Wojciech Zaremba: OpenAI Codex, GPT-3, Robotics, and the Future of AI
Guests: Wojciech Zaremba
Full transcript · Read the full conversation in order, with its original context.
Contents18 chapters
§01Transcript · 00:00–09:46§02Transcript · 09:50–20:35§03Transcript · 20:40–30:04§04Transcript · 30:06–40:20§05Transcript · 40:24–50:09§06Transcript · 50:12–1:00:34§07Transcript · 1:00:38–1:10:00§08Transcript · 1:10:04–1:20:01§09Transcript · 1:20:03–1:30:15§10Transcript · 1:30:17–1:39:40§11Transcript · 1:39:42–1:48:30§12Transcript · 1:48:32–1:58:02§13Transcript · 1:58:05–2:06:42§14Transcript · 2:06:43–2:16:02§15Transcript · 2:16:04–2:25:41§16Transcript · 2:25:44–2:34:37§17Transcript · 2:34:39–2:44:51§18Transcript · 2:44:56–2:51:26§01Transcript · 00:00–09:46
The following is a conversation with Wojciech Zaremba,
co-founder of OpenAI,
which is one of the top organizations in the world
doing artificial intelligence research and development.
Wojciech is the head of language and cogeneration teams,
building and doing research on GitHub Copilot,
OpenAI Codecs, and GPT-3,
and who knows, 4, 5, 6, N,
and N+1, and he also previously led OpenAI's robotic efforts.
These are incredibly exciting projects to me
that deeply challenge and expand our understanding
of the structure and nature of intelligence.
The 21st century, I think, may very well be remembered
for a handful of revolutionary AI systems
and their implementations.
GPT, Codecs, and applications of language models
and transformers in general to the language
and visual domains may very well be
at the core of these AI systems.
To support this podcast, please check out our sponsors.
They're listed in the description.
This is the Lex Friedman Podcast,
and here is my conversation with Wojciech Zaremba.
You mentioned that Sam Altman asked
about the Fermi Paradox,
and the people at OpenAI had really sophisticated,
interesting answers, so that's when you knew
this is the right team to be working with.
So let me ask you about the Fermi Paradox, about aliens.
Why have we not found overwhelming evidence
for aliens visiting Earth?
- I don't have a conviction in the answer,
but rather kind of probabilistic perspective
on what might be a, let's say, possible answer.
It's also interesting that the question itself
even can touch on your typical question
of what's the meaning of life,
because if you assume that we don't see aliens
because they destroy themselves,
that kind of upweights the focus
on making sure that we won't destroy ourselves.
At the moment, the place where I am actually
with my belief, and these things also change over the time,
is I think that we might be alone in the universe,
which actually makes life more,
or let's say consciousness life, more kind of valuable,
and that means that we should more appreciate it.
- Have we always been alone?
So what's your intuition about our galaxy, our universe?
Is it just sprinkled with graveyards
of intelligent civilizations, or are we truly,
is life, intelligent life, truly unique?
- At the moment, my belief that it is unique,
but I would say I could also, you know,
there was some footage released with UFO objects,
which makes me actually doubt my own belief.
- Yes.
- Yeah, I can tell you one crazy answer that I have heard.
- Yes.
- So apparently, when you look actually
at the limits of computation, you can compute more
if the temperature of the universe would drop down.
So one of the things that aliens might want to do
if they are truly optimizing to maximize amount of compute,
which, you know, maybe can lead to,
or let's say simulations or so,
it's instead of wasting current entropy of the universe,
because, you know, we, by living,
we are actually somewhat wasting entropy,
then you can wait for the universe to cool down
such that you have more computation.
So that's kind of a funny answer.
I'm not sure if I believe in it,
but that would be one of the reasons
why you don't see aliens.
It's also possible, see, some people say
that maybe there is not that much point
in actually going to other galaxies if you can go inwards.
So there is no limits of what could be an experience
if we could, you know, connect machines to our brains,
while there are still some limits
if we want to explore universe.
- Yeah, there could be a lot of ways to go inwards too.
Once you figure out some aspect of physics
we haven't figured out yet,
maybe you can travel to different dimensions.
I mean, travel in three-dimensional space
may not be the most fun kind of travel.
There may be like just a huge amount
of different ways to travel.
And it doesn't require a spaceship going slowly
in 3D space to space time.
- It also feels, you know, one of the problems
is that speed of light is low and the universe is vast.
And it seems that actually most likely
if we want to travel very far,
then we would instead of actually sending spaceships
with humans that weigh a lot,
we would send something similar
to what Uri Miller is working on.
These are like a huge sail, which is at first powered,
there is a shot of laser from an air
and it can propel it to quarter of speed of light.
And sail itself contains a few grams of equipment.
And that might be the way to actually
transport matter through universe.
But then when you think what would it mean for humans,
it means that we would need to actually
put their 3D printer and, you know,
3D print a human on other planet.
I don't know, play them YouTube or let's say,
or like a 3D print like huge human right away
or maybe a womb or so.
- With our current techniques of archeology,
if a civilization was born and died
long enough ago on earth,
we wouldn't be able to tell.
And so that makes me really sad.
And so I think about earth in that same way.
How can we leave some remnants if we do destroy ourselves?
How can we leave remnants for aliens
in the future to discover?
Like, here's some nice stuff we've done.
Like Wikipedia and YouTube,
do we have it like in a satellite orbiting earth
with a hard drive?
Like how do we say, how do we back up human civilization
for the good parts or all of it is good parts
so that it can be preserved longer than our bodies can?
That's kind of a, it's a difficult question.
It also requires the difficult acceptance
of the fact that we may die.
And if we die, we may die suddenly as a civilization.
- So let's see.
I think it kind of depends on the cataclysm.
We have observed in other parts of universe
that birds of gamma rays,
these are high energy rays of light
that actually can apparently kill entire galaxy.
So there might be actually nothing even to,
nothing to protect us from it.
I'm also, and I'm looking actually at the past civilization.
So it's like Aztecs or so,
they disappeared from the surface of the earth.
And one can ask, why is it the case?
And the way I'm thinking about it is,
you know, that definitely they had some problem
that they couldn't solve.
And maybe there was a flood
and all of a sudden they couldn't drink,
there was no potable water and they all died.
And I think that so far,
the best solution to such problems is I guess technology.
So, I mean, if they would know that you can just boil water
and then drink it after,
then that would save their civilization.
And even now, when we look actually at the current pandemic,
it seems that once again, actually science comes to rescue
and somehow science increases size of the action space.
And I think that's a good thing.
- Yeah, but nature has a vastly larger action space.
- But still it might be a good thing for us
to keep on increasing action space.
(laughing)
- Okay, looking at past civilizations, yes.
But looking at the destruction of human civilization,
perhaps expanding the action space will add actions
that are easily acted upon, easily executed,
and as a result, destroy us.
- So let's see.
I was pondering why actually even we have negative impact
on the globe, because, you know,
if you ask every single individual,
they would like to have clean air,
they would like healthy planet,
but somehow it actually, it's not the case
that as a collective, we are not going in this direction.
I think that there exists very powerful system
to describe what we value, that's capitalism.
It assigns actually monetary values to various activities.
At the moment, the problem in the current system
is that there are some things which we value,
there is no cost assigned to it.
So even though we value clean air,
or maybe we also value lack of destruction
on let's say internet or so,
at the moment, these quantities, you know,
companies, corporations can pollute them for free.
So in some sense, I wish, or like,
that's I guess purpose of politics
to align the incentive systems.
And we are kind of maybe even moving in this direction.
The first issue is even to be able to measure
the things that we value,
then we can actually assign the monetary value to them.
- Yeah, and that's, so it's getting the data
and also probably through technology,
enabling people to vote and to move money around
in a way that is aligned with their values.
And that's very much a technology question.
§02Transcript · 09:50–20:35
So like having one president and Congress and voting
that happens every four years or something like that,
that's a very outdated idea.
There could be some technological improvements
to that kind of idea.
- So I'm thinking from time to time about these topics,
but it's also feels to me that it's a little bit like,
it's hard for me to actually make correct predictions
what is the appropriate thing to do.
I extremely trust Sam Altman, our CEO on these topics.
He, I'm more on the side of being, I guess,
naive hippie that, yeah.
(laughing)
- That's your life philosophy.
Well, I think self-doubt and I think hippie implies optimism.
Those two things are pretty good way to operate.
- I mean, still it is hard for me to actually understand
how the politics works or like how this,
like exactly how the things would play out.
And Sam is really excellent with it.
- What do you think is rarest in the universe?
You said we might be alone.
What's hardest to build is another engineering way
to ask that.
Life, intelligence or consciousness.
So like you said that we might be alone,
which is the thing that's hardest to get to?
Is it just the origin of life?
Is it the origin of intelligence?
Is it the origin of consciousness?
- So let me at first explain you my kind of mental model,
what I think is needed for life to appear.
So I imagine that at some point
there was this primordial soup of amino acids
and maybe some proteins in the ocean.
And some proteins were turning into some other proteins
through reaction.
And you can almost think about this cycle
of what turns into what as there is a graph
essentially describing which substance
turns into some other substance.
And essentially life means that all the sudden
in the graph has been created that cycle
such that the same thing keeps on happening
over and over again.
That's what is needed for life to happen.
And in some sense, you can think almost
that you have this gigantic graph
and it needs like a sufficient number of edges
for the cycle to appear.
Then from perspective of intelligence and consciousness,
my current intuition is that they might be quite intertwined.
First of all, it might not be that it's like a binary thing
that you have intelligence or consciousness.
It seems to be more like a continuous component.
Let's see, if we look for instance
on the even networks recognizing images,
people are able to show that the activations
of these networks correlate very strongly
with activations in visual cortex of some monkeys.
The same seems to be true about language models.
Also, if you for instance, look,
if you train agent in 3D world,
at first, it barely recognizes what is going on.
Over the time it kind of recognizes foreground
from background, over the time it kind of knows
where there is a foot and it just follows it.
Over the time, it actually starts having a 3D perception.
So it is possible for instance,
to look inside of the head of an agent
and ask what would it see if it looks to the right.
And the crazy thing is, initially when the agents
are barely trained, that these predictions are pretty bad.
Over the time, they become better and better.
You can still see that if you ask what happens
when the head is turned by 360 degrees,
for some time, they think that the different thing appears.
And then at some stage, they understand actually
that the same thing supposed to appear.
So they get like a understanding of 3D structure.
It's also very likely that they have inside some level
of like a symbolic reasoning,
like they're particularly symbols for other agents.
So when you look at Dota agents, they collaborate together.
And they have some anticipation of if they would win battle,
they have some expectations with respect to other agents.
I might be too much anthropomorphizing
how the things look for me.
But then the fact that they have a symbol for other agents
makes me believe that at some stage
as they are optimizing for skills,
they would have also symbol to describe themselves.
This is like a very useful symbol to have.
And this particularity, I would call it
like a self-consciousness or self-awareness.
And still it might be different from the consciousness.
So I guess the way how I'm understanding
the word consciousness, let's say the experience
of drinking a coffee, or let's say experience of being a bat,
that's the meaning of the word consciousness.
It doesn't mean to be awake.
Yeah, it feels, it might be also somewhat related
to memory and recurrent connections.
So it's kind of, okay, if you look at anesthetic drugs,
they might be, like they essentially,
they disturb brain waves such that
maybe memory is not formed.
- And so there's a lessening of consciousness
when you do that.
- Correct.
- And so that's one way to intuit
what is consciousness.
There's also kind of another element here.
It could be that it's this kind of self-awareness module
that you described, plus the actual subjective experience
is a storytelling module that tells us a story
about what we're experiencing.
- The crazy thing, so let's say, I mean, in meditation,
they teach people not to speak story inside of the head.
And there is also some fraction of population
who doesn't have actually narrator.
I know people who don't have a narrator,
and they have to use external people
in order to kind of solve tasks
that require internal narrator.
So it seems that it's possible to have the experience
without the talk.
- What are we talking about
when we talk about the internal narrator?
Is that the voice when you're reading a book?
- Yeah, I thought that that's what you are referring to.
- Well, I was referring more on the not an actual voice.
I meant there's some kind of subjective experience
feels like it's fundamentally about storytelling
to ourselves.
It feels like the feeling is a story
that is much simpler abstraction
than the raw sensory information.
So there feels like it's a very high level abstraction
that is useful for me to feel like entity in this world.
Most useful aspect of it is that because I'm conscious,
I think there's an intricate connection to me
not wanting to die.
So it's a useful hack to really prioritize not dying.
Those seem to be somewhat connected.
So I'm telling the story of it's richly feels
like something to be me,
and the fact that me exists in this world,
I wanna preserve me.
So that makes it a useful agent hack.
- So I will just refer maybe to the fact
that I will just refer maybe to the first part,
as you said about that kind of story
of describing who you are.
I was thinking about that even,
so obviously I like thinking about consciousness.
I like thinking about AI as well,
and I'm trying to see analogies of these things in AI,
what would it correspond to?
So, open AI train,
a model called GPT,
which can generate pretty amusing text on arbitrary topic.
And one way to control GPT
is by putting into prefix at the beginning of the text,
some information, what would be the story about?
You can have even chat with GPT
by saying that the chat is with Lex or Elon Musk or so,
and GPT would just pretend to be you or Elon Musk or so.
And it almost feels that this story
that we give ourselves to describe our life,
it's almost like things that you put into context of GPT.
- Yeah, and it generates the,
but the context we provide to GPT is multimodal.
- It's more so, GPT itself is multimodal.
GPT itself hasn't learned actually
from experience of single human,
but from the experience of humanity, it's a chameleon.
You can turn it into anything.
And in some sense, by providing context,
it behaves as the thing that you wanted it to be.
It's interesting that people have stories of who they are.
And as you said, these stories,
they help them to operate in the world.
But it's also interesting,
I guess various people find it out through meditation or so,
that there might be some patterns that you have learned
when you were a kid
that actually are not serving you anymore.
And you also might be thinking that that's who you are,
and that's actually just a story.
- Yeah, so it's a useful hack,
but sometimes it gets us into trouble.
It's a local optima.
- It's a local optima.
- You wrote that Stephen Hawking,
he tweeted, Stephen Hawking asked,
what breathes fire into equations?
Which meant what makes given mathematical equations
realize the physics of a universe?
Similarly, I wonder what breathes fire into computation?
