown that your algorithms perform better, using smaller amounts of data, often called sparse data.

In the real world, there’s a long tail of many cases where there isn’t a lot of data

So deep learning works well when you have tons of data for common examples and high frequency things. But in the real world, in most domains, there’s a long tail of things where there isn’t a lot of data. So while neural nets may be good at low level perception, they aren’t as good at understanding integrated wholes. So tell us more about your approach, and how your training in cognitive neuroscience has informed it?

GM: My training was with Steve Pinker. And through that training I became sensitive to the fact that human children are very good at learning language, phenomenally good, even when they’re not that good at other things. Of course I read about that as a graduate student, now I have some human children, I have a four-year-old and a two-and-a-half year old. And it’s just amazing how fast they learn.

AMLG: The best AI’s you’ve ever seen.

GM: The best AI’s I’ve ever seen. Actually my son shares a birthday with Rodney Brooks, who’s one of the great roboticists, I think you know him well. For a while I was sending Rodney an e-mail message every year saying “happy birthday. My son is now a year old. I think he can do this and your robots can’t.” It was kind of a running joke between us.

AMLG: And now he’s vastly superior to all of the robots

GM: And I didn’t even bother this year. The four year olds of this world, what they can do in terms of motor control and language is far ahead of what robots can do. And so I started thinking about that kind of question really in the early 90s. and I’ve never fully figured out the answer but part of the motivation for my company was, hey we have these systems now that are pretty good at learning if you have gigabytes of data and that’s great work if you can get it, and you can get it sometimes.

So speech recognition, if you’re talking about white males asking search queries in a quiet room, you can get as much labelled data, which is critical, for these systems as you want. This is how somebody says something and this is the word written out. But my kids don’t need that. They don’t have labelled data, they don’t have gigabytes of labelled data, they just kind of watch the world and figure all this stuff out.

Andrew Ng, who was running Google Brain in 2012. Google scientists connected 16,000 processors and presented it with 10 million images from YouTube. The neural network taught itself to recognize cats. [Source: NYT]

AMLG: My favorite example is the Google brain, which had a neural net of 16,000 processors. They showed it 10 million random youtube videos over a week and it learned to recognize cats. I’m sure your four-year-old learned that in like two days.

GM: Well and if it learned to recognize cats the way that that system, which would be like they recognize a cat face when it’s square on, straight in front of them, but not anywhere else? I would bring the kid to the neurologist. I would say there’s something wrong my kid is not really understanding what a cat is and I’ve shown him 16 million videos. The system even when you give it all that data is not really learning the same thing as a human child would, it’s not as resourceful.

AMLG: So why wouldn’t everyone do this sparse data approach? because most people, most companies don’t have a lot of data and it seems to make sense to model on the human brain. Why isn’t it a more popular approach?

GM: Well there’s a few different things. One is that we don’t know how the brain does it. So I can construct an argument for the fact that the brain does, but doesn’t mean how the brain does it.

Another example is, I know the brain has short term memory and I know a lot of the properties of it from psychological experiments, but I don’t know what the neural realization of short term memory is. You’d think that we would by now, but 100 to 200 years into neuroscience, we still don’t actually know the mechanism for short term memory that allows me to have you say, “meet me at the top of the elevator” and I find you at the top of the elevator. Two is, if you’re Google, maybe that’s not actually what your first choice project should be anyway. You are sitting on mountains of data. You’re sitting on planets full of data, and there’s a lot of stuff you can do with that, why work on the

AMLG: Play to your advantage.

GM: Play to your advantage. So there’s a particular project that you could call the cat detector that you just referred to which was actually the first project of Google brain, which is a particular kind of unsupervised learning, it was on the front page of the New York Times. It got an enormous amount of press. There’s something worth expanding there. So most of what’s going on is supervised learning. So we have an input. We have an output. In that case maybe I have an English sentence, I have a French translation. And we know a lot about supervised learning now — that’s what deep learning does. Unsupervised learning is actually what the cat detector did. So what was interesting about it, but which so far is a failure, I’m not saying it will always be a failure, was that it didn’t have any labels for the videos at all. It just kind of tried to sort out what it saw. And along the way it built this cat detector. It was not told, these are cats. So that was pretty cool. But the truth is it wasn’t very robust.

