An AI Expert Explains The Hype

Actually true. There are no science fiction stories where they create robots and the robots are just artists. Like it's... I've actually... [laughter] you don't... and and [snorts] like if the AI made a mistake, that would be a plot point. Like [clears throat] somebody had broken its dilithium crystals and it was... [laughter] you know, like that would not be the normal way that that system worked.

My pleasure to be here.

Before we get there, let me just say I'm often miscast as a hater and I'm thrilled that you understand the nuance in my position. Like that's that's rare, right? Um, because I don't hate hate AI. I actually...

I can tell in a minute like who's actually heard the position with the nuance that it has.

I mean the problems in this field start with the fact that artificial intelligence is a marketing term. It's not clearly defined. It was actually invented in the 50s by John McCarthy who incidentally taught my father Fortran and then left in the middle of the semester but that's another story for another day. Um, uh, and...

It's it's a weird thing but it's a fact. I don't think I ever brought it up in an interview before, [laughter] but you set me at ease because you listen. Um, and so so I actually met McCarthy once myself. That's more direct name uh name dropping, but anyway. Um, so you know, he made this term, but it wasn't clear exactly what it referred to, but it it was a general notion, right? And the general notion was how do you get machines to do the kinds of things that people do? And McCarthy who was one of the founders, the real godfathers of AI, unlike people this term has applied to lately, um, raised for example the common sense reasoning problem. He McCarthy was into logic and classical AI, symbolic AI, people sometimes call it good old-fashioned AI, and he realized that getting machines to do the kinds of things that people do was in general hard because machines lacked common sense. It was actually not that hard to get them to reason about formal logic even in the 1950s.

And you know, they did some interesting math stuff in the 50s. Everybody's excited about what AI is doing now in math. And it's interesting, but people often forget that General Problem Solver was doing this stuff in the 1950s. Not quite as well, but I mean, you know, obviously we should be doing better now 70 years later. Um, but the general notion of AI was, you know, doing stuff that people do and then that's a broad definition. So it could include like doing arithmetic, right? Is a calculator AI? Depends on your definition. Um, there's a whole lot of things that have been jokes in the field that like something is called AI until it becomes sort of standardized and then they just call it engineering. So like is [laughter] GPS navigation AI? I mean it is, right? GPS, you know, when I do route or when my my Apple iPhone does route planning for me to get me from place A to place B, that's actually a form of intelligence if you have a broad enough definition of intelligence. And I'm comfortable with that. I don't like to get in fights about definitions for the most part. Um, so I would say [clears throat] GPS navigation system is a form of intelligence.

A lot of the debate right now, excuse me, [clears throat] is really about not whether this thing or that thing is AI. It's really about this notion of AGI. That's artificial general intelligence. And it's really the generality that I think we're kind of arguing about in a way. So, we've had narrow artificial intelligence since the 50s. You know, there were geometry theorem proving systems in the 1950s. Nobody thought that those were going to take anybody's job. Well, actually, I should walk that back. Some people even in the 50s were worried about like when jobs would disappear, but not many people were worried in the 1950s. New Yorker had an article in 1957 about an early neural network and they thought it was going to kind of take over the world.

Yeah. So, you know, people have been and I mean the whole notion of robots was they were going to take over the world, right?

There is for for robots in particular and then I'll come back to the AI part. For robots in particular for me the line is actually pretty clear: is it autonomous or not? So I mean this again you can have whatever definitions you like but for for me the definition of a robot is that it should be autonomous. So if you have a robot arm that, you know, keeps doing this task, but a human is overseeing and it's just doing that one task, I don't really want to call that a robot. I want to call a robot something that on its own can act in the world, make decisions and so forth. We can come back to that. Um, and it relates to the other thing which is so we've had narrow AI for a long time. Deep Blue, you know, beat Kasparov in 1997. Watson, you know, beat Ken Jennings in Jeopardy in what was it, 2009, something like that. So, um, but each of these were narrow.

