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 are fundamentally good.
And in fact, a lot of doomers are 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
in the walk by open sourcing many of their biggest models, including Lama 2 and eventually
Lama 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 Ruhmert Podcast.
To support it, please check out our sponsors in the description.
And now, dear friends, 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 auto-aggressive LLMs are not the way we're going to make progress towards
superhuman intelligence.
These are the large language models, like GPT-4, like Lama 2 and 3 soon, and so on.
How do they work, and why are they not going to take us all the way?
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, you know, if you expect a system to become intelligent just, you know, without having
the possibility of doing those things, you're making a mistake.
That is not to say that auto-aggressive 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 texts on the internet, right?
That's typically on the order of 10 to the 13 tokens.
Each token is typically two bytes.
So that's two 10 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, right, 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 four-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 four-year-old versus two 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 would be good to maybe push against some of the intuition behind what you're saying.
So it is true there are 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 in 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 LLM's 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, you know, a physical reality.
It could be simulated, 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, a modern task of, you know, grabbing something, etc.
We plan or 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.
And 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, you know, 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, right?
This is the old Moravec paradox from the pioneer of robotics, Hans Moravec, who said, you know,
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, you know, grabbing an object, we can't do with computers.
And, you know, 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, like, you know, what are we missing?
What type of learning or reasoning architecture or whatever are we missing that basically prevent us from, you know, having level 5 start-over in 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 kinds of tricks to get an LLM to basically digest visual 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, it's been working along those lines for quite a long time.
And now you see those systems, right?
I mean, there are LLMs that can, that have some vision extension.
But they're basically hacks in the sense that those things are not, like, 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, right?
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, etc.
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, like, 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 say 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 do.
Is that obvious that we don't?
Like, you're saying your thinking is the same in French as it is in English.
Yeah, pretty much.
Pretty much.
Or is this, like, how flexible are you?
Like, if there's a probability of distribution.
Well, 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, right.
Or much worse.
Is there an abstract representation of puns?
Like, is your humor an abstract representation?
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?
Or figure out, like, a reaction you want to cause and then figure out how to say it, right?
So that it causes that reaction.
But that's, like, really close to language.
But think about, like, a mathematical concept or, you know, imagining something you want to build out of wood or something like this, right?
The kind of thinking you're doing has absolutely nothing to do with language, really.
Like, it's not like you have necessarily, like, an internal monologue in any particular language.
You're, you know, imagining mental models of the thing, right?
I mean, if I ask you to, like, 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, you know, 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, you know, uttered words, as opposed to an output being, you know, muscle actions, right?
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 like, it's a bit like the, you know, subconscious actions where you don't, like, you're distracted, you're doing something, you're completely concentrated, and someone comes to you and, you know, ask 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 eternal world model.
So it really goes to the, I think the fundamental question is, can you build a really complete world model?
Not complete, but 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 a 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 plus one?
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 is 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 frame 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 the 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, you know, 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, you know, fine texture of the carpet and the couch and the painting on the wall.
So 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, you know, 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.
Like, all the systems that attempt to predict missing parts of an image or video, you know, from a corrupted version of it, basically.
So, right, 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, what is 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, you know, a variant of denoising autoencoders.
Something called MAE, developed by some of my colleagues at FAIR, massed autoencoder.
So, it's basically like the, you know, 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 net to reconstruct.
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 label data, with textual descriptions of images, etc., 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.
Okay, 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 it 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 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 transformed 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 it's joint embedding, and there is this predictor that predicts the representation of the good guy 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 contrastive learning, where, and the idea of contrastive 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.
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 representations 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 we know are different.
You turn them only with images that are 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, we've argued about the G in AGI.
Yes.
I get it.
I get it.
Well, we'll probably continue to argue about it.
It's great.
You like, I mean, this, because you like French, and I mean, 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, right?
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 JAPA, 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 JAPA 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 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.
And so we can get away without doing the joint embedding, without, you know, 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 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 then 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 concept 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 talked 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.
You know, they don't have language.
And they understand the world much better than any LLM.
And 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 like, 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 IJEPA and BYOL certainly are.
And there's another one also called Dino or Dino, also produced 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 it 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 the 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, Dino and IJEPA, 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,
you know, crop the image.
And in some cases, in the case of IJEPA, you don't need to do any of this.
You just mask some parts of it, right?
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 IJEPA.
It 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 Dino, you need to know it's an image because you need to do things like,
you know, geometry transformation and blurring and things like that that are really image
specific.
A more recent version of this that we have is called VJEPA.
So it's basically the same idea as IJEPA, 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, you know, like a whole segment of each frame in the video over the entire video.
And that tube is like statically positioned throughout the frames?
It's literally a straight tube?
The tube 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 that'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 results that seem to indicate that the representation allows us,
allow our system to tell whether the video is physically possible or completely impossible
because some object disappeared or an object, you know, suddenly jumped from one location
to another or, or change shape or something.
So it's able to capture some physical, some physics-based constraints about the reality
represented in the video?
Yeah.
About the appearance and the disappearance of objects?
Yeah.
That's really new.
Oh, 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.
I mean, this is going to take a while before we get to that point, but there are systems
already, you know, 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'll 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 probably predict what's going
to happen.
So now what you have is an internal model that says, here is my idea of 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 plus one, t plus delta t, t plus
two 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.
Right?
So you can have a number of objectives, right?
If, you know, I can predict that if I have an object like this, right, and I open my hand,
it's going to fall, right?
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, you know, 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, right?
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, you know, 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, you know, rocket trajectories have been planned since computers have been
around, so since the early 60s, essentially.
So yes, for model predictive control, but you also often talk about hierarchical planning.
Yeah.
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 will 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 have a taxi, which you
can do in New York?
Okay.
Now I have another sub-goal, go down on the street.
What 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?
Like, you know, you can imagine going down, all the way down to basically what amounts
to millisecond by millisecond muscle control.
Okay.
And 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 of what's going to happen.
You know, how long it's going to take to catch a taxi or to go to the airport with traffic,
you know.
I mean, 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
re-planning 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 also 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.
So 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?
So, 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, right?
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 of abstraction do you think you'll start?
Because, like, I can imagine almost every single part of that an LLM would 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 so I can't say that an LLM 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, like, if you want to go down to, like, how do you, you know, climb down the stairs
or just stand up from your chair in terms of words, like, you can't do it.
You need, 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, the 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.
They're not, people, most people don't invent the, you know, plans, they, by themselves,
they, you know, we have some ability to do this, of course, obviously, but most plans
that people use are plans that have been trained on, like, they've seen other people use those
plans or they've been told how to do things, right?
Like, you can't invent how you, like, take a person who's never heard of airplanes and
tell them, like, how do you go from New York to Paris and probably not going to be able
to kind of, you know, 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 we need JEPA's form.
So, 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 auto-aggressive 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 auto-aggressive 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-aggressive LLMs, or that LLMs in general, not just the
auto-aggressive 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, you know, started, it didn't start with BERT, but it was really kind of
a good demonstration with this.
So, the idea that, you know, 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 where, you know, you, 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 have 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, you know,
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, like, you know,
with work from, you know, Google Meta, OpenAI, etc., you know, going back to, you know, the
GPT kind of work general pre-trained transformers.
You mean like GPT-2?
Like, 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, you know,
place it in the GPT timeline, there would be around GPT-2, yeah.
Well, I just, because you said it, you're 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?
Hans Moravec would say the Moravec paradox still applies.
Okay.
Okay.
Okay, we can pass through.
You don't think he would be really impressed?
No, of course, everybody would be impressed.
But, you know, it's not a question of being impressed or not.
It's a question of knowing what the limit of those systems can do.
Like, 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, you know, how we get there.
And, you know, and I'm not seeing this.
I'm seeing this from basically, you know, 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 piece of a set of inputs without training
the system for any particular task, right?
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, you know, 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, you know, what we used to call unsupervised learning and
sort of revive the idea of unsupervised learning in the early 2000s with Yosha Benjo and Jeff
Hinton.
Then discovered that supervised learning actually works pretty well if you can collect enough
data.
And so the whole idea of, you know, 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, you know, starting in 2014, basically
when we started FAIR and really pushing for, like, 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, you know, 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 that progress using
self-supervised learning for NLP, combining this with, you know, 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 labeled 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 just 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 kind of speech
units that are discrete.
But it's called text-less NLP.
We used to call it this way.
But yeah, so that, I mean, incredible success there.
And then, you know, 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.
And we tried many times.
We published lots of papers on it.
You know, they kind of sort of work, but not really great.
It 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 is 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 like, 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 jumper 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 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.
Like, 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, 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 the common experience of the low level, which LLMs currently do not
have.
You know, 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.
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 in 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 have, 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.
Yeah.
But that's not there.
And that's not there in LLMs.
LLMs are purely trained from text.
So then the other statement you made, 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 the conversations we have, 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 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 if I accidentally knock this over, you'll probably make fun of me.
And 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.
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 four-year-old and 10 to the 15 bytes,
you know, going through vision, just vision, right?
There is a similar bandwidth, you know, of touch and a little less through audio.
And then text doesn't, language doesn't come in until like, 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, don't really have any influence on the world.
They can only observe, right?
And you accumulate like 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 rest, you stay
within the set of correct answers 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?
