The Problem with AI

Despite the stunning advances in AI over the past decade, the so-called "deep learning" AI technology prevalent today has under-appreciated limitations and even poses societal dangers. Our speakers are world-renowned AI experts and AI "dissenters" who believe we need an AI that's both more accountable and better able to produce common sense results.

[MUSIC PLAYING]

BRIAN: The Problem With AI. Moderator, Ned Desmond. Speakers, Dave Ferrucci and Gary Marcus.

NED DESMOND: Thank you for that introduction, Brian. Let me warn everyone at the outset that this is the one session at Sight Tech Global in which we are not talking about accessibility, at least, not directly. Instead, our topic is the debate within artificial intelligence circles about deep learning. The overwhelmingly dominant model for AI today. In the view of many AI experts, this deep learning approach with its reliance on massive amounts of data and immense computing power is the unquestioned future of AI. Our guests don’t agree. They have serious doubts about deep learning’s potential to really advance AI or even overcome its own well-known limitations. You might say our guests are dissidents in a world dominated by deep learning AI. And they are dangerously well-credentialed dissidents. So let me introduce them. Gary Marcus is a leading voice in artificial intelligence. He’s a scientist, bestselling author, and serial entrepreneur, founder of Robust AI and Geometric AI. He is well-known for his challenges to contemporary AI anticipating many of the current limitations decades in advance and for his research in human language development and cognitive neuroscience. Gary is emeritus professor of Psychology and Neuroscience at NYU and author of five books, including his most recent, Rebooting AI, one of Forbes seven must-read books in AI. Dave Ferrucci is an award-winning artificial intelligence researcher who started and led the IBM Watson team from its inception through its landmark jeopardy success in 2011. Dave has worked for more than 25 years in AI. And his passion is to design computers that fluently think, learn, and communicate. That inspired him to found Elemental Cognition in 2015. Elemental Cognition’s mission is to accelerate and improve the quality and explicability of AI decision-making in the face of growing complexity, shortages of accessible expertise, and overwhelming data. Dave graduated from Rensselaer Polytechnic Institute with a PhD in Computer Science and has over a hundred patents and publications. Dave and Gary know each other well. And they’ll drive this conversation for the most part. But let me get them started with a question that goes to the heart of this AI debate. Dave and Gary, what is deep learning AI good for, and what are its limitations?

GARY MARCUS: I think you wanted Dave to go first since you said it in that order.

DAVE FERRUCCI: Yeah, I’m happy to do that. Deep learning is, from my perspective, evidence of incredible invention, really. And it’s fabulous at detecting patterns and large quantities of data. It’s fabulous at doing pattern recognition based on how the data occur. It’s changing the face of AI. It has changed the face of AI because we can use it to make predictions to extrapolate from data in a way it to learn the patterns without a lot of human program. And in the past, this was extremely difficult to do. So it’s remarkable in that regard. What it’s lacking is the ability to understand the reasons why that data occurs the way it does. It doesn’t build an understanding or the causal models that we need to make more thoughtful rational decisions. Some complain about the biases. But deep learning is designed exactly to find the biases in the data and to mimic them. That’s what it’s designed to do. And by design, it’s very good at that. And where that allows us to make predictions we couldn’t otherwise make or to extrapolate the data or to find functions in the data that we otherwise would be impossible or take forever to find is phenomenal. Where we want partners to help us make rational, thoughtful decisions that we can explain, it doesn’t do– it’s not even designed to do that, really.

GARY MARCUS: I completely agree with Dave. This is not going to be a debate where we go back and forth and disagree about a lot, although, we may find some things to disagree about. The way I would put it is that these systems are skilled mimics. They’re, for example, really good at making fluent language. But they’re not in touch with reality, and they can’t reason about reality. And that ties to this notion of causation versus correlation that Dave was hinting at. So we get problems, like, when they generate fluent text, there’s often misinformation there. So one example that I recall from the sequel to GPT-3 called InstructGPT is somebody asked it about eating socks after meditating. And the system fluently said some experts believe that it’s good to eat socks after meditating because it brings you out of the meditative trance or something like that. Well, it sounds fluent. It sounds like something a human would say. But it’s not something that is found in an actual source. It’s all unsourced. It’s just putting together words in sequences like AutoPredict might do in your text completion, predictive text. And that–

DAVE FERRUCCI: Sorry, go ahead, Gary.