Yann LeCun — who’s my colleague at NYU — and I disagree about many things. We have public disagreements it’s kind of a fun game for us. This is one thing that we actually agree on, that we need to do unsupervised learning better. We disagree about how to get there. But we would both say the world doesn’t give you supervision for everything. Sometimes they’re neat tricks you can use, like if I’m trying to predict the future, I can have a video and I can predict from time one what time two is going to be like. There are tricks you can use, but in the end, most of what children do is unsupervised learning. We parents don’t give supervised training examples that much and when we do the kids ignore them. I say “went” when my kid says “go'ed” and then he says “go’ed”, he doesn’t care that I said “went.” He’s speaking his own language not my language.

AMLG: How has parenting influenced your thinking about AI ?

Gary’s son is really good at logical reasoning

GM: I mean I’m even more impressed with how much better the kids are than the machines. I’m also impressed with how slow kids are to learn certain things, like how many times do I need to correct my child and it has no effect. But on the whole as slow as kids are to learn certain things, like how to read time, as compared to machines, at the core of things, which I would say are understanding language, being able to maneuver around the world, being able to predict the world, kids are still way way better than machines. I wish I could say, having watched my kids I know exactly how. I don’t know exactly how. I have some clues and hints, but it’s a pretty hard problem.

AMLG: You have said before when talking about the brain, that humans are a rationalizing animal, that we come up with things, post hoc rationalizations for what we do. Beyond the sort of Daniel Kahneman irrational thinking and quirks in judgement, we make pretty terrible decisions, like genocide and jihad. You’ve said that the brain’s not amazingly well engineered, it’s not elegant in the way that say physics is elegant, that it’s in fact fairly crummy despite evolution wanting to make it better. So do we really want to model AI on the human brain when a) we don’t really know how it works and b) its kind of a mess of a machine?

GM: I don’t think we want to slavishly copy it. But there are some hints in there that we haven’t figured out. In any given domain it’s actually an open question. So it turns out the best way to play chess at least for now is not anything at all like a human being. The best way to play chess is do these enormous move trees. I go there you go there I go there. People try to do that but they can only go a few moves deep. A computer takes the same algorithm and it can go 20 moves deep and it can do it a billion times a second and it’s just better for them to do that than all the pattern recognition that people do.

Some things are going to be like chess, where the best way to do it has nothing to do with how people do it, and it could be that the best way to do everything is nothing to do with people. But right now we have no ways to read unstructured text, which is to say like read a newspaper story or a novel. We have no way to have a machine take an arbitrary problem and take a whack at it and do a good job. You can kind of cheat a little and say, I’m going to limit it to Atari games that are low resolution that I can play a billion times and the rules will never change. But that’s not the same thing as like, I can tell an undergraduate, I want you to write a term paper on something that you like about developmental psychology and they turn that into something. They don’t they don’t have 16,000 training examples of writing papers about developmental psychology. They go out in the literature they read new studies they form ideas. We don’t know how to build a machine to do that.

AMLG: Do you think we’re missing some of the use cases, do you think maybe we’re coming from the wrong direction, focused on what can we do, rather than what should we be doing?

GM: The famous proverb there is, give a man a hammer and he thinks everything’s a nail. Deep learning is a great hammer and people have found a whole bunch of nails and then they’ve also taken a whole bunch of screws that really shouldn’t be nailed and try to bang bang bang to get it in there because they have that hammer and they don’t want to take the time to learn how to build a screwdriver.

AMLG: Right. The other interesting thing is that you’re following a tradition of professors going from academia to corporations. Rob Fergus and Yann LeCun, both colleagues of yours at NYU, have figured out an arrangement where they have one foot in Facebook and one in NYU. I mean it’s obviously more lucrative to run AI at Facebook, but arguably more interesting too, since you can make more progress with the weight of those resources behind you right. What’s your take on the flood of talent from the classroom to the private sector, do you think there’s a vacuum being created in academia?

GM: We may live in a tragedy of the commons time where what’s good for a lot of individuals is not good for the group. So yes, for people who are good at this field right now, there’s a strong reason to be in industry rather than in academy, or at least to have a joint appointment — I still have an appointment at NYU, I’m on leave — and that strong argument starts with data. There’s no question that having a lot of data helps and what you see in industry is just completely different from what’s available in the academy. In industry, they can commission a thousand people to do this Amazon Turk stuff to get the data that they need, or they have it from a billion customers.