Like, 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, you know, 80% or whatever of questions that people will ask
by, you know, 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, you know, the enormous set of prompts that you have not covered during
training.
And that set is enormous.
Like, 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 has 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 that'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, like, 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, like, you put a, essentially a random sequence
of characters in the prompt.
And that's enough to kind of throw the system into a mode where, you know, 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, you know, go outside of its conditioning,
right?
So that's a very clear demonstration of it.
But, of course, you know, that goes outside of what it's designed to do, right?
If you actually stitch together reasonably grammatical sentences, is that, is it that easy to break
it?
Yeah.
Some people have done things like, you write a sentence in English, right, that has an,
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, you know, 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 has been kind of a giant lookup table, right, essentially,
which is not really what you want.
You want systems that can reason, certainly that can plan.
So the type of reasoning that takes place in LLM is very, very primitive.
And the reason you can tell is 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 like, you know, it's the size of the prediction network, you know, with its 36 layers
or 92 layers or whatever it is, multiplied by the 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 flaw of LLMs?
Yes.
Or does it mean that...
There's more part to that question.
Now you're just behaving like an LLM, immediately answering.
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 this mental 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, right?
So system one is the type of tasks that you can accomplish without deliberately, consciously
thinking about how you do them.
You just do them.
You've done them enough that you can just do it subconsciously, right?
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.
Right?
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 where you play
against another experienced chess player, you think about all kinds of options, right?
You think about it for a while, right?
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 role model, that's system
two.
This is what LLMs currently cannot do.
So 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, you know, what used
to be called probabilistic models or graphical models and things of that type.
So basically the principle is like this.
You know, 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, you know, 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 constructed by the LLM.
Well, so really what you need to do would be to not search over possible strings of text
that minimize that energy.
But what you would do is do this in abstract representation space.
So in sort of the space of abstract thoughts, you would elaborate a thought, right?
Using this process of minimizing the output of your model, which is just a scalar, it's
an optimization process, right?
So now the way the system produces its answer is through optimization by, you know, minimizing
an objective function, basically, right?
And this is, we're talking about inference.
We're not talking about training, right?
The system has been trained already.
So now we have an abstract representation of the thought of the answer, representation
of the answer.
We feed that to, basically, an autorexative decoder, which can be very simple, that turns
this into a text that expresses this thought.
Okay?
So that, in my opinion, is the blueprint of future dialogue systems.
They will think about their answer, plan their answer by optimization before turning it
into text.
And that is Turing complete.
Can you explain exactly what the optimization problem there is?
Like, what's the objective function?
Just linger on it.
You kind of briefly described it.
But over what space are you optimizing?
The space of representations.
It goes abstract representation.
Abstract representation.
So you have an abstract representation inside the system.
You have a prompt.
The prompt goes to an encoder, produces a representation, perhaps goes to a predictor that predicts a
representation of the answer, of the proper answer.
But that representation may not be a good answer because there might be some complicated
reasoning you need to do, right?
So then you have another process that takes the representation of the answers and modifies
so as to minimize a cost function that measures to what extent the answer is a good answer for the question.
Now, we sort of ignore the fact for, I mean, the issue for a moment of how you train that system to measure whether an answer is a good answer for a question.
But suppose such a system could be created.
Right.
But what's the process, this kind of search-like process?
It's an optimization process.
You can do this if the entire system is differentiable.
That scalar output is the result of, you know, running through some neural net, running the answer, the representation of the answer through some neural net.
Then by gradient descent, by backpropagating gradients, you can figure out, like, how to modify the representation of the answer so as to minimize that.
So that's still a gradient-based...
It's gradient-based inference.
So now you have a representation of the answer in abstract space.
Now you can turn it into text.
Mm-hmm.
Right?
And the cool thing about this is that the representation now can be optimized through gradient descent,
but also is independent of the language in which you're going to express the answer.
Right.
So you're operating in the abstract representation.
I mean, this goes back to the joint embedding.
Right.
That is better to work in the space of, I don't know, to romanticize the notion, like, space of concepts versus the space of concrete sensory information.
Right.
Okay.
But can this do something like reasoning, which is what we're talking about?
Well, not really.
Only in a very simple way.
I mean, basically, you can think of those things as doing the kind of optimization I was talking about, except they optimize in the discrete space, which is the space of possible sequences of tokens.
And they do this optimization in a horribly inefficient way, which is generate a lot of hypotheses and then select the best ones.
Right.
And that's incredibly wasteful in terms of computation, because you basically have to run your LLM for, like, every possible, you know, generated sequence.
And it's incredibly wasteful.
So it's much better to do an optimization in continuous space where you can do gradient descent, as opposed to, like, generate tons of things and then select the best.
You just iteratively refine your answer to go towards the best.
Right.
That's much more efficient.
But you can only do this in continuous spaces with differentiable functions.
You're talking about the reasoning, like, ability to think deeply or to reason deeply.
How do you know what is an answer that's better or worse based on deep reasoning?
Right.
So then we're asking the question of, conceptually, how do you train an energy-based model, right?
So an energy-based model is a function with a scalar output, just a number.
You give it two inputs, X and Y.
And it tells you whether Y is compatible with X or not.
X you observe.
Let's say it's a prompt, an image, a video, whatever.
And Y is a proposal for an answer, a continuation of the video, you know, whatever.
And it tells you whether Y is compatible with X.
And the way it tells you that Y is compatible with X is that the output of that function
will be zero if Y is compatible with X.
It would be a positive number, non-zero, if Y is not compatible with X.
Okay.
How do you train a system like this at a completely general level?
Is you show it pairs of X and Ys that are compatible, a question and the corresponding answer.
And you train the parameters of the big neural net inside to produce zero.
Okay.
Now, that doesn't completely work because the system might decide, well, I'm just going to
say zero for everything.
So now you have to have a process to make sure that for a wrong Y, the energy would be larger
than zero.
And there you have two options.
One is contrastive method.
So contrastive method is you show an X and a bad Y.
And you tell the system, well, that's, you know, give a high energy to this, like push
up the energy, right?
Change the weights in the neural net that computes the energy so that it goes up.
So that's contrastive methods.
The problem with this is if the space of Y is large, the number of such contrastive samples
you're going to have to show is gigantic.
But people do this.
They do this when you train a system with RLHF, basically what you're training is what's
called a reward model, which is basically an objective function that tells you whether
an answer is good or bad.
And that's basically exactly what this is.
So we already do this to some extent.
We're just not using it for inference.
We're just using it for training.
There is another set of methods, which are non-contrastive, and I prefer those.
And those non-contrastive methods basically say, okay, the energy function needs to have
low energy on pairs of X, Ys that are compatible, that come from your training set.
How do you make sure that the energy is going to be higher everywhere else?
And the way you do this is by having a regularizer, a criterion, a term in your cost function that
basically minimizes the volume of space that can take low energy.
And the precise way to do this is all kinds of different specific ways to do this, depending
on the architecture.
But that's the basic principle.
So that if you push down the energy function for particular regions in the X, Y space, it
will automatically go up in other places because there's only a limited volume of space that
can take low energy, okay, by the construction of the system or by the regularizer, regularizing
function.
We've been talking very generally, but what is a good X and a good Y?
What is a good representation of X and Y?
Because we've been talking about language, and if you just take language directly, that
presumably is not good.
And so there has to be some kind of abstract representation of ideas.
Yeah.
So you, I mean, you can do this with language directly by just, you know, X is a text and
Y is a continuation of that text.
Yes.
Or X is a question, Y is the answer.
But you're saying that's not going to take, I mean, that's going to do what LLMs are doing.
Well, no, it depends on how you, how the internal structure of the system is built.
If the internal structure of the system is built in such a way that inside of the system,
there is a latent variable, let's call it Z, that you can manipulate so as to minimize
the output energy.
Then that Z can be viewed as a representation of a good answer that you can translate into
a Y that is a good answer.
So this kind of system could be trained in a very similar way.
Very similar way, but you have to have this way of preventing collapse, of ensuring that,
you know, there is high energy for things you don't train it on.
And currently it's very implicit in LLM.
It's done in a way that people don't realize is being done, but it is being done.
It's due to the fact that when you give a high probability to a word, automatically you give
low probability to other words because you only have a finite amount of probability to
go around right there, to sum to one.
So when you minimize the cross-entropy or whatever, when you train your LLM to produce the, to predict
the next word, you're increasing the probability your system will give to the correct word,
but you're also decreasing the probability it will give to the incorrect words.
Now, indirectly, that gives a low probability to, a high probability to sequences of words
that are good and low probability to sequences of words that are bad, but it's very indirect.
And it's not, it's not obvious why this actually works at all, but because you're not doing
it on a joint probability of all the symbols in a, in a sequence.
You're just doing it kind of, you sort of factorize that probability in terms of conditional
probabilities over successive tokens.
So how do you do this for visual data?
So we've been doing this with Ojepa architectures, basically.
The joint of betting.
Ojepa.
So there, the compatibility between two things is, you know, here's, here's an image or a
video, here's a corrupted, shifted or transformed version of that image or video or masked.
Okay.
Okay.
And then, uh, the energy of the system is the prediction error of the representation,
uh, the, the predicted representation of the good thing versus the actual representation
of the good thing.
Right.
So, so you run the corrupted image to the system, predict the representation of the, the good
input uncorrupted, and then compute the prediction error.
