Ilya Sutskever: Deep Learning
全文文稿 · 按对话顺序保留完整内容，深入理解上下文。
本篇目录12 章节
§01Transcript · 00:00–08:03§02Transcript · 08:06–16:24§03Transcript · 16:27–24:24§04Transcript · 24:26–32:36§05Transcript · 32:37–40:49§06Transcript · 40:52–49:50§07Transcript · 49:51–58:23§08Transcript · 58:25–1:06:04§09Transcript · 1:06:06–1:15:06§10Transcript · 1:15:08–1:23:49§11Transcript · 1:23:54–1:33:24§12Transcript · 1:33:26–1:37:21§01Transcript · 00:00–08:03
The following is a conversation with Ilya Sutskever,
co-founder and chief scientist of OpenAI,
one of the most cited computer scientists in history
with over 165,000 citations.
And to me, one of the most brilliant and insightful minds
ever in the field of deep learning.
There are very few people in this world
who I would rather talk to and brainstorm with
about deep learning, intelligence, and life in general
than Ilya, on and off the mic.
This was an honor and a pleasure.
This conversation was recorded
before the outbreak of the pandemic.
For everyone feeling the medical, psychological,
and financial burden of this crisis,
I'm sending love your way.
Stay strong, we're in this together, we'll beat this thing.
This is the Artificial Intelligence Podcast.
If you enjoy it, subscribe on YouTube,
review it with Five Stars and Apple Podcast,
support it on Patreon, or simply connect with me on Twitter
at Lex Friedman, spelled F-R-I-D-M-A-N.
As usual, I'll do a few minutes of ads now
and never any ads in the middle
that can break the flow of the conversation.
I hope that works for you
and doesn't hurt the listening experience.
This show is presented by Cash App,
the number one finance app in the App Store.
When you get it, use code LEXPODCAST.
Cash App lets you send money to friends, buy Bitcoin,
invest in the stock market with as little as $1.
Since Cash App allows you to buy Bitcoin,
let me mention that cryptocurrency
in the context of the history of money is fascinating.
I recommend "A Scent of Money"
as a great book on this history.
Both the book and audiobook are great.
Debits and credits on ledgers
started around 30,000 years ago.
The US dollar created over 200 years ago,
and Bitcoin, the first decentralized cryptocurrency,
released just over 10 years ago.
So given that history,
cryptocurrency is still very much in its early days
of development, but it's still aiming to,
and just might, redefine the nature of money.
So again, if you get Cash App
from the App Store or Google Play and use the code LEXPODCAST,
you get $10, and Cash App will also donate $10 to FIRST,
an organization that is helping advance robotics
and STEM education for young people around the world.
And now, here's my conversation with Ilya Sutskever.
You were one of the three authors,
with Alex Krashevsky, Jeff Hinton,
of the famed Alex Ned paper
that is arguably the paper
that marked the big catalytic moment
that launched the deep learning revolution.
At that time, take us back to that time,
what was your intuition about neural networks,
about the representational power of neural networks?
And maybe you could mention,
how did that evolve over the next few years,
up to today, over the 10 years?
- Yeah, I can answer that question.
At some point in about 2010 or 2011,
I connected two facts in my mind.
Basically, the realization was this.
At some point, we realized that we can train very large,
I shouldn't say very, you know,
they were tiny by today's standards,
but large and deep neural networks,
end to end with back propagation.
At some point, different people obtained this result.
I obtained this result.
The first moment in which I realized
that deep neural networks are powerful
was when James Martens invented
the Hessian Free Optimizer in 2010.
And he trained a 10-layer neural network,
end to end, without pre-training, from scratch.
And when that happened, I thought, this is it.
Because if you can train a big neural network,
a big neural network can represent very complicated function.
Because if you have a neural network with 10 layers,
it's as though you allow the human brain
to run for some number of milliseconds.
Neuron firings are slow,
and so in maybe 100 milliseconds,
your neurons only fire 10 times.
So it's also kind of like 10 layers.
And in 100 milliseconds,
you can perfectly recognize any object.
So I thought, so I already had the idea then
that we need to train a very big neural network
on lots of supervised data, and then it must succeed,
because we can find the best neural network.
And then there's also theory
that if you have more data than parameters,
you won't overfit.
Today, we know that actually this theory is very incomplete
and you won't overfit
even if you have less data than parameters.
But definitely, if you have more data than parameters,
you won't overfit.
- So the fact that neural networks
were heavily over-parameterized wasn't discouraging to you?
So you were thinking about the theory
that the number of parameters,
the fact there's a huge number of parameters is okay?
It's gonna be okay?
- I mean, there was some evidence before
that it was okay-ish,
but the theory was that if you had a big dataset
and a big neural net, it was going to work.
The over-parameterization just didn't really figure much
as a problem.
I thought, well, with images,
you're just gonna add some data augmentation
and it's gonna be okay.
- So where was any doubt coming from?
- The main doubt was, can we train a big,
will we have enough compute to train a big enough neural net?
- With backpropagation.
- Backpropagation, I thought, would work.
The thing which wasn't clear
was whether there would be enough compute
to get a very convincing result.
And then at some point, Alex Kirzhevsky
wrote these insanely fast CUDA kernels
for training convolutional neural nets.
And that was, bam, let's do this.
Let's get an image in it,
and it's gonna be the greatest thing.
- Was most of your intuition from empirical results
by you and by others?
So like, just actually demonstrating
that a piece of program can train
a 10-layer neural network?
Or was there some pen and paper or marker and whiteboard
thinking intuition?
'Cause you just connected a 10-layer large neural network
to the brain, so you just mentioned the brain.
So in your intuition about neural networks,
does the human brain come into play as a intuition builder?
- Definitely.
I mean, you gotta be precise with these analogies
between artificial neural networks and the brain.
But there's no question that the brain is a huge source
of intuition and inspiration for deep learning researchers
since all the way from Rosenblatt in the '60s.
Like, if you look at, the whole idea of a neural network
is directly inspired by the brain.
You had people like McCallum and Pitts who were saying,
"Hey, you got these neurons in the brain.
"And hey, we recently learned about the computer
"and automata, can we use some ideas
"from the computer and automata to design
"some kind of computational object
"that's going to be simple, computational,
"and kind of like the brain?"
And they invented the neuron.
So they were inspired by it back then.
Then you had the convolutional neural network
from Fukushima, and then later, Jan Lekan,
who said, "Hey, if you limit the receptive fields
"of a neural network, it's gonna be especially suitable
"for images," as it turned out to be true.
So there was a very small number of examples
where analogies to the brain were successful.
And I thought, well, probably an artificial neuron
is not that different from the brain
if you squint hard enough.
So let's just assume it is and roll with it.
- So we're now at a time where deep learning
is very successful, so let us squint less
and say, let's open our eyes and say,
what to you is an interesting difference
between the human brain, now I know you're probably
not an expert, neither a neuroscientist
or a neurobiologist, but loosely speaking,
what's the difference between the human brain
and artificial neural networks that's interesting to you
for the next decade or two?
- That's a good question to ask.
What is an interesting difference between the brain
and our artificial neural networks?
So I feel like today, artificial neural networks,
so we all agree that there are certain dimensions
in which the human brain vastly outperforms our models.
But I also think that there are some ways
in which our artificial neural networks
have a number of very important advantages over the brain.
Looking at the advantages versus disadvantages
is a good way to figure out what is the important difference.
So the brain uses spikes, which may or may not be important.
- That's a really interesting question.
Do you think it's important or not?
That's one big architectural difference
§02Transcript · 08:06–16:24
between artificial neural networks.
- It's hard to tell, but my prior is not very high
and I can say why.
There are people who are interested
in spiking neural networks and basically,
what they figured out is that they need to simulate
the non-spiking neural networks in spikes.
And that's how they're gonna make them work.
If you don't simulate the non-spiking neural networks
in spikes, it's not going to work
because the question is, why should it work?
And that connects to questions around back propagation
and questions around deep learning.
You've got this giant neural network.
Why should it work at all?
Why should the learning rule work at all?
It's not a self-evident question, especially if you,
let's say if you were just starting in the field
and you read the very early papers,
you can say, "Hey," people are saying,
"Let's build neural networks."
That's a great idea because the brain is a neural network,
so it would be useful to build neural networks.
Now let's figure out how to train them.
It should be possible to train them probably, but how?
And so the big idea is the cost function.
That's the big idea.
The cost function is a way of measuring the performance
of the system according to some measure.
By the way, that is a big, actually, let me think.
Is that, one, a difficult idea to arrive at
and how big of an idea is that,
that there's a single cost function?
- Sorry, let me take a pause.
Is supervised learning a difficult concept to come to?
- I don't know.
All concepts are very easy in retrospect.
- Yeah, that's what, it seems trivial now, but I,
'cause the reason I ask that, and we'll talk about it,
'cause is there other things?
Is there things that don't necessarily have a cost function,
maybe have many cost functions,
or maybe have dynamic cost functions,
or maybe a totally different kind of architectures?
'Cause we have to think like that
in order to arrive at something new, right?
- So the only, so the good examples of things
which don't have clear cost functions are GANs.
- Right. - In a GAN, you have a game.
So instead of thinking of a cost function
where you wanna optimize,
where you know that you have an algorithm gradient descent,
which will optimize the cost function,
and then you can reason about the behavior of your system
in terms of what it optimizes.