Watson turned out to be narrower than we realized because of a sort of sneaky thing. It's not like they cheated, but they discovered along the way that, um, 99% of the answers of Jeopardy uh questions are actually titles of Wikipedia pages. So it looked like it was this massive general intelligence, but really it was basically doing information retrieval across the Wikipedia pages. It was still a beautiful piece of research. It was an amazing piece of work. In my own podcast a few years ago, I interviewed um David Ferrucci who directed it. It was it was really an impressive thing, but I would still call it narrow AI. I mean really it was it was a good information retrieval system tuned to a particular way of writing clues and things like that, literally tuned to that. And what IBM learned is it was not a general intelligence. Probably some executive eye in the company was like, 'Wow, we won in Jeopardy. We can do anything with this.' And I think Ferrucci was actually pushing back saying, you know, hold your horses. But IBM like sold it as a general intelligence that was going to next solve cancer, which we all know if we remember the story, failed to do that.

So the question before us is how do you get to an AI system that is general enough that you can throw whatever problem you want at it and expect a reasonable answer. Something that nobody thought about in those days was that it had to be right. You know we all people in my generation grew up with calculators. You type in a question and you get a correct answer. One of the things I did not foresee was that we would get the kind of intelligence that was general. You can ask an LLM anything, but it's not very reliable. And so, right...

It's actually true. There are no science fiction stories where they create robots and the robots are just artists. Like it's I've actually... [laughter]

Exactly. Or bad bad mystery novels, right? Like you know, in the future there will become a time when AI writes all the bad mystery novels. Wow. Like I mean like nobody ever even thought about that, but that's where we actually are. Right.

That one's I gave you a long answer to last one. This one is a short answer. [laughter] No.

I mean, well, we have something that's different from a calculator, but but you know, we obviously don't have a genie if machine can do cannot do a lot of Let me say that again. Obviously, we don't have a genie if the current AI systems actually aren't as good as the best human experts at a lot of things, right? I mean, I I assume that any genie can do anything that a person could do. Invent me a new theory of physics. If a person can do it, the genie ought to be able to do it, right? Um, and there are lots of things that current AI systems actually can't do, especially with reliability. I have a bet with Miles Brundage, which is basically about AGI, though we don't fully um explicitly say that, but we placed this bet the last day of December in 2024, if I recall correctly. Um, you can find it on my Substack. Look for Brundage Marcus bet um at GaryMarcus.com.

And I laid out 10 things. Most of this had come from my earlier writing around what AGI is. And they're just examples, but they're examples like you should be able to watch a movie and answer questions about what's going on. Like explain the plot points. Like why is this guy going out of the spaceship right now? You know, or you know, why did he just insult his wife? Like what was going on with that? Did he mean it? Was he joking? You know, a human in at least with a like ordinary movie can answer those kinds of questions, right? Movies when you watch them are kind of exercises in filling out filling in in your head the things that aren't said often times and the you know, right? And so you ought to be able to do that. I actually proposed that particular one in 2014 in a New Yorker article called I think it's After the Turing Test or Beyond the Turing Test. And I I use this there an example like a kid should be able to watch The Simpsons and knowing you're supposed to laugh like an AI system that could watch 24 frames a second audio and video and actually say yeah that's the funny part that would impress me right but we don't have that yet.

Another was write a novel. Um, the way we framed it there was was a Pulitzer-caliber novel. So we all know not always in any given year does the best novel win, but Pulitzer-caliber is like you ought to be nominated for it and you know some people would make a case that that you're a shoo-in, but somebody just David Korn just actually did a version of this. He had an AI write a novel and then he went through it and he's like it wasn't terrible partly because of the novelty. I read it, but it had this problem and that problem with you know characterization and and so forth. It was clearly not a Pulitzer-level novel. Um, you know you ought to be able to ask your genie write me a novel you know where Sam Altman is the lead and he gets his just desserts and it should be interesting and fun. Right.

Well, we may actually see [laughter] that. I I've been joking for a while actually about all the drama. When you got fired should have been a Netflix series and fired and rehired.

Right? No, no, but but but watching it and being able to [clears throat] pick out the plot points or...