That's the energy of the system.
So this system will tell you, this is a good, you know, if this is a good image and this
is a corrupted version, it will give you zero energy.
If those two things are effectively, one of them is a corrupted version of the other,
give you a high energy.
If the, if the two images are completely different.
And hopefully that whole process gives you a really nice compressed representation of,
of, uh, reality, of visual reality.
And we know it does because then we use those for our presentations as input to a classification
system or something.
And it works.
And that classification system works really nicely.
Okay.
Well, so to summarize, you recommend in a, in a, in a spicy way that only Yelena Kuhn can,
you recommend that we abandon generative models in favor of joint embedding architectures.
Yes.
Abandon auto-aggressive generation.
Yes.
Abandon probable, this feels like court testimony, uh, abandon probabilistic models in favor
of energy-based models, as we talked about.
Abandon contrastive methods in favor of regularized methods.
And, uh, let me ask you about this.
You've been for a while a critic of reinforcement learning.
Yes.
So, uh, the last recommendation is that we abandon RL in favor of model predictive control,
as you were talking about, and only use RL when planning doesn't yield the predicted outcome.
And, uh, we use RL in that case to adjust the world model or the critic.
Yes.
So, uh, you mentioned, uh, RLHF, reinforcement learning with human feedback.
Uh, why do you still hate, uh, reinforcement learning?
I don't hate reinforcement learning.
And I think it should not be, uh, abandoned completely, but I think it's used to be minimized
because it's incredibly inefficient in terms of samples.
And so the, the proper way to train a system is to first have it learn, uh, good representations
of the world and world models from mostly observation, maybe a little bit of interactions.
And then steer based on that.
If the representation is good, then the adjustments should be minimal.
Yeah.
And now there's two things you can use.
If you've learned a world model, you can use the world model to plan a sequence of actions
to arrive at a particular objective.
You don't need RL unless the way you measure whether you succeed might be inexact.
Your idea of, you know, whether you were going to fall from your bike might be wrong or whether
the person you're fighting with MMA was going to do something and then do something else.
Um, so there, uh, so there's two ways you can be wrong.
Either your, your objective function does not reflect the actual objective function you
want to optimize, or your world model is inaccurate, right?
So you didn't, you, the prediction you were making about what was going to happen in the
world is inaccurate.
So if you want to adjust your world model while you are operating the world or your objective
function, that is basically in the realm of RL.
This is what RL deals with, uh, to some extent, right?
So adjust your world model.
And the way to adjust your world model, even in advance, uh, is to explore parts of the space
where your world model, where you know that your world model is inaccurate.
That's called curiosity, basically, or play, right?
When you play, you kind of explore parts of the space space that, um, you know, you don't
want to do in for real because it might be dangerous, but, uh, but you can adjust your
world model, uh, without killing yourself, basically.
Um, so that's what you want to use RL for when, when, when it comes time to learning a
particular task, you already have all the good representations, you already have your
world model, but you want, you need to adjust it for the situation at hand.
That's when you use RL.
What do you think RLHF works so well?
This enforcement learning with human feedback, why did it have such a transformational effect
on large language models?
So what, what's had the transformational effect is human feedback.
There is many ways to use it.
And some of it is just purely supervised.
Actually, it's not really reinforced by learning.
So it's the, it's the HF.
It's the HF.
Uh, and then there is various ways to use human feedback, right?
So you can, uh, you can ask humans to rate answers, uh, multiple answers that are produced
by a world model.
And, uh, and, and then what you do is you train an objective function to predict that
rating.
And then you can use that objective function to predict, you know, whether an answer is
good.
And you can back propagate gradient through this to fine tune your system so that it only
produces high, highly rated answers.
Okay.
So that's one way.
So that's like, in RL, that means, uh, training what's called a reward model, right?
Uh, so something that, you know, basically a small neural net that estimates to what extent
an answer is good, right?
It's very similar to the objective I was, I was talking about, or talking about earlier
for planning, except now it's not used for planning.
It's, it's used for fine tuning your system.
I think it would be much more efficient to use it for planning, but, um, but, but, uh,
currently it's used to, uh, fine tune the parameters of the system.
Now there, there's several ways to do this.
Um, you know, some of, some of them are supervised.
You just, you know, ask a human person, like, what is a good answer for this?
Right.
Then you just type the answer.
Um, uh, I mean, there's, there's lots of ways that those systems are being adjusted.
Now, a lot of people have been very critical of the recently released Google's Gemini 1.5
for essentially, in my words, I could say super woke.
Woke in the, uh, negative connotation of that word.
Uh, there's some almost hilariously absurd things that it does, like it modifies history,
uh, like generating images of a, um, black George Washington, or, um, perhaps more seriously,
something that you commented on Twitter, which is refusing to comment on or generate images
images of, um, or even descriptions of, uh, Tiananmen Square or the, uh, the Tank Man, one
of the most sort of legendary protest images in history.
Of course, these images are highly censored by the Chinese government, and therefore,
everybody started asking questions of what is the process of, uh, designing these LLMs,
what is, what is, what is, what is the role of censorship in these, and all that kind of
stuff.
So you, uh, commented on Twitter saying that open source is the answer.
Yeah.
Essentially.
So, um, can you explain?
I, I actually made that comment on just about every social network I can, and I've, I, I have,
uh, I've made that point multiple times in, in various, uh, forums.
Um, uh, here's my, my point of view on this.
Uh, people can complain that AI systems are biased, and they generally are biased by the
distribution of the training data that they've been trained on, um, that reflects biases in
society, um, and that is potentially offensive to some people or potentially not.
Uh, and, and some techniques to de-bias then become offensive to some people, um, because
of, you know, historical, uh, incorrectness and things like that.
Um, and so you can ask the question, you can ask two questions.
The first question is, is it possible to produce an AI system that is not biased?
And the answer is absolutely not.
And it's not because of technological, uh, challenges, although there are, uh, technological
challenges to that.
It's because bias is in the eye of the beholder.
Um, different people may have different ideas about what constitutes bias, um, you know, for
a lot of, uh, a lot of things.
I mean, there are facts that are, you know, indisputable, but there are a lot of opinions
or, or things that can be expressed in different ways.
Uh, and so, uh, you cannot have an unbiased system.
That's just an impossibility.
Um, and so what's the, what's the answer to this?
And the, the answer is the same answer that we found in liberal democracy about the press.
The press needs to be free and, uh, diverse.
We have free speech for a good reason is because, uh, we don't want all of our information to
be, uh, to come from a unique source, um, because that's opposite to the whole idea of democracy
and, uh, you know, progress of ideas and even science, right?
In, in science, people have to argue for different opinions and, and science makes progress when
people disagree and they come up with an answer and, you know, a consensus forms, right?
And it's true in all democracies around the world.
So there is a, uh, a future which is already happening where every single one of our interaction
with the digital world will be mediated by AI, AI systems, AI assistance, right?
We're going to have smart glasses.
You can already buy them from Meta, the Ray-Ban Meta, where, um, you know, you can talk to
them and they are connected with an LLM and you can get answers on any question you have,
or you can be looking at a monument and there is a camera in the, in the system that in,
in the glasses, you can ask it, like, what can you tell me about this, uh, building or
this monument?
You can be looking at a menu in a foreign language and the thing will translate it for
you, or you can, we can do real-time translation if we speak different languages.
So a lot of our interactions with the digital world are going to be mediated by those systems
in the near future.
Um, you know, increasingly the, uh,
search engines that we're going to use are not going to be search engines.
They're going to be, uh, dialogue systems that will just ask a question and it will
answer and then point you to perhaps, uh, appropriate reference for it.
But here is the thing.
We cannot afford those systems to come from a handful of companies on the West Coast of
the U S because those systems will constitute the repository of all human knowledge.
And we cannot have that be controlled by a small number of people, right?
It has to be diverse for the same reason, the process has to be diverse.
So how do we get a diverse set of AI assistants?
Um, it's very expensive and difficult to train a base model, right?
A base LLM at the moment, you know, in the future, it might be something different, but
at the moment that's an LLM, uh, so only a few companies can do this properly.
And if some of those top systems are open source, anybody can use them.
Anybody can fine tune them.
Um, if we put in place some systems that allows any group of people, whether they are, um, individual
citizens, groups of citizens, government organizations, NGOs, um, companies, whatever, to take those
open source, uh, systems, AI systems, and fine tune them for their own purpose on their
own data, they were going to have a very large diversity of, uh, different AI systems that
are specialized for all of those things.
Right?
So I'll tell you, I talked to the French government quite a bit and the French government
will not accept that the digital diet of all their citizens be controlled by three companies
on the West Coast of the U.S.
That's just not acceptable.
It's a danger to democracy, regardless of how well-intentioned those companies are.
Right?
Um, and so, uh, and it's also a danger to local culture, to values, to language.
Right?
I was talking with, um, uh, the, uh, founder of Infosys in India.
Um, he's funding a project to fine tune Lama2, the open source model produced by, by Meta,
so that Lama2 speaks all 22 official languages in India.
Um, it's very important for people in India.
I was talking to a former colleague of mine, Mustafa Sisse, who used to be a scientist at
FAIR, and then moved back to Africa.
I created a research lab for Google in Africa, and now is, has a new startup called Kera.
