Yann Lecun 谈“专有 AI 系统权力集中的危险”

- I see the danger of thisconcentration of power

through proprietary AI systems

as a much bigger dangerthan everything else.

What works against this

is people who think thatfor reasons of security,

we should keep AI systemsunder 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 smallnumber of companies

through proprietary systems.

- I believe that peopleare 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 arefundamentally good. (laughing)

And in fact a lot of doomers are doomers

because they don't think thatpeople are fundamentally good.

- The following is aconversation with Yann 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 ofopen sourcing AI development,

and have been walking the walk

by open sourcing manyof their biggest models,

including LLaMA 2 and eventually LLaMA 3.

Also, Yann 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 AGIwill 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 somewhata controversial position.

And so it's been fun

seeing Yann get into a lot of intense

and fascinating discussions online

as we do in this very conversation.

This is the Lex Fridman podcast.

To support it,

please check out oursponsors in the description.

And now, dear friends, here's Yann LeCun.

You've had some strong statements,

technical statements

about the future of artificialintelligence recently,

throughout your careeractually but recently as well.

You've said that autoregressive LLMs

are not the way we'regoing to make progress

towards superhuman intelligence.

These are the large language models

like GPT-4, like LLaMA2 and 3 soon and so on.

How do they work

and why are they not goingto take us all the way?

- For a number of reasons.

The first is that there isa number of characteristics

of intelligent behavior.

For example, the capacityto understand the world,

understand the physical world,

the ability to rememberand retrieve things,

persistent memory,

the ability to reasonand the ability to plan.

Those are four essential characteristic

of intelligent systems or entities,

humans, animals.

LLMs can do none of those,

or they can only do themin a very primitive way.

And they don't reallyunderstand the physical world,

they don't really have persistent memory,

they can't really reason

and they certainly can't plan.

And so if you expect thesystem to become intelligent

just without having thepossibility of doing those things,

you're making a mistake.

That is not to say thatautoregressive LLMs are not useful,

they're certainly useful.

That they're not interesting,

that we can't build

a whole ecosystem ofapplications around them,

of course we can.

But as a path towardshuman 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 onenormous amounts of text.

Basically the entirety

of all publicly availabletext on the internet, right?

That's typically on theorder of 10 to the 13 tokens.

Each token is typically two bytes.

So that's two 10 to the13 bytes as training data.

It would take you or me 170,000 years

to just read through this ateight hours a day. (laughs)

So it seems like an enormousamount of knowledge, right?

That those systems can accumulate.

But then you realize it'sreally not that much data.

If you talk todevelopmental psychologists,

and they tell you a 4-year-old

has been awake for 16,000hours in his or her life,

and the amount of information

that has reached thevisual cortex of that child

in four years

is about 10 to 15 bytes.

And you can compute this

by estimating that the optical nerve

carry about 20 megabytesper second, roughly.

And so 10 to the 15 bytes for a 4-year-old

versus two times 10 to the 13 bytes

for 170,000 years worth of reading.

What that tells you isthat through sensory input,

we see a lot more information

than we do through language.

And that despite our intuition,

most of what we learnand most of our knowledge

is through our observation and interaction

with the real world,

not through language.

Everything that we learn inthe first few years of life,

and certainly everythingthat animals learn

has nothing to do with language.

- So it would be good

to maybe push againstsome of the intuition

behind what you're saying.

So it is true there'sseveral orders of magnitude

more data coming into thehuman mind, much faster,

and the human mind is able tolearn very quickly from that,

filter the data very quickly.

Somebody might argue

your comparison betweensensory 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 waywe stitch them together,

it already contains a lot of information.

So is it possible that language alone

already has enough wisdomand knowledge in there

to be able to, from thatlanguage construct a world model

and understanding of the world,

an understanding of the physical world

that you're saying LLMs lack?

- So it's a big debate among philosophers

and also cognitive scientists,

like whether intelligence needsto 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 physical reality,

it could be simulated

but the environment is just much richer

than what you can express in language.

Language is a very approximaterepresentation or percepts

and or mental models, right?