With a GAN, you say, "I have a game,
"and I'll reason about the behavior of the system
"in terms of the equilibrium of the game."
But it's all about coming up with these mathematical objects
that help us reason about the behavior of our system.
- Right, that's really interesting.
Yeah, so GAN is the only one, it's kind of a,
the cost function is emergent from the comparison.
- It's, I don't know if it has a cost function.
I don't know if it's meaningful
to talk about the cost function of a GAN.
It's kind of like the cost function of biological evolution
or the cost function of the economy.
It's, you can talk about regions
to which it will go towards, but I don't think,
I don't think the cost function analogy is the most useful.
- So if evolution doesn't, that's really interesting.
So if evolution doesn't really have a cost function,
like a cost function based on its,
something akin to our mathematical conception
of a cost function, then do you think cost functions
in deep learning are holding us back?
Yeah, so you just kind of mentioned that cost function
is a nice first profound idea.
Do you think that's a good idea?
Do you think it's an idea we'll go past?
So self-play starts to touch on that a little bit
in reinforcement learning systems.
- That's right.
Self-play and also ideas around exploration
where you're trying to take action
that surprise a predictor.
I'm a big fan of cost functions.
I think cost functions are great
and they serve us really well.
And I think that whenever we can do things
with cost functions, we should.
And you know, maybe there is a chance
that we will come up with some,
yet another profound way of looking at things
that will involve cost functions in a less central way.
But I don't know, I think cost functions are, I mean,
I would not bet against cost functions.
- Is there other things about the brain
that pop into your mind that might be different
and interesting for us to consider
in designing artificial neural networks?
So we talked about spiking a little bit.
- I mean, one thing which may potentially be useful,
I think people, neuroscientists have figured out
something about the learning rule of the brain,
or I'm talking about spike time independent plasticity,
and it would be nice if some people
would just study that in simulation.
- Wait, sorry, spike time independent plasticity?
- Yeah, that's right. - What's that?
- STD, it's a particular learning rule
that uses spike timing to figure out how to,
to determine how to update the synapses.
So it's kind of like, if a synapse fires into the neuron
before the neuron fires, then it's strengthen the synapse.
And if the synapse fires into the neurons
shortly after the neuron fired, then it weakens the synapse.
Something along this line.
I'm 90% sure it's right, so if I said something wrong here,
don't, don't get too angry.
- But you sound brilliant while saying it.
But the timing, that's one thing that's missing.
The temporal dynamics is not captured.
I think that's like a fundamental property of the brain,
is the timing of the signals.
- Well, you have recurrent neural networks.
But you think of that as this,
I mean, that's a very crude, simplified,
what's that called?
There's a clock, I guess, to recurrent neural networks.
It seems like the brain is the general,
the continuous version of that, the generalization,
where all possible timings are possible,
and then within those timings is contained some information.
You think recurrent neural networks,
the recurrence in recurrent neural networks
can capture the same kind of phenomena
as the timing that seems to be important for the brain,
in the firing of neurons in the brain?
- I mean, I think recurrent neural networks are amazing,
and they can do, I think they can do anything
we'd want them to, we'd want a system to do.
Right now, recurrent neural networks
have been superseded by transformers,
but maybe one day they'll make a comeback,
maybe they'll be back, we'll see.
- Let me, on a small tangent, say,
do you think they'll be back?
So, so much of the breakthroughs recently
that we'll talk about on natural language processing
and language modeling has been with transformers
that don't emphasize recurrence.
Do you think recurrence will make a comeback?
- Well, some kind of recurrence, I think, very likely.
Recurrent neural networks for,
as they're typically thought of for processing sequences,
I think it's also possible.
- What is, to you, a recurrent neural network?
And generally speaking, I guess,
what is a recurrent neural network?
- You have a neural network which maintains
a high-dimensional hidden state.
And then when an observation arrives,
it updates its high-dimensional hidden state
through its connections, in some way.
- So do you think, you know,
that's what, like, expert systems did, right?
Symbolic AI, the knowledge-based,
growing a knowledge base is maintaining a hidden state,
which is its knowledge base,
and is growing it by sequentially processing.
Do you think of it more generally in that way?
Or is it simply, is it the more constrained form
of a hidden state with certain kind of gating units
that we think of as today with LSTMs and that?
- I mean, the hidden state is technically
what you described there, the hidden state
that goes inside the LSTM or the RNN or something like this.
But then what should be contained, you know,
if you want to make the expert system analogy, I'm not,
I mean, you could say that the knowledge
is stored in the connections,
and then the short-term processing
is done in the hidden state.
- Yes, could you say that?
- Yes.
- So, sort of, do you think there's a future
of building large-scale knowledge bases
within the neural networks?
- Definitely.
(Lex laughing)
- So, we're gonna pause in that confidence,
'cause I wanna explore that.
But let me zoom back out and ask,
back to the history of ImageNet.
Neural networks have been around for many decades,
as you mentioned.
What do you think were the key ideas
that led to their success, that ImageNet moment,
§03Transcript · 16:27–24:24
and beyond the success in the past 10 years?
- Okay, so the question is,
to make sure I didn't miss anything,
the key ideas that led to the success of deep learning
over the past 10 years.
- Exactly, even though the fundamental thing
behind deep learning has been around for much longer.
- So, the key idea about deep learning,
or rather, the key fact about deep learning
before deep learning started to be successful,
is that it was underestimated.
People who worked in machine learning
simply didn't think that neural networks could do much.
People didn't believe that large neural networks
could be trained.
People thought that, well, there was a lot of debate
going on in machine learning
about what are the right methods and so on.
And people were arguing,
because there was no way to get hard facts.
And by that I mean, there were no benchmarks
which were truly hard,
that if you do really well on them,
then you can say, "Look, here's my system."
That's when you switch from...
That's when this field becomes a little bit more
of an engineering field.
So, in terms of deep learning,
to answer the question directly,
the ideas were all there.
The thing that was missing was a lot of supervised data
and a lot of compute.
Once you have a lot of supervised data and a lot of compute,
then there is a third thing which is needed as well,
and that is conviction.
Conviction that if you take the right stuff,
which already exists,
and apply and mix it with a lot of data
and a lot of compute,
that it will in fact work.
And so that was the missing piece.
It was, you had the...
You needed the data,
you needed the compute,
which showed up in terms of GPUs,
and you needed the conviction to realize
that you need to mix them together.
- So that's really interesting.
So, I guess the presence of compute
and the presence of supervised data
allowed the empirical evidence to do the convincing
of the majority of the computer science community.
So I guess there's a key moment
with Jitendra Malik and Alex,
Alyosha Efros,
who were very skeptical, right?
And then there's a Jeffrey Hinton
that was the opposite of skeptical.
And there was a convincing moment,
and I think emission had served as that moment.
- That's right.
- And that represented this kind of,
where the big pillars of computer vision community
kinda, the wizards got together,
and then all of a sudden there was a shift.
And it's not enough for the ideas to all be there
and the compute to be there,
it's for it to convince the cynicism that existed.
It's interesting that people just didn't believe
for a couple of decades.
- Yeah, well, but it's more than that.
It's kind of, when put this way,
it sounds like, well, you know,
those silly people who didn't believe
what were they missing.
But in reality, things were confusing
because neural networks really did not work on anything.
And they were not the best method
on pretty much anything as well.
And it was pretty rational to say,
yeah, this stuff doesn't have any traction.
And that's why you need to have these very hard tasks
which produce undeniable evidence.
And that's how we make progress.
And that's why the field is making progress today
because we have these hard benchmarks
which represent true progress.
And this is why we are able to avoid endless debate.
- So incredibly, you've contributed
some of the biggest recent ideas in AI
in computer vision, language, natural language processing,
reinforcement learning, sort of everything in between.
Maybe not GANs.
There may not be a topic you haven't touched.
And of course, the fundamental science of deep learning.
What is the difference to you between vision, language,
and as in reinforcement learning, action,
as learning problems?
And what are the commonalities?
Do you see them as all interconnected?
Are they fundamentally different domains
that require different approaches?
- Okay, that's a good question.
Machine learning is a field with a lot of unity,
a huge amount of unity.
In fact-- - What do you mean by unity?
Like overlap of ideas?
- Overlap of ideas, overlap of principles.
In fact, there's only one or two or three principles
which are very, very simple.
And then they apply in almost the same way,
in almost the same way to the different modalities
to the different problems.
And that's why today, when someone writes a paper
on improving optimization of deep learning and vision,
it improves the different NLP applications
and it improves the different
reinforcement learning applications.
Reinforcement learning, so I would say that computer vision
and NLP are very similar to each other.
Today, they differ in that they have
slightly different architectures.
We use transformers in NLP
and we use convolutional neural networks in vision.
But it's also possible that one day this will change
and everything will be unified with a single architecture.
Because if you go back a few years ago
in natural language processing,
there were a huge number of architectures
for every different tiny problem had its own architecture.
Today, there's just one transformer
for all those different tasks.
And if you go back in time even more,
you had even more and more fragmentation
and every little problem in AI
had its own little subspecialization
and sub, you know, little set of collection of skills,
people who would know how to engineer the features.
Now it's all been subsumed by deep learning.
We have this unification.