And what he's trying to do is basically have LLM that speaks the local languages in Senegal
so that people can have access to medical information, because they don't have access to doctors.
It's a very small number of doctors per capita in the, in Senegal.
Um, I mean, you can't have any of this unless you have open source platforms.
So with open source platforms, you can have AI systems that are not only diverse in terms of
political opinions or things of that type, but in terms of, uh, uh, language, culture,
value systems, political opinions, um, technical abilities in various domains.
And you can have an industry, an ecosystem of companies that fine tune those open source
systems for vertical applications in industry, right?
You, you have, I don't know, a publisher has thousands of books and they want to build a
system that allows a customer to just ask a question about any, about the content of any
of their books.
You need to train on their proprietary data, right?
Um, you have a company, we have one within Meta, it's called MetaMate, and it's basically
an LLM that can answer any question about internal, uh, stuff about, about the company.
Um, very useful.
A lot of companies want this, right?
A lot of companies want this, not just for their employees, but also for their customers
to take care of their customers.
So the only way you're going to have an AI industry, the only way you're going to have
AI systems that are not uniquely biased is if you have open source platforms on top of
which, uh, any group can, uh, build specialized systems.
So the, the direction of, of inevitable direction of history is that the vast majority of AI systems
will be built on top of open source platforms.
So that's a beautiful vision.
So meaning like a company like Meta or Google or so on should take only minimal fine tuning
steps after the building the foundation pre-trained model, as few steps as possible.
Basically.
Can Meta afford to do that?
No.
So I don't know if you, you know this, but companies are supposed to make money somehow
and, uh, open source is, is, is like giving away, I don't know, Mark made a video, Mark
Zuckerberg, uh, very sexy video talking about 350,000 NVIDIA H100s.
The, the, the, the math of that is just for the GPUs, that's a hundred billion, um, plus the
infrastructure for training, everything.
So I'm no business guy, but how do you make money on that?
So the division you paint is a really powerful one, but how is it possible to make money?
Okay.
So you have several business models, right?
The business model that, uh, Meta is built around is, um, you offer a service and the,
the financing of that service is, uh, either through ads or through business customers.
So for example, if you have an LLM that, uh, you know, can help a mom and pop pizza place,
um, by, you know, talking to the customers through WhatsApp.
And so the customers can just order a pizza and the system will just, you know, ask them
like, what topping do you want or what size, blah, blah, blah.
Um, the business will pay for that.
Okay.
That's a model.
Um, and otherwise, you know, if it's a system that is on the more kind of classical services,
it can be, uh, ad supported or, you know, there's several models.
But the point is, uh, if you have a big enough, uh, uh, potential customer base and you need
to build that system anyway for them, it doesn't hurt you to actually distribute it in open source.
Again, I'm no business guy, but if you release the open source model, then other people can
do the same kind of task and compete on it.
Basically provide fine-tuned models for businesses.
Sure.
Is the bet that Meta is making, by the way, I'm a huge fan of all this, but is the bet
that Meta is making, it's like, we'll do a better job of it.
Well, no, the, the bet is, is more, we have, we already have a huge, uh, user base and customer
base.
Ah, right.
Right.
So it's going to be useful to them.
Whatever we offer them is going to be useful.
And there is a way to derive revenue from this.
Uh, and it doesn't hurt that, you know, we provide that system or the base, the base model,
right?
The foundation model, uh, in open source for others to build applications on top of it
too.
So if those applications turn out to be useful for our customers, we can just buy it from
them.
Um, uh, it could be that they will improve the platform.
In fact, we see this already.
Um, I mean, there is, you know, literally millions of downloads of, uh, Lama 2 and thousands
of people who have, you know, provided ideas about how to make it better.
Um, so, you know, this, this, this clearly accelerates progress to make the system available
to, uh, uh, uh, sort of a, a wide, uh, community of people.
And, and there is literally thousands of businesses who are building applications with it.
So, um, so our ability to, Meta's ability to derive revenue from this technology is not
impaired, uh, by the distribution of it, of base models in open source.
The fundamental criticism that Gemini is getting is that, as you pointed out on the West
Coast, just to, just to clarify, we're currently in the East Coast, where I would suppose Meta
AI headquarters would be.
So there are strong words about the West Coast, but, uh, I guess the issue that happens is,
I think it's fair to say that most tech people have, uh, a political affiliation with the
left wing.
They're, they lean left.
And so the problem that people are criticizing Gemini with is that there's, in that debiasing
process that you mentioned, that their ideological lean becomes obvious.
Uh, is this something that could be escaped?
You're saying open source is the only way.
Have, have you witnessed this kind of ideological lean that makes engineering difficult?
No, I don't think it has to do, I don't think the issue has to do with the political
leaning of the people designing those systems.
It has to do with the, uh, acceptability or political leanings of the, their customer
base or audience, right?
So a big company cannot afford to offend too many people.
So they're going to make sure that whatever product they put out is safe, whatever that
means.
And, and it's very possible to overdo it.
And it's also very possible to, it's impossible to do it properly for everyone.
You're not going to satisfy everyone.
So that's what I said before.
You cannot have a system that is unbiased, that is perceived as unbiased by everyone.
It's going to be, you know, you, you push it in one way, one set of people are going
to see it as biased and then you push it the other way.
And another set of people is going to see it as biased.
And then in addition to this, there's the issue of, if you push the system, perhaps a little
too far in one direction, it's going to be non-factual, right?
You're going to have, you know, uh, you know, black Nazi, uh, soldiers in, uh, yeah.
So we should, we should mention image generation of, of, uh, black Nazi soldiers, which is not
factually accurate.
Right.
And can be offensive for some people as well.
Right.
So, uh, uh, so, you know, it's going to be impossible to kind of produce systems that
are unbiased for everyone.
So the only solution that I see is diversity and diversity in the full meaning of that word
diversity of in every possible way.
Yeah.
Uh, Mark Andreessen just tweeted today.
Let me do a TLDR.
The conclusion is only startups and open source can avoid the issue that he's highlighting with
big tech.
He's asking, can big tech actually field generative AI products?
One, ever escalating demands from internal activists, employee mobs, crazed executives,
broken boards, pressure groups, extremist regulators, government agencies, the press in quotes, experts
and everything, uh, corrupting the output to constant risk of generating a bad answer or
drawing a bad picture or rendering a bad video.
Who knows what is going to say or do at any moment?
Three, legal exposure, product liability, slander, election law, many other things and so on.
Anything that makes Congress mad for continuous attempts to tighten grip, unacceptable output,
degrade the model, like how good it actually is, uh, in terms of usable and pleasant to use
and effective and all that kind of stuff.
And five, publicity of bad text, images, video, actual puts those examples into the training
data for the next version and so on.
So he just highlights how difficult this is from all kinds of people being unhappy.
He said you can't create a system that makes everybody happy.
Yes.
Uh, so if you're going to do the fine tuning yourself and keep a closed source, essentially
the problem there is then trying to minimize the number of people who are going to be unhappy.
Yeah.
Um, and you're saying like the only, that, that almost impossible to do, right?
And that's the better ways to do open source.
Basically.
Yeah.
I mean, he's, Mark is right about, uh, a number of things that he lists that, uh, indeed
scare, um, large companies, uh, you know, certainly congressional investigations is one
of them, legal liability, uh, you know, uh, making things that, uh, get people to, you know,
hurt themselves or hurt others.
Like, you know, um, big companies are really careful about not, um, producing things of this
type.
And, um, uh, because they have, you know, they don't want to hurt anyone, first of all,
and then second, they want to preserve their business.
So, um, it's essentially impossible for systems like this that can inevitably formulate political
opinions and, you know, opinions about various things that may be political or not, but that
people may disagree about, about, you know, moral issues and, you know, um, things about
like questions about religion and things like that, right?
Or, or cultural issues that people from different communities would disagree with in the first
place.
Um, so there's only kind of a relatively small number of things that people will, uh, sort
of agree on, you know, basic principles, but beyond that, if you, if you want those systems
to be useful, they will necessarily have to, uh, offend a number of people inevitably.
And so open source is just better.
And then diversity is better, right?
And open source enables diversity.
That's right.
Open source enables diversity.
That's, that's going to be a fascinating world where if it's true that the open source
world, if meta leads the way and creates this kind of open source foundation model world,
there's going to be like governments will have a find new model and, and, and then potentially,
uh, uh, you know, people that vote left and right will have their own model and preference
to be able to choose.
And it will potentially divide us even more, but that's on us humans.
We get to figure out basically the technology enables humans to human more effectively and
all the difficult ethical questions that humans raise will just, it'll, um, leave it up to
us to figure it out.
Yeah.
I mean, there are some limits to what, you know, the same way there are limits to free
speech.
There has to be some limit to the kind of stuff that those systems might, uh, be authorized
to, um, to produce, um, you know, some guardrails.
So, I mean, that's one thing I've been interested in, which is, uh, in the type of architecture
that we were discussing before, where the output of a system is a result of an inference
to satisfy an objective.
That objective can include guardrails and, uh, we can put guardrails in open source systems.
I mean, if we eventually have systems that are built with this blueprint.
Uh, we can put guardrails, uh, in those systems that guarantee that there is sort of a minimum
set of guardrails that make the system non-dangerous and non-toxic, et cetera.