GARY MARCUS: So as Dave is saying, it doesn’t really understand what’s going on. Here’s an easy thing to remember. The phrase deep learning invokes a notion of theoretical depth, of conceptual depth. But it really just means how many layers you have in a neural network. And the level of comprehension that we want such that a system that is talking about socks and meditation actually understands the potential relation or nonrelation between those is far outside the bounds of the deep learning systems by themselves. I don’t think either Dave or I want to get rid of deep learning. But I think we both think there are serious holes in what it can do. And I think we both think that a lot of people out there see it as a universal solvent, that it’s not. It’s a component in a larger system that I think is fine. But the larger system needs to be able to reason to have common sense kinds of things that I’m thinking about that Dave is thinking about at his company.

DAVE FERRUCCI: Yeah, I think– so a couple of good points you make, Gary. So first of all, I like that point about deep learning. I think the notion that deep, when we hear that, it evokes deep understanding. But in fact, as you said, it comes from how many layers in the neural network and the power of that neural network to find functions in the database and the relationships in the data. And so I think that that’s definitely could contribute to the expectations around it that is deeply comprehending, which is a very different thing than finding those statistical patterns. I like the other point too is like, one of the metaphors I use, this notion of a super parrot. We think about parrots and how they can listen to us. And they can start guessing or actually saying the next thing that we might say. But do we imagine they really understand the depth of the mental models that we have, the meaning we associate with those words in those experiences just because they can make that prediction. And the other thing is, is what I find really useful when I think about AI, I like to think about human cognition and how humans think because I think humans think in different ways. And the way we think creates this model for us to think about how AI works. So one thing we do is we make statistical predictions. We just extrapolate from the data. In fact, even in language generation, yeah, I’m sure we’ve all have experienced this moment even in ourselves or in others where we’re listening to a conversation or engaged in a debate of sorts. And we formulate a sentence. And then someone challenges us on, wait a second. That doesn’t make any sense. And we said, Oh, you know what. You’re right. I didn’t really think about it, right? But my brain was able to formulate the words, make a grammatically correct structure that sounded good. But when I really think about it, I wasn’t making the connections. It really didn’t reflect an understanding. Our brains work both ways. Our brains work as with statistical language models. But they also work as rational thought. Machines that can think about the meaning and the logical implications, the mental models.

GARY MARCUS: You can think about Danny Kahneman’s famous distinction between system one and system two, which I like to call the reflexive system. It’s automatic. It’s fast. It’s probably too crude a distinction, but we start with that. So you have this reflexive system, and then you have the deliberative system, which is what Dave is saying, that makes rational decisions. I don’t like to call it the rational system because it’s not always right. But it’s at least trying. It’s like the Supreme Court. It’s deliberating. Maybe it doesn’t come to the right conclusion, but it is trying to make those logical connections between–

DAVE FERRUCCI: Well, transparent. It can be. It should be transparent in how it’s making the judgments its making. It can explain them in some rational and deductive process.

GARY MARCUS: So take my socks and meditation example, right? The system can’t actually tell you why it has said– some experts believe that it’s OK to eat socks after meditation because blabbity blah. It can spit out a string of words, but you couldn’t pursue it. There’s no internal thing in the system that you could inspect the way with a classical AI system or just a classical computer programming system. You could say, well, the reason it’s made this mistake is because this value in this database is set to the wrong thing. It’s this big, abstract collection of predicting numbers about sequences in text. It’s not interpretable. It’s not that there’s something in there that represents a database of like, this expert believes this and this expert believes that. It can’t say, well, I found this on some web page because it didn’t find it on a web page. It just made it up. And it’s not grounded in reality. The deliberative system is at least grounded in our model of reality. I was thinking about the parrots a second ago in thinking is that fair to the parrots? Because in a certain way, they are like parrots because parrots when they use language don’t really much understand the language, although, there’s some interesting work on that. But when the parrot thinks about the world and like, flies through a room, it has a representation of the entities in the world that it’s interacting with. So it is grounded. You could fool the parrot. You could put it in a VR system. But the parrot has some representation of reality that it’s working over that it’s analyzing. And system two reasoning should be about that. But deep learning alone doesn’t really do that.