I mean, there's a lot oftasks that we accomplish

where we manipulate a mentalmodel 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 moderate task of grabbingsomething, et cetera,

we plan our action sequences,

and we do this

by essentially imagining the result

of the outcome of sequence ofactions 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 interactionwith the physical world.

So a lot of my colleagues

who are more interested inthings like computer vision

are really on that camp

that AI needs to be embodied, essentially.

And then other peoplecoming from the NLP side

or maybe some other motivation

don't necessarily agree with that.

And philosophers are split as well.

And the complexity of theworld is hard to imagine.

It's hard to representall the complexities

that we take completely forgranted in the real world

that we don't even imaginerequire intelligence, right?

This is the old Moravec's paradox

from the pioneer ofrobotics, Hans Moravec,

who said, how is it that with computers,

it seems to be easy to dohigh level complex tasks

like playing chess and solving integrals

and doing things like that,

whereas the thing we take forgranted that we do every day,

like, I don't know,learning to drive a car

or grabbing an object,

we can't do with computers. (laughs)

And we have LLMs thatcan pass the bar exam,

so they must be smart.

But then they can'tlaunch a drive in 20 hours

like any 17-year-old.

They can't learn to clearout the dinner table

and fill out the dishwasher

like any 10-year-oldcan learn in one shot.

Why is that?

Like what are we missing?

What type of learning

or reasoning architectureor whatever are we missing

that basically prevent us

from having level five self-driving cars

and domestic robots?

- Can a large language modelconstruct 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 lotof people are working on.

So the answer,

the short answer is no.

And the more complex answer is

you can use all kind of tricks

to get an LLM to basicallydigest visual representations

of images or video oraudio 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 waysto train vision systems,

either supervised,unsupervised, self-supervised,

all kinds of different ways.

That will turn any image intoa high level representation.

Basically, a list of tokens

that are really similarto the kind of tokens

that a typical LLM takes as an input.

And then you just feed that to the LLM

in addition to the text,

and you just expectthe LLM during training

to kind of be able touse those representations

to help make decisions.

I mean, there's beenwork along those lines

for quite a long time.

And now you see those systems, right?

I mean, there are LLMs thathave some vision extension.

But they're basically hacks

in the sense that those things

are not like trained to handle,

to really understand the world.

They're not trainedwith video, for example.

They don't really understandintuitive physics,

at least not at the moment.

- So you don't think

there's something special toyou 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 gonna be able to do this

with the type of LLMs thatwe are working with today.

And there's a number of reasons for this,

but the main reason is

the way LLMs are trained isthat you take a piece of text,

you remove some of the wordsin that text, you mask them,

you replace them by black markers,

and you train a gigantic neural net

to predict the words that are missing.

And if you build this neuralnet in a particular way

so that it can only look at words

that are to the left of theone it's trying to predict,

then what you have is a system

that basically is trying to predict

the next word in a text, right?

So then you can feed it a text, a prompt,

and you can ask it topredict the next word.

It can never predictthe next word exactly.

And so what it's gonna do

is produce a probability distribution

of all the possiblewords in the dictionary.

In fact, it doesn't predict words,

it predicts tokens thatare 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 computea distribution over them.

Then what the system does

is that it picks a wordfrom that distribution.

Of course, there's a higherchance of picking words

that have a higher probabilitywithin that distribution.

So you sample from that distribution

to actually produce a word,

and then you shift thatword into the input.

And so that allows the system now

to predict the second word, right?

And once you do this,

you shift it into the input, et cetera.

That's called autoregressive prediction,

which is why those LLMs

should be called autoregressive LLMs,

but we just call them at LLMs.

And there is a differencebetween this kind of process

and a process by whichbefore producing a word,

when you talk.

When you and I talk,

you and I are bilinguals.

We think about what we're gonna say,

and it's relatively independent

of the language in which we're gonna say.

When we talk about like, I don't know,

let's say a mathematicalconcept or something.

The kind of thinking that we're doing

and the answer thatwe're planning to produce

is not linked to whetherwe're gonna say it

in French or Russian or English.

- Chomsky just rolled hiseyes, but I understand.

So you're saying thatthere's a bigger abstraction

that goes before language-

- [Yann] Yeah.- And maps onto language.

