Yann Lecun: Meta AI, Open Source, Limits of LLMs, AGI & the Future of AI | Lex Fridman Podcast #416

I see the danger of this concentration of power through proprietary AI systems as a much bigger danger than everything else. What works against this is people who think that for reasons of security we should keep AI systems under lock and key because it's too dangerous to put it in the hands of everybody. That would lead to a very bad future in which all of our information diet is controlled by a small number of companies through proprietary systems.

I believe that people are fundamentally good and so if AI, especially open source AI, can make them smarter, it just empowers the goodness in humans.

So I share that feeling, okay? I think people have found me to be good and in fact a lot of doomers doomers because they don't think that people are fundamentally good.

The following is a conversation with Jan LeCun, his third time on this podcast. He is the Chief AI Scientist at Meta professor at NYU, Turing Award winner and one of the seminal figures in the history of artificial intelligence. He and Meta AI have been big proponents of open sourcing AI development and have been walking the walk by open sourcing many of their biggest models, including Llama 2 and eventually Llama 3. Also, Jan has been an outspoken critic of those people in the AI community who warn about the looming danger and existential threat of AGI. He believes the AGI will be created one day, but it will be good. It will not escape human control, nor will it dominate and kill all humans. At this moment of rapid AI development, this happens to be somewhat a controversial position and so it's been fun seeing Jan get into a lot of intense and fascinating discussions online, as we do in this very conversation. This is the Lex Ruben podcast to support it. Please check out our sponsors in the description. And now, dear friends, and here's Jan Lecun. You've had some strong statements, technical statements, about the future of artificial intelligence recently throughout your career actually. But recently as well, you've said that autoregressive LLMs are not the way we're going to make progress towards superhuman intelligence. These are the large language models, like GPT4, like llama 2 and 3, soon and so on. How do they work and why are they not going to take us all away?

For a number of reasons. The first is that there is a number of characteristics of intelligent behavior. For example, the capacity to understand the world, understand the physical world, the ability to remember and retrieve things, persistent memory, the ability to reason, and the ability to plan. Those are four essential characteristics of intelligent systems or entities. Humans, animals, LLMs can do none of those, or they can only do Them in a very primitive way. And they don't really understand the physical world, they don't really have persistent memory, they can't really reason, and they certainly can't plan. And so if you expect a system to become intelligent just without having the possibility of doing those things, you're making a mistake. That is not to say that autoregressive LLMs are not useful. They're certainly useful. That they're not interesting, that we can't build a whole ecosystem of applications around them. Of course we can. But as a path towards human level intelligence, they're missing essential components. And then there is another tidbit or fact that I think is very interesting. Those LLMs are trained on enormous amounts of text, basically the entirety of all publicly available text on the Internet, right? That's typically on the order of 10 to the 13 tokens. Each token is typically 2 bytes. So that's 210 to the 13 bytes as training data. It would take you or me 170,000 years to just read through this at eight hours a day. So it seems like an enormous amount of knowledge that those systems can accumulate. But then you realize it's really not that much data. If you talk to developmental psychologists and they tell you a 4 year old has been awake for 16,000 hours in his or her life, and the amount of information that has reached the visual cortex of that child in four years is about 10 to the 15 bytes. And you can compute this by estimating that the optical nerve carry about 20 megabytes per second, roughly. And so 10 to the 15 bytes for a 4 year old versus 2 times 10 to the 13 bytes for 170,000 years worth of reading. What that tells you is that through sensory input we see a lot more information than we do through language. And that despite our intuition, most of what we learn and most of our knowledge is through our observation and interaction with the real world, not through language. Everything that we learn in the first few years of life, and certainly everything that animals learn, has nothing to do with language.

So it'd be good to maybe push against some of the intuition behind what you're saying. So it is true, there's several orders of magnitude more data coming into the human mind much faster. And the human mind is able to learn very quickly from that, filter the data very quickly. You know, somebody might argue your comparison between sensory data versus language. That language is already very compressed. It already contains a lot more information than the bytes it takes to store them if you compare it to visual data. So there's A lot of wisdom and language. There's words and the way we stitch them together, it already contains a lot of information. So is it possible that language alone already has enough wisdom and knowledge in there to be able to, from that language, construct a world model and understanding of the world, an understanding of the physical world that you're saying all lambs lack?

So it's a big debate among philosophers and also cognitive scientists, like whether intelligence needs to be grounded in reality. I'm clearly in the camp that yes, intelligence cannot appear without some grounding in some reality. It doesn't need to be a physical reality, it could be simulated, but. But the environment is just much richer than what you can express in language. Language is a very approximate representation of our percepts and our mental models. Right? I mean, there's a lot of tasks that we accomplish where we manipulate a mental model of the situation at hand. And that has nothing to do with language. Everything that's physical, mechanical, whatever, when we build something, when we accomplish a task, modern task of grabbing something, et cetera, we plan our action sequences and we do this by essentially imagining the result of the outcome of a sequence of actions that we might imagine. And that requires mental models that don't have much to do with language. And that's, I would argue, most of our knowledge is derived from that interaction with the physical world. So a lot of my colleagues who are more interested in things like computer vision are really on that camp, that AI needs to be embodied essentially. And then other people coming from the NLP side or maybe some other motivation don't necessarily agree with that. And philosophers are split as well. And the complexity of the world is hard to imagine. It's hard to represent all the complexities that we take completely for granted in the real world that we don't even imagine require intelligence. This is the old Moravec paradox from the pioneer of robotics, hence Moravec, who said, how is it that with computers it seems to be easy to do high level complex tasks like playing chess and solving integrals and doing things like that, Whereas the thing we take for granted that we do every day, like, I don't know, learning to drive a car or grabbing an object, we can do as computers and we have LLMs that can pass the bar exam, so they must be smart, but then they can't learn to drive in 20 hours like any 17 year old. They can't learn to clear up the dinner table and fill up the dishwasher like any 10 year old can learn in one shot. Why is that? What are we missing? What type of learning or reasoning architecture or whatever are we missing that basically prevent us from having level 5 shuttering cars and domestic robots?