DAVE FERRUCCI: But I think one of the things you bring up does, particularly with the parrot example, is when we think about how we assign meaning and what that evokes ultimately is relativistic, meaning that you and I have to assign similar meetings. We have to have a shared understanding in order to communicate. In order for me to produce an argument that you can agree with or that you can evaluate or probe, we have to have a shared understanding around the meaning of the words and the phrases and so forth. And when we think about AI, we have to think about what do we expect from the machine? Do we expect the machine to project a shared– [AUDIO OUT]

GARY MARCUS: Nope, we may have lost Dave for a second.

DAVE FERRUCCI: –anything like the understanding that we’re reducing right now. In fact, quite deliberately is designed to look at the artifact much more as a pattern of numbers or words and to extrapolate from that pattern to say, here’s what this pattern alone would suggest as opposed to build– as opposed to build that shared understanding.

GARY MARCUS: I actually think that deep learning is dangerous in a certain sense. And the reason I think is dangerous hinges on the kinds of things Dave is just talking about. So we tend to attribute a shared understanding to these systems that don’t actually have it. So when he talks about the socks and meditation and stuff, you assume that any system, any person that talks about socks and meditating knows what it’s talking about. But it doesn’t. It’s actually just saying the statistics of this word following this word is this other thing. It’s very difficult for us as human beings to recognize how superficial the understanding of those systems are.

DAVE FERRUCCI: This is a human cognition challenge, right?

GARY MARCUS: It is. It is.

DAVE FERRUCCI: [INAUDIBLE] Shutter that way as well.

GARY MARCUS: That’s right. I mean, it goes back to Eliza in 1965. You had a psychiatrist. All it was doing was keyword matching, but some people fell for it. So we’re not built to discriminate these things. If we’re going to make AI available that has conversations with people and that has control over, I don’t know, air traffic grids or home robots or whatever, it’s really important that those systems understand the world as we do. So if you don’t want your home robot to sometimes mistake what you say is metaphorical is being literal, for example, because you want that system to have an understanding of the world.

DAVE FERRUCCI: I think we have to be careful that a trash deep learning. I mean, it can be used in dangerous ways, I agree. But when we think holistically about intelligence, it plays a very, very significant role. And it’s like, you don’t want to throw the baby out with the bathwater, plays a very, very significant role, which is to look at the data and to generate hypotheses that are consistent with how the data occurs because those hypotheses now can feed systems that can model in reason in ways that do build and advance a shared understanding. So this goes back to your point about Daniel Kahneman and the system one and system two. When you put these things together, you have more of the holistic view of AI in my opinion.

GARY MARCUS: Yeah, I mean, I don’t think we should throw the baby out with the bathwater either. I’ll clarify that I think the real problem is expecting too much from the stuff that’s more like system one and expecting it to do the system two. So I think we agree that you want to use this stuff. They’re great techniques for doing certain kinds of pattern recognition. You shouldn’t throw them away, but you also have to be realistic about what they can do and what they can’t do. And that requires training and understanding some of the psychological phenomena that you’re really trying to capture with your machines. And so no, we shouldn’t be throwing away deep learning. But we have to recognize we need more. Part of what Dave is doing in his company is to try to look at the other side of that divide, try to say, well, how do we represent in a machine commonsense knowledge? How do we get a machine to do argumentation? So for example, lots of things in the world, they might be true. They might be false. We often don’t know. We want to be able to evaluate different arguments. I think part of what Dave is working on, how do you get machines to do that?

DAVE FERRUCCI: Yeah, an element to cognition of view is we want machines to always be able to explain why to help us solve problems but not to make black box predictions. It’s OK to say, I’m thinking this because here’s my understanding of the world. And by the way, part of that understanding includes a statistical prediction from the data. But as long as we can model that holistic argument, we equip humans with the ability to say, here’s my understanding. Here’s why I’m making this decision versus that decision as opposed to extrapolate and say this is just a correlation. I don’t really know what’s going on. I mean, that’s again, encompassed in a more holistic intelligence. So my big thing is I keep learning big important part of the answer, amazing advance, critical advance. How do we include it in a holistic architecture that really addresses the partner intelligence we want when we think about AI?