- Right.

It's certainly true for alot of thinking that we do.

- Is that obvious that we don't?

Like you're saying yourthinking is 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 aprobability distribution?

(both laugh)

- Well, it depends whatkind of thinking, right?

If it's like producing puns,

I get much better in Frenchthan English about that (laughs)

or much worse-

- Is there an abstractrepresentation of puns?

Like is your humor an abstract...

Like when you tweet

and your tweets aresometimes a little bit spicy,

is there an abstract representationin your brain of a tweet

before it maps onto English?

- There is an abstract representation

of imagining the reactionof a reader to that text.

- Oh, you start with laughter

and then figure out howto make that happen?

- Figure out like areaction you wanna cause

and then figure out how to say it

so that it causes that reaction.

But that's like really close to language.

But think about likea mathematical concept

or imagining something youwant to build out of wood

or something like this, right?

The kind of thinking you're doing

has absolutely nothing todo with language, really.

Like it's not like you have necessarily

like an internal monologuein any particular language.

You're imagining mentalmodels 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 a more abstractlevel of representation

in which we do most of our thinking

and we plan what we're gonna say

if the output is uttered words

as opposed to an outputbeing muscle actions, right?

We plan our answer before we produce it.

And LLMs don't do that,

they just produce oneword after the other,

instinctively if you want.

It's a bit like the subconsciousactions where you don't...

Like you're distracted.

You're doing something,

you're completely concentrated

and someone comes to youand asks you a question.

And you kind of answer the question.

You don't have time tothink about the answer,

but the answer is easy

so you don't need to pay attention

and 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'saccumulated a lot of knowledge,

so it can retrieve some things,

but it's going to just spitout one token after the other

without planning the answer.

- But you're making it soundjust one token after the other,

one token at a time generationis bound to be simplistic.

But if the world model issufficiently sophisticated,

that one token at a time,

the most likely thing itgenerates as a sequence of tokens

is going to be a deeply profound thing.

- Okay.

But then that assumes that those systems

actually possess an internal world model.

- So it really goes to the...

I think the fundamental question is

can you build a reallycomplete world model?

Not complete,

but one that has a deepunderstanding of the world.

- Yeah.

So can you build thisfirst of all by prediction?

- [Lex] Right.

- And the answer is probably yes.

Can you build it by predicting words?

And the answer is most probably no,

because language isvery poor in terms of...

Or weak or low bandwidth if you want,

there's just not enough information there.

So building world modelsmeans observing the world

and understanding why the worldis evolving the way it is.

And then the extracomponent of a world model

is something that can predict

how the world is going to evolve

as a consequence of anaction you might take, right?

So one model really is,

here is my idea of the stateof 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 representeverything about the world,

it just needs to represent

enough that's relevant forthis 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 modelthat's 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 you to predictthe reminder of the video.

Basically predict what's gonna happen.

- One frame at a time.

Do the same thing as sort ofthe autoregressive LLMs do,

but for video.

- Right.

Either one frame at a time ora group of frames at a time.

But yeah, a large videomodel, if you want. (laughing)

The idea of doing this

has been floating around for a long time.

And at FAIR,

some colleagues and I

have been trying to dothis for about 10 years.

And you can't really do thesame trick as with LLMs,

because LLMs, as I said,

you can't predict exactlywhich word is gonna follow

a sequence of words,

but you can predict thedistribution of the words.

Now, if you go to video,

what you would have to do

is predict the distribution

of all possible frames in a video.

And we don't really knowhow to do that properly.

We do not know how torepresent 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 discreet.

Video is high dimensional and continuous.

A lot of details in this.

So if I take a video of this room,

and the video is a camera panning around,

there is no way I can predict

everything that's gonna bein the room as I pan around,

the system cannot predictwhat's gonna be in the room

as the camera is panning.

Maybe it's gonna predict,

this is a room where there'sa light and there is a wall

and things like that.

It can't predict what thepainting of the wall looks like

or what the texture ofthe couch looks like.

Certainly not the texture of the carpet.

So there's no way it canpredict all those details.

So the way to handle this

is one way to possibly to handle this,

which we've been working for a long time,

is to have a model that haswhat's called a latent variable.