And so I expect vision to become unified
with natural language as well.
Or rather, I shouldn't say expect, I think it's possible.
I don't wanna be too sure because I think
on the convolutional neural net,
it's very computationally efficient.
RL is different.
RL does require slightly different techniques
because you really do need to take action.
You really do need to do something about exploration.
Your variance is much higher.
But I think there is a lot of unity even there.
And I would expect, for example,
that at some point there will be some
broader unification between RL and supervised learning,
where somehow the RL will be making decisions
to make the supervised learning go better.
And it will be, I imagine one big black box
and you just throw everything, you know,
you shovel things into it and it just figures out
what to do with whatever you shovel at it.
- I mean, reinforcement learning has some aspects
of language and vision combined almost.
There's elements of a long-term memory
that you should be utilizing and there's elements
of a really rich sensory space.
So it seems like the, it's like the union of the two
or something like that.
- I'd say something slightly differently.
I'd say that reinforcement learning is neither,
but it naturally interfaces and integrates
with the two of them.
- You think action is fundamentally different?
So yeah, what is interesting about,
what is unique about policy of learning to act?
- Well, so one example, for instance,
is that when you learn to act,
you are fundamentally in a non-stationary world
because as your actions change,
the things you see start changing.
You experience the world in a different way.
And this is not the case for the more traditional
static problem where you have some distribution
and you just apply a model to that distribution.
- You think it's a fundamentally different problem
or is it just a more difficult,
it's a generalization of the problem of understanding?
- I mean, it's a question of definitions almost.
There is a huge, I mean, no,
there's a huge amount of commonality for sure.
You take gradients, you try, you take gradients,
we try to approximate gradients in both cases.
In some, in the case of reinforcement learning,
you have some tools to reduce the variance
of the gradients, you do that.
There's lots of commonality.
You use the same neural net in both cases.
You compute the gradient, you apply Adam in both cases.
So, I mean, there's lots in common for sure,
but there are some small differences
§04Transcript · 24:26–32:36
which are not completely insignificant.
It's really just a matter of your point of view,
what frame of reference you,
how much do you want to zoom in or out
as you look at these problems.
- Which problem do you think is harder?
So people like Noam Chomsky believe
that language is fundamental to everything.
So it underlies everything.
Do you think language understanding
is harder than visual scene understanding or vice versa?
- I think that asking if a problem is hard
is slightly wrong.
I think the question is a little bit wrong
and I want to explain why.
- So what does it mean for a problem to be hard?
Okay, the non-interesting, dumb answer to that
is there's a benchmark
and there's a human level performance on that benchmark.
And how is the effort required
to reach the human level benchmark?
- So from the perspective of how much
until we get to human level on a very good benchmark.
- Yeah, I understand what you mean by that.
So what I was going to say that a lot of it depends on,
you know, once you solve a problem, it stops being hard.
And that's always true.
And so whether something is hard or not
depends on what our tools can do today.
So, you know, you say today, true human level,
language understanding and visual perception
are hard in the sense that there is no way
of solving the problem completely in the next three months.
So I agree with that statement.
Beyond that, I'm just,
my guess would be as good as yours, I don't know.
- Okay, so you don't have a fundamental intuition
about how hard language understanding is.
- Well, I think, I know I changed my mind.
I'd say language is probably going to be harder.
I mean, it depends on how you define it.
Like if you mean absolute top-notch,
100% language understanding, I'll go with language.
- And so-
- But then if I show you a piece of paper
with letters on it, is that, you see what I mean?
It's like you have a vision system,
you say it's the best human level vision system.
I show you, I open a book and I show you letters.
Will it understand how these letters form
into word and sentences and meaning?
Is this part of the vision problem?
Where does vision end and language begin?
- Yeah, so Chomsky would say it starts at language.
So vision is just a little example
of the kind of structure and fundamental hierarchy
of ideas that's already represented in our brain somehow
that's represented through language.
But where does vision stop and language begin?
That's a really interesting question.
- So one possibility is that it's impossible
to achieve really deep understanding
in either images or language
without basically using the same kind of system.
So you're going to get the other for free.
- I think it's pretty likely that yes,
if we can get one, our machine learning
is probably that good that we can get the other.
But I'm not 100% sure.
And also, I think a lot of it really does depend
on your definitions.
- Definitions of?
- Like perfect vision.
Because reading is vision, but should it count?
- Yeah, to me, so my definition is
if a system looked at an image
and then a system looked at a piece of text
and then told me something about that
and I was really impressed.
- That's relative.
You'll be impressed for half an hour
and then you're gonna say, well,
I mean, all the systems do that,
but here's the thing they don't do.
- Yeah, but I don't have that with humans.
Humans continue to impress me.
- Is that true?
- Well, the ones, okay, so I'm a fan of monogamy,
so I like the idea of marrying somebody,
being with them for several decades.
So I believe in the fact that yes,
it's possible to have somebody continuously giving you
pleasurable, interesting, witty, new ideas, friends.
Yeah, I think so.
They continue to surprise you.
- The surprise, it's that injection of randomness
seems to be a nice source of, yeah,
continued inspiration, like the wit, the humor.
I think, yeah, that would be,
it's a very subjective test,
but I think if you have enough humans in the room.
- Yeah, I understand what you mean.
Yeah, I feel like I misunderstood
what you meant by impressing you.
I thought you meant to impress you with its intelligence,
with how well it understands an image.
I thought you meant something like,
I'm gonna show it a really complicated image
and it's gonna get it right, and you're gonna say, wow,
that's really cool.
Our systems of January 2020 have not been doing that.
- Yeah, no, I think it all boils down to
like the reason people click like on stuff on the internet,
which is like, it makes them laugh.
So it's like humor or wit or insight.
- I'm sure we'll get that as well.
- So forgive the romanticized question,
but looking back to you,
what is the most beautiful or surprising idea
in deep learning or AI in general you've come across?
- So I think the most beautiful thing about deep learning
is that it actually works.
And I mean it, because you got these ideas,
you got the little neural network,
you got the back propagation algorithm.
And then you got some theories as to, you know,
this is kind of like the brain.
So maybe if you make it large,
if you make the neural network large
and you train it on a lot of data,
then it will do the same function that the brain does.
And it turns out to be true, that's crazy.
And now we just train these neural networks
and you make them larger and they keep getting better.
And I find it unbelievable.
I find it unbelievable that this whole AI stuff
with neural networks works.
- Have you built up an intuition of why
are there little bits and pieces of intuitions,
of insights of why this whole thing works?
- I mean, some definitely.
While we know that optimization,
we now have good, you know,
we've had lots of empirical,
you know, huge amounts of empirical reasons
to believe that optimization should work
on most problems we care about.
- Do you have insights of why,
so you just said empirical evidence.
Is most of your,
sort of empirical evidence kind of convinces you,
it's like evolution is empirical.
It shows you that, look, this evolutionary process
seems to be a good way to design organisms
that survive in their environment.
But it doesn't really get you to the insights
of how the whole thing works.
- I think a good analogy is physics.
You know how you say, hey,
let's do some physics calculation
and come up with some new physics theory
and make some prediction.
But then you got around the experiment.
You know, you got around the experiment, it's important.
So it's a bit the same here,
except that maybe sometimes the experiment
came before the theory.
But it still is the case.
You know, you have some data
and you come up with some prediction.
You say, yeah, let's make a big neural network.
Let's train it.
And it's going to work much better than anything before it.
And it will in fact continue to get better
as you make it larger.
And it turns out to be true.
That's amazing when a theory is validated like this.
You know, it's not a mathematical theory.
It's more of a biological theory almost.
So I think there are not terrible analogies
between deep learning and biology.
I would say it's like the geometric mean
of biology and physics.
That's deep learning.
- The geometric mean of biology and physics.
I think I'm going to need a few hours
to wrap my head around that.
'Cause just to find the geometric,
just to find the set of what biology represents.
- Well, biology, in biology,
things are really complicated.
The theories are really, really,
it's really hard to have good predictive theory.
And in physics, the theories are too good.
In physics, people make these super precise theories
which make these amazing predictions.
And in machine learning, we're kind of in between.
- Kind of in between, but it'd be nice
if machine learning somehow helped us discover
the unification of the two
§05Transcript · 32:37–40:49
as opposed to sort of the in between.
But you're right.
You're kind of trying to juggle both.
So do you think there are still beautiful
and mysterious properties in neural networks
that are yet to be discovered?
- Definitely.
I think that we are still massively
underestimating deep learning.
- What do you think it'll look like?
Like what?
- If I knew, I would have done it.
But if you look at all the progress
from the past 10 years, I would say most of it,
I would say there've been a few cases
where things that felt like really new ideas showed up.
But by and large, it was every year we thought,
okay, deep learning goes this far.
Nope, it actually goes further.
And then the next year, okay, now this is big deep learning.
We are really done.
Nope, it goes further.
It just keeps going further each year.
So that means that we keep underestimating,
we keep not understanding it.
It has surprising properties all the time.
- Do you think it's getting harder and harder
to make progress?
- Need to make progress?
- It depends on what you mean.
I think the field will continue to make
very robust progress for quite a while.
I think for individual researchers,
especially people who are doing research,
it can be harder because there is a very large number
of researchers right now.