Can a large language model construct a world model that does know how to drive and does know how to fill a dishwasher, but just doesn't know how to deal with visual data at this time? So it can operate in a space of concepts.

So yeah, that's what a lot of people are working on. So the answer, the short answer is no. And the more complex answer is you can use all kind of tricks to get an LLM to basically digest visual representations of representations of images or video or audio for that matter. And a classical way of doing this is you train a vision system in some way. And we have a number of ways to train vision systems, either supervised, semi supervised, self supervised, all kinds of different ways that will turn any image into a high level representation. Basically a list of tokens that are really similar to the kind of tokens that typical LLM takes as an input, and then you just feed that to the LLM in addition to the text and you just expect the LLM to kind of during training to kind of be able to use those representations to help make decisions. I mean, there's been work along those lines for quite a long time. And now you see those systems, right? I mean, there are LLMs that have some vision extension, but they're basically hacks in the sense that those things are not trained end to end, to handle, to really understand the world. They're not trained with video, for example. They don't really understand intuitive physics, at least not at the moment.

So you don't think there's something special to you about intuitive physics, about sort of common sense reasoning, about the physical space, about physical reality. That to you is a giant leap that LLMs are just not able to do.

We're not going to be able to do this with the type of LLMs that we are working with today. And there's a number of reasons for this. But the main reason is the way LLMs are trained is that you take a piece of text, you remove some of the words in that text, you mask them, you replace them by blank markers, and you train a gigantic neural net to predict the words that are missing. And if you build this neural net in a particular way so that it can only look at words that are to the left of the one it's trying to predict, then what you have is a system that basically is trained to predict the next word in a text. So Then you can feed it a text, a prompt, and you can ask it to predict the next word. It can never predict the next word. Exactly. And so what it's going to do is produce a probability distribution of all the possible words in your dictionary. In fact, it doesn't predict words. It predicts tokens that are kind of subword units. And so it's easy to handle the uncertainty in the prediction there because there's only a finite number of possible words in the dictionary, and you can just compute the distribution over them. Then what the system does is that it picks a word from that distribution. Of course, there's a higher chance of picking words that have a higher probability within that distribution. So you sample from that distribution to actually produce a word, and then you shift that word into the input. And so that allows the system now to predict the second word. Right. And once you do this, you shift it into the input, et cetera. That's called autoregressive prediction, which is why those LLMs should be called autoregressive LLMs, but we just call them LLMs. And there is a difference between this kind of process and a process by which, before producing a word, when you talk, when you and I talk, you and I are bilinguals, we think about what we're going to say, and it's relatively independent of the language in which we're going to say it. When we talk about, I don't know, let's say, a mathematical concept or something, the kind of thinking that we're doing and the answer that we're planning to produce is not linked to whether we're going to see it in French or Russian or English.

Chomsky just rolled his eyes, but I understand. So you're saying that there's a bigger abstraction that goes before language and maps onto language.

Right. It's certainly true for a lot of thinking that we. That we do.

Is that obvious that we don't. Like, you're saying your thinking is same in French as it is in English?

Yeah, pretty much.

Yeah, pretty much. Or is this, like, how. How flexible are you? Like, if. If there's a probability distribution?

Well, it. It depends what kind of thinking. Right. If it's just. If it's like producing puns, I get much better in French than English about that.

No, but so. Right.

Much worse.

Is there an abstract representation of puns? Like, is your humor an abstract. Like when you tweet and your tweets are sometimes a little bit spicy, Is there an abstract representation in your brain of a tweet? Before it maps onto English, there is.

An abstract representation of imagining the reaction of a reader to that text.

Or you start with laughter and then figure out how to make that happen.

Figure out a reaction you want to cause and then figure out how to say it so that it causes that reaction. But that's really close to language. But think about mathematical concept or imagining something you want to build out of wood or something like this. The kind of thinking you're doing has absolutely nothing to do with language, really. It's not like you have necessarily an internal monologue in any particular language. You're imagining mental models of the thing. Right. I mean, if I ask you to imagine what this water bottle will look like if I rotate it 90 degrees, that has nothing to do with language. And so clearly there is a more abstract level of representation in which we do most of our thinking and we plan what we're going to say. If the output is uttered words, as opposed to an output being muscle actions, we plan our answer before we produce it. And LLMs don't do that. They just produce one word after the other, instinctively, if you want. It's a bit like the subconscious actions where you don't, like, you're distracted, you're doing something, you're completely concentrated, and someone comes to you and asks you a question and you kind of answer the question. You don't have time to think about the answer, but the answer is easy, so you don't need to pay attention. You sort of respond automatically. That's kind of what an LLM does, right? It doesn't think about its answer really. It retrieves it because it's accumulated a lot of knowledge, so it can retrieve some things, but it's going to just spit out one token after the other without planning the answer.

But you're making it sound just one token after the other, one token at a time. Generation is bound to be simplistic. But if the world model is sufficiently sophisticated that one token at a time, the most likely thing it generates is a sequence of tokens is going to be a deeply profound thing.

Okay, but then that assumes that those systems actually possess an internal world model.

So really goes to the. I think the fundamental question is, can you build a really complete world model? Not complete, but a one that has a deep understanding of the world?

Yeah. So can you build this first of all by prediction?

Right.

And the answer is probably yes. Can you build it by predicting words? And the answer is most probably no, because language is very poor in terms of weak or low bandwidth. If you Want, there's just not enough information there. So building world models means observing the world and understanding why the world is evolving the way it is. And then the extra component of a world model is something that can predict how the world is going to evolve as a consequence of an action you might take. Right? So world model really is. Here is my idea of the state of the world at time T. Here is an action I might take. What is the predicted state of the world at time T +1. Now that state of the world does not need to represent everything about the world. It just needs to represent enough that's relevant for this planning of the action, but not necessarily all the details. Now here is the problem. You're not going to be able to do this with generative models. So a generative model has trained on video and we've tried to do this for 10 years. You take a video, show a system a piece of video and then ask it to predict the reminder of the video. Basically predict what's going to happen one.