So the data are phenomenal, and the compute resources are phenomenal. Especially for people who are doing deep learning, they need lots of GPUs to do their work and it’s hard to get that. The government’s not going to give people the money that they need. So it’s very tempting if you have talent in this field to go off into industry. I’m not saying it’s the best long-term solution for the world. But at the individual level it’s the rational thing to do, unless you know society changes how it structures things.

AMLG: I’ve heard Zuckerberg personally calls major AI hires to recruit them, and we’re seeing pretty incredible starting salaries of seven figures and huge signing bonuses. And from the VC side, a lot of founders raising money sprinkle “AI” into their pitch deck to get a valuation bump and raise quicker. Do you think there’s a sort of inflation going on here, or is it merited?

AMLG: Well I think it’s a very useful tool right now and not everybody knows how to do it so that drives the price up. We live in an interesting moment in history where if you have a slightly better solution to a problem, you might be able to scale it up to a massive amount. Is it worth it for Google to sell ads 2% better? It’s worth a ton of money. So it’s not irrational from Google’s perspective to spend that amount of money. Then if you want to build your own video captioning company or something like that you’re competing with Google. That’s a fact.

Geometric was co-founded by Gary Marcus, Zoubin Ghahramani, Doug Bemis and Ken Stanley in 2015

AMLG: So how does a new AI startup compete with market prices. How did you manage that? I mean, you built a company in 18 months with 15 of the top names in the field and some of them had offers from Google. How do you compete with the big guys for talent, how do you survive as a startup with the costs of GPUs and cost of legal around patents, and all of that? How do you build a standalone business?

GM: I mean it’s challenging. One side of what you said is the money to start a company is pretty freely available in a way that probably it wouldn’t be in even another industry right now or in a different time. So we didn’t have a lot of trouble raising money, which helped. But the main thing is I got some great people from the beginning that I knew well and who trusted me. And who I felt really good about. Together we had names that were strong, and talent breeds talent so if you have some good people then you can get some more. And we had some pretty interesting ideas, which unfortunately I’m not allowed to talk about as they’re now the property of Uber —

The exponential curve of Moore’s Law suggests it’s going to take us until 2025 to build a computer with the processing power of the human brain. Source

AMLG: Do you think having the top talent in industry will help us get to general intelligence faster? What other factors contribute to how fast we get to AGI — does the number of people working on AI matter? What do you think the trajectory towards AGI over the next 10 to 20 years will look like, is it steady or lumpy progress?

GM: Certainly having more bodies on the ground is going to get us there faster. I am concerned that right now the commercial exigencies lead towards short-term things that aren’t solving long-term problems. So I don’t know that right around the corner we’re going to get to the kind of AGI that I would be satisfied with. But I think it will happen in the next century and I certainly don’t see any reason why it should take longer than that. It might happen sooner. There are many variables that are at stake. One thing that might happen is that in about three years people realize that we’re still not making enough progress on natural language understanding. That even with all this deep learning, we have poor results that work like 80% of the time, which is fine for some problems. But if you had a domestic robot that only understood a fifth of your instructions you would return it.

AMLG: Or behead it.

GM: Hopefully before it beheads you. I mean it would be a serious problem. So at some point society is going to be pressed to make it’s natural language understanding stuff better. We certainly aren’t going to get to machines that really understand science if they can’t read. I mean this is ridiculous.

In Jan 2017, a TV segment about Amazon Echo accidentally set of a slew of other Alexas to attempt shopping sprees

AMLG: Speaking of ridiculous, did you see the Alexa dollhouse chain reaction? There was a TV piece about Amazon Echo, and they’re filming, and the kid “says, Alexa, order me a dollhouse.” Then this segment was playing and every Alexa that was listening to the program started ordering dollhouses.

GM: I think it’s proof of one of the points that I like to make which is how superficial the systems are. If you program something to respond to a keyword, rather than a understanding of the scene and what’s going on, that’s going to happen. If you’re just looking for this word Alexa and you don’t know anything about social context how could it not happen. Of course these things are going to happen because we’re not trying to build systems that build internal representations of the world that they’re in, the relation between the people that are in it and the things going on around them. Until we do, you are just relying on things like keywords. You’re going to see this over and over again.