You know, basic things that everybody would agree on.
Um, and, and then, you know, the, the fine tuning that people will add or the additional
guardrails that people will add will kind of cater to their, um, community, whatever it
is.
And, and, yeah, the fine tuning will be more about the gray areas of what is hate speech,
what is dangerous, and all that kind of stuff.
I mean, you've...
Or different value systems.
Still value systems.
I mean, like, uh, but still, even with the objectives of how to build a bioweapon, for
example, I think something you've commented on, or at least there's a paper where a collection
of researchers is trying to understand the social impacts of these LLMs.
And I guess one threshold is nice.
It's like, does the LLM make it any easier than a, than a search would, like a Google search
would?
Right.
So the increasing, uh, number of studies on this seems to point to the fact that it
doesn't help.
So having an LLM doesn't help you, uh, design or build a bioweapon or a chemical weapon if
you already have access to, uh, you know, a search engine and a library.
Uh, and, and so the, the sort of increased information you get or the ease with which
you get it doesn't really help you.
Um, that's the first thing.
The second thing is it's one thing to have a list of instructions of how to make a, a
chemical weapon, for example, or bioweapon.
Um, it's another thing to actually build it and it's much harder than you might think
and then LLM will not help you with that.
Um, in fact, you know, nobody in the world, not even like, you know, countries use bioweapons
because most of the time they have no idea how to protect their own populations against
it.
So, um, so it's too dangerous actually to kind of ever use.
Um, and it's in fact banned by, uh, uh, international treaties.
Um, chemical weapons is different.
It's also banned by treaties, uh, but, um, uh, but it's the same problem.
It's difficult to use in situations that doesn't turn against the perpetrators, but we could
ask Elon Musk, like I can, I can give you a very precise list of instructions of how you
build a rocket engine.
And even if you have a team of 50 engineers that are really experienced building it, you're
still going to have to blow up a dozen of them before you get one that works.
Um, and you know, it's the same with, uh, you know, the chemical weapons or bioweapons
or things like this.
So it requires expertise, you know, in the, in the real world that another line is not
going to help you with.
And it requires even the common sense expertise that we've been talking about, which is how
to take, uh, language-based instructions and materialize them in the physical world requires
a lot of knowledge that's not in the instructions.
Yeah, exactly.
A lot of biologists have posted on this actually in response to those things saying like, do
you realize how hard it is to actually do the lab work?
Okay, I can know this is not trivial.
Yeah.
And that's Hans Marvick comes, comes to light once again.
Uh, just to linger on LLAMA, you know, Mark announced that LLAMA 3 is coming out eventually.
I don't think there's a release date, but what, what are you most excited about?
First of all, LLAMA 2 that's already out there and maybe the future, LLAMA 3, 4, 5, 6,
10, just, uh, the future of the open source under Meta.
Well, a number of things.
So, uh, there's going to be like various versions of, of LLAMA that are, uh, you know,
improvements of previous LLAMAs, bigger, better, multimodal, things like that.
And then in future generations, systems that are capable of planning, that really understand
how the world works, uh, maybe are trained from video.
So they have some world model, maybe, you know, capable of the type of reasoning and planning
I was talking about earlier.
Like how long is that going to take?
Like when is the research that is doing, going in that direction, going to sort of feed into
the product line, if you want, of LLAMA?
I don't know.
I can't tell you.
And there is, you know, a few breakthroughs that we have to basically, uh, go through before
we can get there.
But you'll be able to monitor our progress because we publish our research, right?
So, you know, if last week we published the VJEPA work, which is sort of a first step towards
training systems for video.
Um, and then the next step is going to be world models based on kind of this type of idea,
training, training from video.
Uh, there's similar work at, uh, at DeepMind also and, um, uh, taking place people and also
at UC Berkeley on, uh, world models from video.
A lot of people are working on this.
I think a lot of good ideas are coming, uh, are appearing.
My bet is that those systems are going to be JEPA-like.
They're not going to be generative models.
Um, and, uh, we'll see what the future will tell.
Um, there's really good work at, uh, um, a gentleman called Danny Jarhoffner, who is
not DeepMind, who, who's worked on kind of models of this type that learn representations
and then use them for planning or learning, uh, tasks by reinforcement training.
Um, and a lot of work at Berkeley by, um, uh, Peter Abiel, Samuel Levine, a bunch of other
people of that type, uh, um, I'm collaborating with actually in the context of some, uh, grants,
uh, with my NYU hat, um, and then collaborations also through Meta, uh, because the, the lab
at Berkeley is associated with Meta in some way.
So with FAIR.
So I, I think, uh, it's very exciting.
You know, I, I think I'm super excited about, uh, I haven't been that excited about like
the direction of machine learning and AI, you know, since, uh, you know, 10 years ago when
FAIR was started and before that, um, 30 years ago, we were working on 35 on, on convolutional
nets and, and, and the early days of neural nets.
So, um, um, I'm super excited because I see a path towards potentially human level intelligence,
uh, with, you know, systems that can, uh, understand the world, remember, plan, reason.
Um, there, there is some, some set of ideas to make progress there that might have a chance
of working.
Um, and I'm really excited about this.
What I like is that, you know, it, uh, uh, somewhat we, we get onto like a good direction
and perhaps succeed before my, uh, brain turns to a white sauce or, or before I need to retire.
Yeah.
Yeah.
Uh, you're also excited by, are you, is it beautiful to you just the amount of GPUs involved?
So the, the, the, the whole training process on this much compute is just zooming out, just
looking at earth and humans together have built these computing devices and are able to
train this one brain.
Then, then we then open source, like giving birth to this open source brain trained on this
gigantic compute system.
And there's just the details of how to train on that, how to build the infrastructure and
the, the hardware, the cooling, all of this kind of stuff.
Uh, are you just still, the most of your excitement is in the, the theory aspect of it.
The, uh, meaning like the software.
Well, I used to be a hardware guy many years ago.
Yes.
Yes.
That's right.
Decades ago.
Hardware has improved a little bit.
Changed a little bit.
Yeah.
I mean, certainly scale is necessary, but not sufficient.
Absolutely.
So we certainly need computation.
I mean, we're, we're still far in terms of compute power, uh, from, you know, what we
would need to match the compute power of the human brain.
Um, you know, this may occur in the next couple of decades, but, um, but we're still
some ways away and certainly in terms of power efficiency, we're really far.
Um, so there's a lot of progress to make in, uh, in, in, in hardware.
And, uh, you know, right now, a lot of progress is, is, is, is not, I mean, there's a bit coming
from Silicon technology, but a lot of it coming from architectural innovation and quite a bit
coming from, uh, uh, like more efficient ways of, you know, implementing the architectures
that have become popular, basically combination of transformers and convenets.
And, uh, so, uh, you know, there's still some ways to go until, uh, we're going to saturate.
We're going to have to come up with like new, new principles, new fabrication technology,
new, uh, basic components, um, perhaps, you know, based on sort of different principles
than those classical digital CMOS.
Interesting.
So you think in order to build AMI, I mean, we need, we potentially might need some hardware
innovation too.
Well, if we want to make it, um, ubiquitous, yeah, certainly, because we're going to have
to reduce the, you know, compute, the power consumption, a GPU today, right, is half a
kilowatt to a kilowatt.
Human brain is about 25 watts.
Uh, and the GPU is way below the power of human brain.
You need, you know, something like a hundred thousand or a million to match it.
So, uh, so, you know, we are off by a huge factor.
You often say that AGI is not coming soon, meaning like not.
Not this year, not the next few years, potentially farther away.
What's your basic intuition behind that?
So, first of all, it's not going to be an event, right?
The idea somehow, which, you know, is popularized by science fiction and Hollywood that, you know,
somehow somebody is going to discover the secret, the secret to AGI or human level AI or AMI,
whatever you want to call it.
And then, you know, turn on a machine and then we have AGI.
That's just not going to happen.
It's not going to be an event.
It's going to be gradual progress.
Are we going to have systems that can learn from video how the world works and learn good representations?
Yeah.
Before we get them to the scale and performance that we observe in humans, it's going to take quite a while.
It's not going to happen in one day.
Are we going to get systems that can have large amount of associative memory so they can remember stuff?
Yeah, but same.
It's not going to happen tomorrow.
I mean, there is some basic techniques that need to be developed.
We have a lot of them.
But, like, you know, to get this to work together with a full system is another story.
Are we going to have systems that can reason and plan, perhaps along the lines of objective-driven AI architectures that I described before?
Yeah.
But, like, before we get this to work properly, it's going to take a while.
So, and before we get all those things to work together, and then on top of this have systems that can learn, like, hierarchical planning, hierarchical representations, systems that can be configured for a lot of different situations at hands the way the human brain can.
You know, all of this is going to take, you know, at least a decade and probably much more because there are a lot of problems that we're not seeing right now that we have not encountered.
And so we don't know if there is an easy solution within this framework.
So, you know, it's not just around the corner.
I mean, I've been hearing people for the last 12, 15 years claiming that, you know, AGI is just around the corner and being systematically wrong.
And I knew they were wrong when they were saying it.
I called it bullshit.
Why do you think people have been calling, first of all, I mean, from the beginning, from the birth of the term artificial intelligence,
there has been an eternal optimism that's perhaps unlike other technologies?