I think that if you have a lot of compute,
then you can make a lot of very interesting discoveries,
but then you have to deal with the challenge
of managing a huge compute cluster to run your experiments.
It's a little bit harder.
- So I'm asking all these questions
that nobody knows the answer to,
but you're one of the smartest people I know,
so I'm gonna keep asking.
So let's imagine all the breakthroughs that happen
in the next 30 years in deep learning.
Do you think most of those breakthroughs can be done
by one person with one computer?
Sort of in the space of breakthroughs,
do you think compute and large efforts will be necessary?
- I mean, I can't be sure.
When you say one computer, you mean how large?
(Lex laughing)
- You're clever.
I mean one GPU.
- I see.
I think it's pretty unlikely.
I think it's pretty unlikely.
I think that there are many...
The stack of deep learning is starting to be quite deep.
If you look at it, you've got all the way from the ideas,
the systems to build the data sets,
the distributed programming,
the building the actual cluster, the GPU programming,
putting it all together.
So now the stack is getting really deep,
and I think it can be quite hard for a single person
to become, to be world-class
in every single layer of the stack.
- What about what like Vladimir Vapnik really insists on
is taking MNIST and trying to learn from very few examples.
So being able to learn more efficiently.
Do you think there'll be breakthroughs in that space
that may not need the huge compute?
- I think there will be a large number of breakthroughs
in general that will not need a huge amount of compute.
So maybe I should clarify that.
I think that some breakthroughs will require a lot of compute
and I think building systems which actually do things
will require a huge amount of compute.
That one is pretty obvious.
If you want to do X and X requires a huge neural net,
you gotta get a huge neural net.
But I think there will be lots of,
I think there is lots of room for very important work
being done by small groups and individuals.
- Can you maybe sort of on the topic of the science
of deep learning, talk about one of the recent papers
that you released, the deep double descent.
Where bigger models and more data hurt.
I think it's a really interesting paper.
Can you describe the main idea?
- Yeah, definitely.
So what happened is that some, over the years,
some small number of researchers noticed that
it is kind of weird that when you make
the neural network larger, it works better
and it seems to go in contradiction with statistical ideas.
And then some people made an analysis showing
that actually you got this double descent bump.
And what we've done was to show that double descent occurs
for pretty much all practical deep learning systems.
And that it'll be also, so can you step back?
What's the X axis and the Y axis of a double descent plot?
Okay, great.
So you can look, you can do things like,
you can take a neural network
and you can start increasing its size slowly
while keeping your dataset fixed.
So if you increase the size of the neural network slowly,
and if you don't do early stopping,
that's a pretty important detail.
Then when the neural network is really small,
you make it larger,
you get a very rapid increase in performance.
Then you continue to make it larger.
And at some point performance will get worse.
And it gets the worst exactly at the point
at which it achieves zero training error,
precisely zero training loss.
And then as you make it larger, it starts to get better again.
And it's kind of counterintuitive
because you'd expect deep learning phenomena
to be monotonic.
And it's hard to be sure what it means,
but it also occurs in the case of linear classifiers.
And the intuition basically boils down to the following.
When you have a large dataset and a small model,
then small, tiny random, so basically what is overfitting?
Overfitting is when your model is somehow very sensitive
to the small, random, unimportant stuff in your dataset.
- In the training dataset.
- In the training dataset, precisely.
So if you have a small model and you have a big dataset,
and there may be some random thing,
some training cases are randomly in the dataset
and others may not be there,
but the small model is kind of insensitive
to this randomness because there is pretty much
no uncertainty about the model.
When the dataset is large.
So, okay, so at the very basic level to me,
it is the most surprising thing
that neural networks don't overfit every time,
very quickly, before ever being able to learn anything.
There are a huge number of parameters.
So here is, so there is one way, okay,
so maybe, so let me try to give the explanation
and maybe that will be, that will work.
So you've got a huge neural network.
Let's suppose you've got a,
you have a huge neural network,
you have a huge number of parameters.
Now let's pretend everything is linear,
which is not, let's just pretend.
Then there is this big subspace
where your neural network achieves zero error.
And SGD is going to find approximately the point--
- Stochastic gradient, that's right.
- Approximately the point with the smallest norm
in that subspace.
- Okay.
- And that can also be proven to be insensitive
to the small randomness in the data
when the dimensionality is high.
But when the dimensionality of the data
is equal to the dimensionality of the model,
then there is a one-to-one correspondence
between all the datasets and the models.
So small changes in the dataset
actually lead to large changes in the model
and that's why performance gets worse.
So this is the best explanation, more or less.
- So then it would be good for the model
to have more parameters, so to be bigger than the data.
- That's right, but only if you don't early stop.
If you introduce early stop in your regularization,
you can make the double descent bump
almost completely disappear.
- What is early stop?
- Early stop is when you train your model
and you monitor your validation performance.
And then if at some point,
validation performance starts to get worse,
you say, "Okay, let's stop training.
"We are good, we are good enough."
- So the magic happens after that moment,
so you don't wanna do the early stopping.
- Well, if you don't do the early stopping,
you get the very pronounced double descent.
- Do you have any intuition why this happens?
- Double descent?
Oh, sorry, early stopping?
- No, the double descent.
- Well, yeah, so I try, let's see.
The intuition is basically, is this,
that when the dataset has as many degrees of freedom
as the model, then there is a one-to-one correspondence
between them and so small changes to the dataset
§06Transcript · 40:52–49:50
lead to noticeable changes in the model.
So your model is very sensitive to all the randomness.
It is unable to discard it,
whereas it turns out that when you have
a lot more data than parameters
or a lot more parameters than data,
the resulting solution will be insensitive
to small changes in the dataset.
- Oh, so it's able to, that's nicely put,
discard the small changes, the randomness.
- Exactly, the spurious correlation which you don't want.
- Jeff Hinton suggested we need to throw back propagation.
We already kind of talked about this a little bit,
but he suggested that we need to throw away
back propagation and start over.
I mean, of course, some of that is a little bit
wit and humor, but what do you think,
what could be an alternative method
of training neural networks?
- Well, the thing that he said precisely is that
to the extent that you can't find back propagation
in the brain, it's worth seeing if we can learn something
from how the brain learns, but back propagation
is very useful and we should keep using it.
- Oh, you're saying that once we discover
the mechanism of learning in the brain
or any aspects of that mechanism,
we should also try to implement that in neural networks?
- If it turns out that we can't find back propagation
in the brain.
- If we can't find back propagation in the brain.
Well, so I guess your answer to that is
back propagation is pretty damn useful,
so why are we complaining?
- I mean, I personally am a big fan of back propagation.
I think it's a great algorithm because it solves
an extremely fundamental problem, which is
finding a neural circuit subject to some constraints.
I don't see that problem going away,
so that's why I really, I think it's pretty unlikely
that we'll have anything which is going to be
dramatically different.
It could happen, but I wouldn't bet on it right now.
- So let me ask a sort of big picture question.
Do you think neural networks can be made to reason?
- Why not?
- Well, if you look, for example, at AlphaGo or AlphaZero,
the neural network of AlphaZero plays Go,
which we all agree is a game that requires reasoning,
better than 99.9% of all humans.
Just the neural network, without the search,
just the neural network itself.
Doesn't that give us an existence proof
that neural networks can reason?
- To push back and disagree a little bit,
we all agree that Go is reasoning.
I think I agree.
I don't think it's a trivial, so obviously,
reasoning, like intelligence, is a loose,
gray area term, a little bit.
Maybe you disagree with that.
But yes, I think it has some of the same elements
of reasoning.
Reasoning is almost akin to search, right?
There's a sequential element of step-wise consideration
of possibilities, and sort of building on top
of those possibilities in a sequential manner
until you arrive at some insight.
So yeah, I guess playing Go is kind of like that.
And when you have a single neural network doing that
without search, it's kind of like that.
So there's an existence proof in a particular
constrained environment that a process akin to
what many people call reasoning exists.
But more general kind of reasoning.
- So off the board.
- There is one other existence proof.
- Oh boy, which one?
Us humans?
- Yes.
- Okay.
All right, so do you think the architecture
that will allow neural networks to reason
will look similar to the neural network architectures
we have today?
- I think it will.
I think, well, I don't wanna make
two overly definitive statements.
I think it's definitely possible that
the neural networks that will produce
the reasoning breakthroughs of the future
will be very similar to the architectures that exist today.
Maybe a little bit more recurrent,
maybe a little bit deeper.
But these neural nets are so insanely powerful.
Why wouldn't they be able to learn to reason?
Humans can reason, so why can't neural networks?
So do you think the kind of stuff we've seen
neural networks do is a kind of just weak reasoning?
So it's not a fundamentally different process?
Again, this is stuff nobody knows the answer to.
- So when it comes to our neural networks,
the thing which I would say is that
neural networks are capable of reasoning.
But if you train a neural network on a task
which doesn't require reasoning, it's not going to reason.
This is a well-known effect where the neural network
will solve exactly the, it will solve the problem
that you pose in front of it in the easiest way possible.
- Right, that takes us to the,
to one of the brilliant sort of ways
you've described neural networks, which is,
you've referred to neural networks
as the search for small circuits.
And maybe general intelligence
as the search for small programs,
which I found as a metaphor very compelling.
Can you elaborate on that difference?