And the latent variableis 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 agood job at predicting pixels,

including the fine textureof the carpet and the couch

and the painting on the wall.

That has been a completefailure, essentially.

And we've tried lots of things.

We tried just straight neural nets,

we tried GANs,

we tried VAEs,

all kinds of regularized auto encoders,

we tried many things.

We also tried those kind of methods

to learn good representationsof images or video

that could then be used as input

for example, an imageclassification system.

And that also has basically failed.

Like all the systems thatattempt to predict missing parts

of an image or a video

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 reconstructthe complete video or image

from the corrupted version.

And then hope that internally,

the system will develop goodrepresentations of images

that you can use for object recognition,

segmentation, whatever it is.

That has been essentiallya complete failure.

And it works really well for text.

That's the principle thatis used for LLMs, right?

- So where's the failure exactly?

Is it that it is very difficult to form

a good representation of an image,

like a good embedding

of all the importantinformation in the image?

Is it in terms of the consistency

of image to image to image toimage that forms the video?

If we do a highlight reelof 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 tellyou exactly what doesn't work

because there is somethingelse that does work.

So the thing that does not work

is training the system tolearn 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 slewof techniques for this

that are variant of thenusing auto encoders.

Something called MAE,

developed by some ofmy colleagues at FAIR,

masked autoencoder.

So it's basically like theLLMs or things like this

where you train thesystem by corrupting text,

except you corrupt images.

You remove patches from it

and you train a giganticneural network to reconstruct.

The features you get are not good.

And you know they're not good

because if you now trainthe same architecture,

but you train it tosupervise with label data,

with textual descriptionsof images, et cetera,

you do get good representations.

And the performance onrecognition tasks is much better

than if you do this selfsupervised free training.

- So the architecture is good.

- The architecture is good.

The architecture of the encoder is good.

Okay?

But the fact that you train thesystem to reconstruct images

does not lead it to produce

long good generic features of images.

- [Lex] When you train itin a self supervised way.

- Self supervised by reconstruction.

- [Lex] Yeah, by reconstruction.

- Okay, so what's the alternative?

(both laugh)

The alternative is joint embedding.

- What is joint embedding?

What are these architecturesthat you're so excited about?

- Okay, so now insteadof training a system

to encode the image

and then training it toreconstruct the full image

from a corrupted version,

you take the full image,

you take the corruptedor transformed version,

you run them both through encoders,

which in general areidentical but not necessarily.

And then you train a predictoron top of those encoders

to predict the representationof the full input

from the representationof the corrupted one.

Okay?

So joint embedding,

because you're taking the full input

and the corrupted versionor transformed version,

run them both through encoders

so you get a joint embedding.

And then you're saying

can I predict therepresentation of the full one

from the representationof the corrupted one?

Okay?

And I call this a JEPA,

so that means joint embeddingpredictive architecture

because there'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 or six years ago,

we didn't have particularly good answers

for how you train those things,

except for one calledcontrastive learning.

And the idea of contrastive learning

is you take a pair of images

that are, again, an imageand a corrupted version

or degraded version somehow

or transformed versionof the original one.

And you train the predicted representation

to be the same as that.

If you only do this,

this system collapses.

It basically completely ignores the input

and produces representationsthat are constant.

So the contrastive methods avoid this.

And those things have beenaround since the early '90s,

I had a paper on this in 1993,

is you also show pairs of imagesthat you know are different

and then you push away therepresentations from each other.

So you say not only dorepresentations of things

that we know are the same,

should be the same or should be similar,

but representation of thingsthat 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 overthe 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 tothose contrastive methods.

What has changed in thelast three, four years

is now we have methodsthat are non-contrastive.

So they don't require thosenegative contrastive samples

of images that we know are different.

You train them only with images

that are different versions

or different views of the same thing.

And you rely on some other tweaks

to prevent the system from collapsing.

And we have half a dozendifferent methods for this now.

- So what is the fundamental difference

between joint embeddingarchitectures and LLMs?

So can JEPA take us to AGI?

Whether we should say thatyou don't like the term AGI

and we'll probably argue,

I think every singletime I've talked to you

we've argued about the G in AGI.