GARY MARCUS: I’m going to give you a little insight into Dave because I’ve been talking to him a bunch lately and thinking about some of his earlier work, too– [INTERPOSING VOICES]

GARY MARCUS: And something just clicked understanding Dave. So Dave right now is telling you about the question of why. He’s saying a good system needs to be able to answer why. And one of the innovations around jeopardy was basically giving an answer to how sure are you? So you had a panel you could see– and he can tell you more obviously– but in jeopardy with Watson would say how sure it was of its answers. And so that’s another piece of like, metacognition that’s really important in doing good cognition. We want to know how sure you are of yourself. You want to know why you got that answer. You don’t want to just spit out an answer and maybe it’s right, maybe it’s wrong. You want to be able to do something.

DAVE FERRUCCI: Interesting, like, interesting. Even the Watson system would be able to– well, it couldn’t give you a causal model where it describe the places in the events and how they’re related and why something happened or something else happened. It did allow you to interrogate why it preferred one answer over the other in terms of concepts that humans could relate to. So it can say things, like, I’ve analyzed this passage from a temporal perspective. It was aligned with– this answer was better than that answer because it was more aligned with the question from a geospatial perspective, from a classification perspective. So it wasn’t a completely opaque model. It actually broke the problem down into independent features. Each of those features were trained with machine learning but to classify them in ways that at least you can get what we called an evidence profile of why you thought this answer was better than that answer.

GARY MARCUS: The ultimate thing that Dave and I would like to see is what he just described as causal models. So like, let’s say I’m a domestic robot in your house, like, Elon Musk’s optimist is supposed to be, looks like C-3PO or whatever. And now, you say something like, can you clean up the living room? And now it has to infer from some unstated things, what exactly you mean by that. What do you mean by clean up the living room? So for example, you want literally everything out of the living room? Do you want to cut up the couch and put it in the closet? Probably not because the couch is of value to you. And so the machine has to understand what you’re thinking is important. What’s in there physically. It’s not enough to just have like some statistical [INAUDIBLE].

DAVE FERRUCCI: I’m going to put my deep learning hat on. And I’m going to argue and say, with enough data, it can associate cleaning the room with an expected result.

GARY MARCUS: Well, I mean, that’s the interesting question about outliers, right? I mean, this is what’s going on with driving is the thesis was with enough data, we’ll be able to make all our unprotected left turns and so forth. And Kate Metts wrote in The Times just today about a ride in a Tesla. There’s clearly still outlier problems where even with all the data that Tesla has collected, the machine is still stymied when things are a little bit different than before.

DAVE FERRUCCI: And I think you have to look at each application because there’s going to be asymmetrical risks with regard to outliers and within some cone of expectations. But there’s always this answer however unsatisfying it might be. Well, look, the last 300 million times I clean the room, you were happy with this result. So right. I mean, it’s really based on, again, extrapolating from the data. And then of course, as I said, there’s this always are these outliers. And then how risky are those outliers? If it’s a life and death situation, those outliers matter more than if it’s [INAUDIBLE].

GARY MARCUS: This is why the biggest financial successes of AI have all been around advertising because if you pick the wrong advertisement, there’s essentially no cost to it. Nobody ever dies because you showed the wrong commercial, whereas, driverless cars, it’s still all invaded after 25 years because the cost of error is much higher. And so Dave is absolutely right. These are empirical questions. In speech recognition, we’re pretty satisfied with 99% correct. It’s not 100% correct. Sometimes it’s annoying when you dictate something in your car and it’s noisy. But yeah, most of the time it helps you. The cost of error isn’t too high. And mostly, you don’t have to deal with so much with the outlier conditions. So every domain, like when you’re thinking about this commercially, you have to ask, what’s the cost of error? How variable are the data going to be today in the next– my own feeling is that domestic robots are a serious problem because people have very different houses from one another, very different things about what is important to them and so forth. And my guess is you are going to have those outlier cases in spades. It turns out, the driverless cars has hinged more on outliers than I think people expected. They thought, well, the roads are all basically the same. It’s going to be fine, and it wasn’t. So some domains are going to be OK, and some aren’t. And sometimes you’re going to be able to take a domain and bend it to your will. You could build a new city from scratch around driverless cars changing how the pedestrians interacted with the cars and probably make them work. If you want to make it work, [INAUDIBLE] hard.