- Yeah, so the thing which I said precisely was that
if you can find the shortest program
that outputs the data at your disposal,
then you will be able to use it
to make the best prediction possible.
And that's a theoretical statement
which can be proved mathematically.
Now, you can also prove mathematically that it is,
that finding the shortest program
which generates some data is not a computable operation.
No finite amount of compute can do this.
So then with neural networks,
neural networks are the next best thing
that actually works in practice.
We are not able to find the best,
the shortest program which generates our data,
but we are able to find, you know, a small,
but now that statement should be amended,
even a large circuit which fits our data in some way.
- Well, I think what you meant by the small circuit
is the smallest needed circuit.
- Well, the thing which I would change now,
back then I really haven't fully internalized
the over-parameterized results.
The things we know about over-parameterized neural nets,
now I would phrase it as a large circuit
whose weights contain a small amount of information,
which I think is what's going on.
If you imagine the training process of a neural network
as you slowly transmit entropy
from the dataset to the parameters,
then somehow the amount of information in the weights
ends up being not very large,
which would explain why they generalize so well.
- So that's, the large circuit might be one that's helpful
for the generalization.
- Yeah, something like this.
- But do you see it important to be able to try
to learn something like programs?
- I mean, if we can, definitely.
I think it's kind of, the answer is kind of yes,
if we can do it.
We should do things that we can't do it.
It's the reason we are pushing on deep learning,
the fundamental reason, the root cause
is that we are able to train them.
So in other words, training comes first.
We've got our pillar, which is the training pillar.
And now we are trying to contort our neural networks
around the training pillar.
We gotta stay trainable.
This is an invariant we cannot violate.
And so being trainable means starting from scratch,
knowing nothing, you can actually pretty quickly
converge towards knowing a lot or even slowly.
But it means that given the resources at your disposal,
you can train the neural net
and get it to achieve useful performance.
- Yeah, that's a pillar we can't move away from.
- That's right, because if you can,
and whereas if you say, hey, let's find the shortest program,
well, we can't do that.
So it doesn't matter how useful that would be.
We can do it, so we won't.
- So do you think, you kind of mentioned
that neural networks are good at finding small circuits
or large circuits.
Do you think then the matter of finding small programs
is just the data?
- No.
- So the, sorry, not the size or character,
the type of data.
Sort of ask giving it programs.
- Well, I think the thing is that right now,
finding, there are no good precedents
of people successfully finding programs really well.
And so the way you'd find programs
is you'd train a deep neural network to do it basically.
- Right.
- Which is the right way to go about it.
- But there's not good illustrations of that.
- It hasn't been done yet,
§07Transcript · 49:51–58:23
but in principle, it should be possible.
- Can you elaborate a little bit?
What's your insight in principle?
And put another way, you don't see why it's not possible.
- Well, it's kind of like more, it's more a statement of,
I think that it's unwise to bet against deep learning.
And if it's a cognitive function
that humans seem to be able to do,
then it doesn't take too long for some deep neural net
to pop up that can do it too.
- Yeah, I'm there with you.
I've stopped betting against neural networks at this point
because they continue to surprise us.
What about long-term memory?
Can neural networks have long-term memory
or something like knowledge basis?
So being able to aggregate important information
over long periods of time that would then serve
as useful sort of representations of state
that you can make decisions by.
So have a long-term context
based on what you make in the decision.
- So in some sense, the parameters already do that.
The parameters are an aggregation of the day,
of the neural, of the entirety of the neural experience.
And so they count as the long, as long-term knowledge.
And people have trained various neural nets
to act as knowledge bases and, you know,
investigated with invest,
people have investigated language models as knowledge basis.
So there is work, there is work there.
- Yeah, but in some sense, do you think in every sense,
do you think there's a, it's all just a matter
of coming up with a better mechanism
of forgetting the useless stuff
and remembering the useful stuff?
'Cause right now, I mean, there's not been mechanisms
that do remember really long-term information.
- What do you mean by that precisely?
- Precisely, I like the word precisely.
So I'm thinking of the kind of compression of information
the knowledge bases represent, sort of creating a,
now, I apologize for my sort of human-centric thinking
about what knowledge is, 'cause neural networks
aren't interpretable necessarily
with the kind of knowledge they have discovered.
But a good example for me is knowledge bases,
being able to build up over time something like
the knowledge that Wikipedia represents.
It's a really compressed, structured,
(scoffs)
knowledge base.
Obviously not the actual Wikipedia or the language,
but like a semantic web,
the dream that semantic web represented.
So it's a really nice compressed knowledge base,
or something akin to that in a non-interpretable sense
as neural networks would have.
- Well, the neural networks would be non-interpretable
if you look at their weights,
but their outputs should be very interpretable.
- Okay, so yeah, how do you make very smart neural networks
like language models interpretable?
- Well, you ask them to generate some text,
and the text will generally be interpretable.
- Do you find that the epitome of interpretability,
like can you do better?
'Cause you can't, okay, I would like to know
what does it know and what doesn't it know?
I would like the neural network to come up with examples
where it's completely dumb,
and examples where it's completely brilliant.
And the only way I know how to do that now
is to generate a lot of examples and use my human judgment.
But it would be nice if a neural network
had some self-awareness about it.
- Yeah, 100%.
I'm a big believer in self-awareness,
and I think neural net self-awareness
will allow for things like the capabilities,
like the ones you described,
like for them to know what they know
and what they don't know,
and for them to know where to invest
to increase their skills most optimally.
And to your question of interpretability,
there are actually two answers to that question.
One answer is, you know, we have the neural net,
so we can analyze the neurons,
and we can try to understand what the different neurons
and different layers mean.
And you can actually do that,
and OpenAI has done some work on that.
But there is a different answer,
which is that, I would say,
that's the human-centric answer,
where you say, you know, you look at a human being,
you can't read, you know,
how do you know what a human being is thinking?
You ask them, you say, hey, what do you think about this?
What do you think about that?
And you get some answers.
The answers you get are sticky,
in the sense you already have a mental model.
You already have an, yeah,
mental model of that human being.
You already have an understanding of,
like a big conception of what it,
of that human being, how they think,
how what they know, how they see the world,
and then everything you ask, you're adding onto that.
And that stickiness seems to be,
that's one of the really interesting qualities
of the human being, is that information is sticky.
You don't, you seem to remember the useful stuff,
aggregate it well, and forget most of the information
that's not useful.
That process, but that's also pretty similar
to the process that neural networks do.
It's just that neural networks are much crappier
at this time.
It doesn't seem to be fundamentally that different.
But just to stick on reasoning for a little longer,
you said, why not?
Why can't I reason?
What's a good, impressive feat,
benchmark to you of reasoning
that you'll be impressed by
if neural networks were able to do?
Is that something you already have in mind?
- Well, I think writing really good code.
I think proving really hard theorems.
Solving open-ended problems with out-of-the-box solutions.
- And sort of theorem type mathematical problems.
- Yeah, I think those ones are a very natural example
as well.
If you can prove an unproven theorem,
then it's hard to argue, you don't reason.
And so by the way, and this comes back to the point
about the hard results.
If you've got a hard, if you have,
machine learning, deep learning as a field is very fortunate
because we have the ability to sometimes produce
these unambiguous results.
And when they happen, the debate changes,
the conversation changes.
We have the ability to produce conversation changing results.
- Conversation, and then of course, just like you said,
people kind of take that for granted
and say that wasn't actually a hard problem.
Well, I mean, at some point,
you'll probably run out of hard problems.
Yeah, that whole mortality thing is kind of a sticky problem
that we haven't quite figured out.
Maybe we'll solve that one.
I think one of the fascinating things
in your entire body of work,
but also the work at OpenAI recently,
one of the conversation changers has been
in the world of language models.
Can you briefly kind of try to describe the recent history
of using neural networks in the domain of language and text?
- Well, there's been lots of history.
I think the Elman network was a small,
tiny recurrent neural network applied to language
back in the '80s.
So the history is really, you know, fairly long at least.
And the thing that started,
the thing that changed the trajectory
of neural networks and language
is the thing that changed the trajectory
of all deep learning, and that's data and compute.
So suddenly you move from small language models,
which learn a little bit.
And with language models in particular,
there's a very clear explanation
for why they need to be large to be good,
because they're trying to predict the next word.
So when you don't know anything,
you'll notice very, very broad strokes,
surface level patterns,
like sometimes there are characters
and there is a space between those characters.
You'll notice this pattern.
And you'll notice that sometimes there is a comma
and then the next character is a capital letter.
You'll notice that pattern.
Eventually you may start to notice
that there are certain words occur often.
You may notice that spellings are a thing.
You may notice syntax.
And when you get really good at all these,
you start to notice the semantics.
You start to notice the facts.
But for that to happen,
the language model needs to be larger.
- So let's linger on that,
'cause that's where you and Noam Chomsky disagree.
So you think we're actually taking incremental steps,
sort of larger network, larger compute,
§08Transcript · 58:25–1:06:04
we'll be able to get to the semantics,
be able to understand language
without what Noam likes to sort of think of
as a fundamental understandings
of the structure of language,
like imposing your theory of language
onto the learning mechanism.
So you're saying the learning,
you can learn from raw data,
the mechanism that underlies language.
- Well, I think it's pretty likely,
but I also wanna say that I don't really
know precisely what Chomsky means when he talks about him.