- [Yann] Yes.

- I get it, I get it, I get it. (laughing)

Well we'll probablycontinue to argue about it.

It's great.

Because you're like French,

and ami is I guess friend in French-

- [Yann] Yes.

- And AMI stands for advancedmachine intelligence-

- [Yann] Right.

- But either way, canJEPA take us to that,

towards that advancedmachine 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 thatare 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 ofresources spent in the system

to actually predict all thosepixels, all the details.

In a JEPA, you're not tryingto predict all the pixels,

you're only trying to predict

an abstract representationof the inputs, right?

And that's much easier in many ways.

So what the JEPA system

when it's being trained is trying to do,

is extract as much informationas 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.

Like for example, if youhave 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 thetree 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'll tell you there's moving leaves,

but it's not gonna keep the details

of exactly what's going on.

And so when you do the predictionin 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 abstractrepresentation of the world

where what can be modeledand predicted is preserved

and the rest is viewed as noise

and eliminated by the encoder.

So it kind of lifts thelevel of abstraction

of the representation.

If you think about this,

this is something we doabsolutely all the time.

Whenever we describe a phenomenon,

we describe it at a particularlevel of abstraction.

And we don't always describeevery natural phenomenon

in terms of quantum field theory, right?

That would be impossible, right?

So we have multiple levels of abstraction

to describe what happens in the world.

Starting from quantum field theory

to like atomic theory andmolecules in chemistry,

materials,

all the way up to kind ofconcrete objects in the real world

and things like that.

So we can't just only modeleverything at the lowest level.

And that's what the ideaof JEPA is really about.

Learn abstract representationin 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 canget away without doing this

because language is alreadyto some level abstract

and already has eliminateda lot of information

that is not predictable.

And so we can get away withoutdoing the joint embedding,

without lifting the abstraction level

and by directly predicting words.

- So joint embedding.

It's still generative,

but it's generative in thisabstract representation space.

- [Yann] Yeah.

- And you're saying language,

we were lazy with language

'cause we already got theabstract representation for free

and now we have to zoom out,

actually think aboutgenerally intelligent systems,

we have to deal with the full mess

of physical of reality, of reality.

And you do have to do this step

of jumping from the full,rich, detailed reality

to an abstract representationof that reality

based on what you can then reason

and all that kind of stuff.

- Right.

And the thing is thoseself supervised algorithms

that learn by prediction,

even in representation space,

they learn more concept

if the input data you feedthem 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 moreredundancy in the structure

in perceptual inputs,sensory input like vision,

than there is in text,

which is not nearly as redundant.

This is back to thequestion you were asking

a few minutes ago.

Language might representmore information really

because it's already compressed,

you're right about that.

But that means it's also less redundant.

And so self supervisedonly will not work as well.

- Is it possible to join

the self supervisedtraining on visual data

and self supervisedtraining on language data?

There is a huge amount of knowledge

even though you talk down aboutthose 10 to the 13 tokens.

Those 10 to the 13 tokens

represent the entirety,

a large fraction of whatus humans have figured out.

Both the shit talk on Reddit

and the contents of allthe books and the articles

and the full spectrum ofhuman intellectual creation.

So is it possible tojoin 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 whatpeople are doing at the moment

with vision language model.

We're basically cheating.

We are using language as a crutch

to help the deficienciesof our vision systems

to kind of learn good representationsfrom images and video.

And the problem with this

is that we might improve ourvision language system a bit,

I mean our language modelsby feeding them images.

But we're not gonna get to the level

of even the intelligence

or level of understanding of the world

of a cat or a dog whichdoesn't have language.

They don't have language

and they understand the worldmuch better than any LLM.

They can plan really complex actions

and sort of imagine theresult of a bunch of actions.

How do we get machines to learn that

before we combine that with language?

Obviously, if we combinethis with language,

this is gonna be a winner,

but before that we have to focus

on like how do we get systemsto learn how the world works?

- So this kind of joint embeddingpredictive architecture,

for you, that's gonna be able to learn

something like common sense,

something like what a cat uses

to predict how to mess withits owner most optimally

by knocking over a thing.