Is it a Marovac paradox?
Is it the explanation for why people are so optimistic about AGI?
I don't think it's just Marovac's paradox.
Marovac's paradox is a consequence of realizing that the world is not as easy as we think.
So, first of all, intelligence is not a linear thing that you can measure with a scalar, with a single number.
You know, can you say that humans are smarter than orangutans?
In some ways, yes.
But in some ways, orangutans are smarter than humans in a lot of domains.
That allows them to survive in the forest, for example.
So IQ is a very limited measure of intelligence.
Do you know intelligence is bigger than what IQ, for example, measures?
Well, IQ can measure, you know, approximately something for humans, because humans kind of, you know, come in relatively kind of uniform form, right?
But it only measures one type of ability that, you know, may be relevant for some tasks, but not others.
And, but then if you are talking about other intelligent entities for which the, you know, the basic things that are easy to them is very different, then it doesn't mean anything.
So intelligence is a collection of skills and an ability to acquire new skills efficiently, right?
And the collection of skills that an intelligent, particular intelligent entity possesses or is capable of learning quickly is different from the collection of skills of another one.
And because it's a multidimensional thing, the set of skills is a high-dimensional space, you can't measure, you can compare, you cannot compare two things as to whether one is more intelligent than the other.
It's multidimensional.
So you push back against what are called AI doomers a lot.
Can you explain their perspective and why you think they're wrong?
Okay, so AI doomers imagine all kinds of catastrophe scenarios of how AI could escape or control and basically kill us all.
And that relies on a whole bunch of assumptions that are mostly false.
So the first assumption is that the emergence of superintelligence could be an event.
That at some point we're going to have, we're going to figure out the secret and we'll turn on a machine that is superintelligent.
And because we've never done it before, it's going to take over the world and kill us all.
That is false.
It's not going to be an event.
We're going to have systems that are like as smart as a cat, have all the characteristics of, you know, human level intelligence.
But their level of intelligence would be like a cat or a parrot, maybe, or something.
And then we're going to walk our way up to kind of make those things more intelligent.
And as we make them more intelligent, we're also going to put some guardrails in them and learn how to kind of put some guardrails so they behave properly.
And we're not going to do this with just one.
It's not going to be one effort, but it's going to be lots of different people doing this.
And some of them are going to succeed at making intelligent systems that are controllable and safe and have the right guardrails.
And if some other goes rogue, then we can use the good ones to go against the rogue ones.
So it's going to be my, you know, smart AI police against your rogue AI.
So it's not going to be like, you know, we're going to be exposed to like a single rogue AI that's going to kill us all.
That's just not happening.
Now, there is another fallacy, which is the fact that because the system is intelligent, it necessarily wants to take over.
And there is several arguments that make people scared of this, which I think are completely false as well.
So one of them is, you know, in nature, it seems to be that the more intelligent species are the one that end up dominating the other and even, you know, extinguishing the others, sometimes by design, sometimes just by mistake.
And so, you know, there is sort of thinking by which you say, well, if AI systems are more intelligent than us, surely they're going to eliminate us, if not by design, simply because they don't care about us.
And that's just preposterous for a number of reasons.
First reason is they're not going to be a species.
They're not going to be a species that competes with us.
They're not going to have the desire to dominate because the desire to dominate is something that has to be hardwired into an intelligent system.
It is hardwired in humans.
It is hardwired in baboons, in chimpanzees, in wolves, not in orangutans.
The species in which this desire to dominate or submit or attain status in other ways is specific to social species.
Non-social species like orangutans don't have it, right?
And they are as smart as we are, almost, right?
And to you, there's not significant incentive for humans to encode that into the AI systems.
And to the degree they do, there will be other AIs that sort of punish them for it.
I'll compete them over it.
Well, there's all kinds of incentive to make AI systems submissive to humans, right?
I mean, this is the way we're going to build them, right?
And so then people say, oh, but look at LLMs.
LLMs are not controllable.
And they're right.
LLMs are not controllable.
But objective-driven AI, so systems that derive their answers by optimization of an objective,
means they have to optimize this objective.
And that objective can include guardrails.
One guardrail is obey humans.
Another guardrail is don't obey humans if it's hurting other humans.
I've heard that before somewhere.
I don't remember.
Yes.
Maybe in a book.
Yeah.
But speaking of that book, could there be unintended consequences also from all of this?
No, of course.
So this is not a simple problem, right?
I mean, designing those guardrails so that the system behaves properly is not going to
be a simple issue for which there is a silver bullet, for which you have a mathematical proof
that the system can be safe.
It's going to be a very progressive, iterative design system where we put those guardrails
in such a way that the system behaves properly.
And sometimes they're going to do something that was unexpected because the guardrail wasn't
right, and we're going to correct them so that they do it right.
The idea somehow that we can't get it slightly wrong because if we get it slightly wrong,
we'll die, is ridiculous.
We're just going to go progressively.
And it's just going to be the analogy I've used many times is turbojet design.
How did we figure out how to make turbojets so unbelievably reliable, right?
I mean, those are incredibly complex pieces of hardware that run at really high temperatures
for 20 hours at a time sometimes.
And we can, you know, fly halfway around the world on a two-engine jetliner at near the
speed of sound.
Like, how incredible is this?
It's just unbelievable, right?
And did we do this because we invented, like, a general principle of how to make turbojets
safe?
No.
It took decades to kind of fine-tune the design of those systems so that they were safe.
Is there a separate group within General Electric or Snekma or whatever that is specialized
in turbojet safety?
No.
The design is all about safety because a better turbojet is also a safer turbojet.
So a more reliable one.
It's the same for AI.
Like, do you need, you know, specific provisions to make AI safe?
No, you need to make better AI systems and they will be safe because they are designed
to be more useful and more controllable.
So let's imagine a system, AI system, that's able to be incredibly convincing and can convince
you of anything.
I can at least imagine such a system.
And I can see such a system be weapon-like because it can control people's minds.
We're pretty gullible.
We want to believe a thing.
You can have an AI system that controls it.
And you could see governments using that as a weapon.
So do you think if you imagine such a system, there's any parallel to something like nuclear
weapons?
No.
So why is that technology different?
So you're saying there's going to be gradual development.
Yeah.
There's going to be, I mean, it might be rapid, but there'll be iterative.
And then we'll be able to kind of respond and so on.
So that AI system designed by Vladimir Putin or whatever, or his minions, you know, is going
to be trying to talk to every American to convince them to vote for, you know, whoever pleases Putin
or whatever, or, you know, or rile people up against each other, as they've been trying
to do, they're not going to be talking to you.
They're going to be talking to your AI assistant, which is going to be as smart as theirs, right?
That AI, because as I said, in the future, every single one of your interaction with the
digital world will be mediated by your AI assistant.
So the first thing you're going to ask is, is this a scam?
Like, is this thing like telling me the truth?
Like, it's not even going to be able to get to you because it's only going to talk to your
AI assistant.
Your AI assistant is not even going to, it's going to be like a spam filter, right?
You're not even seeing the email, the spam email, right?
It's automatically put in a folder that you never see.
It's going to be the same thing.
That AI system that tries to convince you of something is going to be talking to your
assistant, which is going to be at least as smart as it.
And it's going to say, this is spam, you know, it's not even going to bring it to your attention.
So to you, it's very difficult for any one AI system to take such a big leap ahead to
where you can convince even the other AI systems.
So like, there's always going to be this kind of race where nobody's way ahead.
That's the history of the world.
History of the world is, you know, whenever there is a progress someplace, there is a countermeasure.
And, and, you know, it's a, it's a cat and mouse game.
This is why mostly, yes, but this is why nuclear weapons are so interesting because that was such
a powerful weapon that it matters who got it first.
That, you know, you could imagine Hitler, Stalin, Mao getting the weapon first and that,
that having a different kind of impact on the world and the United States getting the weapon
first.
Yeah.
To you, nuclear weapons is like, you, you don't imagine a breakthrough discovery and then
Manhattan project like effort for AI.
No.
As I said, it's not going to be an event.
It's going to be, you know, continuous progress.
And, and whenever, you know, one breakthrough occurs, it's going to be widely disseminated
really quickly.
Yeah.
Probably first within industry.
I mean, this is not a domain where, you know, government or military organizations are
particularly innovative and they're in fact way behind.
And so this is going to come from industry and, and this kind of information disseminates
extremely quickly.
We've seen this over the last few years, right?
Where you have a new, like, you know, even take AlphaGo, this was reproduced within three
months, even without like particularly detailed information, right?
Yeah.
This is an industry that's not good at secrecy.
No, but even, even if there is, just the fact that you know that something is possible
makes you like realize that it's worth investing the time to actually do it.
You, you may be the second person to do it, but you know, you'll, you'll do it.
Uh, and, you know, same for, you know, all the innovations, uh, you know, self-supervised
learning transformers, decoder only architectures, LLMs.
I mean, those things, you don't need to know exactly the details of how they work to know
that, you know, it's possible, uh, because it's deployed and then it's getting reproduced.
And then, you know, people who work for those companies move, they go from one company to
another and, you know, the information disseminates.
What makes the success of the, the U S tech industry and Silicon Valley in particular is
exactly that is because the information circulates really, really quickly and this, you know,
disseminates, uh, very quickly.