§03Transcript · 20:40–30:04
What makes given computation conscious?
Okay, so how do we engineer consciousness?
How do you breathe fire and magic into the machine?
- So it seems clear to me
that not every computation is conscious.
I mean, you can, let's say,
just keep on multiplying one matrix over and over again,
and it might be gigantic matrix,
you can put a lot of computation,
I don't think it would be conscious.
So in some sense, the question is,
what are the computations which could be conscious?
I mean, so one assumption is
that it has to do purely with computation,
that you can abstract away matter,
and other possibilities,
it's very important was the realization of computation
that it has to do with some force fields or so,
and they bring consciousness.
At the moment, my intuition is
that it can be fully abstracted away.
So in case of computation, you can ask yourself,
what are the mathematical objects or so
that could bring such properties?
So for instance, if we think about the models, AI models,
that what they truly try to do,
or like models like GPT,
is they try to predict next word or so,
and this turns out to be equivalent to compressing text.
And because in some sense,
compression means that you learn the model of reality,
and you have just to remember where are your mistakes,
the better you are in predicting the...
And in some sense, when we look at our experience,
also when you look, for instance, at the car driving,
you know in which direction it will go,
you are good like in prediction.
And it might be the case
that the consciousness is intertwined with compression.
It might be also the case that self-consciousness
has to do with compressor trying to compress itself.
So, okay, I was just wondering,
what are the objects in mathematics
or computer science which are mysterious
that could have to do with consciousness?
And then I thought, you know, you see in mathematics,
there is something called Gadiel theorem,
which means if you have sufficiently
complicated mathematical system,
it is possible to point the mathematical system
back on itself.
In computer science,
there is something called Halping problem.
It's somewhat similar construction.
So I thought that, you know, if we believe
that under assumption that consciousness
has to do with compression,
then you could imagine that
as you keep on compressing things,
then at some point it actually makes sense
for the compressor to compress itself.
- Meta compression. - Yeah.
- Consciousness is meta compression.
- That's an idea.
And in some sense, you know, the crazy--
- I love it.
- Thank you.
- So, but do you think, if we think of a Turing machine,
a universal Turing machine,
can that achieve consciousness?
So is there something beyond our traditional definition
of computation that's required?
- So it's a specific computation.
And I said, this computation has to do with compression.
And the compression itself,
maybe other way of putting it is like,
you're internally creating the model of reality
in order, it's like you try inside to simplify reality
in order to predict what's gonna happen.
And that also feels somewhat similar
to how I think actually about my own conscious experience.
So clearly I don't have access to reality.
The only access to reality is through, you know,
cable going to my brain.
And my brain is creating a simulation of reality,
and I have access to the simulation of reality.
- Are you by any chance aware of the Hutter Prize,
Marcus Hutter?
He made this prize for compression of Wikipedia pages.
And there's a few qualities to it.
One, I think has to be perfect compression, which makes,
I think that little quirk makes it much less applicable
to the general task of intelligence.
'Cause it feels like intelligence
is always going to be messy.
Like perfect compression feels like it's not the right goal,
but it's nevertheless a very interesting goal.
So for him, intelligence equals compression.
And so the smaller you make the file,
given a large Wikipedia page,
the more intelligent the system has to be.
- Yeah, that makes sense.
So you can make perfect compression if you store errors.
And I think that actually what he meant
is you have algorithm plus errors.
By the way, Hutter is a,
he was PhD advisor of Sian Leck,
who is a DeepMind co-founder.
- Yeah, yeah.
So there's an interesting, and now he's a DeepMind,
there's an interesting network of people.
He's one of the people that I think seriously
took on the task of what would an AGI system look like?
I think for a longest time,
the question of AGI was not taken seriously
or rather rigorously.
And he did just that.
Like mathematically speaking,
what would the model look like?
If you remove the constraints of it having to be,
having to have a reasonable amount of memory,
reasonable amount of running time complexity,
computation time, what would it look like?
And essentially it's a half math,
half philosophical discussion of how would it,
like a reinforcement learning type of framework
look like for an AGI?
- Yeah, so he developed the framework
even to describe what's optimal
with respect to reinforcement learning.
Like there is a theoretical framework,
which is, as you said, under assumption,
there is infinite amount of memory and compute.
There was actually one person before,
his name is Solomonov.
Huter extended Solomonov's work to reinforcement learning,
but there exists a theoretical algorithm,
which is optimal algorithm to build intelligence.
And I can actually explain the algorithm.
- Yes.
Let's go, let's go, let's go.
- So the task itself, you can-
- Can I just pause how absurd it is
for brain in a skull trying to explain
the algorithm for intelligence?
Just go ahead.
- It is pretty crazy.
It is pretty crazy that the brain itself
is actually so small and it can ponder.
- How to design algorithms
that optimally solve the problem of intelligence.
Okay.
All right, so what's the algorithm?
- So let's see.
So first of all, the task itself is described as
you have infinite sequence of zeros and ones, okay?
You read n bits and you're about to predict n plus one bit.
So that's the task.
And you could imagine that every task
could be casted as such a task.
So if for instance, you have images and labels,
you can just turn every image into a sequence
of zeros and ones, then label, you concatenate labels
and that's actually,
and you could start by having training data first,
and then afterwards you have test data.
So theoretically, any problem could be casted
as a problem of predicting zeros and ones
on this infinite tape.
So let's say you read already n bits
and you want to predict n plus one bit.
And I will ask you to write every possible program
that generates these n bits, okay?
So, and you can have, you choose programming language.
It can be Python or C++.
And the difference between programming languages
might be, there is a difference by constant,
asymptotically your predictions will be equivalent.
So you read n bits, you enumerate all the programs
that produce these n bits in their output.
And then in order to predict n plus one bit,
you actually weight the programs according to their length.
And there is like a specific formula how you weight them.
And then the n plus one bit prediction
is the prediction from each of these program
according to that weight.
- Like statistically.
- Statistically, yeah.
- You pick, so the smaller the program,
the more likely you are to pick its output.
So that algorithm is grounded in the hope
or the intuition that the simple answer is the right one.
- It's a formalization of it.
- Yeah.
- It also means like if you would ask the question
after how many years would Sun explode?
You can say, it's more likely the answer is two
to some power because they're shorter program.
- Yeah.
- And then other.
§04Transcript · 30:06–40:20
- Well, I don't have a good intuition
about how different the space of short programs
are from the space of large programs.
Like what is the universe where short programs
like run things?
- So as I said, the things have to agree with n bits.
So even if you have, you need to start,
okay, if you have very short program
and they're like still some,
if it's not perfectly prediction of n bits,
you have to start errors.
What are the errors?
And that gives you the full program that agrees on n bits.
- Oh, so you don't agree perfectly with the n bits
and you store errors.
- That's like a longer program, slightly longer program
because it contains these extra bits of errors.
- That's fascinating.
What's your intuition about the programs
that are able to do cool stuff
like intelligence and consciousness?
Are they perfectly, like, is there if then statements
in them?
So like, is there a lot of exceptions that they're storing?
- So you could imagine if there would be tremendous amount
of if statements, then they wouldn't be that short
in case of neural networks.
You could imagine that what happens is they,
when you start with an uninitialized neural network,
it stores internally many possibilities
how the problem can be solved.
And SGD is kind of magnifying some paths
which are slightly similar to the correct answer.
So it's kind of magnifying correct programs.
And in some sense, SGD is a search algorithm
in the program space.
And the program space is represented by, you know,
kind of the wiring inside of the neural network.
And there's like an insane number of ways
how the features can be computed.
- Let me ask you the high level basic question
that's not so basic.
What is deep learning?
Is there a way you'd like to think of it
that is different than like a generic textbook definition?
- The thing that I hinted just a second ago
is maybe the closest to how I'm thinking these days
about deep learning.
So the statement is neural networks
can represent some programs.
Seems that various modules that we are actually adding up
to like, you know, we want networks to be deep
because we want multiple steps of the computation.
And deep learning provides the way
to represent space of programs, which is searchable.
And it's searchable with stochastic gradient descent.
So we have an algorithm to search over
humongous number of programs
and gradient descent kind of bubbles up
the things that tend to give correct answers.
- So a neural network with fixed weights,
that's optimized.
Do you think of that as a single program?
- So there is a work by Christopher Olach,
where he, so he works on interpretability
of neural networks.
And he was able to identify inside of the neural network,
for instance, a detector of a wheel for a car
or the detector of a mask for a car.
And then he was able to separate them out
and assemble them together using a simple program
for the detector, for a car detector.
- That's like, if you think of traditionally
defined programs, that's like a function within a program
that this particular neural network was able to find.
And you can tear that out, just like you can copy
and paste from Stack Overflow.
So any program is a composition of smaller programs.
- Yeah, I mean, the nice thing about the neural networks
is that it allows the things to be more fuzzy
than in case of programs.
In case of programs, you have this like a branching
this way or that way.
And the neural networks, they have an easier way
to be somewhere in between or to share things.
- What to use the most beautiful or surprising idea
in deep learning, in the utilization
of these neural networks, which by the way,
for people who are not familiar,
neural networks is a bunch of, what would you say?
It's inspired by the human brain.
There's neurons, there's connection between those neurons.
There's inputs and there's outputs and there's millions
or billions of those neurons.
And the learning happens by adjusting the weights
on the edges that connect these neurons.
- Thank you for giving definition.
I supposed to do it, but I guess you have enough empathy
to listeners to actually know that that might be useful.
- No, that's like, so I'm asking Plato of like,
what is the meaning of life?
He's not gonna answer.
You're being philosophical and deep and quite profound
talking about the space of programs,
which is very interesting, but also for people
who are just not familiar, what the hell we're talking about
when we talk about deep learning.
Anyway, sorry, what is the most beautiful
or surprising idea to you in all the time
you've worked at deep learning?
And you worked on a lot of fascinating projects,
applications of neural networks.
It doesn't have to be big and profound.
It can be a cool trick.
- Yeah, I mean, I'm thinking about the trick,
but like it's still amusing to me that it works at all.
That let's say that the extremely simple algorithm,
stochastic gradient descent, which is something
that I would be able to derive on the piece of paper
to high school student when put at the scale
of thousands of machines actually can create the behaviors
which we called kind of human like behaviors.
- So in general, any application of stochastic gradient
descent in neural networks is amazing to you.
So, or is there a particular application
in natural language, reinforcement learning?
And also, what do you attribute that success to?
Is it just scale?
What profound insight can we take from the fact that
the thing works for gigantic sets of variables?
- I mean, the interesting thing is these algorithms,
they were invented decades ago
and people actually gave up on the idea.
And back then they thought that we need
profoundly different algorithms
and they spent a lot of cycles on very different algorithms.
And I believe that we have seen that various innovations
that say like transformer or dropout
or so they can vastly help.
But it's also remarkable to me that this algorithm
from '60s or so, or I mean, you can even say
that the gradient descent was invented by Leibniz
in I guess 18th century or so.
That actually is the core of learning.
In the past, people are, it's almost like
out of the maybe an ego, people are saying
that it cannot be the case that such a simple algorithm
is there, could solve complicated problems.
So they were in search for the other algorithms.
And as I'm saying, like, I believe that actually
we are in the game where there is,
there are actually frankly three levers.
There is compute, there are algorithms and there is data.
And if we want to build intelligent systems,
we have to pull all three levers
and they are actually multiplicative.
It's also interesting, so you ask, is it only compute?
People internally, they did the studies
to determine how much gains they were coming
from different levers.
And so far we have seen that more gains came
from compute than algorithms.
But also we are in the world that in case of compute,
there is a kind of exponential increase in funding.
And at some point it's impossible to invest more.
It's impossible to invest $10 trillion.
We are speaking about, let's say all taxes in US.
- But you're talking about money
that could be innovation in the compute.
- That's true as well.
So, I mean, there are like a few pieces.
So one piece is human brain is an incredible supercomputer.
And they're like, it has,
a hundred trillion parameters.
Or like if you try to count various quantities in the brain,
they're like a neuron synapses,
that small number of neurons, there is a lot of synapses.
It's unclear even how to map synapses
to parameters of neural networks,
but it's clear that there are many more.
So it might be the case that our networks
are still somewhat small.
It also might be the case that they are more efficient
than brain or less efficient by some huge factor.
I also believe that there will be,
you know, at the moment we are at the stage
that these neural networks, they require a thousand X
or like a huge factor of more data than humans do.
And it will be a matter of,
there will be algorithms that vastly
decrease sample complexity, I believe so.
But the place where we are heading today
is there are domains which contains million X more data.
And even though computers might be
1000 times slower than humans in learning,
that's not the problem.
Like for instance, I believe that it should be possible
to create superhuman therapies by,
and there are like even simple steps of doing it.
And you know, the core reason is there is just,
§05Transcript · 40:24–50:09
machine will be able to read way more
transcripts of therapies,
and then it should be able to speak simultaneously
with many more people.
And it should be possible to optimize it all in parallel.
- Now you're touching on something I deeply care about
and think is way harder than we imagine.
What's the goal of a therapist?
- What's the goal of therapist?
So, okay, so one goal, now this is terrifying to me,
but there's a lot of people that contemplate suicide,
self-affirmed depression,
and they could significantly be helped with therapy.
And the idea that an AI algorithm
might be in charge of that,
it's like a life and death task.
The stakes are high.
So one goal for a therapist, whether human or AI,
is to prevent suicide ideation, to prevent suicide.
How do you achieve that?
- So let's see.
So to be clear, I don't think that the current models
are good enough for such a task
because it requires insane amount of understanding, empathy,
and the models are far from this place, but it's-
- But do you think that understanding empathy,
that signal is in the data?
- I think there is some signal in the data, yes.
I mean, there are plenty of transcripts of conversations
and it is possible from it to understand personalities.
It is possible from it to understand
if conversation is friendly, amicable, antagonistic.
It is, I believe that, you know,
given the fact that the models that we train now,
they can have, they are chameleons,
that they can have any personality.
They might turn out to be better in understanding
personality of other people than anyone else.
And they should- - Be empathetic.
- To be empathetic.
- Yeah, interesting.
But I wonder if there's some level
of multiple modalities required
to be able to be empathetic of the human experience,
whether language is not enough, to understand death,
to understand fear, to understand childhood trauma,
to understand wit and humor required
when you're dancing with a person
who might be depressed or suffering,
both humor and hope and love and all those kinds of things.
So there's another underlying question,
which is self-supervised versus supervised.