Frame at a time. Do the same thing as sort of the auto aggressive LLMs do, but for video, right?

Either one friend at a time or a group of friends at a time. But yeah, a large video model if you want. The idea of doing this has been floating around for a long time and at fair, some of my colleagues and I have been trying to do this for about 10 years. And you can't really do the same trick as with LLMs, because LLMs, as I said, you can't predict exactly which word is going to follow a sequence of words, but you can predict a distribution over words. Now if you go to video, what you would have to do is predict the distribution over all possible frames in a video. And we don't really know how to do that properly. We do not know how to represent distributions over high dimensional continuous spaces in ways that are useful. And there lies the main issue. And the reason we can do this is because the world is incredibly more complicated and richer in terms of information than text. Text is discrete, video is high dimensional and continuous. A lot of details in this. So if I take a video of this room and the video is a camera panning around, there is no way I can predict everything that's going to be in the room as I pan around. The system cannot predict what's going to be in the room as the camera is panning. Maybe it's going to predict this is a room where there's a light and there is a wall and things like that. It can't Predict what the painting on the wall looks like or what the texture of the couch looks like. Certainly not the texture of the carpet. So there's no way I can predict all those details. So the way to handle this is one way possibly to handle this, which we've been working for a long time, is to have a model that has what's called a latent variable. And the latent variable is fed to a neural net and it's supposed to represent all the information about the world that you don't perceive yet, and that you need to augment the system for the prediction, to do a good job at predicting pixels, including the fine texture of the carpet and the couch and the painting on the wall. That has been a complete failure, essentially. And we've tried lots of things. We tried just straight neural nets, we tried GANs, we tried VAEs, all kinds of regularized autoencoders. We tried many things. We also tried those kind of methods to learn good representations of images or video that could then be used as input to, for example, an image classification system. And that also has basically failed all the systems that attempt to predict missing parts of an image or video from a corrupted version of it, basically. So take an image or a video, corrupt it or transform it in some way, and then try to reconstruct the complete video or image from the corrupted version, and then hope that internally the system will develop good representations of images that you can use for object recognition, segmentation, whatever it is. That has been essentially a complete failure. And it works really well for text. That's the principle that is used for LLMs. Right?

So where is the failure exactly? Is it that it's very difficult to form a good representation of an image, like a good embedding of all the important information in the image? Is it in terms of the consistency of image to image to image to image that forms the video? Like, where are the. If we do a highlight reel of all the ways you failed, what's that look like?

Okay, so the reason this doesn't work is, first of all, I have to tell you exactly what doesn't work, because there is something else that does work. So the thing that does not work is training the system to learn representations of images by training it to reconstruct a good image from a corrupted version of it. Okay, that's what doesn't work. And we have a whole slew of techniques for this that are variant of denoising autoencoders, something called mae, developed by some of my colleagues at Fair Maxed autoencoder so it's basically like the LLMs or things like this, where you train the system by corrupting text, except you corrupt images, you remove patches from it, and you train a gigantic neural network. The features you get are not good. And you know they're not good because if you now train the same architecture, but you train it supervised, with labeled data, with textual descriptions of images, et cetera, you do get good representations. And the performance on recognition tasks is much better than if you do this self supervised pre training.

So the architecture is good, the architecture.

Is good, the architecture of the encoder is good. But the fact that you train the system to reconstruct images does not lead it to produce, to learn good generic features of images when you train in.

A self supervised way.

Self supervised by reconstruction.

Yeah, by reconstruction.

Okay, so what's the alternative? The alternative is joint embedding.

What is joint embedding? What are these architectures that you're so excited about?

Okay, so now instead of training a system to encode the image and then training it to reconstruct the full image from a corrupted version, you take the full image, you take the corrupted or transformed version, you run them both through encoders, which in general are identical, but not necessarily, and then you train a predictor on top of those encoders to predict the representation of the full input from the representation of the corrupted one. Okay, so joint embedding, because you're taking the full input and the corrupted version or transform version, run them both through encoders. So you get a joint embedding and then you're saying, can I predict the representation of the full one from the representation of the corrupted one? Okay, And I call this a jepa. So that means joint embedding, predictive architecture, because there's joint embedding and there is this predictor that predicts the representation of the good guy from, from the bad guy. And the big question is, how do you train something like this? And until five years ago, six years ago, we didn't have particularly good answers for how you train those things, except for one called contractive learning. And the idea of contractive learning is you take a pair of images that are again an image and a corrupted version or degraded version, somehow or transformed version of the original one, and you train the predicted representation to be the same as that. If you only do this, the system collapses. It basically completely ignores the input and produces representations that are constant. So the contrastive methods avoid this. And those things have been around since the early 90s. I had a paper on this in 1993 is you also show pairs of images that you know are different and then you push away the representations from each other. So you say not only do representations of things that we know are the same should be the same or should be similar, but representation of things that we know are different should be different. And that prevents the collapse. But it has some limitation and there's a whole bunch of techniques that have appeared over the last six, seven years that can revive this type of method, some of them from fair, some of them from Google and other places. But there are limitations to those contrastive methods. What has changed in the last three, four years is now we have methods that are non contrastive, so they don't require those negative contrastive samples of images that are, that we know are different. You train them only with images that are, you know, different versions or different views of the same thing. And you rely on some other tricks to prevent the system from collapsing. And we have half a dozen different methods for this now.

So what is the fundamental difference between joint embedding architectures and LLMs? So can JAPA take us to AGI, whether we should say that you don't like the term AGI? And we'll probably argue, I think every single time I've talked to you with argued about the G&AGI. Yes, I get it, I get it. Well, we'll probably continue to argue about it. It's great. You, you like me this because, because you like French and AMIS is I guess friend in French.

Yes.

And AMI stands for Advanced Machine Intelligence.

Right.

But either way, can JAPA take us to that towards that advanced Machine intelligence?