DAVE FERRUCCI: Here’s a funny reality with regard to deep learning is that I’ve actually experienced people complaining about deep learning systems giving them the wrong answers. And then the data scientist says, well, but this is what the data reflects. And the answer is well, I don’t really think those answers are right. They’re not upholding how we view the problem, how we reason about it, what we think the right judgments and values are. So go and change the data so that the machine learning does the right thing.

GARY MARCUS: Well, here’s a problem related to that that’s really hard to solve has not yet been solved. We now have these generative models, like Dolly that you mentioned, and stable diffusion, which are fabulous. They’re also incredibly biased. So if you type in CEO, you’re guaranteed to get a white male, for example. And there’s a brand new paper that I just tweeted about. I’m trying to think who the authors are. It goes in excruciating detail showing that for many, many examples you wouldn’t even immediately think about, you get all of this kind of stereotyping. And there, the system is a direct reflection of the data that are out there. And that’s a reflection of what we want the world to be like.

DAVE FERRUCCI: Exactly.

GARY MARCUS: Not actually even a reflection. Here’s the interesting part– of the actual world but of the data they’re in the world. So take CEOs as being white males. That’s partly because CEOs are overrepresented in movies that are then in the database. So if you actually took representation, you would see like, I think her name’s Safra Catz of Oracle should be in your top 50 pictures of CEO. But she won’t show up there because she’s not in the news as much as Elon Musk. And so you get more representation of Elon Musk because people talk about him. But sometimes the systems don’t even quite reality.

DAVE FERRUCCI: So there’s the reality. Then there is the data. Then there’s a reality that the data might be biased and not reflect what you want the system to actually do. And so then when you look at applications of AI or of deep learning to sit back and say, what do I really expect the system to do? Is the data– the data that I’m training the system with representative of the phenomenon– what the behavior that I desire? And, of course, it’s silly when I could write down the behavior I desire. Why would I go muck with the data if I could just write it down, a program?

GARY MARCUS: Well, and that’s the real problem is we don’t yet have– well, this is the kind of thing I think Dave is trying to work on. Systems where we can fluidly even represent what our desires are and mix that with the data. So you really want an answer. There’s a mix of data and let’s say human value. So human values would be like don’t discriminate.

DAVE FERRUCCI: Human values, human judgment, decision processes. I should make it very easy to capture those to evolve those but then to inform them what’s actually going on, the patterns that were in the actual data and the empirical data. And those are the two sides of intelligence.

GARY MARCUS: That is the Holy Grail going forward is to be able to connect the human values that you might express explicitly. They could also be laws, like just the driver’s handbook for California. Things that are represented in language and then relate that to the data. Right now, we have– we’re really good at the data. And we’re pretty good at doing symbolic reasoning on stuff that’s innately specified in some particular architecture and not really good at bringing those together it takes from our domains.

DAVE FERRUCCI: Exactly. Right. That’s the vision. Right.

NED DESMOND: And how do we make it better, gentlemen?

DAVE FERRUCCI: Do we answer the first– do we answer the first question, then?

GARY MARCUS: I was going to ask the same.

NED DESMOND: You did an excellent job on the first question. So can deep learning become better? Or is it going to take fundamentally new AI architectures?

DAVE FERRUCCI: Yeah, to add, that’s the million-dollar question. That’s the interesting philosophical question, I should say. I don’t know if it’s the million-dollar question.

GARY MARCUS: I think it’s the trillion-dollar question.

NED DESMOND: [LAUGHS]

GARY MARCUS: You found something to disagree about.