Another example I had was was just being able to read an article and answer questions about it. Like some of it's artistic and some of it's not. I had a range of things because humans can do a range of things. I tried to capture that. One of them that I can't personally do. We had some things that anybody can do and some things that not anybody can do but experts can.

That is part of the source of the the controversy, right? So, so a lot of people have a very simple-minded idea of what intelligence is coming maybe from the the IQ test which they don't actually most people don't even actually understand but you know you get a score right you get a one-dimensional score you get you know 100 or 130 or 72 or whatever right um and so a lot of people think it's just this one thing but anybody who has ever studied it more or less has come to the conclusion that is multi-dimensional. You the first hint of that might be like essay SAT verbal and SAT math are correlated but they're not perfectly correlated and somebody might be strong in one and weak in the other and so forth. And then you have people like Howard Gardner and Robert Sternberg who had different takes on different kinds of intelligence. So there's kinesthetic intelligence which is like you know what you do with your hands and and understanding the physical world and then there's mathematical intelligence and so forth. So like to a cognitive psychologist which is part of what I am and what my training was, it's just obvious that many things go into intelligence and it's pretty hard to define um intelligence.

So you can have this argument. You can say well a chess computer because it's good at chess is intelligent. If you want have that definition that's fine. Another definition would be like the ability to do adaptive problem solving um and to encounter something you haven't seen before and figure out a good way to cope with that. And I think that's an important strand of intelligence. I think one of the weakest things with current AI techniques is they're not necessarily good at problems that they don't have an enormous amount of data with. You know, they tend to be very reliant on the specifics of what they've been trained on. [clears throat] I think they still, 30 years after I started making this point, still have trouble kind of thinking outside the box. They're better at it, but it's partly because the box has gotten bigger because they put in all the training data. There's still something missing in terms of their ability to cope with novel and unfamiliar things.

So, uh I'll answer that in a couple parts. The first is to say the phrase you're probably looking for was mine originally. I would take responsibility, which was autocomplete on steroids, which is true. Um, they are actually basically autocomplete on steroids. What they're built to do is to predict the next word in the sequence. It turns out that that gives them some degree of generality. It does not make them mimic humans. It's not necessarily the case that AI must mimic humans. It really ought to be reliable. I think that's a reasonable expectation that you ought to be able to um count on it to some, you know, degree. Um and it turns out that doing this process of predicting next tokens, as they would call it in the trade, actually lets you solve a bunch of problems. It doesn't let you solve all problems, but it it turns out to have some general utility.

That's right. And so the original ChatGPT was pretty much not 100% and they never fully attested to it a a pure large language model which did which was autocomplete on steroids. It wasn't very good. And what happened over time is that they've actually added in a lot of gadgets to try to make it more accurate. This is most true in things like math and coding. And part of what makes it hard for outsiders to understand is the industry doesn't want you to know what's inside the box. And the pure large language models really are autocomplete on steroids. And if you remember the original ChatGPT, it really wasn't very reliable. Everything it did, it did by kind of predicting, you know, what would come next here. And it did that not by having a deep conceptual understanding of what's going on, but just having a lot of text. They basically took everything that was on the internet and you get somewhere with that.

So like an example that you can get just from that is brainstorming. So you can say give me some ideas about how to write um a card to thank my my kid's teacher for helping in school. Well, it turns out there's like 10,000 examples. I'm making up the number. You know, there's thousands and thousands or millions of examples on the internet of how to do that, of how to write a thank you note to a teacher. And so, just by doing predicting next word in context, you can actually get a good answer to that. Now, here's an example of something you can't get from that is you can't actually get one of these systems to follow the rules of chess, which is sort of remarkable because they're trained on millions of games of chess that's in the database that they're trained on. And they're trained on the rules of chess. They can actually recite the rules of chess. You can say, you know, can a queen jump over a bishop? And they'll tell you no, but then they'll play and they will actually have a queen jump over a bishop.