- That's the hope.

In fact, the techniques we'reusing are non-contrastive.

So not only is thearchitecture non-generative,

the learning procedures we'reusing are non-contrastive.

We have two sets of techniques.

One set is based on distillation

and there's a number of methodsthat use this principle.

One by DeepMind called BYOL.

A couple by FAIR,

one called VICReg andanother one called I-JEPA.

And VICReg, I should say,

is not a distillation method actually,

but I-JEPA and BYOL certainly are.

And there's another onealso called DINO or Dino,

also produced at FAIR.

And the idea of those things

is that you take the fullinput, let's say an image.

You run it through an encoder,

produces a representation.

And then you corrupt thatinput or transform it,

run it through essentially whatamounts to the same encoder

with some minor differences.

And then train a predictor.

Sometimes a predictor is very simple,

sometimes it doesn't exist.

But train a predictor topredict a representation

of the first uncorrupted inputfrom 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 thetype I was explaining before

where the system basicallyignores the input.

So that works very well.

The two techniqueswe've developed at FAIR,

DINO and I-JEPA work really well for that.

- So what kind of dataare we talking about here?

- So there's several scenarios.

One scenario is you take an image,

you corrupt it by changingthe cropping, for example,

changing the size a little bit,

maybe changing theorientation, blurring it,

changing the colors,

doing all kinds of horrible things to it-

- But basic horrible things.

- Basic horrible things

that sort of degradethe quality a little bit

and change the framing,

crop the image.

And in some cases, in the case of I-JEPA,

you don't need to do any of this,

you just mask some parts of it, right?

You just basically remove some regions

like a big block, essentially.

And then run through the encoders

and train the entire system,

encoder and predictor,

to predict the representationof the good one

from the representationof the corrupted one.

So that's the I-JEPA.

It doesn't need to know thatit'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 geometry transformation and blurring

and things like that thatare really image specific.

A more recent version of thisthat we have is called V-JEPA.

So it's basically the same idea as I-JEPA

except it's applied to video.

So now you take a whole video

and you mask a whole chunk of it.

And what we mask is actuallykind of a temporal tube.

So like a whole segmentof each frame in the video

over the entire video.

- And that tube is likestatically positioned

throughout the frames?

It's literally just a straight tube?

- Throughout the tube, yeah.

Typically it's 16 frames or something,

and we mask the same regionover the entire 16 frames.

It's a different one forevery video, obviously.

And then again, train that system

so as to predict therepresentation 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 feedthose representations

to a supervised classifier head,

it can tell you what actionis taking place in the video

with pretty good accuracy.

So it's the first time we getsomething of that quality.

- So that's a good test

that a good representation is formed.

That means there's something to this.

- Yeah.

We also preliminary result

that seem to indicate

that the representationallows our system to tell

whether the video is physically possible

or completely impossible

because some object disappeared

or an object suddenly jumpedfrom one location to another

or changed shape or something.

- So it's able to capturesome physics based constraints

about the realityrepresented in the video?

- [Yann] Yeah.

- About the appearance andthe disappearance of objects?

- Yeah.

That's really new.

- Okay, but can this actually

get us to this kind of world model

that understands enough about the world

to be able to drive a car?

- Possibly.

And this is gonna take a while

before we get to that point.

And there are systemsalready, robotic systems,

that are based on this idea.

What you need for this

is a slightly modified version of this

where imagine that you have a video,

a complete video,

and what you're doing to this video

is that you are eithertranslating it in time

towards the future.

So you'll only see thebeginning of the video,

but you don't see the latter part of it

that is in the original one.

Or you just mask the secondhalf of the video, for example.

And then you train this I-JEPA system

or the type I described,

to predict representationof the full video

from the shifted one.

But you also feed thepredictor with an action.

For example, the wheel is turned

10 degrees to the rightor something, right?

So if it's a dash cam in a car

and you know the angle of the wheel,

you should be able topredict to some extent

what's going to happen to what you see.

You're not gonna be ableto predict all the details

of objects that appearin the view, obviously,

but at an abstract representation level,

you can probably predictwhat's gonna happen.