So can you get that from the data
by just reading a huge number of transcripts?
- I actually, so I think that reading
huge number of transcripts is a step one.
It's like the same way as you cannot learn to dance
if just from YouTube.
By watching it, you have to actually try it out yourself.
So I think that here, that's a similar situation.
I also wouldn't deploy the system
in the high stakes situations right away,
but kind of see gradually where it goes.
And obviously, initially, it would have to go
hand in hand with humans.
But at the moment, we are in the situation
that actually there is many more people
who actually would like to have a therapy
or speak with someone than there are therapies out there.
So fundamentally, I was thinking,
what are the things that can vastly
increase people's well-being?
Therapy is one of them.
I think meditation is other one.
I guess maybe human connection is a third one.
And I guess pharmacologically, it's also possible.
Maybe direct brain stimulation or something like that.
But these are pretty much options out there.
Then let's say the way I'm thinking about the AGI endeavor
is by default, that's an endeavor
to increase amount of wealth.
And I believe that we can vastly increase
amount of wealth for everyone.
And simultaneously, so I mean,
these are like two endeavors that make sense to me.
One is like essentially increase amount of wealth.
And second one is increase overall human well-being.
- And those are coupled together.
- And they can, I would say, these are different topics.
One can help another.
- And therapist is a funny word
'cause I see friendship and love as therapy.
I mean, so therapist broadly defined
is just friendship, is a friend.
So like therapist has a very kind of clinical sense to it.
But what is human connection?
Not to get all Camus and Dostoevsky on you,
but life is suffering and we draw,
we seek connection with other humans
as we desperately try to make sense of this world
in a deep, overwhelming loneliness that we feel inside.
- So I think connection has to do with understanding.
And I think that almost like a lack of understanding
causes suffering.
If you speak with someone and you feel ignored,
that actually causes pain.
If you are feeling deeply understood,
that actually they might not even tell you
what to do in life, but like a pure understanding.
- Or just being heard.
Understanding is a kind of, it's a lot,
just being heard, feel like you're being heard.
Like somehow that's a alleviation temporarily
of the loneliness.
That if somebody knows you're here
with their body language, with the way they are,
with the way they look at you, with the way they talk,
you feel less alone for a brief moment.
- Yeah, very much, I agree.
So I thought in the past about a somewhat similar question
to yours, which is what is love?
Rather, what is connection?
And obviously I think about these things
from AI perspective, what would it mean?
So I said that the intelligence has to do
with some compression, which is more or less,
like I can say, almost understanding
of what is going around.
It seems to me that other aspect is there seem
to be reward functions and you can have a reward
for food, for maybe human connection,
for let's say warmth, sex and so on.
And it turns out that the various people
might be optimizing slightly different reward functions.
They essentially might care about different things.
And in case of love, at least the love between two people,
you can say that the boundary between people dissolves
to such extent that they end up optimizing
each other reward functions.
- Yeah, that's interesting.
Celebrate the success of each other.
- In some sense, I would say love means helping others
to optimize their reward functions,
not your reward functions, not the things
that you think are important,
but the things that the person cares about.
You try to help them to optimize it.
- So love is, if you think of two reward functions,
you just, it's like addition.
You combine them together.
- Yeah, pretty much.
- Maybe like with a weight and it depends
like the dynamic of the relationship.
- Yeah, I mean, you could imagine that
if you are fully optimizing someone's reward function
without yours, then maybe are creating codependency
or something like that.
I'm not sure what's the appropriate weight.
But the interesting thing is I even think
that the individual person,
we ourselves, we are actually less of a unified insight.
So for instance, if you look at the donut,
on the one level, you might think,
"Oh, this looks tasty, I would like to eat it."
On other level, you might tell yourself,
"I shouldn't be doing it because I want to gain muscles."
So, and you might do it regardless,
kind of against yourself.
So it seems that even within ourselves,
they're almost like a kind of intertwined personas.
And I believe that the self-love means that the love
between all these personas,
which also means being able to love yourself
when we are angry or stressed or so.
- Combining all those reward functions
of the different selves you have.
- Yeah, and accepting that they are there.
Like, you know, often people,
they have a negative self-talk or they say,
"I don't like when I'm angry."
And like, I try to imagine if there would be
like a small baby Lex, like five years old.
- Who's angry.
- Angry, and then you're like, "You shouldn't be angry.
"Like, stop being angry."
But like, instead actually you want Lex to come over,
give him a hug and just like say, "It's fine.
"Okay, you can be angry as long as you want."
And then he would stop.
- Or maybe not.
- Or maybe not, but you cannot expect it even.
- Yeah, but still that doesn't explain the why of love.
Like, why is love part of the human condition?
Why is it useful to combine the reward functions?
It seems like that doesn't, I mean,
I don't think reinforcement learning frameworks
can give us answers to why.
Even the Hutter framework has an objective function
that's static.
- So we came to existence as a consequence
§06Transcript · 50:12–1:00:34
of evolutionary process.
And in some sense, the purpose of evolution is survival.
And then this complicated optimization objective
baked into us, let's say compression,
which might help us operate in the real world
and it baked into us various reward functions.
- Yeah.
- Then to be clear, at the moment we are operating
in the regime which is somewhat out of distribution
where the even evolution optimize us.
- It's almost like love is a consequence of cooperation
that we've discovered is useful.
- Correct.
In some way, it's even the case if you--
- I just love the idea that love
is like the out of distribution.
- Or it's not out of distribution.
It's like, as you said, it evolved for cooperation.
- Yes.
- And I believe that the, in some sense,
cooperation ends up helping each of us individually.
So it makes sense, evolutionary.
And there is, in some sense, and you know,
love means there is this dissolution of boundaries
that you have a shared reward function.
And we evolved to actually identify ourselves
with larger groups.
So we can identify ourselves, you know, with a family.
We can identify ourselves with a country
to such extent that people are willing
to give away their life for country.
So there is, we are wired actually even for love.
And at the moment, I guess,
maybe it would be somewhat more beneficial
if we would identify ourselves
with all the humanity as a whole.
So you can clearly see when people travel around the world,
when they run into person from the same country,
they say, "Oh, we see TR."
And all of a sudden they find all these similarities.
They find some, they befriend those folks
earlier than others.
So there is like a sense, some sense of the belonging.
And I would say, I think it would be overall
good thing to the world for people to move towards,
I think it's even called open individualism,
move toward the mindset of a larger and larger groups.
- So the challenge there, that's a beautiful vision
and I share it to expand that circle of empathy,
that circle of love towards the entirety of humanity.
But then you start to ask, well, where do you draw the line?
Because why not expand it to other conscious beings?
And then finally for our discussion,
something I think about is why not expand it to AI systems?
Like we start respecting each other when the person,
the entity on the other side has the capacity to suffer,
'cause then we develop a capacity to sort of empathize.
And so I could see AI systems that are interacting
with humans more and more having conscious-like displays.
So like they display consciousness through language
and through other means.
And so then the question is like,
well, is that consciousness?
Because they're acting conscious.
And so the reason we don't like torturing animals
is because they look like they're suffering
when they're tortured.
And if AI looks like it's suffering when it's tortured,
how is that not requiring of the same kind of empathy
from us and respect and rights that animals do
and other humans do?
- I think it requires empathy as well.
I mean, I would like, I guess us or humanity
or so make a progress in understanding what consciousness is
because I don't want just to be speaking
about the philosophy,
but rather actually make a scientific.
There was a time that people thought
that there is a force of life
and the things that have this force, they are alive.
And I think that there is actually a path
to understand exactly what consciousness is.
And in some sense, it might require essentially
putting probes inside of a human brain,
what neural link does.
- So the goal there, I mean, there's several things
with consciousness that make it a real discipline,
which is one is rigorous measurement of consciousness.
And then the other is the engineering of consciousness,
which may or may not be related.
I mean, you could also run into trouble,
like for example, in the United States,
for the Department of DOT, Department of Transportation,
and a lot of different places put a value on human life.
I think DOT's value is $9 million per person.
So in that same way, you can get into trouble
if you put a number on how conscious a being is,
'cause then you can start making policy.
If a cow is 0.1 or like 10% as conscious as a human,
then you can start making calculations
and it might get you into trouble.
But then again, that might be a very good way to do it.
- I would like to move to that place
that actually we have scientific understanding
what consciousness is.
And then we'll be able to actually assign value
and I believe that there is even the path
for the experimentation in it.
So we said that you could put the probes
inside of the brain.
There is actually few other things that you could do
with devices like Neuralink.
So you could imagine that the way even to measure
if AI system is conscious
is by literally just plugging into the brain.
I mean, that assumes that it's kind of easy,
but the plugging into the brain and asking person
if they feel that their consciousness expanded.
This direction, of course, has some issues.
You can say, if someone takes a psychedelic drug,
they might feel that their consciousness expanded
even though that drug itself is not conscious.
- Right, so like you can't fully trust the self-report
of a person saying their consciousness is expanded or not.
Let me ask you a little bit about psychedelics
'cause there've been a lot of excellent research
on different psychedelics, psilocybin, MDMA, even DMT,
drugs in general, marijuana too.
What do you think psychedelics do to the human mind?
It seems they take the human mind
to some interesting places.
Is that just a little hack, a visual hack,
or is there some profound expansion of the mind?
- So let's see.
I don't believe in magic.
I believe in science, in causality.
Still, let's say, and then as I said,
I think that the brain,
our subjective experience of reality is,
we live in the simulation run by our brain,
and the simulation that our brain runs,
they can be very pleasant or very hellish.
Drugs, they are changing some hyperparameters
of the simulation.
It is possible, thanks to change of these hyperparameters,
to actually look back on your experience
and even see that the given things
that we took for granted, they are changeable.
So they allow to have an amazing perspective.
There is also, for instance, the fact that after DMT,
people can see the full movie inside of their head,
gives me further belief that the brain can generate
that full movie, that the brain is actually learning
the model of reality to such extent
that it tries to predict what's gonna happen next.
- Yeah, very high resolution,
so it can replay reality, essentially.
- Extremely high resolution.
Yeah, and it's also kind of interesting to me
that somehow there seems to be some similarity
between these drugs and meditation itself.
And I actually started even these days
to think about meditation as a psychedelic.
- Do you practice meditation?
- I practice meditation.
I mean, I went a few times on the retreats
and it feels after, like after second or third day
of meditation, there is almost like a sense of tripping.
- What does a meditation retreat entail?
- So you wake up early in the morning
and you meditate for extended period of time.
- Alone?
Sorry, Jake, you've been talking.
- Yeah, so it's optimized, even though there are other people
it's optimized for isolation.
So you don't speak with anyone,
you don't actually look into other people's eyes.
And you sit on the chair.
So Vipassana meditation tells you to focus on the breath.
So you try to put all attention into breathing in
and breathing out.
And the crazy thing is that as you focus attention like that
after some time, there's some starts coming back
like some memories that you're completely forgotten.
It almost feels like that you have a mailbox
and then you know, you are just like archiving email
one by one.
And at some point there is like a amazing feeling
of getting to mailbox zero, zero emails.
And it's very pleasant.
It's kind of, it's crazy to me that once you resolve
these inner stories or like inner traumas,
then once there is nothing left,
the default state of human mind is extremely peaceful
and happy.
Extreme, like some sense, it feels that,
it feels at least to me the way how when I was a child
that I can look at any object and it's very beautiful.
I have a lot of curiosity about the simple things.
And that's where usually meditation takes me.
§07Transcript · 1:00:38–1:10:00
- Are you, what are you experiencing?
Are you just taking in simple sensory information
and are just enjoying the rawness
of that sensory information?
So there's no memories or all that kind of stuff.
You're just enjoying being?
- Yeah, pretty much.
I mean, still there is, thoughts are slowing down.
Sometimes they pop up, but it's also somehow
the extended meditation takes you to the space
that they are way more friendly, way more positive.
There is also this thing that we've,
it almost feels that,
it almost feels that we are constantly getting
a little bit of a reward function
and we are just spreading this reward function
on various activities.
But if you stay still for extended period of time,
it kind of accumulates, accumulates, accumulates.
And there is a sense, there is a sense
that some point it passes some threshold
and it feels as drop is falling into kind of ocean
of love and bliss.
And that's like a, this is like a very pleasant,
and that's why I'm saying like,
that corresponds to the subjective experience.
Some people, I guess, in spiritual community,
they describe it that that's the reality.
And I would say, I believe that there are like
all sorts of subjective experience that one can have.
And I believe that for instance, meditation might take you
to the subjective experiences,
which are very pleasant, collaborative.
And I would like a world to move toward
more collaborative place.
Yeah, I would say that's very pleasant
and I enjoy doing stuff like that.
- I wonder how that maps to your mathematical model of love
with the reward function, combining a bunch of things.
It seems like our life then is we're just,
we have this reward function and we're accumulating
a bunch of stuff in it with weights.
It's like multi-objective.
And what meditation is, is you just remove them,
and remove them until the weight on one
or just a few is very high.
And that's where the pleasure comes from.
- Yeah, so something similar to how I'm thinking about this.
So I told you that there is this like,
there is a story of who you are.
And I think almost about it as a text prepended to GPT.
(laughing)
- Yeah.
- And some people refer to it as ego.
Okay, there's like a story.
Who you are, okay?
- So ego is the prompt for GPT-3?
- Yeah. - Or GPT?
- Yes, yes, and that's description of you.
And then with meditation, you can get to the point
that actually you experience things without the prompt.
And you experience things as they are.
You are not biased or the description
how they supposed to be.
That's very pleasant.
And then with respect to the reward function,
it's possible to get to the point
that there is the solution of self.
And therefore you can say that you are having a,
or like your brain attempts to simulate
the reward function of everyone else or like everything.
That's, there is this like a love
which feels like a oneness with everything.
And that's also, you know, very beautiful, very pleasant.
At some point, you might have a lot of altruistic thoughts
during that moment, and then the self always comes back.
- How would you recommend,
if somebody is interested in meditation,
like a big thing to take on as a project,
would you recommend a meditation retreat?
How many days?
What kind of thing would you recommend?
- I think that actually retreat is the way to go.
It almost feels that, as I said,
like meditation is a psychedelic,
but when you take it in the small dose,
you might barely feel it.
Once you get the high dose, actually you're gonna feel it.
- So even cold turkey,
if you haven't really seriously meditated
for a prolonged period of time, just go to a retreat.
- Yeah. - How many days?