This is what I call the problem of world models. They don't actually construct an understanding of how the world works. They just construct a text of how people talk about things. So when you play chess, you see things like pawn to K4, P to K4, pawn to king four, and then you see P to K4 because that's, you know, a common chess opening. And so you can have a library that you are drawing on. You by similarity you do this stuff, but then eventually you say like Qd6 and knight six, you can't actually do it legally on the board position, but it doesn't have a world model of what the board is right now. It just knows that's something that gets said in this context. And so if you push it hard enough, you can often see that its conceptual understanding of the world is actually pretty limited. And yet it can get kind of far with just a pure large language model.

I wrote an article in 2020 about an earlier generation of these things and I think what I said still held holds and held. Um, one thing I said is you need to be able to represent symbolic information, abstract information the way that you can in a computer program, in algebra, in logic, and so forth. And that we needed a a hybrid between neural networks and symbolic AI. And we actually have that now. Um, people started doing that in late 2023. Sorry, in late 2023 when they started adding what they called interpreters. Code interpreters ran classic computer code. They didn't want to talk about it in the AI industry because for various reasons we could get into they like this idea of scale is all you need that you could just build more compute and infrastructure and so forth and we should probably talk about that but things actually did change in the industry.

So that was one of the things I said in 2020 that we would need and I also said that again in 2020 two in my most famous paper which was called Deep Learning Is Hitting a Wall which everybody read the title and nobody read the paper and what the paper said was we can still use deep learning but we need to have this symbolic stuff too in order to make this work which has in fact transpired but what everybody remembers is the title and they think that I said deep learning can't do anything which was not what I meant. Right.

Um, okay. So, that was one. Another thing was having [clears throat] machine-interpretable common sense knowledge about how the world works. We've made some progress on that, but it's still problematic and you can still find around the edges that the systems don't really have that. They kind of give the appearance of it. Um, another thing was better ways of reasoning, abstract reasoning, than we still have. We've not made a lot of progress on that. We have something that looks a little like progress, but I think it's misleading. We could go into it. It's a little bit technical, but we could talk about it. And then the the um fourth thing that I talked about in that paper I called cognitive models, but most people nowadays call world models.

And that my first example to you was chess. Just knowing what's on the board and pieces. And here's another example. I'm watching there. There are four people on this Zoom call or three three people on this Zoom call. I think there was a fourth. Um, [clears throat] I tricked myself the way it's set up. So there there's three people on the call. So I have a model of the world that we're in. I can see, you know, people have microphones, they have bookshelves behind them or curtains behind them, etc. I have a bookshelf. Um, and so I build a model of who's doing what also of I like look at people's facial expressions. Do they understand what I'm saying right now? They don't. Are they laughing? Do they look lost? And so I'm I'm modeling what's going on and I'm trying to adapt my answers to that. And these AI systems don't really do that. They give the illusion of that. There are some complications again about the most recent systems and what they're doing, but we don't have a general way to do that.

So, here's an example of what humans can do that I think is amazing, and it'll come back to your question about mimicry. Um, my kids just watched the movie Project Hail Mary. There's a bunch of things that without trying to give it away, um, there's a bunch of things in Project Hail Mary that at least have never been attested in our world. And so it's a different world than our world as we know it. And my kids are 12 and 13 and they watched the movie. They understood, you know, perfectly well the nature of these interactions um between entities that don't usually interact in our world, etc., etc. So they built a model from the data of watching the movie of how the world works. That's what a world model is.

No AI system can really do that right now. can really build a coherent understanding of a new world which might be in a movie or in in a novel um or it might be a new political situation in the real world that they have to understand but our AI systems just aren't very good at that. Does that mean they have to do it exactly like the human brain? No. There might be a way to do it that's different from human brain. Human brain is the only system we know that can do this well. But I don't think it means that we need to slavishly imitate the human brain. There's lots of problems with the human brain. The yellow book behind me, you probably can't see it, but was called Clueless. It was all about all the problems of the human mind. Like there are lots of problems we have like confirmation bias. We notice evidence for our own theories and and we don't notice is inconsistent with our theories. So I I do want AI to succeed. I don't want it to mimic human beings, but we should learn from the things that people do better.