And so, you know, the, the whole region sort of is ahead because of that circulation of information.
So maybe I just to linger on the psychology of AI doomers, you give, uh, in the classic
Yann LeCun way, a pretty good example of just when a new technology comes to be, you say, uh, engineer says, I invented this new thing.
I call it a ball pen.
And then the Twitter sphere responds, OMG, people could write horrible things with it.
Like misinformation, propaganda, hate speech, ban it now.
Then writing doomers come in akin to the AI doomers.
Imagine if everyone can get a ball pen, this could destroy society.
There should be a law against using ball pen to write hate speech, regulate ball pens now.
And then the pencil industry mogul says, yeah, ball pens are very dangerous.
Unlike pencil writing, which is erasable, ball pen writing stays forever.
Government should require a license for a pen manufacturer.
I mean, this does seem to be part of, um, human psychology when, when it comes up against new technology.
So what, what deep insights can you speak to about this?
Well, there is a, a natural fear of, uh, new technology and the impact it can have on society.
And people have kind of instinctive reaction to, um, you know, the world they know being threatened by major transformations, um, that are either cultural phenomena or technological, um, revolutions.
And they fear for their culture, they fear for their, they fear for their, you know, the future of their children, um, and, uh, their way of life, right?
So, so any change, um, is feared.
And, and you see this, you know, along history, like any technological revolution or cultural phenomenon was always accompanied by, uh, you know, groups or reaction in the media.
Uh, that, that, that basically attributed the, all the problems, the current problems of society to that particular change, right?
Electricity was going to kill everyone at some point, you know, you, uh, the train was going to be a horrible thing because, you know, you can't breathe past 50 kilometers an hour.
Um, and so there's a wonderful website called a pessimist archive, uh, which has all those newspaper clips of all the horrible things people imagine would, would arrive because of, uh, either, uh, technological, uh, innovation or, uh, a cultural phenomenon.
Um, you know, um, you know, the, is this wonderful examples of, uh, uh, uh, you know, jazz or comic books, uh, being blamed for, uh, unemployment or, or, you know, young people not wanting to work anymore and things like that.
Right.
And, and that has existed for, for centuries.
Um, and it's, you know, knee jerk reactions.
Um, the question is, you know, do we embrace change, uh, or do we resist it?
And what are the real dangers as opposed to the imagined, uh, imagined ones?
So people worry about, I think one thing they worry about with big tech, something we've been talking about over and over, but I think worth mentioning again, they worry about how powerful AI will be.
And they worry about it being in the hands of one centralized power of just a handful of central control.
And so that's the skepticism with big tech.
You can make, these companies can make a huge amount of money and control this technology.
And by so doing, you know, take advantage, uh, abuse the little guy in society.
Well, that's exactly why we need open source platforms.
Yeah.
I just wanted to nail the point home more and more.
Yes.
Um, so let me ask you on your, like I said, you do get a little bit, uh, um, you know, flavorful on the internet.
Uh, Yosha Bach tweeted something that you LOL that, uh, in reference to how 9,000 quote, I appreciate your argument and I fully understand your frustration, but whether the pod bay doors should be opened or closed is a complex and nuanced issue.
So you're at the head of meta AI, um, you know, this is something that really worries me that AI are AI overlords will speak down to us with corporate speak, um, of this nature.
And you sort of resist that with your way of being, um, is this something you can just comment on sort of working at a big company, how you can avoid the overfearing, I suppose, the, um, through caution, create harm.
Yeah.
Again, I think the answer to this is open source platforms and then enabling a widely diverse set of people to build AI assistants that represent the diversity of, uh, cultures, opinions, languages, and value systems across the world.
Um, so that you're not bound to just, uh, you know, be, uh, brainwashed by a particular way of thinking because of, uh, single AI entity.
Um, so, I mean, I, I, I, I think it's really, really important question for society and the problem I'm seeing is, um, is that, um, which is why I've been so vocal and sometimes a little sardonic about it.
Never stop.
Never stop, Jan.
We love it.
It's because I see the danger of this concentration of power through, through proprietary AI systems as a much bigger danger than everything else.
That if we really want, you know, uh, uh, diversity of opinion, uh, AI systems that, you know, in, in the future that where we'll all be interacting through AI systems, we need those to be diverse for the preservation of, uh, uh, diversity of ideas and, you know, creeds and political opinions and, and, and whatever.
Uh, and the preservation of democracy and 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, of everybody, um, because it could be used by terrorists or something.
Um, um, that would lead to, uh, you know, potentially, uh, uh, uh, a very bad future in which all of our information diet is controlled by a small number of, uh, uh, uh, companies who proprietary systems.
Do you trust humans with this technology to, uh, to build systems that are on the whole good for humanity?
Isn't that what democracy and free speech is all about?
I think so.
Do you trust institutions to do the right thing?
Do you trust people to do the right thing?
And, and yeah, there's bad people who are going to do bad things, but they're not going to have superior technology to the good people.
So then it's going to be my good AI against your bad AI, right?
I mean, it's the examples that we were just talking about of, you know, maybe, uh, some rogue country will build, you know, some AI system that's going to try to convince everybody to go into a civil war or something or, or, or elect, uh, favorable, uh, ruler.
And, um, but then they will have to go past our AI systems.
An AI system with a strong Russian accent will be trying to convince our.
And doesn't put any, uh, articles in their sentences.
Um, well, it'll be at the very least absurdly comedic.
Okay.
Uh, so I, uh, since we talked about sort of the, uh, physical reality, I'd love to ask your vision of the future with, with robots in, in this physical reality.
So many of the kinds of intelligence that you've been speaking about would empower robots to be more effective collaborators with us humans.
So, um, since, uh, Tesla's Optimus team has been showing off some progress on humanoid robots, I think it really reinvigorated the whole industry.
And that's, that, I think Boston Dynamics has been leading for a very, very long time.
So now there's all kinds of companies, figure AI, obviously Boston Dynamics.
Unitree.
Unitree.
Uh, but there's like a lot of them.
It's great.
It's great.
I mean, I love it.
Uh, so do you think there'll be, uh, millions of humanoid robots walking around soon?
Um, not soon, but it's gonna, it's gonna happen.
Like the next decade, I think is going to be really interesting in robots.
Like, uh, the, the emergence of the robotics industry has been in the waiting for, you know, 10, 20 years without really emerging other than for like, you know, kind of pre-programmed behavior and stuff like that.
Um, and, uh, and the main issue is, again, the Moravec paradox, like, you know, how do we get the system to understand how the world works and, and kind of, you know, plan actions.
And so we can do it for really specialized tasks.
Um, and, uh, the way Boston Dynamics goes about it is, you know, basically with a lot of, um, handcrafted dynamical models and careful planning, uh, in advance, which is very classical robotics with a lot of innovation, a little bit of perception.
Um, but it's still not like they can't build a domestic robot.
Right.
Um, and, you know, we're still sometimes.
We're a long distance away from completely autonomous, level five driving.
Uh, and we're certainly very far away from having, uh, you know, level five autonomous driving by a system that can train itself by driving 20 hours, like any 17 year old.
Uh, so until we have, uh, again, world models, systems that can train themselves to understand how the world works.
Uh, we're not going to, uh, we're not going to have significant progress in robotics.
So a lot of the people working on robotic hardware at the moment are, are betting or banking on the fact that AI is going to make sufficient progress towards that.
And they're hoping to discover a product in it too.
So there's, uh, before you have a really strong world model, there'll be an almost strong world model.
And, um, people are trying to find a product in a clumsy robot, I suppose, like not a perfectly efficient robot.
So there's the factory setting where, uh, humanoid robots can help automate some aspects of the factory.
I think that's a crazy difficult task because of all the safety required and all this kind of stuff.
I think in the home is more interesting, but then you start to think, I think you mentioned loading the dishwasher, right?
Yeah.
Like, I suppose that's one of the main problems you're working on.
I mean, there's, you know, uh, cleaning up, cleaning the house, uh, clearing up the table after a meal, um, washing the dishes, you know, all those tasks.
You know, cooking, I mean, all the tasks that, you know, in principle could be automated, but are actually incredibly sophisticated, really complicated.
But even just basic navigation around a space full of uncertainty.
That sort of works.
Like you can sort of do this now.
Navigation is fine.
Well, navigation in a way that's compelling to us humans is a different thing.
Yeah.
It's not going to be, you know, necessarily.
I mean, we have demos actually, because, you know, there is a so-called embodied AI group at FAIR and, uh, you know, they've been not building their own robots, but using commercial robots.
Um, and you can, you can tell a robot dog, like, you know, go to the fridge and they can actually open the fridge and they can probably pick up a can in the fridge and stuff like that.
And, and, and bring it to you, you know, so it can navigate, it can grab objects as long as it's been trying to recognize them, which, you know, vision systems work pretty well nowadays.
Um, but, but it's not like a completely, you know, general robot that would be, you know, sophisticated enough to do things like clearing up the dinner table.
So, yeah, to me, that's an exciting future, uh, of getting humanoid robots, robots in general, in the whole, more and more, because that gets, uh, humans to really directly interact with AI systems in the physical space.
And, and so doing, it allows us to philosophically, psychologically explore our relationships with robots.
It can be really, really, really interesting.
So I hope you make progress on the whole, uh, JAPA thing soon.