Well, so it's a first step. Okay, so first of all, what's the difference with generative architectures like LLMs? So LLMs or vision systems that are trained by reconstruction generate the inputs, they generate the original input that is non corrupted, non transformed. Right. So you have to predict all the pixels and there is a huge amount of resources spent in the system to actually predict all those pixels. All the details in a jepa, you're not trying to predict all the pixels, you're only trying to predict an abstract representation of the inputs. Right. And that's much easier in many ways. So what the JETPA system, when it's being trained, is trying to do is extract as much information as possible from the input, but yet only extract information that is relatively easily predictable. Okay, so there's a lot of things in the world that we cannot predict. For example, if you have a self driving, car driving down the street or road, there may be trees around the road and it could be a windy day. So the leaves on the tree are kind of moving in kind of semi chaotic, random ways that you can't predict and you don't care, you don't want to predict. So what you want is your encoder to basically eliminate all those details. It will tell you there's moving leaves, but it's not going to keep the details of exactly what's going on. And so when you do the prediction in representation space, you're not going to have to predict every single pixel of every leaf. And that not only is a lot simpler, but also it allows the system to essentially learn an abstract representation of the world where what can be modeled and predicted is preserved and the rest is viewed as noise and eliminated by the encoder. So it kind of lifts the level of abstraction of the representation. If you think about this, this is something we do absolutely all the time. Whenever we describe a phenomenon, we describe it at a particular level of abstraction. And we don't always describe every natural phenomenon in terms of quantum field theory, that would be impossible. So we have multiple levels of abstraction to describe what happens in the world, starting from quantum field theory to atomic theory and molecules and chemistry, materials, and all the way up to kind of concrete objects in the real world and things like that. So we can't just only model everything at the lowest level. And that's what the idea of JEPA is really about. Learn abstract representation in a self supervised manner and you can do it hierarchically as well. So that I think is an essential component of an intelligent system. And in language we can get away without doing this, because language is already, to some level abstract and already has eliminated a lot of information that is not predictable. So we can get away without doing those joint embedding, without lifting the abstraction level and by directly predicting words.

So joint embedding, it's still generative, but it's generative in this abstract representation space. And you're saying language. We were lazy with language because we already got the abstract representation for free and now we have to zoom out, actually think about generally intelligent systems. We have to deal with a full mess of physical reality, of reality. And you can't, you, you do have to do this step of jumping from the full, rich, detailed reality to a abstract representation of that reality based on which you can reason and all that kind of stuff.

Right. And the thing is, those self supervised algorithms that learn by prediction, even in representation space, they learn more concepts. If the Input data you feed them is more redundant. The more redundancy there is in the data, the more they're able to capture some internal structure of it. And so there, there is way more redundancy and structure in perceptual inputs, sensory input, like vision, than there is in text, which is not nearly as redundant. This is back to the question you were asking a few minutes ago. Language might represent more information, really because it's already compressed. You're right about that. But that means it's also less redundant. And so self supervised learning will not work as well.

Is it possible to join the self supervised training on visual data and self supervised training on language data? There is a huge amount of knowledge. Even though you talk down about those 10 to the 13 tokens, those 10 to the 13 tokens represent the entirety, a large fraction of what us humans have figured out, both the shit talk on Reddit and the contents of all the books and the articles, and the full spectrum of human intellectual creation. So is it possible to join those two together?

Well, eventually, yes. But I think if we do this too early, we run the risk of being tempted to cheat. And in fact, that's what people are doing at the moment. With vision language model, we're basically cheating. We're using language as a crutch to help the deficiencies of our vision systems to kind of learn good representations from images and video. And the problem with this is that we might improve our vision language system a bit, I mean, our language models by feeding them images, but we're not going to get to the level of even the intelligence or level of understanding of the world of a cat or dog, which doesn't have language. They don't have language and they understand the world much better than any LLM. They can plan really complex actions and sort of imagine the result of a bunch of actions. How do we get machines to learn that before we combine that with language? Obviously, if we combine this with language, this is going to be a winner. But before that we have to focus on how do we get systems to learn how the world works.

So this kind of joint embedding, predictive architecture for you that's going to be able to learn something like common sense, something like what a cat uses to predict how to mess with its owner, most optimally by knocking over a thing.

That's the hope. In fact, the techniques we're using are non contrastive. So not only is the architecture non generative, the learning procedures we're using are non contrastive. So we have two sets of techniques One set is based on distillation, and there's a number of methods that use this principle. One by DeepMind called BYOL, a couple by Fair, one called Vicreg, and another one called Ijepa. And Vicreg, I should say, is not a distillation method, actually, but I JEPA and BYOL certainly are. And there's another one, also called DINO or Dyno, also produced from AT fair. And the idea of those things is that you take the full input, let's say an image, you run it through an encoder, produces a representation, and then you corrupt that input or transform it, run it through essentially what amounts to the same encoder with some minor differences, and then train a predictor. Sometimes the predictor is very simple, sometimes doesn't exist. But train a predictor to predict a representation of the first uncorrupted input from the corrupted input, but you only train the second branch. You only train the part of the network that is fed with the corrupted input. The other network you don't train, but since they share the same weight, when you modify the first one, it also modifies the second one. And with various tricks, you can prevent this system from collapsing with the collapse of the type I was explaining before, where the system basically ignores the input. So that works very well. The two techniques we've developed at fair, Deno and I jepa, work really well for that.

So what kind of data are we talking about here?

So there's several scenarios. One scenario is you take an image, you corrupt it by changing the cropping, for example, changing the size a little bit, maybe changing the orientation, blurring it, changing the colors, doing all kinds of horrible things to it, but basic horrible things, basic horrible things that sort of degrade the quality a little bit and change the framing, crop the image. And in some cases, in the case of I jepa, you don't need to do any of this. You just mask some parts of it, you just basically remove some regions, like a big block, essentially, and then, you know, run through the encoders and train the entire system, encoder and predictor, to predict the representation of the good one from the representation of the corrupted one. So that's the I JEPA doesn't need to know that it's an image, for example, because the only thing it needs to know is how to do this masking. Whereas with Deno, you need to know it's an image because you need to do things like geometry transformation and blurring and things like that that are really image specific. A More recent version of this that we have is called V jepa. So it's basically the same idea as I jepa, except it's applied to video. So now you take a whole video and you mask a whole chunk of it. And what we mask is actually kind of a temporal tube. So a whole segment of each frame in the video over the entire video.