Two Google Homes arguing— see full video

AMLG: Are you using Amazon Echo (Alexa) at home or Google Home? Have you got a couple of them set up next to each other arguing back and forth in a never-ending loop?

GM: It doesn’t hold that much interest from me. I have a piece in the Scientific American about replacing the Turing test. Part of the premise is that fooling a person doesn’t really tell us anything, doesn’t really move us towards understanding.

AMLG: So you’ve actually proposed that we come up with a new ultimate test since it hasn’t aged well. As you say this trickery or deception is not true intelligence, that we should dump it in favor of real systems, maybe an AI triathlon. What do you think are the good proposals that have come out of that?

Read the proposals for a new Turing test in Scientific American 03/01/17

GM: There are a bunch of partial answers and I think that’s the way it has to be. I don’t think there is one, it’s not singular it’s going to be plural. Hence the triathlon idea. Intelligence is a multi-dimensional variable. IQ has confused everybody about this because you get a number. But even if you dig into the SAT you realize there’s more than one number, there’s your verbal or your mathematical. Really there’s many different traits that go into intelligent behavior.

The Turing test makes it as if it were true that if you did this one thing — basically fool a person for a few minutes into thinking that you are a person— that that would solve the problem. We now know, you can cheat very easily. It turns out in that case, that the Turing test is just a measure of gullibility, if it’s a pure measure of anything. And we want to look at questions like true natural understanding. Think of it this way; if I hired an 18-year-old as a summer intern, what could I reasonably expect them to do for me. Knowing they have no training on what I do, but they’re just intelligent, they’ve lived in the world and now I give them a week to learn how to do something new. What could they do? You’d like your machines to at least be that smart. Think about the Star Trek computer. You can ask it anything. Of course it’s fiction but why can’t we build that now? We’ve had AI for 60 years. The explosion in microprocessor power, in RAM power, it’s literally exponential —

AMLG: It went out of favor a couple of times, with the AI winters — you’ve observed a few of them, when it was just not cool to say you are working in AI.

GM: Sure. There were a couple of winters. There all kinds of little excuses you could make for the field. But the fact is the field is 60 years along. And I would say that towards the Star Trek computer, the biggest progress we have is like Google Now or Siri and it’s falling disappointingly short relative to that. We know there’s no principled limitation. The human brain is proof of concept that given three or four years of data, you can learn most of what you need about language, or given 18 years of education, you can learn almost anything —

AMLG: There’s also the crazy gap in that we keep seeing all these movie depictions, we’ve been seeing amazing AI for 50 60 years since Hal, which possibly was not so favorable, to Tars one of my favorite ones — in Interstellar which I loved — where you could adjust his humor setting. As you say there’s a gap and we’re pretty far off from that. I want your take on these different depictions of A.I. — Chappie, you’ve written about that. The AI in Her, Ex Machina, which really freaked me out, and there’s another reference to the Turing test. What are your favorite depictions of AI in popular culture?

GM: Right now my favorites are probably “Her” and maybe the series Black Mirror. I guess because they both illuminate a dark side that is very plausible. Not tomorrow, for “Her.” Black Mirror, some of the stuff like is happening now.

AMLG: Which Black Mirror episode are you thinking of?

In “Nosedive” (Black Mirror S3:E1) everyone walks around with a score, and people tailor their lives and interactions to improve their social ratings

GM: There’s a lot of them. There’s there’s one in which you know you evaluate people for everything.

AMLG: Right, they get points for everything. The crazy social media tracking and rating.

GM: Yeah that could happen next week. In fact in China there is a rating on like how good a citizen you are. It’s kind of straight up. It’s not clear whether Black Mirror stole from what’s happened in China or the other way around. But you know, that is happening.

AMLG: Just watch China and put it on TV.

GM: It’s there already. And then there’s another episode — I’m not going say anything more because I don’t want to give anything away — but it reminds me of Second Life. And Second Life is here now. I mean people have forgotten about it, but there are a lot of people that still absorb themselves in that world. In “Her” you have an AI that we can’t build now but you should be able to build someday, an AI that can basically talk about anything, has a good model of human psychology, a good model of the world, is able to interact —

AMLG: What about this idea of attachment? Like in the movie “Her” where he gets attached. I also really liked “Robot and Frank” the one with the old guy.