You said something about imposing
your structure on language.
I'm not 100% sure what he means,
but empirically it seems that
when you inspect those larger language models,
they exhibit signs of understanding the semantics,
whereas the smaller language models do not.
We've seen that a few years ago
when we did work on the sentiment neuron,
we trained a small, you know,
smallish LSTM to predict the next character
in Amazon reviews.
And we noticed that when you increase the size of the LSTM
from 500 LSTM cells to 4,000 LSTM cells,
then one of the neurons starts to represent the sentiment
of the article, of, sorry, of the review.
Now, why is that?
Sentiment is a pretty semantic attribute.
It's not a syntactic attribute.
- And for people who might not know,
I don't know if that's a standard term,
but sentiment is whether it's a positive or negative review.
- That's right.
Like, is the person happy with something
or is the person unhappy with something?
And so here we had very clear evidence
that a small neural net does not capture sentiment
while a large neural net does.
And why is that?
Well, our theory is that at some point
you run out of syntax to models,
you start to gotta focus on something else.
- And with size, you quickly run out of syntax to model,
and then you really start to focus on the semantics,
would be the idea.
- That's right.
And so I don't wanna imply that our models
have complete semantic understanding
because that's not true,
but they definitely are showing signs
of semantic understanding, partial semantic understanding,
but the smaller models do not show those signs.
- Can you take a step back and say, what is GPT-2,
which is one of the big language models
that was the conversation changer
in the past couple of years?
- Yeah, so GPT-2 is a transformer
with one and a half billion parameters
that was trained on about 40 billion tokens of text,
which were obtained from web pages
that were linked to from Reddit articles
with more than three upvotes.
- And what's a transformer?
- The transformer, it's the most important advance
in neural network architectures in recent history.
- What is attention maybe too?
'Cause I think that's an interesting idea,
not necessarily sort of technically speaking,
but the idea of attention
versus maybe what recurring neural networks represent.
- Yeah, so the thing is the transformer
is a combination of multiple ideas simultaneously
of which attention is one.
- Do you think attention is the key?
- No, it's a key, but it's not the key.
The transformer is successful
because it is the simultaneous combination
of multiple ideas.
And if you were to remove either idea,
it would be much less successful.
So the transformer uses a lot of attention,
but attention existed for a few years.
So that can't be the main innovation.
The transformer is designed in such a way
that it runs really fast on the GPU.
And that makes a huge amount of difference.
This is one thing.
The second thing is that transformer is not recurrent.
And that is really important too,
because it is more shallow
and therefore much easier to optimize.
So in other words, it uses attention.
It is a really great fit to the GPU
and it is not recurrent,
so therefore less deep and easier to optimize.
And the combination of those factors make it successful.
So now it makes great use of your GPU.
It allows you to achieve better results
for the same amount of compute.
And that's why it's successful.
- Were you surprised how well transformers worked
and GPT-2 worked?
So you worked on language.
You've had a lot of great ideas
before transformers came about in language.
So you got to see the whole set of revolutions
before and after.
Were you surprised?
- Yeah, a little.
- A little?
- Yeah.
I mean, it's hard to remember
because you adapt really quickly,
but it definitely was surprising.
It definitely was.
In fact, you know what?
I'll retract my statement.
It was pretty amazing.
It was just amazing to see generate this text of this.
And you know, you got to keep in mind
that at that time, you've seen all this progress in GANs,
in improving the samples produced by GANs were just amazing.
You have these realistic faces,
but text hasn't really moved that much.
And suddenly we moved from, you know,
whatever GANs were in 2015
to the best, most amazing GANs in one step.
And that was really stunning.
Even though theory predicted,
yeah, you train a big language model,
of course you should get this.
But then to see it with your own eyes, it's something else.
- And yet we adapt really quickly.
And now there's sort of some cognitive scientists
write articles saying that GPT-2 models
don't truly understand language.
So we adapt quickly to how amazing
the fact that they're able to model the language so well is.
So what do you think is the bar?
- For what?
- For impressing us that it-
- I don't know.
- Do you think that bar will continuously be moved?
- Definitely.
I think when you start to see
really dramatic economic impact, that's when,
I think that's in some sense the next barrier.
Because right now, if you think about the work in AI,
it's really confusing.
It's really hard to know what to make of all these advances.
It's kind of like, okay, you got an advance.
Now you can do more things.
And you got another improvement.
And you got another cool demo.
At some point, I think people who are outside of AI,
they can no longer distinguish this progress anymore.
- So we were talking offline
about translating Russian to English
and how there's a lot of brilliant work in Russian
that the rest of the world doesn't know about.
That's true for Chinese.
It's true for a lot of scientists
and just artistic work in general.
Do you think translation is the place
where we're going to see sort of economic big impact?
- I don't know.
I think there is a huge number of,
I mean, first of all, I would want to,
I want to point out that translation already today is huge.
I think billions of people interact
with big chunks of the internet
primarily through translation.
So translation is already huge
and it's hugely, hugely positive too.
I think self-driving is going to be hugely impactful.
And that's, you know, it's unknown exactly when it happens,
but again, I would not bet against deep learning.
So I--
- So that's deep learning in general, but you think--
- Deep learning for self-driving.
- Yes, deep learning for self-driving.
But I was talking about sort of language models.
- I see.
- Just to check. - I veered off a little bit.
- Just to check.
You're not seeing a connection
between driving and language.
- No, no. - Okay.
- Or rather, both use neural nets.
- That'd be a poetic connection.
I think there might be some, like you said,
there might be some kind of unification
towards a kind of multitask transformers
that can take on both language and vision tasks.
That'd be an interesting unification.
Now let's see, what can I ask about GPT-2 more?
- It's simple, so not much to ask.
§09Transcript · 1:06:06–1:15:06
It's, you take a transform, you make it bigger,
give it more data,
and suddenly it does all those amazing things.
- Yeah, one of the beautiful things is that GPT,
the transformers are fundamentally simple
to explain, to train.
Do you think bigger will continue
to show better results in language?
- Probably.
- Sort of like what are the next steps with GPT-2,
do you think?
- I mean, I think for sure seeing what larger versions
can do is one direction.
Also, I mean, there are many questions.
There's one question which I'm curious about,
and that's the following.
So right now, GPT-2,
so we feed it all this data from the internet,
which means that it needs to memorize
all those random facts about everything in the internet.
And it would be nice if the model
could somehow use its own intelligence
to decide what data it wants to accept
and what data it wants to reject.
Just like people,
people don't learn all data indiscriminately.
We are super selective about what we learn.
And I think this kind of active learning,
I think, would be very nice to have.
- Yeah, listen, I love active learning.
So let me ask, does the selection of data,
can you just elaborate that a little bit more?
Do you think the selection of data is,
like, I have this kind of sense
that the optimization of how you select data,
so the active learning process,
is going to be a place for a lot of breakthroughs,
even in the near future,
because there hasn't been many breakthroughs there
that are public.
I feel like there might be private breakthroughs
that companies keep to themselves,
'cause it's a fundamental problem that has to be solved
if you wanna solve self-driving,
if you wanna solve a particular task.
What do you think about the space in general?
- Yeah, so I think that for something like active learning,
or in fact, for any kind of capability, like active learning,
the thing that it really needs is a problem.
It needs a problem that requires it.
It's very hard to do research about the capability
if you don't have a task,
because then what's going to happen
is that you will come up with an artificial task,
get good results, but not really convince anyone.
- Right, like, we're now past the stage
where getting a result on MNIST,
some clever formulation of MNIST will convince people.
- That's right.
In fact, you could quite easily come up
with a simple active learning scheme on MNIST
and get a 10x speedup, but then, so what?
And I think that with active learning,
the active learning will naturally arise
as problems that require it pop up.
That's my take on it.
- There's another interesting thing
that OpenAI has brought up with GPT-2,
which is when you create
a powerful artificial intelligence system,
and it was unclear what kind of detrimental,
once you release GPT-2,
what kind of detrimental effect it'll have,
because if you have a model
that can generate pretty realistic text,
you can start to imagine that it would be used by bots
in some way that we can't even imagine.
So there's this nervousness about what it's possible to do.
So you did a really kind of brave
and I think profound thing,
which is start a conversation about this.
Like, how do we release
powerful artificial intelligence models to the public?
If we do it all, how do we privately discuss
with other, even competitors,
about how we manage the use of the systems and so on?
So from this whole experience,
you released a report on it,
but in general, are there any insights
that you've gathered from just thinking about this,
about how you release models like this?
- I mean, I think that my take on this
is that the field of AI has been in a state of childhood,
and now it's exiting that state
and it's entering a state of maturity.
What that means is that AI is very successful
and also very impactful,
and its impact is not only large, but it's also growing.
And so for that reason, it seems wise to start thinking
about the impact of our systems before releasing them,
maybe a little bit too soon,
rather than a little bit too late.
And with the case of GPT-2, like I mentioned earlier,
the results really were stunning,
and it seemed plausible.
It didn't seem certain.
It seemed plausible that something like GPT-2
could easily use to reduce the cost of disinformation.
And so there was a question
of what's the best way to release it,
and a staged release seemed logical.
A small model was released,
and there was time to see the...
Many people use these models in lots of cool ways.
There've been lots of really cool applications.