And that tube was like statically positioned throughout the frames.

The tube, yeah, typically is 16 frames or something. And we mask the same region over the entire 16 frames. It's a different one for every video, obviously. And then again, train that system so as to predict the representation of the full video from the partially masked video. And that works really well. It's the first system that we have that learns good representations of video so that when you feed those representations to a supervised classifier head, it can tell you what action is taking place in the video with, you know, pretty good accuracy. So it's the first time we get something of that quality.

So that's a good test that a good representation is formed. That means there's something to this.

Yeah. We have also preliminary result that seem to indicate that the representation allow our system to tell whether the video is physically possible or completely impossible because some object disappeared or an object suddenly jumped from one location to another or change shape or something.

So it's able to capture some physics based constraints about the reality represented in the video about the appearance and the disappearance of objects.

Yeah, that's really new.

Okay, but can this actually get us to this kind of world model that understands enough about the world to be able to drive a car?

Possibly. This is going to take a while before we get to that point. And there are systems already, robotic systems, that are based on this idea. And what you need for this is a slightly modified version of this where imagine that you have a video and a complete video and what you're doing to this video is that you're either translating it in time towards the future. So you only see the beginning of the video, but you don't see the latter part of it that is in the original one. Or you just mask the second half of the video, for example, and then you train a JEPA system of the type I described to predict the representation of the full video from the shifted one. But you also feed the predictor with an action. For example, you know, the wheel is turned 10 degrees to the right or something. Right. So if it's a, you know, a dash cam in a car and you know, the angle of the wheel, you should be able to predict to some extent what's going to happen to what you see. You're not going to be able to predict all the details of, you know, objects that appear in the view, obviously, but at a abstract representation level, you can, you can probably predict what's going to happen. So now what you have is an internal model that says, here is my idea of the state of the world at time T. Here is an action I'm taking. Here is a prediction of the state of the world at time t +1, t +Delta t, t +2 seconds, whatever it is. If you have a model of this type, you can use it for planning. So now you can do what LLMs cannot do, which is planning what you're going to do so as to arrive at a particular outcome or satisfy a particular objective. So you can have a number of objectives. I can predict that if I have an object like this and I open my hand, it's going to fall, and if I push it with a particular force on the table, it's going to move. If I push the table itself, it's probably not going to move with the same force. So we have this internal model of the world in our mind which allows us to plan sequences of actions to arrive at a particular goal. And so now if you have this world model, we can imagine a sequence of actions, predict what the outcome of the sequence of action is going to be, measure to what extent the final state satisfies a particular objective, like moving the bottle to the left of the table, and then plan a sequence of actions that will minimize this objective. At runtime. We're not talking about learning, we're talking about inference time. So this is planning, really. And in optimal control, this is a very classical thing. It's called model predictive control. You have a model of the system you want to control that can predict the sequence of states corresponding to a sequence of commands. And you're planning a sequence of commands so that according to your world model, the end state of the system will satisfy an objective that you fix. This is the way rocket trajectories have been planned since computers have been around since the early 60s, essentially.

So, yes, for model predictive control, but you also often talk about hierarchical planning. Can hierarchical planning emerge from this somehow?

Well, so no, you will have to build a specific architecture to allow for hierarchical planning. So hierarchical planning is absolutely necessary if you want to plan complex actions. If I want to go from, let's say, from New York to Paris, this is the example I use all the time, and I'M sitting in my office at nyu. My objective that I need to minimize is my distance to Paris at a high level, a very abstract representation of my location. I would have to decompose this into two sub goals. First one is go to the airport. Second one is catch a plane to Paris. Okay, so my sub goal is now going to the airport. My objective function is my distance to the airport. How do I go to the airport, where I have to go in the street and hail a taxi, which you can do in New York. Okay, now I have another sub goal. Go down on the street. Well, that means going to the elevator, going down the elevator, walk out the street. How do I go to the elevator? I have to stand up from my chair, open the door of my office, go to the elevator, push the button. How do I get up from my chair? You can imagine going down, all the way down to basically what amounts to millisecond by millisecond muscle control. Obviously, you're not going to plan your entire trip from New York to Paris in terms of millisecond by millisecond muscle control. First. That would be incredibly expensive, but it will also be completely impossible because you don't know all the conditions, what's going to happen, how long it's going to take to catch a taxi or to go to the airport with traffic. You would have to know exactly the condition of everything to be able to do this planning. And you don't have the information. So you have to do this hierarchical planning so that you can start acting and then sort of replanning as you go. And nobody really knows how to do this in AI. Nobody knows how to train a system to learn the appropriate multiple levels of representation so that hierarchical planning works.

Does something like that already emerge? So, like, can you use an LLM, state of the art LLM to get you from New York to Paris by doing exactly the kind of detailed set of questions that you just did, which is, can you give me a list of 10 steps I need to do to get from New York to Paris? And then for each of those steps, can you give me a list of 10 steps, how I make that step happen? And for each of those steps, can you give me a list of 10 steps to make each one of those until you're moving your individual muscles, maybe not whatever you can actually act upon using your own mind?

Right, so there's a lot of questions that are sort of implied by this. Right. So the first thing is LLMs will be able to answer some of those questions down to some level of abstraction under the condition that they've been trained with similar scenarios in their training set.

They would be able to answer all of those questions, but some of them may be hallucinated, meaning non factual.

Yeah, true. I mean they will probably produce some answer, except they're not going to be able to really kind of produce millisecond by millisecond muscle control of how you stand up from your chair. Right. But down to some level of abstraction where you can describe things by words. They might be able to give you a plan, but only under the condition that they've been trained to produce those kind of plans. They're not going to be able to plan for situations where that they never encountered before. They basically are going to have to regurgitate the template that they've been trained on.

But where, like just for the example of New York to Paris, is it going to start getting into trouble? Like at which layer, layer of abstraction do you think you'll start? Because like I can imagine almost every single part of that an LLM will be able to answer somewhat accurately, especially when you're talking about New York and Paris, major cities.