DAVE FERRUCCI: The trillion-dollar question, maybe. All right, the reason why I say it’s like, it’s such an interesting philosophical question is because– and I’m quite frankly agnostic to that right now. But like, I believe that the– that when you think about the behavior that you want– and I’m a big proponent about stepping back and saying, forget about how you do it. What behavior do you really want out of the systems? And I think when you look at that, I think I don’t know the answer to whether or not, deep learn– when you look at that holistic behavior, like the thing we’ve been describing. Something that could analyze data, generate hypotheses, understand the possibilities, being able to meta-reason about the data to what extent does it cover the phenomenon to then fit that to a model, a shared understanding conceptual model that can now reason over so it can give answers and explain those answers or give predictions or can make arguments. It can evidence those arguments. It can understand whether or not some of the assumptions it makes are based on statistical patterns or whether it’s based on deductive arguments, right? This is the rational, intelli– again, intelligence, I think that we can describe. And I don’t think it’s that hard to describe. It’s all over science fiction, right? I mean, it’s– we’ve imagined this for decades and decades. It’s above and beyond in many ways. It corrects for problems with human cognition and some of our cognitive biases, like that’s this vision. So now, we can say can deep learning get there? Maybe. I think– I don’t know, right? I don’t anybody knows yet. I mean, I’d love to hear arguments about it. My sense is I don’t know. I’m [INAUDIBLE] I’m like, maybe. But there’s one thing that’s probably true if deep learning can get there on its own. And that is you have to give it the data to train on. There are a couple of problems. One is that we don’t articulate the data in vast quantities to train a machine about how to really do all of that, how to reason, how to think, how to do this meta-reasoning and this deductive system– deductive reasoning. Where’s the data that trains machines on how to do that? Do we articulate that? Is that in the form we can train machines on? So it’s question number one. It’s really not– there’s some work on that, but that’s the big nut. And then the other nut is when we think and build shared understanding, we react to features that are generated by our what where that are generated by our brain being human. That’s not in the mix right now. And there’s a lot. There’s a lot there. So those are two problems. So on one hand, I’m like, maybe. Those are the two problems I see with machine learn– deep learning alone getting there.

GARY MARCUS: So first of all, technically, I’m agnostic too. I’m agnostic about almost everything in the sense that I’m a scientist. I look at the data, and I try to update based on the data. My view is it’s unlike–

DAVE FERRUCCI: So you’re a Bayesian.

GARY MARCUS: So I’m a Bayesian, exactly. And I don’t see a lot of– I’ve been looking at these particular issues for about 30 years writing about how it is that deep learning systems don’t do these kinds of reasoning things and extrapolation beyond the data. We call it nowadays, distribution shift as one of a first people to write about that in 1998. I don’t see a lot of movement on those problems. All the movement I see is what we call within distribution generalization rather than outside distribution generalization. And I don’t see how you can do the reasoning without it. This doesn’t mean that I think it’s literally impossible to fathom. But it’s like, I keep seeing these kinds of approaches. And they don’t look to me like they’re really working. And they keep breaking down in the places where I would anticipate them. And I start to have some confidence in my own predictions over time.

DAVE FERRUCCI: Well, what do you say– what do you say, Gary, about the view that the human brain is essentially an inductive machine, and we figured out how to do system one in system two?

GARY MARCUS: I mean, I think it’s a neat that we have system one and system two. I don’t think that we have induced the variability to do logical reasoning. I have data from when I used to be an experimental developmental psychologist showing that babies could learn abstract rules at seven months old. And then what happens in developmental psychology is you do an experiment on someone else’s. I can get even younger kids to do it. So somebody replicated my work but did it in newborns. So they showed newborns could represent abstract rules. Another friend of mine, Luca Benati– actually, a couple of different friends of mine have done work showing some basic reasoning, like disjunctive syllogisms. If I know it’s a– or if I know it’s either A or B. And I know it’s not A, then I can–

DAVE FERRUCCI: But when you say innate, you’re basically now hardwiring to the brain. And then you have– go to, well, how did that occur? And that occurred through an evolutionary process is the evolution–

GARY MARCUS: And my–

DAVE FERRUCCI: The evolutionary process probably also is of if you were to give it a shape, it would probably still be inductive.