There haven't been any negative applications we know of,
and so eventually it was released.
But also other people replicated similar models.
- That's an interesting question, though, that we know of.
So in your view, staged release
is at least part of the answer to the question of how do we...
What do we do once we create a system like this?
- It's part of the answer, yes.
- Is there any other insights?
Like, say you don't wanna release the model at all
because it's useful to you for whatever the business is.
- Well, plenty of people don't release models already.
- Right, of course,
but is there some moral, ethical responsibility
when you have a very powerful model to sort of communicate?
Just as you said, when you had GPT-2,
it was unclear how much it could be used for misinformation.
It's an open question,
and getting an answer to that might require
that you talk to other really smart people
that are outside of your particular group.
Please tell me there's some optimistic pathway
for people across the world
to collaborate on these kinds of cases.
Or is it still really difficult from one company
to talk to another company?
- So it's definitely possible.
It's definitely possible to discuss these kind of models
with colleagues elsewhere
and to get their take on what to do.
- How hard is it though?
- I mean...
- Do you see that happening?
- I think that's a place where it's important
to gradually build trust between companies
because ultimately, all the AI developers
are building technology
which is going to be increasingly more powerful.
And so it's...
The way to think about it
is that ultimately we're all in it together.
- Yeah, it's...
I tend to believe in the better angels of our nature,
but I do hope that...
That when you build a really powerful AI system
in a particular domain,
that you also think about
the potential negative consequences of...
Yeah.
It's an interesting and scary possibility
that there'll be a race for AI development
that would push people to close that development
and not share ideas with others.
- I don't love this.
I've been a pure academic for 10 years.
I really like sharing ideas and it's fun.
It's exciting.
- What do you think it takes to...
Let's talk about AGI a little bit.
What do you think it takes to build a system
of human level intelligence?
We talked about reasoning.
We talked about long-term memory,
but in general, what does it take, do you think?
- Well, I can't be sure,
but I think the deep learning plus maybe another small idea.
- Do you think self-play will be involved?
So like you've spoken about the powerful mechanism
of self-play where systems learn by
sort of exploring the world in a competitive setting
against other entities that are similarly skilled as them
and so incrementally improve in this way.
Do you think self-play will be a component
of building an AGI system?
- Yeah, so what I would say to build AGI,
I think is going to be deep learning plus some ideas.
And I think self-play will be one of those ideas.
I think that that is a very...
Self-play has this amazing property that it can surprise us
in truly novel ways.
For example, pretty much every self-play system,
both our Dota bot, I don't know if OpenAI had a release
about multi-agent where you had two little agents
who were playing hide and seek.
And of course, also AlphaZero.
§10Transcript · 1:15:08–1:23:49
They will all produce surprising behaviors.
They all produce behaviors that we didn't expect.
They are creative solutions to problems.
And that seems like an important part of AGI
that our systems don't exhibit routinely right now.
And so that's why I like this area, I like this direction
because of its ability to surprise us.
- To surprise us.
And an AGI system would surprise us fundamentally.
- Yes, and to be precise, not just a random surprise,
but to find a surprising solution to a problem
that's also useful.
- Right.
Now, a lot of the self-play mechanisms have been used
in the game context or at least in a simulation context.
How far along the path to AGI
do you think will be done in simulation?
How much faith, promise do you have in simulation
versus having to have a system that operates
in the real world, whether it's the real world
of digital real-world data or real world,
like actual physical world of robotics?
- I don't think it's an either or.
I think simulation is a tool and it helps.
It has certain strengths and certain weaknesses
and we should use it.
- Yeah, but, okay, I understand that.
That's true, but one of the criticisms of self-play,
one of the criticisms of reinforcement learning
is one of the,
its current power, its current results,
while amazing, have been demonstrated
in a simulated environments
or very constrained physical environments.
Do you think it's possible to escape them,
escape the simulated environments
and be able to learn in non-simulated environments?
Or do you think it's possible to also just simulate
in a photorealistic and physics realistic way,
the real world in a way that we can solve real problems
with self-play in simulation?
- So I think that transfer from simulation
to the real world is definitely possible
and has been exhibited many times
in by many different groups.
It's been especially successful in vision.
Also, OpenAI in the summer has demonstrated a robot hand
which was trained entirely in simulation
in a certain way that allowed
for sim-to-real transfer to occur.
- Is this for the Rubik's Cube?
- Yes, that's right.
- I wasn't aware that was trained in simulation.
- It was trained in simulation entirely.
- Really, so it wasn't in the physics,
the hand wasn't trained?
- No, 100% of the training was done in simulation
and the policy that was learned in simulation
was trained to be very adaptive.
So adaptive that when you transfer it,
it could very quickly adapt to the physical world.
- So the kind of perturbations with the giraffe
or whatever the heck it was,
were those part of the simulation?
- Well, the simulation was generally,
so the simulation was trained to be robust
to many different things,
but not the kind of perturbations we've had in the video.
So it's never been trained with a glove,
it's never been trained with a stuffed giraffe.
- So in theory, these are novel perturbations.
- Correct, it's not in theory, in practice.
- That those are novel perturbations?
Well, that's okay.
That's a clean, small scale, but clean example
of a transfer from the simulated world
to the physical world.
- Yeah, and I will also say that I expect
the transfer capabilities of deep learning
to increase in general.
And the better the transfer capabilities are,
the more useful simulation will become.
Because then you could take,
you could experience something in simulation
and then learn a moral of the story,
which you could then carry with you to the real world.
As humans do all the time when they play computer games.
- So let me ask sort of a embodied question,
staying on AGI for a sec.
Do you think AGI says that we need to have a body?
We need to have some of those human elements
of self-awareness, consciousness,
sort of fear of mortality, sort of self-preservation
in the physical space, which comes with having a body.
- I think having a body will be useful.
I don't think it's necessary.
But I think it's very useful to have a body for sure,
because you can learn a whole new,
you can learn things which cannot be learned without a body.
But at the same time, I think that you can,
if you don't have a body, you could compensate for it
and still succeed.
- You think so?
- Yes.
Well, there is evidence for this.
For example, there are many people
who were born deaf and blind,
and they were able to compensate for the lack of modalities.
I'm thinking about Helen Kaler specifically.
- So even if you're not able to physically interact
with the world, and if you're not able to,
I mean, I actually was getting at,
maybe let me ask on the more particular,
I'm not sure if it's connected to having a body or not,
but the idea of consciousness.
And a more constrained version of that is self-awareness.
Do you think an AGI system should have consciousness?
We can't define consciousness,
whatever the heck you think consciousness is.
- Yeah, hard question to answer,
given how hard it is to define it.
- Do you think it's useful to think about?
- I mean, it's definitely interesting.
It's fascinating.
I think it's definitely possible
that our systems will be conscious.
- Do you think that's an emergent thing
that just comes from,
do you think consciousness could emerge
from the representation that's stored within your networks?
So like that it naturally just emerges
when you become more and more,
you're able to represent more and more of the world?
- Well, I'd say, I'd make the following argument,
which is humans are conscious,
and if you believe that artificial neural nets
are sufficiently similar to the brain,
then there should at least exist artificial neural nets
we should be conscious to.
- You're leaning on that existence proof pretty heavily.
Okay.
- But that's the best answer I can give.
- No, I know, I know, I know.
There's still an open question
if there's not some magic in the brain that we're not,
I mean, I don't mean a non-materialistic magic,
but that the brain might be a lot more complicated
and interesting than we give it credit for.
- If that's the case, then it should show up,
and at some point we will find out
that we can't continue to make progress,
but I think it's unlikely.
- So we talk about consciousness,
but let me talk about another poorly defined concept
of intelligence.
Again, we've talked about reasoning,
we've talked about memory,
what do you think is a good test of intelligence for you?
Are you impressed by the test that Alan Turing formulated
with the imitation game with natural language?
Is there something in your mind
that you will be deeply impressed by
if a system was able to do?
- I mean, lots of things.
There's a certain frontier of capabilities today,
and there exist things outside of that frontier,
and I would be impressed by any such thing.
For example, I would be impressed by a deep learning system
which solves a very pedestrian task,
like machine translation or computer vision task
or something which never makes mistake
a human wouldn't make under any circumstances.
I think that is something
which have not yet been demonstrated,
and I would find it very impressive.
- Yeah, so right now they make mistakes,
they might be more accurate than human beings,
but they still, they make a different set of mistakes.
- So I would guess that a lot of the skepticism
that some people have about deep learning
is when they look at their mistakes and they say,
"Well, those mistakes, they make no sense."
Like if you understood the concept,
you wouldn't make that mistake.
And I think that changing that would inspire me,
that would be, yes, this is progress.
- Yeah, that's a really nice way to put it.
But I also just don't like that human instinct
to criticize a model as not intelligent.
That's the same instinct as we do
when we criticize any group of creatures as the other.
Because it's very possible that GPT-2
is much smarter than human beings at many things.
- That's definitely true.
It has a lot more breadth of knowledge.
- Yes, breadth of knowledge,
and even perhaps depth on certain topics.
- It's kind of hard to judge what depth means,
but there's definitely a sense
in which humans don't make mistakes that these models do.
§11Transcript · 1:23:54–1:33:24
- Yes, the same is applied to autonomous vehicles.