So I mean certainly LLM would be able to solve that problem if you fine tune it for it. You know, just. And, and so I can't say that NLM cannot do this. It can do this if you train it for it. There's no question down to a certain level where things can be formulated in terms of words. But if you want to go down to how do you climb down the stairs or just stand up from your chair in terms of words, you can't do it. That's one of the reasons you need experience of the physical world which is much higher bandwidth than what you can express in words in human language.

So everything we've been talking about on the joint embedding space, is it possible that that's what we need for like interaction with physical reality for on the robotics front and then just the LLMs are the thing that sits on top of it for the bigger reasoning about like the fact that I need to book a plane ticket and I need to know I know how to go to the websites and so on.

Sure. And you know, a lot of plans that people know about that are relatively high level are actually learned. Most people don't invent the plans by themselves. We have some ability to do this, of course, obviously. But most plans that people use are plants that have been trained on, they've seen other people use those plants or they've been Told how to do things that you can't invent. How you take a person who's never heard of airplanes and tell them how do you go from New York to Paris? And they're probably not going to be able to kind of deconstruct the whole plan unless they've seen examples of that before. So certainly LLMs are going to be able to do this. But then how you link this from the low level of actions that needs to be done with things like JEPA that basically lift the abstraction level of the representation without attempting to reconstruct every detail of the situation. That's why you would need JPA for.

I would love to sort of linger on your skepticism around auto aggressive LLMs. So one way I would like to test that skepticism is everything you say makes a lot of sense. But if I apply everything you said today and in general to like, I don't know, 10 years ago, maybe a little bit less. No, let's say three years ago, I wouldn't be able to predict the success of LLMs. So does it make sense to you that autoregressive LLMs are able to be so damn good?

Yes.

Can you explain your intuition? Because if I were to take your wisdom and intuition at face value, I would say there's no way autoregressive LLMs, one token at a time, would be able to do the kind of things they're doing.

No, there's one thing that auto regressive LLMs or that LLMs in general, not just the autoregressive one, but including the BERT style bidirectional ones, are exploiting. And it's self supervised learning. And I've been a very, very strong advocate of self supervised learning for many years. So those things are an incredibly impressive demonstration that self supervised learning actually works. The idea that it didn't start with bert, but it was really kind of a good demonstration with this. So the idea that you take a piece of text, you corrupt it, and then you train some gigantic neural net to reconstruct the parts that are missing, that has been an enormous, produced an enormous amount of benefits. It allowed us to create systems that understand language, systems that can translate hundreds of languages in any direction, systems that are multilingual, so they're not. It's a single system that can be trained to understand hundreds of languages and translate in any direction and produce summaries and then answer questions and produce text. And then there's a special case of it, which is the autoregressive trick where you constrain the system to not elaborate a representation of the text from looking at the entire text, but only predicting a word from the words that are come before. Right. And you do this by constraining the architecture of the network. And that's what you can build an autoregressive LLM from. So there was a surprise many years ago with what's called decoder only LLM. So since systems of this type that are just trained to produce words from the previous one, and the fact that when you scale them up, they tend to really kind of understand more about language, when you train them on lots of data, you make them really big, that was kind of a surprise. And that surprise occurred quite a while back with work From Google Meta OpenAI, et cetera. Going back to the GPT kind of work. General pre trained transformers.

You mean like GPT2? There's a certain place where you start to realize scaling might actually keep giving us an emergent benefit.

Yeah, I mean there were work from various places, but if you want to kind of place it in the GPT timeline, that would be around GPT too.

Yeah, well, I just. Because you said it. It's so charismatic and you said so many words, but self supervised learning. Yes, but again, the same intuition you're applying to saying that autoregressive LLMs cannot have a deep understanding of the world if we just apply that same intuition, does it make sense to you that they're able to form enough of a representation of the world to be damn convincing, essentially passing the original Turing test with flying colors?

Well, we're fooled by their fluency. Right? We just assume that if a system is fluent in manipulating language, then it has all the characteristics of human intelligence. But that impression is false. We're really fooled by it.

What do you think Alan Turing would say without understanding anything, just hanging out.

With it, Alan Turing would decide that a Turing test is a really bad test. Okay. This is what the AI community has decided many years ago, that the Turing test was a really bad test of intelligence.

What would Hans Moravec say about the large language models?

Hence Maravec would say the Maravec paradox still applies. Okay, okay, okay. We can pass.

You don't think he would be really impressed?

No, of course everybody would be impressed. But it's not a question of being impressed or not. It's a question of knowing what the limit of those systems can do. Again, they are impressive. They can do a lot of useful things. There's a whole industry that is being built around them. They're going to make progress. But there is a lot of things they cannot do. And we have to realize what they cannot do and then figure out how we get there. And I'm seeing this from basically 10 years of research on the idea of self supervised learning. Actually that's going back more than 10 years. But the idea of self supervised learning, so basically capturing the internal structure of a set of inputs without training the system for any particular task. Learning representations. You know, the conference I Co founded 14 years ago is called International Conference on Learning Representations. That's the entire issue that deep learning is dealing with. Right. And it's been my obsession for almost 40 years now. So learning representation is really the thing. For the longest time we could only do this with supervised learning. And then we started working on what we used to call unsupervised learning and sort of revived the idea of unsupervised learning in the early 2000s with Joshua Benjo and Jay Finton. Then discovered that supervised learning actually works pretty well if you can collect enough data. And so the whole idea of unsupervised self supervised learning kind of took a backseat for a bit. And then I kind of tried to revive it in a big way starting in 2014, basically when we started fair and really pushing for finding new methods to do self supervised learning both for text and for images and for video and audio. And some of that work has been incredibly successful. I mean the reason why we have multilingual translation system things to do, content moderation on Meta, for example, on Facebook, that are multilingual, that understand whether a piece of text is hate speech or not or something is due to their progress using self supervised learning for nlp. Combining this with transformer architectures and blah blah, blah. But that's the big success of self supervised learning. We had similar success in speech recognition. A system called wave2vec, which is also a joint embedding architecture by the way, trained with contrastive learning. And that system also can produce speech recognition systems that are multilingual with mostly unlabeled data and only need a few minutes of label data to actually do speech recognition. That's amazing. We have systems now based on those combination of ideas that can do real time translation of hundreds of languages into each other. Speech to speech, Speech to speech, even.