GARY MARCUS: Well, I mean, I guess there’s a couple of things. I wrote a whole book about this called The Birth of the Mind in 2004, which was about if things are innate, what are the processes of developmental biology by which they might be there? There is some abstract level at which you can call evolution learning. But the causal mechanisms are quite different from the other things that we associate with learning. So there’s parameter setting in both cases. But in one case, it’s building a genotype that creates a phenotype that does something in advance of experience. And the other, it’s what do you do with the experience?

DAVE FERRUCCI: And that succeeds, right?

GARY MARCUS: And so my view is that the minimum that you need to have in an adequate cognitive system like ours or in an AI is you have to have, let’s say, prewired. I don’t like the word hardwired. But prewired capacities for things, like representing time and space and object and other agents. And if you have those basic things in place– and this is a view also associated with the Philosopher Immanuel Kant and with the Developmental Psychologist Liz Spelke. If you have those basic things in place– also with Noam Chomsky– then you can go learn some other stuff. So I take a pretty strong position that you need some innate nucleus that you then sort things around. And I think pure deep learning doesn’t have that. So there’s a conversation about what deep learning even means and how the term changes over time that we could have. So that was part one. Part two is in terms of the data. I think there’s some data we’re unlikely to get. And it’s further reason to want to have some innate stuff. So one example is in these systems that just are driven by the data from like, conversations and Reddit and so forth, you get weird effects. Like, murder is more common than breathing even though in reality, breathing is more common than murder. But people like to write about murder. And breathing is boring.

DAVE FERRUCCI: Well, it’s– yeah.

GARY MARCUS: And you have that problem. I’ll just say, say what one more second. And then people’s internal mental states, we don’t necessarily write that much about. There’s some fiction where people are really like, third person omniscient and describe it. But in general, there’s not as much data around that. So the paradigms we have work very well for things you can explicitly label a bunch of data. Maybe do some unsupervised learning on the rest. But like, this is a picture of a cat, and you’ve got a label for it. But things like harm. So having the robot not do harm or understanding like, the difference between like, belief and desire and all this kind of stuff is pretty subtle to get from the data-driven paradigm– not saying it’s impossible, but those are some places that like, out of the gate, pose some challenges.

NED DESMOND: Yep, Gentlemen, we’re just– we’re out of time. But could I ask one quick lightning-round question? How do you think educated consumers who are using variations on computer vision and natural language, word processing, all sorts of things related to AI, how should they think about AI? Should they roll along with the developments that touch their lives? Or should they maintain a skeptical stance to what’s going on around them? What would be your very concise word of advice to the general, thoughtful public who is dependent on AI already?

GARY MARCUS: I would say don’t treat AI as magic. Understand that it’s really an evolving set of techniques that have strengths and weaknesses. And you need to get your hands a little dirty to understand will this work with your particular business problems? So we gave examples of like, advertising versus driving versus domestic robots. It’s an empirical question in any given domain, is this going to work for you? It’s not magic. You can’t just assume it’s going to work. You shouldn’t assume it’s not going to work. There is a lot of gold in these hills. There are a lot of ways that you can make things more efficient and so forth. But you really need to dive in and learn how this stuff works if you’re going to make good decisions around it.

DAVE FERRUCCI: I think getting educated and understanding the specifics is probably good advice. But I do worry. I think, in general, I worry about also being aware and where AI is seeping into your world, particularly with regard to recommendation– not recommendation systems. It’s more like, in terms of how it’s managing the content that you are getting pushed. I think it’s just, that’s a big issue for us.

GARY MARCUS: Yeah, I think the worst thing that AI has done so far is to polarize society with content moderation, which is driven by AI. So there are pluses and minuses to AI so far. And that’s one of the biggest minuses.

DAVE FERRUCCI: Understanding the roles playing in content moderation is a very poignant thing.

NED DESMOND: Yep, Gentlemen, that is extremely good counsel. Thank you for your time. Incredibly lucid and helpful conversation, I think, to everyone at Sight Tech Global. Thank you, again.

GARY MARCUS: Thanks for having us. It’s fun for us.

[MUSIC PLAYING]