In the 30 months after it launched, Microsoft’s XiaoIce talked with over 40 million people

GM: I’ve not seen that one. But attachment is already here. There’s this thing in China called Xiaoice (小冰 “Little Ice”) and there are millions of people literally that are as attached to Xiaoice more or less as that guy in “Her” was attached to the Scarlett Johansson the operating system.

That part the Turing test actually shows us something about. We can get attached, we can get fooled. That’s not only a solvable problem but a partly solved one already. People know how to build machines that prey upon the weaknesses of human psychology and allow them to attach. The movie I thought captured well what that would look like you know, had a kind of dark picture and then it’s a signpost for us, hey we don’t want to go there we don’t want our lives to be you know quite that way. I like science fiction that makes us face where we might realistically go, maybe not where we’ll definitely go but where we could plausibly go.

AMLG: Do you think Chappie was realistic?

Read Gary Marcus’ take on the film “Chappie” in the New Yorker, 03/12/15

GM: So Chappie was not particularly realistic but it raised a really interesting issue — basically about where you get your ethics from. People don’t usually think about that in the context of AI, and Chappie wasn’t a great movie, but it really focused on that. This character, this robot starts by knowing nothing about ethics, its creator imbues it with some basic ethics and then it goes out in the world and learns things from people, some of whom are pretty unsavory.

AMLG: Kind of like the movie version of Microsoft’s Tay’s chat bot which just started copying everyone on the internet and swearing.

GM: That’s right. It was ahead of Tay. If the Tay founders had watched Chappie more carefully maybe they would have realized that future Trump voters could influence the way the system developed. That’s kind of what happened. I took Chappie to be a real question to us about, as the systems get more sophisticated, how are we going to build them to have ethics that we’re comfortable with. There’s the same issue in childrearing right. You can inculcate some values in your kids and then they go to school.

AMLG: But it’s not Asimov’s three rules for your kids, they’re actually just going to copy what you do.

GM: Well we do tell those rules to our kids, like “don’t do any harm” and sort of they understand it and sort of they don’t. The problem with Asimov’s rules right now is we have no idea how to program harm. How do you tell a computer what counts as harm? We can’t tell a computer anything other than geometric figures. We can teach a computer to recognize a square, but once it gets a more sophisticated than that, like the cat detector only recognizes some cats it recognize all cats, it’s pretty limited. Harm — there’s no picture for that to say, memorize 10,000 pictures of harm here.

AMLG: Were you always fascinated by this stuff? The human brain, it’s a lifelong thing you’ve been interested in?

GM: Well there was a brief astronomy phase and a dinosaur phase. But since I was eight actually I’ve been pretty interested in it. I first learned what a computer was when I was eight I was given a paper computer and I got it right away —

AMLG: What is a paper computer?

GM: So it’s like a simulation of a computer basically where you would write what’s in the memory addresses of the computer and then manipulate and perform operations on it. I really immediately got it. This is actually documented, though I don’t think anybody can find the footage, but there was a local television station that came to this after-school program for gifted kids where I learned about this and they had me explain on TV, “this is what a paper computer does” and after that I was hooked.

AMLG: I have to say you seem pretty down to earth about all this given the hype around killer robots and whatnot — does anything about technology scare you? Does anything about the future keep you up at night?

GM: Gene drive does. I mean this is where you use CRISPR to create things that don’t follow Mendel’s laws. So the good side is you use this to wipe out malaria. And the bad side is maybe somebody else does something pretty nasty —

AMLG: Technology is just tools, so it can go in any direction. If you could live forever would you?

GM: Don’t know. I would certainly choose life extension. I think once you have kids — and I had kids relatively late in life — you want to see what happens to them. So I would give a lot to live longer.

AMLG: You definitely need to see what happens with your kids as they continue to zoom ahead of Rodney Brooks’ robots. It’s a hilarious bet. But it seems like he’s been doing a decent job.

GM: He’s been doing a good job and my kids have been doing a good job. Let’s see who wins in the long run.

AMLG: Let’s see who wins. Well thanks a lot for coming on Gary, really fun, and I’m excited to see what you do next.

GM: Thanks a lot.