Including which is fascinating, languages that don't have written forms.

That's right.

They're spoken only.

That's right. We don't go through text. It goes directly from speech to speech using an internal representation of speech units that are discrete. But it's called Text Lesson lp. We used to call it this way, but yeah, so I mean, incredible success there. And then for 10 years we tried to apply this idea to learning representations of images by training a system to predict videos, learning intuitive physics by training a system to predict what's going to happen in a video. And tried and tried. And failed. And failed. With generative models, with models that predict pixels, we could not get them to learn good representations of images, we could not get them to learn good representations of videos. We tried many times. We published lots of papers on it. They kind of sort of work, but not really great. They started working. We abandoned this idea of predicting every pixel and basically just doing the joint embedding and predicting in representation space. That works. So there's ample evidence that we're not going to be able to learn good representations of the real world using generative model. So I'm telling people, everybody's talking about generative AI. If you're really interested in human level AI, abandon the idea of generative AI.

Okay, but you really think it's possible to get far with the joint embedding representation? So there's common sense reasoning and then there's high level reasoning. I feel like those are two. The kind of reasoning that LLMs are able to do. Okay, let me not use the word reasoning, but the kind of stuff that LLMs are able to do seems fundamentally different than the common sense reasoning we use to navigate the world. Yeah, it seems like we're going to need both. You're not sure. Would you be able to get with the joint embedding, which is a JABA type of approach, looking at video, would you be able to learn, let's see, well, how to get from New York to Paris or how to understand the state of politics in the world today. Right. These, these are things where various humans generate a lot of language and opinions on in the space of language, but don't visually represent that in any clearly compressible way.

Right. Well, there's a lot of situations that might be difficult for a purely language based system to know. Okay. You can probably learn from reading text, the entirety of the publicly available text in the world that I cannot get from New York to Paris by snapping my fingers. That's not going to work. Right? Yes, but there's probably sort of more complex scenarios of this type of which an LLM may never have encountered and may not be able to determine whether it's possible or not. So that link from the low level to the high level. The thing is that the High level that language expresses is based on a common experience of the low level, which LLMs currently do not have. When we talk to each other, we know we have a common experience of the world. Like, you know, a lot of it is similar and LLMs don't have that.

But see, it's present. You and I have a common experience of the world in terms of the physics of how gravity works and stuff like this and that. Common knowledge of the world, I feel like is there in the language. We don't explicitly express it, but if you have a huge amount of text, you're going to get this stuff that's between the lines. You're gonna, you're going to. In order to form a consistent world model, you're going to have to understand how gravity works, even if you don't have an explicit explanation of gravity. So even though in the case of gravity, there is explicit explanations of gravity and Wikipedia, but the stuff that we think of as common sense reasoning, I feel like to generate language correctly, you're going to have to figure that out. Now you could say as you. There's not enough text. Sorry. Okay, so what, you don't think so?

No, I agree with what you just said, which is that to be able to do high level common sense, to have high level common sense, you need to have the low level common sense to build on top of. And, um, but that's not there. And that's not there in LLMs. LLMs are purely trained from text. So. So then the other statement you made, I would not, I would not agree with the fact that implicit in all languages in the world is the underlying reality. There's a lot about underlying reality which is not expressed in language.

Is that obvious to you?

Yeah, totally.

So like all, all the conversations we had. What? Okay, there's the dark web, meaning whatever, the private conversations like DMS and stuff like this, which is much, much larger probably than what's available. What, what LLMs are trained on.

You don't need to communicate the stuff that is common, but the humor, all of it.

No, you do. Like when you, you don't need to, but it comes through. Like you, like if I accidentally knock this over, you'll probably make fun of me in, in the content of the. You making fun of me will be explanation of the fact that cups fall. And then, you know, gravity works in this way. And then you'll have some very vague information about what kind of things explode when they hit the ground. And then maybe you'll make a joke about entropy or something. Like this. And we'll never be able to reconstruct this again. Like, okay, you'll make a little joke like this and there'll be trillion of other jokes. And from the jokes you can piece together the fact that gravity works and mugs can break and all this kind of stuff you don't need to see. It'll be very inefficient. It's easier for like to knock the thing over.

Yeah.

But I feel like it would be there if you have enough of that data.

I just think that most of the information of this type that we have accumulated when we were babies is just not present in text, in any description, essentially.

And the sensory data is a much richer source for getting that kind of understanding.

I mean, that's the 16,000 hours of wake time of a 4 year old and 10 to the 15 bytes going through vision, just vision. Right. There is a similar bandwidth of touch and a little less through audio. And then text. Language doesn't come in until, you know, a year in life. And by the time you are nine years old, you've learned about gravity, you know about inertia, you know about gravity, you know the stability, you know, you know about the distinction between animate and inanimate objects, you know, by 18 months, you know about like why people want to do things and you help them if they can't. You know, I mean, there's a lot of things that you learn mostly by observation, really, not even through interaction. In the first few months of life, babies don't really have any influence on the world. They can only observe. Right. And you accumulate a gigantic amount of knowledge just from that. So that's what we're missing from current AI systems.

I think in one of your slides you have this nice plot that is one of the ways you show that LLMs are limited. I wonder if you could talk about hallucinations from your perspectives. The why hallucinations happen from large language models and why and to what degree is that a fundamental flaw of large language models?

Right. So because of the autoregressive prediction, every time an LLM produces a token or a word, there is some level of probability for that word to take you out of the set of reasonable answers. And if you assume, which is a very strong assumption, that the probability of such error is that those errors are independent across a sequence of tokens being produced. What that means is that every time you produce a token, the probability that you stay within the set of correct answer decreases and it decreases exponentially.

So there's a strong like you said assumption there that if there's a non zero probability of making a mistake, which there appears to be, then there's going to be a kind of drift.