How many days? - I would start weekend one.
- Weekend, so like two, three days.
- And it's like, it's interesting that first or second day,
it's hard, and at some point it becomes easy.
- There's a lot of seconds in a day.
How hard is the meditation retreat?
Just sitting there in a chair.
- So the thing is actually,
it literally just depends on your own framing.
Like if you are in the mindset
that you are waiting for it to be over,
or you are waiting for a nirvana to happen,
it will be very unpleasant.
And in some sense, even the difficulty,
it's not even in the lack of being able
to speak with others.
Like you are sitting there,
your legs will hurt from sitting.
- In terms of like the practical things,
do you experience kind of discomfort,
like physical discomfort of just sitting,
like your butt being numb,
your legs being sore, all that kind of stuff?
- Yes, you experience it,
and then they teach you to observe it.
The crazy thing is, you at first might have a feeling
toward trying to escape it,
and that becomes very apparent
that that's extremely unpleasant.
And then you just observe it,
and at some point it just becomes, it just is.
It's like, I remember Ilya told me some time ago
that he takes a cold shower,
and his mindset of taking a cold shower
was to embrace suffering.
- Yeah, excellent.
I do the same.
- Is this your style?
- Yeah, it's my style.
I like this.
- So my style is actually,
I also sometimes take cold showers.
It is purely observing how the water goes through my body,
like purely being present,
not trying to escape from there.
- Yeah.
- And I would say then it actually becomes pleasant.
It's not like, "Ugh!"
- Well, that's interesting.
I'm also, that's the way to deal
with anything really difficult,
especially in the physical space,
is to observe it.
To say it's pleasant,
hmm.
I would use a different word.
You're accepting of the full beauty of reality,
I would say, 'cause say pleasant.
But yeah, in some sense it is pleasant.
That's the only way to deal with a cold shower
is to become an observer and to find joy in it.
Same with really difficult physical exercise,
or running for a really long time, endurance events,
just any time you're exhausted, any kind of pain.
I think the only way to survive it is not to resist it,
it's to observe it.
You mentioned Ilya.
Ilya Sutskever.
- He's very, he's our chief scientist,
but also he's a very close friend of mine.
- He co-founded OpenAI with you.
I've spoken with him a few times.
He's brilliant, I really enjoy talking to him.
His mind, just like yours, works in fascinating ways.
Both of you are not able to define deep learning simply.
What's it like having him as somebody
you have technical discussions with
on in-space machine learning, deep learning, AI,
but also life?
What's it like when these two agents
get into a self-play situation in a room?
What's it like collaborating with him?
- So I believe that we have extreme respect to each other.
So I love Ilya's insight,
both like, I guess, about consciousness, life, AI.
- But in terms of the, it's interesting to me
'cause you're a brilliant thinker
in the space of machine learning, like intuition,
digging deep in what works, what doesn't,
why it works, why it doesn't, and so is Ilya.
I'm wondering if there's interesting,
deep discussions you've had with him in the past
or disagreements that were very productive.
- So I can say, I also understood over the time
where are my strengths.
So obviously we have plenty of AI discussions.
And I myself have plenty of ideas,
but I consider Ilya one of the most prolific AI scientists
in the entire world.
And I think that I realized that maybe my super skill
is being able to bring people to collaborate together,
that I have some level of empathy
that is unique in AI world.
And that might come from either meditation, psychedelics,
or let's say I read just hundreds of books on this topic.
So, and I also went through a journey of,
I developed all sorts of algorithms.
So I think that maybe I can,
§08Transcript · 1:10:04–1:20:01
that's my superhuman skill.
Ilya is one of the best AI scientists,
but then I'm pretty good in assembling teams.
And I'm also not holding to people,
like I'm growing people
and then people become managers at OpenAI.
I grew many of them, like a research managers.
- So you find places where you're excellent
and he finds like his deep scientific insights
is where he is.
And you find ways you can, the puzzle pieces fit together.
- Correct.
Like, ultimately, for instance, let's say Ilya,
he doesn't manage people.
That's not what he likes or so.
I like hanging out with people.
By default, I'm an extrovert and I care about people.
- Oh, interesting.
Okay.
Okay, cool.
So that fits perfectly together.
But I mean, I also just like your intuition
about various problems in machine learning.
He's definitely one I really enjoy.
I remember talking to him
about something I was struggling with,
which is coming up with a good model for pedestrians,
for human beings that cross the street
in the context of autonomous vehicles.
And he immediately started to like formulate a framework
within which you can evolve a model for pedestrians,
like through self-play, all that kind of mechanisms.
The depth of thought on a particular problem,
especially problems he doesn't know anything about,
is fascinating to watch.
It makes you realize like,
yeah, the limits that the human intellect might be limitless.
Or it's just impressive to see a descendant of ape
come up with clever ideas.
- Yeah, I mean, so even in the space of deep learning,
when you look at various people,
there are people who invented some breakthroughs once,
but there are very few people who did it multiple times.
And you can think if someone invented it once,
that might be just a sheer luck.
And if someone invented it multiple times,
if a probability of inventing it once is one over a million,
then probability of inventing it twice or three times
would be one over a million squared or to the power of three,
which would be just impossible.
So it literally means that it's given that it's not the luck.
And Ilya is one of these few people
who have a lot of these inventions in his arsenal.
It also feels that, for instance,
if you think about folks like Gauss or Euler,
at first they read a lot of books
and then they did thinking and then they figure out math.
And that's how it feels with Ilya.
You know, at first he read stuff
and then he spent his thinking cycles.
- That's a really good way to put it.
When I talk to him, I see thinking.
He's actually thinking.
Like he makes me realize that there's like deep thinking
that the human mind can do.
Like most of us are not thinking deeply.
Like you really have to put a lot of effort to think deeply.
Like I have to really put myself in a place
where I think deeply about a problem.
It takes a lot of effort.
It's like an airplane taking off or something.
You have to achieve deep focus.
He's just, what is it?
His brain is like a vertical takeoff
in terms of airplane analogy.
So it's interesting.
But I mean, Cal Newport talks about this
as ideas of deep work.
Most of us don't work much at all
in terms of like deeply think about particular problems,
whether it's a math, engineering, all that kind of stuff.
You want to go to that place often
and that's real hard work.
And some of us are better than others at that.
- So I think that the big piece has to do
with actually even engineering your environment
such that it's conducive to that.
So see, both Ilya and I, on the frequent basis,
we kind of disconnect ourselves from the world
in order to be able to do extensive amount of thinking.
So Ilya usually, he just leaves iPad at hand.
He loves his iPad.
And for me, I'm even sometimes,
just going for a few days to different location, to Airbnb,
I'm turning off my phone
and there is no access to me.
And that's extremely important for me
to be able to actually just formulate new thoughts,
to do deep work rather than to be reactive.
And the older I am,
the more of this like random tasks are at hand.
- Before I go on to that thread,
let me return to our friend, GPT.
Let me ask you another ridiculously big question.
Can you give an overview of what GPT-3 is?
Or like you say in your Twitter bio, GPT-N+1,
how it works and why it works?
- So GPT-3 is a humongous neural network.
Let's assume that we know what is neural network,
gave the definition.
And it is trained on the entire internet
just to predict next word.
So let's say it sees part of the article
and the only task that it has at hand,
it is to say what would be the next word?
What would be the next word?
And it becomes really exceptional at the task
of figuring out what's the next word.
So you might ask, why would this be an important task?
Why would it be important to predict what's the next word?
And it turns out that a lot of problems can be formulated
as a text completion problem.
So GPT is purely learning to complete the text.
And you could imagine, for instance,
if you are asking a question,
who is the president of United States?
Then GPT can give you an answer to it.
It turns out that many more things can be formulated
this way, you can format text
in the way that you have sent us in English.
You make it even look like some content of a website
elsewhere, which would be teaching people
how to translate things between languages.
So it would be EN colon, text in English, FR colon.
And then you ask people, and then you ask model to continue.
And it turns out that such a model
is predicting translation from English to French.
The crazy thing is that this model can be used
for way more sophisticated tasks.
So you can format text such that it looks
like a conversation between two people.
And that might be a conversation between you and Elon Musk.
And because the model read all the texts about Elon Musk,
it will be able to predict Elon Musk words
as it would be Elon Musk.
It will speak about colonization of Mars,
about sustainable future and so on.
And it's also possible to even give
arbitrary personality to the model.
You can say, here is a conversation
that we have a friendly AI bot.
And the model will complete the text as a friendly AI bot.
- So, I mean, how do I express how amazing this is?
So just to clarify, a conversation,
generating a conversation between me and Elon Musk,
it wouldn't just generate good examples
of what Elon would say.
It would get the syntax all correct.
So like interview style,
it would say like Elon colon, Lex colon.
It's not just like inklings of semantic correctness.
It's like the whole thing, grammatical, syntactic, semantic.
It's just really, really impressive generalization.
- Yeah, I mean, I also want to provide some caveats.
So it can generate few paragraphs of coherent text,
but as you go to longer pieces,
it actually goes off the rails.
Okay, if you try to write a book,
it won't work out this way.
- What way does it go off the rails, by the way?
Is there interesting ways in which it goes off the rails?
Like what falls apart first?
- So the model is trained on the old existing data
that is out there,
which means that it is not trained on its own mistakes.
So for instance, if it would make a mistake,
then I kept, so to give you an example.
So let's say I have a conversation
with a model pretending that is Elon Musk.
And then I start putting some,
I'm start actually making up things which are not factual.
I would say- - Sounds like Twitter.
(laughing)
But I got you, sorry, yeah.
- Like, I don't know, I would say that Elon is my wife
and the model will just keep on carrying it on.
- As if it's true.
- Yes, and in some sense,
if you would have a normal conversation with Elon,
he would be, "What the fuck?"
- Yeah, there would be some feedback between.
So the model is trained on things that humans have written,
but through the generation process,
there's no human in the loop feedback.
- Correct.
- That's fascinating.
- Makes sense?
- It magnifies, like the errors get magnified and magnified.
- Correct.
- And it's also interesting.
§09Transcript · 1:20:03–1:30:15
I mean, first of all, humans have the same problem.
It's just that we will make fewer errors
and magnify the errors slower.
- I think that actually what happens with humans is
if you have a wrong belief about the world as a kid,
then very quickly you will learn that it's not correct
because you are grounded in reality
and you are learning from your new experience.
- Yes.
But do you think the model can correct itself too?
Won't it through the power of the representation,
and so the absence of Elon Musk being your wife,
information on the internet, won't it correct itself?
- There won't be examples like that.
- So the errors will be subtle at first.
- Subtle at first, and in some sense,
you can also say that the data that is not out there
is the data which would represent how the human learns.
And maybe model would be trained on such a data,
then it would be better off.
- How intelligent is GPT-3, do you think?
Like when you think about the nature of intelligence,
it seems exceptionally impressive.
But then if you think about the big AGI problem,
is this footsteps along the way to AGI?
- So let's see.
Seems that intelligence itself is,
there are multiple axis of it.
And I would expect that the systems that we are building,
they might end up being superhuman on some axis,
and subhuman on some other axis.
It would be surprising to me on all axis simultaneously,
they would become superhuman.
Of course, people ask this question,
is GPT a spaceship that would take us to the moon,
or are we building a ladder to heaven,
that we are just building bigger and bigger ladder?
And we don't know in some sense,
which one of these two--
- Which one is better?
I'm trying to, I like "Stairway to Heaven",
it's a good song,
so I'm not exactly sure which one is better.
But you're saying like the spaceship to the moon
is actually effective?
- Correct, so people who criticize GPT,
they say, "You're guys just building a taller ladder."
And it will never reach the moon.
And at the moment, I would say the way I'm thinking is,
this is like a scientific question.
And I'm also in heart, I'm a builder creator.
And like, I'm thinking, let's try out,
let's see how far it goes.
And so far we see constantly that there is a progress.
- Yeah, so do you think GPT-4,
GPT-5, GPT-N+1,
there'll be a phase shift,
like a transition to a place
where we'll be truly surprised?
Then again, like GPT-3 is already very, truly surprising.
The people that criticize GPT-3 as a, what is it?
Ladder to heaven,
I think too quickly get accustomed to how impressive it is,
that the prediction of the next word
can achieve such depth of semantics,
with accuracy of syntax, grammar, and semantics.
Do you think GPT-4 and 5 and 6
will continue to surprise us?
- I mean, definitely there will be more impressive models.
There is a question, of course,
if there will be a phase shift.
And also even the way I'm thinking about these models
is that when we build these models,
you know, we see some level of the capabilities,
but we don't even fully understand everything
that the model can do.
And actually one of the best things to do
is to allow other people to probe the model,
to even see what is possible.
- Hence the using GPT as an API
and opening it up to the world.
- Yeah, I mean, so when I'm thinking from perspective of,
obviously various people have concerns about AGI,
including myself.
And then when I'm thinking from perspective,
what's the strategy even to deploy these things to the world?
Then the one strategy that I have seen many times working
is the iterative deployment,
that you deploy slightly better versions
and you allow other people to criticize you.
So you actually, or try it out,
you see where are there fundamental issues.
And it's almost, you don't want to be in that situation
that you are holding into powerful system
and there's like a huge overhang,
then you deploy it and it might have a random chaotic impact
on the world.
So you actually want to be in the situation
that you are gradually deploying systems.
- I asked this question of Ilya,
let me ask you this question.
I've been reading a lot about Stalin and power.
If you're in possession of a system that's like AGI,
that's exceptionally powerful,
do you think your character and integrity
might become corrupted?
Like famously power corrupts
and absolute power corrupts absolutely.
- So I believe that you want at some point
to work toward distributing the power.
I think that you want to be in the situation
that actually AGI is not controlled
by a small number of people,
but essentially by a larger collective.
- So the thing is that requires
a George Washington style move.
In the ascent to power,
there's always a moment when somebody gets a lot of power
and they have to have the integrity
and the moral compass to give away that power.
That humans have been good and bad throughout history
at this particular step.
And I wonder, I wonder we like blind ourselves in,
for example, between nations,
a race towards, yeah, AI race between nations.
We might blind ourselves and justify to ourselves
the development of AI without distributing the power
because we want to defend ourselves against China,
against Russia, that kind of logic.
And I wonder how we design governance mechanisms
that prevent us from becoming power hungry
and in the process destroying ourselves.
- So let's see.