So now what you have is an internal model

that says, here is my idea

of the state of the world at time T,

here is an action I'm taking,

here is a prediction

of the state of theworld 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 gonna do

so as you arrive at a particular outcome

or satisfy a particular objective, right?

So you can have a numberof objectives, right?

I can predict that if I havean object like this, right?

And I open my hand,

it's gonna fall, right?

And if I push it with aparticular force on the table,

it's gonna move.

If I push the table itself,

it's probably not gonnamove with the same force.

So we have this internal modelof the world in our mind,

which allows us to plansequences of actions

to arrive at a particular goal.

And so now if you have this world model,

we can imagine a sequence of actions,

predict what the outcome

of the sequence of action is going to be,

measure to what extent the final state

satisfies a particular objective

like moving the bottleto the left of the table.

And then plan a sequence of actions

that will minimize thisobjective 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 thesystem you want to control

that can predict the sequence of states

corresponding to a sequence of commands.

And you are planninga sequence of commands

so that according to your world model,

the end state of the system

will satisfy any objectives that you fix.

This is the way rockettrajectories have been planned

since computers have been around.

So since the early '60s, essentially.

- So yes, for a model predictive control,

but you also often talkabout hierarchical planning.

- [Yann] Yeah.

- Can hierarchical planningemerge from this somehow?

- Well, so no.

You will have to builda specific architecture

to allow for hierarchical planning.

So hierarchical planningis absolutely necessary

if you want to plan complex actions.

If I wanna go from, let'ssay, from New York to Paris,

this 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 abstractrepresentation of my location,

I would have to decomposethis 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 nowgoing to the airport.

My objective function ismy distance to the airport.

How do I go to the airport?

Well, I have to go in thestreet and hail a taxi,

which you can do in New York.

Okay, now I have another sub-goal.

Go down on the street.

Well, that means going to the elevator,

going down the elevator,

walk out to 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 for my chair?

Like you can imagine goingdown all the way down

to basically what amounts

to millisecond bymillisecond muscle control.

Okay?

And obviously you're notgoing to plan your entire trip

from New York to Paris

in terms of millisecond bymillisecond 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 gonna happen.

How long it's gonna take to catch a taxi

or to go to the airport with traffic.

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 thishierarchical planning

so that you can start acting

and then sort of re-planning as you go.

And nobody really knowshow to do this in AI.

Nobody knows how to train a system

to learn the appropriatemultiple 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 alist 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 movingyour 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 LLMswill be able to answer

some of those questions

down to some level of abstraction.

Under the condition thatthey've been trained

with similar scenariosin their training set.

- They would be able toanswer all of those questions.

But some of them may be hallucinated,

meaning non-factual.

- Yeah, true.

I mean they'll probablyproduce some answer.

Except they're not gonna be able

to really kind of produce

millisecond by millisecond muscle control

of how you stand upfrom your chair, right?

But down to some level of abstraction

where you can describe things by words,

they might be able to give you a plan,

but only under the conditionthat they've been trained

to produce those kind of plans, right?

They're not gonna be ableto plan for situations

they never encountered before.

They basically are going tohave to regurgitate the template

that they've been trained on.

- But where, just for theexample of New York to Paris,

is it gonna 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 will be able toanswer somewhat accurately,

especially when you're talkingabout New York and Paris,

major cities.

- So I mean certainly an LLM

would be able to solve that problem

if you fine tune it for it.

- [Lex] Sure.

- And so I can't say thatan LLM cannot do this,

it can't do this if you train it for it,

there's no question,

down to a certain level

where things can beformulated in terms of words.

But like if you wanna go down

to like how do you climb down the stairs

or just stand up from yourchair in terms of words,

like you can't do it.

That's one of the reasons you need

experience of the physical world,

which is much higher bandwidth

than what you can express in words,

in human language.

- So everything we've been talking about

on the joint embedding space,

is it possible that that's what we need

for like the interactionwith physical reality

on the robotics front?

And then just the LLMs are thething that sits on top of it

for the bigger reasoning

about like the fact that Ineed to book a plane ticket

and I need to know know how togo to the websites and so on.

- Sure.

And a lot of plans that people know about

that are relatively highlevel are actually learned.