Yeah, and that drift is exponential. It's like errors accumulate. Right. So the probability that an answer would be nonsensical increases exponentially with the number of tokens.

Is that obvious to you, by the way? Well, so mathematically speaking, maybe. But like, isn't there a kind of gravitational pull towards the truth? Because on average, hopefully the truth is well represented in the training set.

No, it's basically a struggle against the curse of dimensionality. So the way you can correct for this is that you fine tune the system by having it produce answers for all kinds of questions that people might come up with. And people are people, so a lot of the questions that they have are very similar to each other. So you can probably cover 80% or whatever of questions that people will ask by collecting data. And then you fine tune the system to produce good answers for all of those things. And it's probably going to be able to learn that because it's got a lot of capacity to learn. But then there is the enormous set of prompts that you have not covered during training, and that set is enormous. Within the set of all possible prompts, the proportion of prompts that have been used for training is absolutely tiny. It's a tiny, tiny, tiny subset of all possible prompts. And so the system will behave properly on the prompts that it's been either trained, pre trained or fine tuned. But then there is an entire space of things that it cannot possibly have been trained on because it's just the number is gigantic. So whatever training the system has been subject to, to produce appropriate answers, you can break it by finding out a prompt that will be outside of the set of prompts it's been trained on or things that are similar, and then it will just spew complete nonsense.

When you say prompt, do you mean that exact prompt or do you mean a prompt that's like in many parts very different than. Is it that easy to ask a question or to say a thing that hasn't been said before on the Internet?

I mean, people have come up with things where you put essentially a random sequence of characters in the prompt and that's enough to kind of throw the system into a mode where it's going to answer something completely different than it would have answered without this. So that's a way to jailbreak the system, basically get it go outside of its conditioning. Right.

So that's a very clear demonstration of it. But of course that goes outside of what is designed to do. If you actually stitch together reasonably grammatical sentences, is it that easy to break it?

Yeah. Some people have done things like you write a sentence in English or you ask a question in English and it produces a perfectly fine answer, and then you just substitute a few words by the same word in another language and all of a sudden the answer is complete nonsense.

Yeah. So I guess what I'm saying is like, which fraction of prompts that humans are likely to generate are going to break the system.

So the problem is that there is a long tail.

Yes.

This is an issue that a lot of people have realized in social networks and stuff like that, which is there's a very, very long tail of things that people will ask. And you can fine tune the system for the 80% or whatever of the things that most people will ask, and then this long tail is so large that you're not going to be able to fine tune the system for all the conditions. And in the end the system ends up being kind of a giant lookup table. Right. Essentially. Which is not really what you want. You want systems that can reason, certainly they can plan. So the type of reasoning that takes place in LLM is very, very primitive. And the reason you can tell it's primitive is because the amount of computation that is spent per token produced is constant. So if you ask a question and that question has an answer in a given number of token, the amount of computation devoted to computing that answer can be exactly estimated. It's the size of the prediction network with its 36 layers or 92 layers or whatever it is, multiplied by number of tokens. That's it. And so essentially it doesn't matter if the question being asked is simple to answer, complicated to answer, impossible to answer because it's undecidable or something. The amount of computation the system will be able to devote to the answer is constant or is proportional to the number of tokens produced in the answer. Right. This is not the way we work. The way we reason is that when we're faced with a complex problem or a complex question, we spend more time trying to solve it and answer it. Right. Because it's more difficult.

There's a prediction element, there's an iterative element where you're like adjusting your understanding of a thing by going over and over and over. There's a hierarchical element, so on. Does this mean it's a fundamental law of LLMs, but does it mean that there's more Part to that question. Now you're just behaving like an LLM. Immediately answer no, that is just the low level world model, on top of which we can then build some of these kinds of mechanisms, like you said, persistent long term memory or reasoning, so on. But we need that world model that comes from language. Maybe it is not so difficult to build this kind of reasoning system on top of a well constructed world model.

Okay. Whether it's difficult or not, the near future will say, because a lot of people are working on reasoning and planning abilities for dialogue systems. I mean, even if we restrict ourselves to language, just having the ability to plan your answer before you answer in terms that are not necessarily linked with the language you're going to use to produce the answer, Right? So this idea of the semantic model that allows you to plan what you're going to say before you say it, that is very important. I think there's going to be a lot of systems over the next few years that are going to have this capability, but the blueprint of those systems will be extremely different from autoregressive LLMs. So it's the same difference as the difference between what psychologists call System one and System two in humans. So system one is the type of task that you can accomplish without deliberately consciously think about how you do them. You just do them. You've done them enough that you can just do it subconsciously without thinking about them. If you're an experienced driver, you can drive without really thinking about it and you can talk to someone at the same time or listen to the radio. Right? If you are a very experienced chess player, you can play against a non experienced chess player without really thinking either. You just recognize the pattern and you play. That's system one. So all the things that you do instinctively without really having to deliberately plan and think about it, and then there is all the tasks where you need to plan. So if you are a not too experienced chess player, or you are experienced, but you play against another experienced chess player, you think about all kinds of options, you think about it for a while and you're much better if you have time to think about it than you are if you play blitz with limited time. So this type of deliberate planning which uses your internal world model, that's system two. This is what LLMs currently cannot do. How do we get them to do this? Right? How do we build a system that can do this kind of planning or reasoning, that devotes more resources to complex problems than to simple problems? And it's not going to be autoregressive prediction of tokens. It's going to be more something akin to inference of latent variables in what used to be called probabilistic models or graphical models and things of that type. So basically the principle is like this. The prompt is like observed variables and what the model does is that it's basically a measure of it can measure to what extent an answer is a good answer for a prompt. Okay, so think of it as some gigantic neural net but it's got only one output and that output is a scalar number which is let's say zero if the answer is a good answer for the question and a large number if the answer is not a good answer for the question. Imagine you had this model. If you had such a model you could use it to produce good answers. The way you would do is produce the prompt and then search through the space of possible answers for one that minimizes that number. That's called an energy based model.

But that energy based model would need the model.