I have been thinking about this topic quite a bit,
but I also want to admit that once again,
I actually want to rely way more on Sam Altman on it.
He wrote an excellent blog
on how even to distribute wealth.
And he proposed in his blog to tax equity
of the companies rather than profit and to distribute it.
And this is an example of Washington move.
I guess I personally have insane trust in Sam.
He already spent plenty of money
running a universal basic income project.
That gives me, I guess, maybe some level of trust to him,
but I also, I guess, love him as a friend.
- Yeah.
I wonder, because we're sort of summoning
a new set of technologies,
I wonder if we'll be cognizant.
Like you're describing the process of open AI,
but it could also be at other places
like in the US government, right?
Both China and the US are now full steam ahead
on autonomous weapons systems development.
And that's really worrying to me
because in the framework of something being
a national security danger or a military danger,
you can do a lot of pretty dark things
that blind our moral compass.
And I think AI will be one of those things.
In some sense, the mission and the work you're doing
in open AI is like the counterbalance to that.
So you want to have more open AI
and less autonomous weapons systems.
- I like these statements.
Like to be clear, this is interesting
and I'm thinking about it myself,
but this is a place that I put my trust
actually in Sam's hands because it's extremely hard
for me to reason about it.
- Yeah.
- One important statement to make is
it's good to think about this.
- Yeah, no question about it.
No question about it.
- Even like low level, quote unquote engineer.
Like there's such a, I remember I programmed a car,
a RC car.
They went really fast, like 30, 40 miles an hour.
And I remember I was like sleep deprived.
So I programmed it pretty crappily
and it like, the code froze.
So it's doing some basic computer vision
and it's going around on track, but it's going full speed.
And there was a bug in the code that the car just went,
it didn't turn, it went straight full speed
and smashed into the wall.
And I remember thinking the seriousness
with which you need to approach the design
of artificial intelligence systems
and the programming of artificial intelligence systems
is high because the consequences are high.
Like that little car smashing into the wall,
for some reason I immediately thought of like an algorithm
that controls nuclear weapons, having the same kind of bug.
And so like the lowest level engineer
and the CEO of a company all need to have the seriousness
in approaching this problem
and thinking about the worst case consequences.
§10Transcript · 1:30:17–1:39:40
- So I think that is true.
I mean, what I also recognize in myself
and others even asking this question
is that it evokes a lot of fear
and fear itself ends up being actually quite debilitating.
The place where I arrived at the moment
might sound cheesy or so,
but it's almost to build things out of love
rather than fear.
- Yeah.
- Like a focus on how I can maximize the value,
how the systems that I'm building might be useful.
I'm not saying that the fear doesn't exist out there
and like it totally makes sense to minimize it,
but I don't want to be working because I'm scared.
I want to be working out of passion, out of curiosity,
out of the looking forward for the positive future.
- With the definition of love arising
from a rigorous practice of empathy.
So not just like your own conception
of what is good for the world,
but always listening to others.
- Correct, like the love where I'm considering
reward functions of others.
- Others.
To limit to infinity is like a sum,
like one to N where N is 7 billion or whatever it is.
- Not projecting my reward functions on others.
- Yeah, exactly.
- Okay, can we just take a step back
to something else super cool, which is OpenAI Codex?
Can you give an overview of what OpenAI Codex
and GitHub Copilot is, how it works,
and why the hell it works so well?
- So with GPT-3, we noticed that the system,
that system train on all the language out there
started having some rudimentary coding capabilities.
So we're able to ask it to implement addition function
between two numbers,
and indeed it can write Python or JavaScript code for that.
And then we thought we might as well just go full steam ahead
and try to create a system that is actually good
at what we are doing every day ourselves,
which is programming.
We optimize models for proficiency in coding.
We actually even created models
that both have a comprehension of language and code.
And Codex is API for these models.
- So it's first pre-trained on language,
and then I don't know if you can say fine-tuned,
'cause there's a lot of code, but it's language and code.
- It's language and code.
It's also optimized for various things,
like let's say low latency and so on.
Codex is the API, the similar to GPT-3.
We expect that there will be proliferation
of the potential products that can use coding capabilities,
and I can speak about it in a second.
Copilot is a first product developed by GitHub.
So as we're building models,
we wanted to make sure that these models are useful.
And we worked together with GitHub
on building the first product.
Copilot is actually as you code,
it suggests you code completions.
And we have seen in the past,
there are various tools that can suggest
how to view characters of the code or the line of code.
The thing about Copilot is it can generate 10 lines of code.
It's often the way how it works is
you often write in the comment what you want to happen,
because people in comments, they describe what happens next.
So these days when I code,
instead of going to Google to search
for the appropriate code to solve my problem,
I say, "Oh, for this array, could you smooth it?"
And then it imports some appropriate libraries
and say it uses NumPy convolution,
or so that I was not even aware that exists,
and it does the appropriate thing.
- So you write a comment, maybe the header of a function,
and it completes the function.
Of course, you don't know what is the space
of all the possible small programs it can generate.
What are the failure cases?
How many edge cases?
How many subtle errors there are?
How many big errors there are?
It's hard to know, but the fact that it works at all
in a large number of cases is incredible.
It's like a kind of search engine
into code that's been written on the internet.
- Correct, so for instance, when you search things online,
then usually you get to some particular case.
Like if you go to Stack Overflow,
people describe that one particular situation,
and then they seek for a solution.
But in case of Copilot, it's aware of your entire context,
and in context is, "Oh, these are the libraries
"that they are using."
That's the set of the variables that is initialized,
and on the spot, it can actually tell you what to do.
So the interesting thing is,
and we think that Copilot is one possible product
using Codex, but there is a place for many more.
So internally, we tried out to create other fun products.
So it turns out that a lot of tools out there,
let's say Google Calendar or Microsoft Word or so,
they all have internal API to build plugins around them.
So there is a way, a sophisticated way,
to control Calendar or Microsoft Word.
Today, if you want more complicated behaviors
from these programs, you have to add a new button
for every behavior.
But it is possible to use Codex
and tell, for instance, to Calendar,
could you schedule an appointment with Lex
next week after 2 p.m.?
And it writes corresponding piece of code.
And that's the thing that actually you want.
- So interesting.
So what you figure out is there's a lot of programs
with which you can interact through code.
And so there, you can generate that code
from natural language.
That's fascinating.
- And that's somewhat like also closest
to what was the promise of Siri or Alexa.
So previously, all these behaviors,
they were hard-coded.
And it seems that Codex on the fly can pick up the API
of, let's say, a given software.
And then it can turn language into use of this API.
- So without hard-coding, it can find,
it can translate to machine language.
- Correct.
- So for example, this would be really exciting for me,
like for Adobe products like Photoshop,
which I think ActionScript,
I think there's a scripting language
that communicates with them.
Same with Premiere.
- And do you could imagine that that allows even
to do coding by voice on your phone?
So for instance, in the past, as of today,
I'm not editing Word documents on my phone
because it's just the keyboard is too small.
But if I would be able to tell to my phone,
make the header large, then move the paragraphs around,
and that's actually what I want.
So I can tell you one more cool thing,
or even how I'm thinking about Codex.
So if you look actually at the evolution of computers,
we started with very primitive interfaces,
which is a punch card.
And punch card, essentially, you make holes
in the plastic card to indicate zeros and ones.
And during that time, there was a small number
of specialists who were able to use computers.
And by the way, people even suspected
that there is no need for many more people to use computers.
But then we moved from punch cards to,
at first, assembly and C.
And these programming languages,
they were slightly higher level.
They allowed many more people to code.
And they also led to more of a proliferation of technology.
And further on, there was a jump to,
say, from C++ to Java and Python.
And every time it has happened,
more people are able to code,
and we build more technology.
And it's even hard to imagine now
if someone will tell you that you should write code
in assembly instead of, let's say, Python or Java or JavaScript.
And Codex is yet another step
toward kind of bringing computers closer to humans,
such that you communicate with a computer
with your own language,
rather than with a specialized language.
And I think that it will lead to
an increase of number of people who can code.
- Yeah, and the kind of technologies
that those people will create is, it's innumerable.
It could be a huge number of technologies
we're not predicting at all.
'Cause that's less and less requirement
of having a technical mind, a programming mind.
You're not opening it to the world of
other kinds of minds, creative minds, artistic minds,
all that kind of stuff.
- I would like, for instance, biologists who work on DNA
to be able to program
and not to need to spend a lot of time learning it.
And I believe that's a good thing to the world.
And I would actually add, I would add,
so at the moment, I'm managing Codex team
and also language team.
And I believe that there is like
plenty of brilliant people out there,
§11Transcript · 1:39:42–1:48:30
and they should apply.
- Oh, okay, yeah, awesome.
So what's the language in the Codex?
So those are kind of, they're overlapping teams.
So it's like GPT, the raw language,
and then the Codex is like applied to programming.
- Correct, and they are quite intertwined.
There are many more teams involved
making these models extremely efficient and deployable.
For instance, there are people who are working to
make our data centers amazing,
or there are people who work on
putting these models into production,
or even pushing it at the very limit of the scale.
- So all aspects from the infrastructure
to the actual machine learning.
- So I'm just saying there are multiple teams.
While the team working on Codex and language,
I guess I'm directly managing them.
I would love to hire more.
- Yeah, if you're interested in machine learning,
this is probably one of the most exciting problems
and systems to be working on.
'Cause it's actually, it's pretty cool.
Like what the program synthesis,
like generating a program is very interesting,
very interesting problem that has echoes of reasoning
and intelligence in it.
And I think there's a lot of fundamental questions
that you might be able to sneak up to
by generating programs.
- Yeah, one more exciting thing about the programs is that,
so I said that the, you know, in case of language,
that one of the troubles is even evaluating language.
So when the things are made up,
you need somehow either a human to say
that this doesn't make sense.
Or so in case of program, there is this one extra lever
that we can actually execute programs
and see what they evaluate to.
So that process might be somewhat more automated
in order to improve the qualities of generations.
- Oh, that's fascinating.
So like the, wow, that's really interesting.
- So for the language, the, you know,
the simulation to actually execute it, that's a human mind.
- Yeah.
- For programs, there is a computer
on which you can evaluate it.
- Wow, that's a brilliant little insight
that the thing compiles and runs.
That's first.
And second, you can evaluate on a,
like do automated unit testing.
And in some sense, it seems to me
that we'll be able to make a tremendous progress.
You know, we are in the paradigm
that there is way more data.
There is like a transcription of millions
of software engineers.
- Yeah.
Yeah.
So, I mean, you just mean,
'cause I was gonna ask you about reliability.
The thing about programs is you don't know
if they're gonna, like a program
that's controlling a nuclear power plant
has to be very reliable.
- So I wouldn't start with controlling nuclear power plant.
Maybe one day, but that's not actually,
that's not on the current roadmap.
That's not the step one.
- You know, it's the Russian thing.
You just wanna go to the most powerful,
destructive thing right away, run by JavaScript.
But I got you.
So this is a lower impact, but nevertheless,
what you're making me realize,
it is possible to achieve some levels of reliability
by doing testing.
And you could imagine that, you know,
maybe there are ways for a model to write even code
for testing itself and so on.
And there exist ways to create the feedback loops
that the model could keep on improving.
- By writing programs that generate tests.
- For instance.
For instance.
- And that's how we get consciousness
because it's meta compression.
That's what you're gonna write.
That's the comment.
That's the prompt that generates consciousness.
- Compressor of compressors.
You just write that.
Do you think the code that generates consciousness
would be simple?
- So let's see.
I mean, ultimately the core idea behind will be simple,
but there will be also decent amount of engineering involved.
Like in some sense, it seems that, you know,
spreading these models on many machines,
it's not that trivial.
- Yeah.
- And we find all sorts of innovations
that make our models more efficient.
I believe that first models that I guess are conscious
are like a truly intelligent.
They will have all sorts of tricks.
- But then again, there's a, which is certain argument
that maybe the tricks are temporary things.
- Yeah, they might be temporary things.
And in some sense, it's also even important
to know that even the cost of a trick.
So sometimes people are eager to put the trick
while forgetting that there is a cost of maintenance.
- Or like a long-term cost.
- Long-term cost or maintenance,
or maybe even flexibility of code
to actually implement new ideas.
So even if you have something that gives you two X,
but it requires, you know, 1000 lines of code,
I'm not sure if it's actually worth it.
So in some sense, you know,
if it's five lines of code and two X, I would take it.
And we see many of this, but also, you know,
that requires some level of, I guess,
lack of attachment to code that we are willing to remove it.
- Yeah.
So you led the OpenAI robotics team.
Can you give an overview of the cool things
you were able to accomplish?
What are you most proud of?
- So when we started robotics,
we knew that actually reinforcement learning works
and it is possible to solve fairly complicated problems.
Like for instance, AlphaGo is an evidence
that it is possible to build superhuman Go players.
Dota2 is an evidence that it's possible
to build superhuman agents playing Dota.
So I asked myself a question, you know,
what about robots out there?
Could we train machines to solve arbitrary tasks
in the physical world?
Our approach was, I guess,
let's pick a complicated problem that if we would solve it,
that means that we made some significant progress,
the domain, and then we went after the problem.
So we noticed that actually the robots out there,
they are kind of at the moment, optimized per task.
So you can have a robot that it's like,
if you have a robot opening a bottle,
it's very likely that the end factor is that bottle opener.
And in some sense, that's a hack to be able to solve a task,
which makes any task easier.
And I asked myself, so what would be a robot
that can actually solve many tasks?
And we concluded that like human hands have such a quality
that indeed they are, you know,
you have five kind of tiny arms attached.
Individually, they can manipulate
pretty broad spectrum of objects.
So we went after a single hand,
like trying to solve Rubik's cube single-handed.
We picked this task because we thought
that there is no way to hard-code it.
And it's also, we picked a robot
on which it would be hard to hard-code it.
And we went after the solution such that
it could generalize to other problems.
- And just to clarify,
it's one robotic hand solving the Rubik's cube.
The hard part is in the solution to the Rubik's cube
is the manipulation of the, of like having it not fall
out of the hand, having it use the five baby arms to,
what is it, like rotate different parts of the Rubik's cube
to achieve the solution.
- Correct.
- Yeah.
So what was the hardest part about that?
What was the approach taken there?
What are you most proud of?
- Obviously we have like a strong belief
in reinforcement learning.
And, you know, one path it is to do reinforcement learning
in the real world.
Other path is to the simulation.