The same is probably gonna continue being applied
to a lot of artificial intelligence systems.
We find, this is the annoying thing,
this is the process of, in the 21st century,
the process of analyzing the progress of AI
is the search for one case where the system fails
in a big way where humans would not,
and then many people writing articles about it,
and then broadly as the public generally gets convinced
that the system is not intelligent.
And we like pacify ourselves by thinking it's not intelligent
because of this one anecdotal case.
And this seems to continue happening.
- Yeah, I mean, there is truth to that.
Although I'm sure that plenty of people
are also extremely impressed by the system that exists today.
But I think this connects to the earlier point we discussed
that it's just confusing to judge progress in AI.
- Yeah.
- And you have a new robot demonstrating something.
How impressed should you be?
And I think that people will start to be impressed
once AI starts to really move the needle on the GDP.
- So you're one of the people that might be able
to create an AGI system here, not you, but you and OpenAI.
If you do create an AGI system
and you get to spend sort of the evening with it, him, her,
what would you talk about, do you think?
- The very first time?
- First time.
- Well, the first time I would just ask all kinds
of questions and try to get it to make a mistake.
And I would be amazed that it doesn't make mistakes.
And I just keep asking broad questions.
- What kind of questions do you think,
would they be factual or would they be personal,
emotional, psychological?
What do you think?
- All of the above.
(Lex laughing)
- Would you ask for advice?
- Definitely.
(Lex laughing)
I mean, why would I limit myself
talking to a system like this?
- Now, again, let me emphasize the fact
that you truly are one of the people
that might be in the room where this happens.
So let me ask sort of a profound question about,
I just talked to a Stalin historian.
I've been talking to a lot of people who are studying power.
Abraham Lincoln said, "Nearly all men can stand adversity,
"but if you want to test a man's character, give him power."
I would say the power of the 21st century,
maybe the 22nd, but hopefully the 21st,
would be the creation of an AGI system
and the people who have control,
direct possession and control of the AGI system.
So what do you think, after spending that evening
having a discussion with the AGI system,
what do you think you would do?
- Well, the ideal world I'd like to imagine
is one where humanity are like the board members
of a company where the AGI is the CEO.
So it would be, I would like,
the picture which I would imagine
is you have some kind of different entities,
different countries or cities,
and the people that leave their vote
for what the AGI that represents them should do,
and then AGI that represents them goes and does it.
I think a picture like that, I find very appealing.
And you could have multiple,
you would have an AGI for a city, for a country,
and it would be trying to, in effect,
take the democratic process to the next level.
- And the board can almost fire the CEO.
- Essentially, press the reset button, say.
- Press the reset button.
- Rerandomize the parameters.
- Well, let me sort of, that's actually,
okay, that's a beautiful vision, I think,
as long as it's possible to press the reset button.
Do you think it will always be possible
to press the reset button?
- So I think that it's definitely will be possible to build.
So you're talking, so the question
that I really understand from you is,
humans people have control over the AI systems that they build.
- Yes.
- And my answer is, it's definitely possible
to build AI systems which will want
to be controlled by their humans.
- Wow, that's part of their,
so it's not that just they can't help but be controlled,
but that's,
they exist, one of the objectives of their existence
is to be controlled.
In the same way that human parents
generally want to help their children,
they want their children to succeed,
it's not a burden for them,
they are excited to help the children
and to feed them and to dress them and to take care of them.
And I believe with high conviction
that the same will be possible for an AGI.
It will be possible to program an AGI,
to design it in such a way that it will have
a similar deep drive, that it will be delighted to fulfill
and the drive will be to help humans flourish.
- But let me take a step back to that moment
where you create the AGI system.
I think this is a really crucial moment.
And between that moment and the Democratic board members
with the AGI at the head,
there has to be a relinquishing of power.
So as George Washington, despite all the bad things he did,
one of the big things he did is he relinquished power.
He, first of all, didn't want to be president.
And even when he became president,
he didn't keep just serving
as most dictators do for indefinitely.
Do you see yourself being able to relinquish control
over an AGI system, given how much power
you can have over the world?
At first financial, just make a lot of money, right?
And then control by having possession of this AGI system.
- I'd find it trivial to do that.
I'd find it trivial to relinquish this kind of power.
I mean, the kind of scenario you are describing
sounds terrifying to me.
That's all.
I would absolutely not want to be in that position.
- Do you think you represent the majority
or the minority of people in the AI community?
- Well, I mean--
- It's an open question and an important one.
Are most people good is another way to ask it.
- So I don't know if most people are good,
but I think that when it really counts,
people can be better than we think.
- That's beautifully put, yeah.
Are there specific mechanism you can think of
of aligning AI gene values to human values?
Is that, do you think about these problems
of continued alignment as we develop the AI systems?
- Yeah, definitely.
In some sense, the kind of question which you are asking is,
so if I were to translate the question to today's terms,
it would be a question about how to get an RL agent
that's optimizing a value function which itself is learned.
And if you look at humans, humans are like that
because the reward function, the value function of humans
is not external, it is internal.
- That's right.
- And there are definite ideas
of how to train a value function.
Basically an objective,
and as objective as possible perception system
that will be trained separately to recognize,
to internalize human judgments on different situations.
And then that component would then be integrated
as the base value function
for some more capable RL system.
You could imagine a process like this.
I'm not saying this is the process,
I'm saying this is an example
of the kind of thing you could do.
- So on that topic of the objective functions
of human existence,
what do you think is the objective function
that's implicit in human existence?
What's the meaning of life?
- Oh.
(sighs)
I think the question is wrong in some way.
I think that the question implies
that there is an objective answer
which is an external answer.
You know, your meaning of life is X.
I think what's going on is that we exist
and that's amazing.
And we should try to make the most of it
and try to maximize our own value and enjoyment
of our very short time while we do exist.
- It's funny 'cause action does require
an objective function.
It's definitely there in some form,
but it's difficult to make it explicit
and maybe impossible to make it explicit,
I guess is what you're getting at.
And that's an interesting fact of an RL environment.
- Well, I was making a slightly different point
is that humans want things
and their wants create the drives that cause them to,
our wants are our objective functions,
our individual objective functions.
We can later decide that we want to change,
that what we wanted before is no longer good
§12Transcript · 1:33:26–1:37:21
and we want something else.
- Yeah, but they're so dynamic.
There's gotta be some underlying sort of Freud,
there's things, there's like sexual stuff,
there's people who think it's the fear of death
and there's also the desire for knowledge
and all these kinds of things, procreation,
sort of all the evolutionary arguments.
It seems to be,
there might be some kind of fundamental objective function
from which everything else emerges,
but it seems like it's very difficult to make it explicit.
- I think that probably is an evolutionary objective
function, which is to survive and procreate
and make your children succeed.
That would be my guess,
but it doesn't give an answer to the question
of what's the meaning of life.
I think you can see how humans are part of this big process,
this ancient process, we exist on a small planet
and that's it.
So given that we exist, try to make the most of it
and try to enjoy more and suffer less as much as we can.
- Let me ask two silly questions about life.
One, do you have regrets,
moments that if you went back, you would do differently?
And two, are there moments that you're especially proud of
that made you truly happy?
- So I can answer both questions.
Of course, there's a huge number of choices
and decisions that I've made
that with the benefit of hindsight, I wouldn't have made them
and I do experience some regret,
but I try to take solace in the knowledge
that at the time I did the best I could.
And in terms of things that I'm proud of,
I'm very fortunate to have done things I'm proud of
and they made me happy for some time,
but I don't think that that is the source of happiness.
- So your academic accomplishments, all the papers,
you're one of the most cited people in the world,
all of the breakthroughs I mentioned
in computer vision and language and so on,
what is the source of happiness and pride for you?
- I mean, all those things are a source of pride for sure.
I'm very grateful for having done all those things
and it was very fun to do them,
but happiness comes, but you know, you can,
happiness, well, my current view is that happiness
comes from our, to a very large degree,
from the way we look at things.
You know, you can have a simple meal
and be quite happy as a result,
or you can talk to someone and be happy as a result as well.
Or conversely, you can have a meal and be disappointed
that the meal wasn't a better meal.
So I think a lot of happiness comes from that,
but I'm not sure, I don't wanna be too confident.
(laughs)
- Being humble in the face of the uncertainty
seems to be also a part of this whole happiness thing.
Well, I don't think there's a better way to end it
than meaning of life and discussions of happiness.
So Ilya, thank you so much.
You've given me a few incredible ideas.
You've given the world many incredible ideas.
I really appreciate it and thanks for talking today.
- Yeah, thanks for stopping by, I really enjoyed it.
- Thanks for listening to this conversation
with Ilya Sutskever,
and thank you to our presenting sponsor, Cash App.
Please consider supporting the podcast
by downloading Cash App and using the code LEXPODCAST.
If you enjoy this podcast, subscribe on YouTube,
review it with five stars on Apple Podcast,
support on Patreon, or simply connect with me on Twitter
at Lex Friedman.
And now let me leave you with some words
from Alan Turing on machine learning.
Instead of trying to produce a program
to simulate the adult mind,
why not rather try to produce one
which simulates the child's?
If this were then subjected
to an appropriate course of education,
one would obtain the adult brain.
Thank you for listening and hope to see you next time.
(upbeat music)
(upbeat music)