Most people don't inventthe plans by themselves.

We have some ability to dothis, of course, obviously,

but most plans that people use

are plans that have been trained on.

Like they've seen otherpeople use those plans

or they've been toldhow to do things, right?

That you can't invent howyou like take a person

who's never heard of airplanes

and tell them like, how doyou go from New York to Paris?

They're probably not going to be able

to kind of deconstruct the whole plan

unless they've seenexamples of that before.

So certainly LLMs aregonna be able to do this.

But then how you link thisfrom the low level of actions,

that needs to be donewith 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 JEPAs for.

- I would love to sort oflinger on your skepticism

around autoregressive LLMs.

So one way I would liketo test that skepticism is

everything you say makes a lot of sense,

but if I apply everythingyou 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 topredict the success of LLMs.

So does it make sense to you

that autoregressive LLMsare able to be so damn good?

- [Yann] Yes.

- Can you explain your intuition?

Because if I were to takeyour wisdom and intuition

at face value,

I would say there's noway autoregressive LLMs

one token at a time,

would be able to do the kindof things they're doing.

- No, there's one thingthat autoregressive LLMs

or that LLMs in general, notjust the autoregressive ones,

but including the BERTstyle bidirectional ones,

are exploiting and itsself supervised running.

And I've been a very, very strong advocate

of self supervised running for many years.

So those things are an incrediblyimpressive demonstration

that self supervisedlearning actually works.

The idea that started...

It didn't start with BERT,

but it was really kind of agood demonstration with this.

So the idea that you take apiece of text, you corrupt it,

and then you train somegigantic 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 systemsthat understand language,

systems that can translate

hundreds of languages in any direction,

systems that are multilingual.

It's a single system

that can be trained tounderstand hundreds of languages

and translate in any direction

and produce summaries

and then answer questionsand produce text.

And then there's a special case of it,

which is the autoregressive trick

where you constrain the system

to not elaborate arepresentation 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 thearchitecture of the network.

And that's what you can buildan autoregressive LLM from.

So there was a surprise many years ago

with what's called decoder only LLM.

So systems of this type

that are just trying to producewords from the previous one.

And the fact that when you scale them up,

they tend to really kind ofunderstand 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 occurredquite a while back.

Like with work from Google,Meta, OpenAI, et cetera,

going back to the GPT

kind of general pre-trained transformers.

- You mean like GPT-2?

Like there's a certain place

where you start to realize

scaling might actually keepgiving us an emergent benefit.

- Yeah, I mean there werework from various places,

but if you want to kind ofplace it in the GPT timeline,

that would be around GPT-2, yeah.

- Well, 'cause you said it,

you're so charismatic andyou said so many words,

but self supervised learning, yes.

But again, the sameintuition you're applying

to saying that autoregressive LLMs

cannot have a deepunderstanding 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 in the world

to be damn convincing,

essentially passing theoriginal Turing test

with flying colors.

- Well, we're fooled bytheir fluency, right?

We just assume that if a system is fluent

in manipulating language,

then it has all the characteristicsof human intelligence.

But that impression is false.

We're really fooled by it.

- Well, what do you thinkAlan Turing would say?

Without understanding anything,

just hanging out with it-

- Alan Turing would decide

that a Turing test is a really bad test.

(Lex chuckles)

Okay.

This is what the AI communityhas decided many years ago

that the Turing test was areally bad test of intelligence.

- What would Hans Moravec say

about the large language models?

- Hans Moravec would say

the Moravec's paradox still applies.

- [Lex] Okay.

- Okay?

Okay, we can pass-

- You don't think hewould be really impressed.

- No, of course everybodywould be impressed.

(laughs)

But it is not a questionof being impressed or not,

it is a question of knowing

what the limit of those systems can do.

Again, they are impressive.

They can do a lot of useful things.

There's a whole industry thatis being built around them.

They're gonna make progress,

but there is a lot ofthings they cannot do.

And we have to realize what they cannot do

and then figure out how we get there.

And I'm not saying this...

I'm saying this frombasically 10 years of research

on the idea of self supervised running,

actually that's goingback more than 10 years,

but the idea of self supervised learning.