In some sense, the tricky part about the real world
is at the moment, our models, they require a lot of data.
There is essentially no data.
And I did, we decided to go through the path
of the simulation.
And in simulation, you can have infinite amount of data.
The tricky part is the fidelity of the simulation.
And also can you in simulation represent everything
that you represent otherwise in the real world.
And, you know, it turned out that, you know,
because there is lack of fidelity, it is possible to,
what we arrived at is training a model
that doesn't solve one simulation,
§12Transcript · 1:48:32–1:58:02
but it actually solves the entire range of simulations,
which vary in terms of like, what's the,
exactly the friction of the cube or weight or so.
And the single AI that can solve all of them
ends up working well with the reality.
- How do you generate the different simulations?
- So, you know, there's plenty of parameters out there.
We just pick them randomly.
And in simulation model just goes for thousands of years
and keeps on solving Rubik's cube in each of them.
And the thing is that neural network that we used,
it has a memory and as it presses, for instance,
the side of the cube, it can sense,
oh, that's actually this side was difficult to press.
I should press it stronger.
And throughout this process kind of learn
it's even how to solve this particular instance
of the Rubik's cube by given mass.
It's kind of like, you know, sometimes when you go to a gym
and after bench press, you try to lift the glass
and you kind of forgot and your head goes like,
up right away because kind of you got used
to maybe different weight and it takes a second to adjust.
- Yeah.
- And this kind of a memory, the model gain
through that process of interacting
with the cube in the simulation.
- I appreciate you speaking to the audience
with the bench press, all the bros in the audience
probably working out right now.
There's probably somebody listening to this
actually doing bench press.
So maybe put the bar down and pick up the water bottle
and you'll know exactly what Czech is talking about.
Okay, so what was the hardest part
of getting the whole thing to work?
- So the hardest part is at the moment
when it comes to physical work,
when it comes to robots, they require maintenance.
It's hard to replicate a million times.
It's also, it's hard to replay things exactly.
I remember this situation that one guy at our company,
he had like a model that performs way better
than other models in solving RubyFube.
And, you know, we kind of didn't know what's going on,
why is that?
And it turned out that, you know,
he was running it from his laptop that had better CPU
or better, maybe local GPU as well.
And because of that, there was less of a latency
and the model was the same.
And that actually made solving RubySquib more reliable.
So in some sense, there might be some saddlebacks like that
when it comes to running things in the real world.
Even hinting on that, you could imagine
that the initial models, you would like to have models
which are insanely huge neural networks
and you would like to give them even more time for thinking.
And when you have these real time systems,
then you might be constrained actually
by the amount of latency.
And ultimately I would like to build a system
that it is worth for you to wait five minutes
because it gives you the answer
that you are willing to wait for five minutes.
- So latency is a very unpleasant constraint
under which to operate.
- Correct.
And also there is actually one more thing
which is tricky about robots.
There is actually not much data.
So the data that I'm speaking about would be a data
of first person experience from the robot
and like gigabytes of data like that.
If we would have gigabytes of data like that
of robots solving various problems,
it would be very easy to make a progress on robotics.
And you can see that in case of text or code,
there is a lot of data, like a first person perspective
data on the writing code.
- Yeah, so you had this,
you mentioned this really interesting idea
that if you were to build like a successful robotics company
so OpenAs mission is much bigger than robotics.
This is one of the things you've worked on.
But if it was a robotics company,
that you wouldn't so quickly dismiss supervised learning.
- Correct.
- You would build a robot that was perhaps what,
like an empty shell, like dumb,
and they would operate under teleoperation.
So you would invest, that's just one way to do it.
Invest in human supervision,
like direct human control of the robots as it's learning.
And over time, add more and more automation.
- That's correct.
So let's say that's how I would build
a robotics company today.
If I would be building a robotics company,
which is, you know, spend $10 million or so
recording human trajectories, controlling a robot.
- After you find a thing that the robot should be doing
that there's a market fit for,
like that you can make a lot of money with that product.
- Correct, correct.
- Yeah.
- So I would record data
and then I would essentially train supervised
learning model on it.
That might be the path today.
Long term, I think that actually what is needed
is to train powerful models over video.
So you have seen maybe a models
that can generate images like DALI.
And people are looking into models generating videos.
They're like, "Why are these algorithmic questions
even how to do it?"
And it's unclear if there is enough compute
for this purpose.
But I suspect that the models that,
which would have a level of understanding of video,
same as GPT has a level of understanding of text,
could be used to train robots to solve tasks.
They would have a lot of common sense.
- If one day, I'm pretty sure one day,
there will be a robotics company,
by robotics company I mean the primary source of income
is from robots that is worth over $1 trillion.
What do you think that company will do?
I think self-driving cars now.
- It's interesting 'cause my mind went to personal robotics,
robots in the home.
It seems like there's much more market opportunity there.
I think it's very difficult to achieve.
I mean, this might speak to something important,
which is I understand self-driving much better
than I understand robotics in the home.
So I understand how difficult it is
to actually solve self-driving.
To a level, not just the actual computer vision
and the control problem
and just the basic problems of driving,
but creating a product that would undeniably be,
that will cost less money,
like it will save you a lot of money,
like orders of magnitude less money
that could replace Uber drivers, for example.
So car sharing that's autonomous,
that creates a similar or better experience
in terms of how quickly you get from A to B
or just whatever.
The pleasantness of the experience,
the efficiency of the experience,
the value of the experience,
and at the same time, the car itself costs cheaper.
I think that's very difficult to achieve.
I think there's a lot more low-hanging fruit in the home.
- That could be.
I also want to give you a perspective
on how challenging it would be at home
or maybe kind of depends on the exact problem
that you'd be solving.
If we're speaking about these robotic arms
and hence these things,
they cost tens of thousands of dollars or maybe 100K.
And maybe obviously,
maybe there would be economy of scale,
these things would be cheaper,
but actually for any household to buy it,
the price would have to go down to maybe a thousand bucks.
- Yeah.
I personally think that,
so self-driving car provides a clear service.
I don't think robots in the home,
there'll be a trillion dollar company
will just be all about service.
Meaning it will not necessarily be about like a robotic arm
that's helps you, I don't know, open a bottle
or wash the dishes or any of that kind of stuff.
It has to be able to take care of that whole,
the therapist thing you mentioned.
I think that's, of course,
there's a line between what is a robot and what is not.
Like, does it really need a body?
But some AI system with some embodiment, I think.
- So the tricky part,
when you think actually what's the difficult part is
when the robot has,
like when there is a diversity of the environment
with which the robot has to interact, that becomes hard.
So on one spectrum, you have industrial robots,
as they are doing over and over the same thing,
it is possible to some extent to prescribe the movements
and with very small amount of intelligence,
the movement can be repeated millions of times.
There are also various pieces of industrial robots
where it becomes harder and harder.
Like for instance, in case of Tesla,
might be a matter of putting a rug inside of a car.
And because the rug kind of moves around,
§13Transcript · 1:58:05–2:06:42
it's not that easy, it's not exactly the same every time.
There's a big the case
that you need actually humans to do it.
While welding cars together, it's a very repetitive process.
Then in case of self-driving itself,
the difficulty has to do with the diversity
of the environment, but still the car itself,
the problem that you are solving
is you try to avoid even interacting with things.
You are not touching anything around
'cause touching itself is hard.
And then if you would have in the home robot
that has to touch things,
and if these things, they change the shape,
if there is a huge variety of things to be touched,
then that's difficult.
If you are speaking about the robot,
which there is head that is smiling in some way
with cameras that doesn't touch things,
that's relatively simple.
- Okay, so to both agree and to push back.
So you're referring to touch like soft robotics,
like the actual touch.
But I would argue that you could formulate
just basic interaction between like non-contact interaction
is also a kind of touch.
And that might be very difficult to solve.
That's the basic, not disagreement,
but that's the basic open question to me
with self-driving cars and disagreement with Elon,
which is how much interaction is required
to solve self-driving cars?
How much touch is required?
You said that in your intuition, touch is not required.
In my intuition to create a product
that's compelling to use,
you're going to have to interact with pedestrians,
not just avoid pedestrians, but interact with them.
When we drive around in major cities,
we're constantly threatening everybody's life
with our movements.
And that's how they respect us.
There's a game they're ready to go on with pedestrians.
And I'm afraid you can't just formulate autonomous driving
as a collision avoidance problem.
- So I think it goes beyond,
like a collision avoidance is the first order approximation,
but then at least in case of Tesla,
they are gathering data from people driving their cars.
And I believe that's an example of supervised learning data
that they can train their models on,
and they are doing it,
which can give a model this like another level of behavior
that is needed to actually interact with the real world.
- Yeah, it's interesting how much data
is required to achieve that.
What do you think of the whole Tesla autopilot approach,
the computer vision based approach with multiple cameras,
and as a data engine,
it's a multitask, multi-headed neural network,
and is this fascinating process of,
similar to what you're talking about
with the robotics approach,
which is, you deploy neural network
and then there's humans that use it,
and then it runs into trouble in a bunch of places
and that stuff is sent back.
So like the deployment discovers a bunch of edge cases
and those edge cases are sent back for supervised annotation
thereby improving the neural network.
And that's deployed again,
it goes over and over until the network becomes really good
at the task of driving, becomes safer and safer.
What do you think of that kind of approach to robotics?
- I believe that's the way to go.
So in some sense, even when I was speaking about,
collecting trajectories from humans,
that's like a first step,
and then you deploy the system
and then you have humans revising all the issues.
And in some sense,
like at this approach converges to system
that doesn't make mistakes
because for the cases where there are mistakes,
you got their data, how to fix them,
and the system will keep on improving.
- So there's a very, to me, difficult question
of how hard that, how long that converging takes,
how hard it is.
The other aspect of autonomous vehicles
is probably applies to certain robotics applications
is society, right?
They put, as the quality of the system converges,
so one, there's a human factors perspective of psychology
of humans being able to supervise those,
even with teleoperation, those robots.
And the other society willing to accept robots.
Currently society is much harsher on self-driving cars
than it is on human driven cars
in terms of the expectation of safety.
So the bar is set much higher than for humans.
And so if there's a death in an autonomous vehicle,
that's seen as a much more,
much more dramatic than a death
in a human driven vehicle.
Part of the success of deployment of robots
is figuring out how to make robots part of society,
both on the, just the human side,
on the media journalist side,
and also on the policy government side.
And that seems to be,
maybe you can put that into the objective function
to optimize, but that is definitely a tricky one.
And I wonder if that is actually the trickiest part
for self-driving cars or any system that's safety critical.
It's not the algorithm,
it's the society accepting it.
- Yeah, I would say,
I believe that the part of the process of deployment
is actually showing people
that they're given things can be trusted.
- Yeah.
- And, you know, trust is also like a glass
that is actually really easy to crack it.
- Yeah.
- And damage it.
And I think that's actually very common
with innovation,
that there is some resistance toward it.
- Yeah.
- And it's just a natural progression.
So in some sense, people will have to keep on proving
that indeed the systems are worth being used.
And I would say,
I also found out that often the best way to convince people
is by letting them experience it.
- Yeah, absolutely.
That's the case with Tesla Autopilot, for example.
That's the case with,
yeah, with basically robots in general.
It's kind of funny to hear people talk about robots.
Like there's a lot of fear,
even with like legged robots.
But when they actually interact with them,
there's joy.
I love interacting with them.
And the same with the car.
With the robot,
if it starts being useful,
I think people immediately understand.
And if the product is designed well, they fall in love.
You're right.
- It's actually even similar
when I'm thinking about Copilot,
the GitHub Copilot.
There was a spectrum of responses that people had.
And ultimately,
the important piece was to let people try it out.
And then many people just loved it.
- Especially like programmers.
- Yeah, programmers.
But like some of them, they came with a fear.
- Yeah.
- But then you try it out
and you think, actually, that's cool.
And you can try to resist the same way as,
you could resist moving from punch cards
to let's say, C++ or so.
And it's a little bit futile.
- So we talked about generation of program,
generation of language,
even self-supervised learning in the visual space
for robotics and then reinforcement learning.
What do you, in like this whole beautiful spectrum of AI,
do you think is a good benchmark,
a good test to strive for
to achieve intelligence?
That's a strong test of intelligence.
You know, it started with Alan Turing and the Turing test.
Maybe you think natural language conversation
is a good test?
- So, you know, it would be nice if, for instance,
machine would be able to solve Riemann hypothesis in math.
That would be, I think that would be very impressive.
- So theorem proving, is that to you,
proving theorems is a good,
oh, like one thing that the machine did,
you would say, damn.
- Exactly.
- Okay.
- That would be quite impressive.
I mean, the tricky part about the benchmarks is,
you know, as we are getting closer with them,
we have to invent new benchmarks.
There is actually no ultimate benchmark out there.
- Yeah, see, my thought with the Riemann hypothesis
would be the moment the machine proves it,
we would say, okay, well, then the problem was easy.
- That's what happens.
And I mean, in some sense,
§14Transcript · 2:06:43–2:16:02
that's actually what happens over the years in AI,
that like, we get used to things very quickly.
- You know something, I talked to Rodney Brooks,
I don't know if you know who that is.
He called AlphaZero a homework problem.
'Cause he was saying like, there's nothing special about it.
It's not a big leap.
And I didn't, well, he's coming from one of the aspects
that we referred to is,
he was part of the founding of iRobot,
which deployed now tens of millions of robot in the home.
So if you see robots that are actually in the homes of people
as the legitimate instantiation of artificial intelligence,
then yes, maybe an AI that plays a silly game
like Go and chess is not a real accomplishment,
but to me, it's a fundamental leap.
But I think we as humans then say, okay,
well, then that game of chess or Go wasn't that difficult
compared to the thing that's currently unsolved.
So my intuition is that from perspective of the evolution
of these AI systems,
we'll at first see the tremendous progress in digital space.
And the main thing about digital space
is also that you can, everything is,
that there is a lot of recorded data,
plus you can very rapidly deploy things
to billions of people.
While in case of physical space,
the deployment part takes multiple years.
You have to manufacture things
and delivering it to actual people is very hard.
So I'm expecting that the first
and that prices in digital space of goods,
they would go down to, let's say marginal costs are to zero.
- And also the question is how much of our life
will be in digital?
Because it seems like we're heading towards
more and more of our lives being in the digital space.
So like innovation in the physical space
might become less and less significant.