Well, I mean, I hope, I hope things kind of, you know, work as, uh, as planned.
Um, I mean, again, we've been kind of working on this idea of self-supervised learning of, uh, from video for, for 10 years and, and, you know, only made significant progress in the last two or three.
And actually you've, you've mentioned that there's a lot of interesting breakups that can happen without having access to a lot of compute.
Yeah.
So if you're interested in doing a PhD and this kind of stuff, there's a lot of possibilities still.
Yeah.
To do innovative work.
So like, what advice would you give to a undergrad that's looking to, uh, go to grad school and do a PhD?
So basically I've listed them already, uh, this idea of how do you train a world model by observation?
Mm-hmm.
And you don't have to train necessarily on gigantic data sets or, I mean, it could turn that to be necessary to actually train on large data sets, to have emergent properties like, like we have with LLMs.
But I think there is a lot of good ideas that can be done without necessarily scaling up.
Then there is how do you do planning with a learned world model?
Well, if the world the system evolves in is not the physical world, but it's the world of, let's say, the internet or, you know, uh, some sort of, uh, world of where an action consists in doing a search in a search engine or interrogating a database or running a simulation or calling a calculator or solving a differential equation.
How do you get a system to actually plan a sequence of actions to, you know, give the solution to a problem?
And so the question of planning is not just a question of planning physical actions.
It could be, you know, planning actions to use tools for a dialogue system or for any kind of intelligent system.
And, um, there's some work on this, but not like, uh, not a huge amount.
Some work at FAIR, uh, um, one called Toolformer, which, uh, was a couple of years ago and some more recent work on planning.
Uh, but, um, but I don't think we have like a good solution for any of that.
Then there is the question of hierarchical planning.
So the example I, I mentioned of, you know, planning a trip from New York to Paris, that's hierarchical, but almost every action that we take involves hierarchical planning in some, in some sense.
And we really have absolutely no idea how to do this.
There's like this zero demonstration of hierarchical planning, uh, in AI where the various levels of representations that are necessary have been learned.
We can do like two level hierarchy, hierarchical planning when we design the two, the two levels.
So for example, you have like a dog-like robot, right?
You want it to go from the living room to the kitchen.
You can plan a path that avoids the obstacle, and then, um, you can send this to a lower level planner that figures out how to move the legs to kind of follow that trajectories, right?
So that works, but that two level planning is designed by hand, right?
Um, we specify what the proper levels of abstraction, the representation at each level of abstraction has, have to be.
How do you learn this?
How do you learn that hierarchical representation of action plans?
Right, we, you know, with CovNets and deep learning, we, we can train the system to learn hierarchical representations of percepts.
Mm-hmm.
What is the equivalent when, what you're trying to represent are action plans?
For action plans, yeah.
So you want, you want basically a robot dog or a humanoid robot that turns on and travels from New York to Paris all by itself.
For example.
All right.
They might have some, uh, trouble at the, at the TSA, but yeah.
No, but even doing something fairly simple, like a household task.
Sure.
Like, you know, uh, cooking or something.
Yeah, that, there's a lot involved.
It's a super complex task.
We take, and once again, we take it for granted.
What hope do you have for, um, the future of humanity?
We're talking about so many exciting technologies, so many exciting possibilities.
What gives you hope when you look out over the next 10, 20, 50, 100 years?
If you look at social media, there's a lot of, there's, there's wars going on.
There's division.
Uh, there's hatred, all this kind of stuff.
That's also part of humanity.
But amidst all that, what gives you hope?
I love that question.
Uh, we can make humanity smarter with AI.
Okay.
I mean, AI basically will amplify human intelligence.
It's as if every one of us will have a staff of smart AI assistants.
They might be smarter than us.
They'll do our bidding.
Okay.
Perhaps execute a task in ways that are much better than we could do ourselves, because
they'll be smarter than us.
And so it's like everyone would be the, the boss of a staff of super smart virtual people.
So we shouldn't feel threatened by, by this any more than we should feel threatened by being
the manager of a group of people, some of whom are more intelligent than us.
I certainly have a lot of experience with this, of, uh, you know, having people working
with me who are smarter than me.
Um, that's actually a wonderful thing.
So, uh, having machines that are smarter than us, that assist us in our, all of our tasks,
our daily lives, whether it's professional or personal, I think would be a absolutely wonderful
thing because intelligence is the most, uh, is the commodity that is most in demand.
That that's really what, I mean, all the mistakes that humanity makes is because of lack of intelligence,
really, or lack of knowledge, which is, you know, related.
So, um, making people smarter, which is, can only be better.
I mean, for the same reason that, you know, public education is a good thing and books are
a good thing and the internet is also a good thing intrinsically.
And even social networks are a good thing if you run them properly.
It's difficult, but you know, you can, um, uh, because, you know, it, it's.
It helps the communication of information and knowledge and the transmission of knowledge.
So AI is going to make humanity smarter.
And the analogy I've been using is the fact that perhaps an equivalent event in the history
of humanity to what might be provided by generalization of AI assistant is the invention of the printing,
the printing press.
It made everybody smarter, the fact that people could, uh, have access to, um, to books.
Books were a lot cheaper than they were before.
And so a lot more people had an incentive to learn to read, which wasn't the case before.
Um, and people became smarter.
It, it enabled the enlightenment, right?
There wouldn't be an enlightenment without the printing press.
It enabled, uh, philosophy, rationalism, uh, escape from religious doctrine, uh, democracy, science.
Uh, and certainly without this, it wouldn't be, there wouldn't have been the American revolution
or the French revolution.
And so it would still be under a feudal, uh, regimes perhaps.
Um, and so it completely transformed the, the world because people became smarter and kind of learned, learned about things.
Now, it also created 200 years of essentially religious conflicts in Europe, right?
Because the first thing that people read was the Bible and, uh, realized that perhaps there was a different interpretation of the Bible than what the priests were telling them.
And so that created the Protestant movement and created the rift.
And in fact, the Catholic school, the Catholic church didn't like the idea of the printing press, but they had no choice.
And so it had some bad effects and some, some good effects.
I don't think anyone today would say that the invention of the printing press had a overall negative effect, despite the fact that it created 200 years of religious conflicts in Europe.
So now compare this, and I, I thought, uh, I was very proud of myself to come up with this analogy, uh, but realized someone else, uh, came with the same idea before me, um, compare this with what happened in the Ottoman Empire.
The Ottoman Empire banned the printing press for 200 years.
Uh, and it didn't ban it, uh, for all languages, only for Arabic.
Like, you could actually print books in Latin or Hebrew or whatever in the Ottoman Empire, just not in Arabic.
And, uh, I thought it was because the rulers just wanted to preserve the control over the population and the dogma, religious dogma and everything.
But after talking with the, uh, UAE minister of AI, uh, Omar Al-Olamar, um, he told me, no, there was another reason.
Uh, and the other reason was that, uh, it was to preserve the cooperation of calligraphers, right?
There was like, uh, an art form, which is, you know, writing those beautiful, uh, you know, Arabic, uh, poems or whatever religious text in, in this thing.
And it was very powerful cooperation of scribes, basically that kind of, you know, run a big chunk of the, uh, empire and, you know, we couldn't put them out of business.
So they, you know, ban the pitching press in part to protect that business.
Now, what's the analogy for AI today?
Like, who are we protecting by banning AI?
Like, who are the people who are asking that AI be regulated to protect their, their jobs?
And of course, you know, there's, it's a, it's a, it's a real question of what is going to be the effect of, uh, you know, technological transformation like AI on the, on the job market and the labor market.
And there are economists who are much more expert at this than I am, but when I talk to them, they, they tell us, you know, we're not going to run out of job.
This is not, this is not going to cause mass unemployment.
This, this is just going to be gradual, uh, shift of different professions.
The professions that are going to be hot 10 or 15 years from now, we have no idea today what they're going to be.
The same way, if we go back 20 years in the past, like who could have thought 20 years ago that like the hottest job, even like five, 10 years ago was mobile app developer.
Like smartphones weren't invented.
But most of the jobs of the future might be in, in the metaverse.
Well, it could be.
Yeah.
But the point is you can't possibly predict, but you're right.
I mean, you made a lot of strong points and I believe that people are fundamentally good.
And so if AI, especially open source AI can, um, make them smarter, it just empowers the goodness in humans.
So I, I share that feeling.
Okay.
I think people are fundamentally good.
Uh, and in fact, a lot of doomers are doomers because they don't think that people are fundamentally good.
Uh, and they either don't trust people or they don't trust the institution to do the right thing so that people behave properly.
Well, I think both you and I believe in humanity and I think I speak for a lot of people in saying, thank you for pushing the open source movement.
I'm pushing to making both research and AI open source, making it available to people and also the models themselves making that open source.
So thank you for that and, uh, thank you for speaking your mind in such colorful and beautiful ways on the internet.
I hope you never stop.
You know, one of the most fun people I know and get to be a fan of.
So, yeah, thank you for speaking to me once again and thank you for being you.
Thank you, Dex.
Thanks for listening to this conversation with Jan LeCun.
To support this podcast, please check out our sponsors in the description.
And now let me leave you with some words from Arthur C. Clark.
The only way to discover the limits of the possible is to go beyond them and to the impossible.
Thank you for listening and hope to see you next time.