Like, why do you need to drive anywhere
if most of your life is spent in virtual reality?
- I still would like to, at least at the moment,
my impression is that I would like to have a physical contact
with other people and that's very important to me.
We don't have a way to replicate it in the computer.
It might be the case that over the time it will change.
- Like in 10 years from now,
why not have like an arbitrary infinite number of people
you can interact with?
Some of them are real, some are not
with arbitrary characteristics that you can define
based on your own preferences.
- I think that's maybe where we are heading
and maybe I'm resisting the future.
- Yeah.
I'm telling you, if I got to choose,
if I could live in Elder Scrolls Skyrim
versus the real world,
I'm not so sure I would stay with the real world.
- Yeah, I mean, the question is,
will VR be sufficient to get us there
or do you need to plug electrodes in the brain?
And it would be nice if these electrodes
wouldn't be invasive.
- Yeah.
Or at least like provably non-destructive.
But in the digital space,
do you think we'll be able to solve the Turing test,
the spirit of the Turing test,
which is, do you think we'll be able to achieve
compelling natural language conversation between people?
Like have friends that are AI systems on the internet.
- I totally think it's doable.
- Do you think the current approach of GPT
will take us there? - Yes.
So there is the part of at first
learning all the content out there.
And I think that still system should keep on learning
as it speaks with you.
- Yeah.
- And I think that should work.
The question is how exactly to do it.
And obviously we have people at OpenAI
asking these questions and kind of at first pre-training
on all existing content is like a backbone
and is a decent backbone.
- Do you think AI needs a body
connecting to our robotics question
to truly connect with humans?
Or can most of the connection be in the digital space?
- So let's see.
We know that there are people who met each other online
and they felt in love.
- Yeah.
- So it seems that it's conceivable to establish connection
which is purely through internet.
Of course, it might be more compelling
the more modalities you add.
- So it would be like you're proposing like a Tinder
but for AI.
You like swipe right and left
and half the systems are AI and the other is humans
and you don't know which is which.
- That would be our formulation of Turing test.
The moment AI is able to achieve more swipe right or left
or whatever, the moment it's able to be more attractive
than other humans, it passes the Turing test.
- Then you would pass the Turing test in attractiveness.
- That's right.
Well, no, like attractiveness just to clarify.
- That would be conversation.
- Not just visual, right, right, right.
It's also attractiveness with wit and humor
and whatever makes conversations pleasant for humans.
Okay, all right.
So you're saying it's possible to achieve
in a digital space.
- In some sense, I would almost ask the question,
why wouldn't that be possible?
- Right.
Well, I have this argument with my dad all the time.
He thinks that touch and smell are really important.
- So they can be very important
and I'm saying the initial systems, they won't have it.
Still, I wouldn't, like there are people being born
without these senses and I believe that they can still
fall in love and have meaningful life.
- Yeah, I wonder if it's possible to go close
to all the way by just training on transcripts
of conversations.
Like I wonder how far that takes us.
- So I think that actually still you want images.
Like I would like, so I don't have kids,
but like I could imagine having AI tutor,
it has to see kids drawing some pictures on the paper.
- And also facial expressions, all that kind of stuff.
We use, dogs and humans use their eyes
to communicate with each other.
I think that's a really powerful mechanism
of communication, body language too,
that words are much lower bandwidth.
- And for body language, we still,
we can have a system that displays an image
of its or facial expression on the computer.
It doesn't have to move, you know, mechanical pieces or so.
So I think that there is like kind of a progression.
You can imagine that text might be the simplest to tackle,
but this is not a complete human experience at all.
You expand it to, let's say, images,
both for input and output.
And what you describe is actually the final,
I guess, frontier, what makes us human,
the fact that we can touch each other or smell or so.
And it's the hardest from perspective
of data and deployment.
And I believe that these things might happen gradually.
- Are you excited by that possibility,
this particular application of human to AI
friendship and interaction?
- So let's see.
- Like, would you, do you look forward to a world,
you said you're living with a few folks
and you're very close friends with them.
Do you look forward to a day where one or two
of those friends are AI systems?
- So if the system would be truly wishing me well,
rather than being in the situation that it optimizes
for my time to interact with the system.
- The line between those is, it's a gray area.
- I think that's the distinction between love and possession.
And these things, they might be often correlated
for humans, but it's like, you might find that
there are some friends with whom you haven't spoke
for months.
And then, you pick up the phone,
it's as the time hasn't passed.
It's they are not holding to you.
And I wouldn't like to have AI system that,
you know, it's trying to convince me to spend time with it.
I would like the system to optimize for what I care about
and help me in achieving my own goals.
- But there's some, I mean, I don't know.
There's some manipulation, there's some possessiveness,
there's some insecurities, there's fragility.
All those things are necessary to form a close friendship
over time, to go through some dark shit together,
some bliss and happiness together.
I feel like there's a lot of greedy self-centered behavior
within that process.
- My intuition, but I might be wrong,
is that human-computer interaction
doesn't have to go through computer being greedy,
possessive, and so on.
It is possible to train systems,
maybe that they actually, you know,
they are, I guess, prompted or fine-tuned or so,
to truly optimize for what you care about.
And you could imagine that, you know,
that the way how the process would look like
§15Transcript · 2:16:04–2:25:41
is at some point, we as humans,
we look at the transcript of the conversation
or like an entire interaction, and we say,
"Actually, here, there was more loving way to go about it."
And we supervise system toward being more loving.
Or maybe we train the system such that
it has a reward function toward being more loving.
- Yeah, or maybe the possibility of the system
being an asshole and manipulative and possessive
every once in a while is a feature, not a bug.
Because some of the happiness that we experience
when two souls meet each other,
when two humans meet each other,
is a kind of break from the assholes in the world.
And so you need assholes in AI as well,
because like, it'll be like a breath of fresh air
to discover an AI that the three previous AIs you had-
- Are too friendly?
- Are, no, or cruel or whatever.
It's like some kind of mix.
And then this one is just right.
But you need to experience the full spectrum.
Like, I think you need to be able to engineer assholes.
- So, let's see.
- Because there's some level to us being appreciated,
to appreciate the human experience,
we need the dark and the light.
- So that kind of reminds me.
I met a while ago at the meditation retreat,
one woman, beautiful, beautiful woman.
And she had a crutch, okay?
She had trouble walking on one leg.
I asked her what has happened.
And she said that five years ago,
she was in Maui, Hawaii,
and she was eating a salad
and some snail fell into the salad.
And apparently there are neurotoxic snails over there.
And she got into coma for a year.
And apparently there is a high chance of even just dying,
but she was in the coma.
At some point, she regained partially consciousness.
She was able to hear people in the room.
People behave as she wouldn't be there.
At some point, she started being able to speak,
but she was mumbling,
like barely able to express herself.
At some point, she got into wheelchair.
Then at some point,
she actually noticed that she can move her toe.
And then she knew that she will be able to walk.
And then, you know, that's where she was five years after.
And she said that since then,
she appreciates the fact that she can move her toe.
And I was thinking,
do I need to go through such experience
to appreciate that I can move my toe?
- Wow, that's really good story.
A really deep example, yeah.
- And in some sense,
it might be the case that we don't see light
if we haven't went through the darkness.
But I wouldn't say that we should.
- We shouldn't assume that that's the case.
We may be able to engineer shortcuts.
- Yeah, Ilya had this, you know,
belief that maybe one has to go for a week or six months
to some challenging camp.
- Yeah.
- To just experience, you know, a lot of difficulties.
And then comes back and actually everything is bright.
Everything is beautiful.
- I'm with Ilya on this.
It must be a Russian thing.
Where are you from originally?
- I'm Polish.
- Polish, okay.
I'm tempted to say that explains a lot,
but yeah, there's something about the Russian,
the necessity of suffering.
I believe suffering or rather struggle is necessary.
- I believe that struggle is necessary.
I mean, in some sense,
you even look at the story of any superhero, the movie.
It's not that it was like everything goes easy, easy, easy.
- I like how that's your ground truth
is the story of superheroes.
Okay, you mentioned that you used to do research at night
and go to bed at like 6 a.m. or 7 a.m.
I still do that often.
What sleep schedules have you tried
to make for a productive and happy life?
Like, is there some interesting wild sleeping patterns
that you engaged that you found
that works really well for you?
- I tried at some point decreasing number of hours of sleep,
like gradually, like half an hour every few days to this.
You know, I was hoping to just save time.
That clearly didn't work for me.
Like at some point there's like a phase shift
and I felt tired all the time.
You know, there was a time that I used to work
during the nights.
The nice thing about the nights is that no one disturbs you.
And even I remember when I was meeting for the first time
with Greg Brockman, he's CTO and chairman of OpenAI.
Our meeting was scheduled to 5 p.m.
and I overstepped for the meeting.
- Overslept for the meeting at 5 p.m., yeah.
Now you sound like me, that's hilarious, okay, yeah.
- And at that moment, in some sense,
my sleeping schedule also has to do with the fact
that I'm interacting with people.
I sleep without an alarm.
- So, yeah, the team thing you mentioned,
the extrovert thing, because most humans operate
during a certain set of hours,
you're forced to then operate at the same set of hours.
But I'm not quite there yet.
I found a lot of joy, just like you said,
working through the night, because it's quiet,
because the world doesn't disturb you.
And there's some aspect,
counter to everything you're saying,
there's some joyful aspect to sleeping
through the mess of the day,
because people are having meetings and sending emails
and there's drama, meetings.
I can sleep through all the meetings.
- You know, I have meetings every day
and they prevent me from having sufficient amount of time
for focused work.
And then I modified my calendar
and I said that I'm out of office Wednesday, Thursday
and Friday every day,
and I'm having meetings only Monday and Tuesday.
And that vastly, positively influenced my mood,
that I have literally like three days for fully focused work.
- Yeah, so there's better solutions to this problem
than staying awake all night.
Okay, you've been part of development
of some of the greatest ideas in artificial intelligence.
What would you say is your process
for developing good novel ideas?
- You have to be aware that clearly
there are many other brilliant people around.
So you have to ask yourself a question,
why the given idea, let's say,
wasn't tried by someone else.
And in some sense, it has to do with,
you know, kind of simple, it might sound simple,
but like thinking outside of the box.
And what do I mean here?
So for instance, for a while, people in academia,
they assumed that you have a fixed data set
and then you optimize the algorithms
in order to get the best performance.
And that was so ingrained assumption
that no one thought about training models on anti-internet.
Or like that maybe some people thought about it,
but it felt to many as unfair.
And in some sense, that's almost like,
it's not my idea or so,
but that's an example of breaking a typical assumption.
So you want to be in the paradigm
that you are breaking a typical assumption.
- In the context of the AI community,
getting to pick your data set is cheating.
- Correct, and in some sense,
so that was assumption that many people had out there.
And then if you free yourself from assumptions,
then they are likely to achieve something
that others cannot do.
And in some sense, if you are trying to do
exactly the same things as others,
it's very likely that you're going to have the same results.
- Yeah, but there's also that kind of tension,
which is asking yourself the question,
why haven't others done this?
Because, I mean, I get a lot of good ideas,
but I think probably most of them suck
when they meet reality.
- So actually, I think the other big piece
is getting into habit of generating ideas,
training your brain towards generating ideas,
and not even suspending judgment of the ideas.
So in some sense, I noticed myself
that even if I'm in the process of generating ideas,
if I tell myself, oh, that was a bad idea,
then that actually interrupts the process,
and I cannot generate more ideas,
because I'm actually focused on the negative part,
why it won't work.
But I created also environment in the way
that it's very easy for me to store new ideas.
So for instance, next to my bed, I have a voice recorder,
and it happens to me often,
like I wake up during the night, and I have some idea.
In the past, I was writing them down on my phone,
but that means turning on the screen,
§16Transcript · 2:25:44–2:34:37
and that wakes me up, or like pulling a paper,
which requires turning on the light.
These days, I just start recording it.
- What do you think, I don't know if you know
who Jim Keller is.
- I know Jim Keller.
- He's a big proponent of thinking harder
on a problem right before sleep,
so that he can sleep through it, and solve it in his sleep,
or like come up with radical stuff in his sleep.
He was trying to get me to do this.
- So it happened from my experience perspective,
it happened to me many times during the high school days
when I was doing mathematics,
that I had the solution to my problem as I woke up.
At the moment, regarding thinking hard
about the given problem is,
I'm trying to actually devote substantial amount of time
to think about important problems,
not just before the sleep.
I can organizing amount of the huge chunks of time,
such that I'm not constantly working on the urgent problems,
but I actually have time to think about the important one.
- So you do it naturally.
But his idea is that you kind of prime your brain
to make sure that that's the focus.
Oftentimes people have other worries in their life
that's not fundamentally deep problems.
They're like, I don't know,
just stupid drama in your life,
and even at work, all that kind of stuff.
He wants to kind of pick the most important problem
that you're thinking about, and go to bed on that.
- I think that's why, I mean,
the other thing that comes to my mind is also,
I feel the most fresh in the morning.
So during the morning,
I try to work on the most important things,
rather than just being pulled by urgent things
or checking email or so.
- What do you do with the,
'cause I've been doing the voice recorder thing too,
but I end up recording so many messages,
it's hard to organize.
- I have the same problem.
Now I have heard that Google Pixel
is really good in transcribing text,
and I might get a Google Pixel
just for the sake of transcribing text.
- Yeah, people listening to this,
if you have a good voice recorder suggestion
that transcribes, please let me know.
Some of it is, this has to do with OpenAI Codex too.
Some of it is simply the friction.
I need apps that remove that friction
between voice and the organization
of the resulting transcripts and all that kind of stuff.
But yes, you're right, absolutely.
During, for me, it's walking, sleep too,
but walking and running, especially running,
get a lot of thoughts during running,
and there's no good mechanism for recording thoughts.
- So one more thing that I do,
I have a separate phone, which has no apps.
Maybe it has like Audible or let's say Kindle.
No one has this phone number,
this kind of my meditation phone.
And I try to expand the amount of time
that that's the phone that I'm having.
It has also Google Maps if I need to go somewhere.
And I also use this phone to write down ideas.
- Ah, that's a really good idea.
That's a really good idea.
- Often actually what I end up doing
is even sending a message from that phone to the other phone.
So that's actually my way of recording messages
or I just put them into notes.
- I love it.
What advice would you give to a young person,
high school, college, about how to be successful?
You've done a lot of incredible things in the past decade.
So maybe you have some--
- Something, there might be something.
- There might be something.
- I mean, it might sound like simplistic or so,
but I would say literally just follow your passion,
double down on it.
And if you don't know what's your passion,
just figure out what could be a passion.
So the step might be an exploration.
When I was in elementary school was math and chemistry.
And I remember for some time I gave up on math
because my school teacher, she told me that I'm dumb.
And I guess maybe an advice would be just ignore people
if they tell you that you're dumb.
- You're dumb.
You mentioned something offline about chemistry
and explosives.
What was that about?
- So let's see.
(laughing)
So a story goes like that.
I got into chemistry, maybe I was like a second grade
of my elementary school, third grade.
I started going to chemistry classes.
I really love building stuff.
And I did all the experiments that they describe
in the book, like how to create oxygen with vinegar
and baking soda or so.
So I did all the experiments.
And at some point I was, so what's next?
What can I do?
And explosives, they also, it's like you have a clear
reward signal, if the thing worked or not.
So I remember at first I got interested
in producing hydrogen.
That was kind of funny experiment from school.
You can just burn it.
And then I moved to nitroglycerin.
So that's also relatively easy to synthesize.
I started producing essentially dynamite
and detonating it with a friend.
I remember there was a, there was at first like maybe
two attempts that I went with a friend to detonate
what we built and it didn't work out.
And like a third time he was like, ah, it won't work.
Like let's don't waste time.
And now we were, I was carrying this,
you know, that tube with dynamite, I don't know,
pound or so, dynamite in my backpack.
We're like riding on the bike to the edges of the city.
(laughing)
- Yeah.
- And--
- Attempt number three.
This was be attempt number three.
- Attempt number three.
And now we dig a hole to put it inside.
It actually had the, you know, electrical detonator.
We draw a cable behind the tree.
I even, I never, I haven't ever seen like a explosion before.
So I thought that there will be a lot of sound.
But you know, we were like laying down
and I'm holding the cable and the battery.
At some point, you know, we kind of like a three to one
and I just connected it and it felt like
at the ground shake.
It was like more like a sound.
And then the soil started kind of lifting up
and started falling on us.
- Yeah, wow.
- And then the other friend said,
let's make sure the next time we have helmets.
(laughing)
But also, you know, I'm happy that nothing happened to me.
It could have been the case that I lost the limbo or so.
- Yeah, but that's childhood of an engineering mind
with a strong reward signal of an explosion.
I love it.
And there's some aspect of chemists,
the chemists I know, like my dad,
with plasma chemistry, plasma physics,
he was very much into explosives too.
It's a worrying quality of people that work in chemistry
that they love.
I think it is that exactly,
is the strong signal that the thing worked.
- There is no doubt.
- There's no doubt.
There's some magic.
It's almost like a reminder that physics works,
that chemistry works.
It's cool.
It's almost like a little glimpse at nature
that you yourself engineer.
That's why I really like artificial intelligence,
especially robotics, is you create a little piece of nature.
- And in some sense, even for me with explosives,
the motivation was creation rather than destruction.
- Yes, exactly.
In terms of advice, I forgot to ask
about just machine learning and deep learning.
For people who are specifically interested
in machine learning,
how would you recommend they get into the field?
- So I would say re-implement everything.
And also there is plenty of courses.
- So like from scratch?
- So on different levels of abstraction in some sense,
but I would say re-implement something from scratch,
re-implement something from a paper,
re-implement something from podcasts
that you have heard about.
I would say that's a powerful way to understand things.
So it's often the case that you read the description
and you think you understand,
but you truly understand once you build it,
then you actually know what really mattered
in the description.
- Is there a particular topics that you find people
§17Transcript · 2:34:39–2:44:51
just fall in love with?
So I've seen,
I tend to really enjoy reinforcement learning
because it's much easier to get to a point
where you feel like you created something special,
like fun games kind of things.
- It's rewarding.
- It's rewarding, yeah.
As opposed to like re-implementing from scratch,
more like supervised learning kind of things.
- So if someone would optimize for things to be rewarding,
then it feels that the things that are somewhat generative,
they have such a property.
So you have, for instance, adversarial networks,
or you have just even generative language models.
And you can even see,
internally we have seen this thing with our releases.
So we released recently two models.
There is one model called DALI that generates images,
and there is other model called CLIP
that actually you provide various possibilities,
what could be the answer to what is on the picture,
and it can tell you which one is the most likely.
And in some sense, in case of the first one, DALI,
it is very easy for you to understand
that actually there is magic going on.
And in case of the second one,
even though it is insanely powerful,
and people from vision community,
as they started probing it inside,
they actually understood how far it goes.
It's difficult for person at first to see
how well it works.
And that's the same, as you said,
that in case of supervised learning models,
you might not kind of see,
or it's not that easy for you to understand the strength.
- Even though you don't believe in magic, to see the magic.
- To see the magic, yeah.
- The generative, that's really brilliant.
So anything that's generative,
'cause then you are at the core of the creation.
You get to experience creation without much effort,
unless you have to do it from scratch.
- And it feels that humans are wired.
There is some level of reward for creating stuff.
Of course, different people have a different weight
on this reward.
- Yeah, in the big objective function.
- In the big objective function, of a person.
- Of a person.
You wrote that beautiful
is what you intensely pay attention to.
Even a cockroach is beautiful if you look very closely.
Can you expand on this?
What is beauty?
- So what I wrote here actually corresponds
to my subjective experience that I had
through extended periods of meditation.
It's pretty crazy that at some point,
the meditation gets you to the place
that you have really increased focus, increased attention.
And then you look at the very simple objects
that were all the time around you.
You can look at the table or on the pen or at the nature.
And you notice more and more details
and it becomes very pleasant to look at it.
And once again, it kind of reminds me of my childhood.
Like just pure joy of being.
It's also, I have seen even the reverse effect
that by default, regardless of what we possess,
we very quickly get used to it.
And you can have a very beautiful house
and if you don't put sufficient effort,
you're just gonna get used to it
and it doesn't bring any more joy
regardless of what you have.
- Yeah.
Well, I actually,
I find that material possessions get in the way
of that experience of pure joy.
So I've always, I've been very fortunate
to just find joy in simple things.
Just like you're saying, just like, I don't know,
objects in my life, just stupid objects,
like this cup, like thing, you know, just objects.
Sounds, okay, I'm not being eloquent,
but literally objects in the world.
They're just full of joy 'cause it's like,
I can't believe, one, I can't believe
that I'm fortunate enough to be alive
to experience these objects.
And then two, I can't believe humans are clever enough
to build these objects.
The hierarchy of pleasure that that provides is infinite.
- I mean, even if you look at the cup of water,
so you see first like a level of like a reflection of light,
but then you think, no man, there's like trillions
upon trillions of particles bouncing against each other.
There is also the tension on the surface
that if the back could like stand on it
and move around, and you think it also has this
like a magical property that as you decrease temperature,
it actually expands in volume, which allows for the,
you know, legs to freeze on the surface
and at the bottom to have actually not freeze,
which allows for like a crazy.
- Yeah.
- You look in detail at some objects
and you think actually, you know, this table,
that was just the figment of someone's imagination
at some point.
And then there was like a thousands of people involved
to actually manufacture it and put it here.
And by default, no one cares.
- And then you can start thinking about evolution,
how it all started from single cell organisms
that led to this table.
- And these thoughts, they give me life appreciation.
- Yeah, exactly.
- And even lack of thoughts, just the pure raw signal
also gives the life appreciation.
- See, the thing is, and then that's coupled for me
with the sadness that the whole ride ends
and perhaps is deeply coupled in that,
the fact that this experience, this moment ends,
gives it an intensity
that I'm not sure I would otherwise have.
So in that same way, I try to meditate on my own death often.
Do you think about your mortality?
Are you afraid of death?
- So fear of death is like one of the most fundamental fears
that each of us has.
We might be not even aware of it.
It requires to look inside to even recognize
that it's out there.
There is still, let's say, this property of nature
that if things would last forever,
then they would be also boring to us.
The fact that the things change in some way
gives any meaning to them.
I also found out that it seems to be very healing
to people to have these short experiences,
like I guess, psychedelic experiences
in which they experience death of self,
in which they let go of this fear
and then maybe can even increase
the appreciation of the moment.
It seems that many people,
they can easily comprehend the fact
that the money's finite
while they don't see that time is finite.
I have this discussion with Ilya from time to time.
He's saying, "Man, life will pass very fast.
"At some point, I will be 40, 50, 60, 70, and then it's over."
This is true, which also makes me believe
that every single moment, it is so unique
that should be appreciated.
And this also makes me think
that I should be acting on my life
because otherwise it will pass.
I also like this framework of thinking from Jeff Bezos
on regret minimization,
that I would like, if I will be at that deathbed,
to look back on my life
and not regret that I haven't done something.
It's usually, you might regret that you haven't tried.
I'm fine with failing.
- But I haven't tried.
What's the Nietzsche eternal occurrence?
Try to live a life that,
if you had to live it infinitely many times,
that would be the, you'd be okay with that kind of life.
So try to live it optimally.
- I can say that it's almost like
unbillable to me where I am in my life.
I'm extremely grateful for actually people whom I met.
I would say, I think that I'm decently smart and so on,
but I think that actually to great extent
where I am has to do with the people who I met.
- Would you be okay if after this conversation you died?
- So if I'm dead, then it kind of,
I don't have a choice anymore.
So in some sense, there's like a plenty of things
that I would like to try out in my life.
I feel that I'm gradually going one by one
and I'm just doing them.
I think that the list will be always infinite.
- Yeah.
So might as well go today.
- Yeah, I mean, to be clear, I'm not looking forward to die.
I would say if there is no choice, I would accept it.
But like in some sense, if there would be a choice,
if there would be possibility to live,
I would fight for living.
- I find it's more honest and real to think about
dying today at the end of the day.
That seems to me, at least to my brain,
more honest slap in the face
as opposed to I still have 10 years today.
§18Transcript · 2:44:56–2:51:26
Then I'm much more about appreciating the cup and the table
and so on, and less about silly worldly accomplishments
and all those kinds of things.
- We have in the company a person who say at some point
found out that they have cancer.
And that also gives huge perspective with respect
to what matters now.
And often people in situations like that,
they conclude that actually what matters
is human connection.
- And love, and people conclude also if you have kids,
kids is family.
You, I think tweeted,
"We don't assign the minus infinity reward to our death.
Such a reward would prevent us from taking any risk.
We wouldn't be able to cross a road
in fear of being hit by a car."
So in the objective function, you mentioned fear of death
might be fundamental to the human condition.
- So as I said, let's assume that there are
like reward functions in our brain.
And the interesting thing is even realization
how different reward functions can play with your behavior.
As a matter of fact, I wouldn't say that you should assign
infinite negative reward to anything
because that messes up the math.
- The math doesn't work out.
- It doesn't work out.
And as you said, even, you know,
government or some insurance companies,
you said they assign $9 million to human life.
And I'm just saying it with respect to,
that might be a hard statement to ourselves,
but in some sense that there is a finite value
of our own life.
I'm trying to put it from perspective of being less,
of being more egoless and realizing fragility
of my own life.
And in some sense, the fear of death
might prevent you from acting
because anything can cause death.
- Yeah, and I'm sure actually,
if you were to put death in the objective function,
there's probably so many aspects to death and fear of death
and realization of death and mortality.
There's just whole components of finiteness
of not just your life, but every experience and so on
that you're gonna have to formalize mathematically.
- And also, you know, that might lead to
you spending a lot of compute cycles
on this like a,
deliberating this terrible future
instead of experiencing now.
And then in some sense,
it's also kind of unpleasant simulation
to run in your head.
- Yeah.
(laughing)
- Do you think there's an objective function
that describes the entirety of human life?
So, you know, usually the way you ask that
is what is the meaning of life?
Is there a universal objective functions
that captures the why of life?
- So, yeah, I mean,
I suspected that they will ask this question,
but it's also a question that I ask myself many, many times.
See, I can tell you a framework that I have these days
to think about this question.
So I think that fundamentally meaning of life
has to do with some of our reward functions
that we have in brain,
and they might have to do with,
let's say, for instance, curiosity or human connection,
which might mean understanding others.
It's also possible for a person
to slightly modify their reward function.
Usually they mostly stay fixed,
but it's possible to modify reward function.
And you can pretty much choose.
So in some sense, reward functions,
optimizing reward functions,
they will give you a life satisfaction.
- Is there some randomness in the function?
- I think when you are born, there is some randomness.
Like you can see that some people, for instance,
they care more about building stuff.
Some people care more about caring for others.
Some people, there are all sorts of default reward functions.
And then in some sense, you can ask yourself,
what is the satisfying way for you
to go after this reward function?
And you just go after this reward function.
And some people also ask,
are these reward functions real?
I almost think about it as, let's say,
if you would have to discover mathematics,
in mathematics, you are likely to run into various objects,
like complex numbers or differentiation,
some other objects.
And these are very natural objects that arise.
And similarly, the reward functions
that we are having in our brain,
they are somewhat very natural,
that there is a reward function for understanding,
like a comprehension, curiosity, and so on.
So in some sense, they are in the same way natural
as they're natural objects in mathematics.
- Interesting.
So you know, there's the old sort of debate,
is mathematics invented or discovered?
You're saying reward functions are discovered.
So nature provides-
- So nature provided some,
you can still, let's say, expand it throughout the life.
Some of the reward functions, they might be futile.
Like for instance, there might be a reward function,
maximize amount of wealth.
- Yeah.
- And this is more like a learned reward function.
But we know also that some reward functions,
if you optimize them, you won't be quite satisfied.
- Well, I don't know which part of your reward function
resulted in you coming today,
but I am deeply appreciative that you did
spend your valuable time with me.
Wojciech, it was really fun talking to you.
You're brilliant, you're a good human being,
and it's an honor to meet you and an honor to talk to you.
Thanks for talking today, brother.
- Thank you, Lex, a lot.
I appreciated your questions, curiosity.
I had a lot of fun being here.
- Thanks for listening to this conversation
with Wojciech Czeremba.
To support this podcast,
please check out our sponsors in the description.
And now, let me leave you with some words
from Arthur C. Clarke,
who is the author of "2001, A Space Odyssey."
It may be that our role on this planet
is not to worship God, but to create him.
Thank you for listening, and I hope to see you next time.
(upbeat music)
(upbeat music)