back to indexWojciech Zaremba: OpenAI Codex, GPT-3, Robotics, and the Future of AI | Lex Fridman Podcast #215
link |
The following is a conversation with Wojciech Zaremba,
link |
cofounder of OpenAI,
link |
which is one of the top organizations in the world
link |
doing artificial intelligence research and development.
link |
Wojciech is the head of language and cogeneration teams,
link |
building and doing research on GitHub copilot,
link |
OpenAI codecs, and GPT3, and who knows,
link |
four, five, six, n, and n plus one.
link |
And he also previously led OpenAI's robotic efforts.
link |
These are incredibly exciting projects to me
link |
that deeply challenge and expand our understanding
link |
of the structure and nature of intelligence.
link |
The 21st century, I think,
link |
may very well be remembered
link |
for a handful of revolutionary AI systems
link |
and their implementations.
link |
GPT codecs and applications of language models
link |
and transformers in general
link |
to the language and visual domains
link |
may very well be at the core of these AI systems.
link |
To support this podcast, please check out our sponsors.
link |
They're listed in the description.
link |
This is the Lex Friedman podcast,
link |
and here is my conversation with Wojciech Zaremba.
link |
You mentioned that Sam Altman asked
link |
about the Fermi Paradox,
link |
and other people at OpenAI
link |
had really sophisticated interesting answers.
link |
So that's when you knew
link |
this is the right team to be working with.
link |
So let me ask you about the Fermi Paradox, about aliens.
link |
Why have we not found overwhelming evidence
link |
for aliens visiting Earth?
link |
I don't have a conviction in the answer,
link |
but rather kind of probabilistic perspective
link |
on what might be, let's say, possible answers.
link |
It's also interesting that the question itself,
link |
even can't touch on your typical question
link |
of what's the meaning of life,
link |
because if you assume that we don't see aliens
link |
because they destroy themselves,
link |
that kind of upgrades the focus
link |
on making sure that we won't destroy ourselves.
link |
But at the moment, the place where I am actually
link |
with my belief, and these things also change over the time,
link |
is I think that we might be alone in the universe,
link |
which actually makes life more, or let's say consciousness,
link |
life more kind of valuable.
link |
And that means that we should more appreciate it.
link |
Have we always been alone?
link |
So what's your intuition about our galaxy, our universe?
link |
Is it just sprinkled with graveyards
link |
of intelligent civilizations,
link |
or is life, intelligent life truly unique?
link |
At the moment, my belief that it is unique.
link |
But I would say I could also,
link |
there was like some footage released with UFO objects,
link |
which makes me actually doubt my own belief.
link |
Yeah, I can tell you one crazy answer that I have heard.
link |
So apparently when you look actually
link |
at the limits of computation,
link |
you can compute more if the temperature
link |
of the universe would drop down.
link |
So one of the things that aliens might want to do
link |
if they are truly optimizing to maximize amount of compute,
link |
which maybe can lead to more, let's say simulations or so,
link |
it's instead of wasting current entropy of the universe,
link |
because we by living,
link |
we are actually somewhat wasting entropy,
link |
then you can wait for the universe to cool down,
link |
such that you have more computation.
link |
So that's kind of a funny answer.
link |
I'm not sure if I believe in it,
link |
but that would be one of the reasons
link |
why you don't see aliens.
link |
It's also possible, see some people say that,
link |
maybe there is not that much point
link |
in actually going to other galaxies,
link |
if you can go inwards.
link |
So there is no limits of what could be an experience
link |
if we could connect machines to our brains,
link |
while there are still some limits
link |
if we want to explore the universe.
link |
Yeah, there could be a lot of ways to go inwards too.
link |
Once you figure out some aspect of physics
link |
we haven't figured out yet,
link |
maybe you can travel to different dimensions.
link |
I mean, travel in three dimensional space
link |
may not be the most fun kind of travel.
link |
There may be like just a huge amount
link |
of different ways to travel.
link |
It doesn't require a spaceship going slowly
link |
in 3D space to space time.
link |
It also fills in one of the problems
link |
is that speed of light is low and the universe is vast.
link |
And it seems that actually most likely
link |
if we want to travel very far,
link |
then we would instead of actually sending spaceships
link |
with humans that wait a lot,
link |
we would send something similar
link |
to what Uri Miller is working on.
link |
These are like a huge sail which is at first powered.
link |
There is a shot of laser from an air
link |
and it can propel it to quarter of speed of light.
link |
And the sail itself contains a few grams of equipment.
link |
And that might be the way to actually transport matter
link |
through the universe.
link |
But then when you think what would it mean for humans,
link |
it means that we would need to actually
link |
put their 3D printer and 3D print a human on another planet.
link |
I don't know, play them YouTube or let's say,
link |
or like a 3D print like huge human right away
link |
or maybe a womb or so, yeah.
link |
With our current techniques of archeology,
link |
if a civilization was born and died long enough to go
link |
on Earth, we wouldn't be able to tell.
link |
And so that makes me really sad.
link |
And so I think about Earth in that same way.
link |
How can we leave some remnants if we do destroy ourselves?
link |
How can we leave remnants for aliens
link |
in the future to discover?
link |
Like here's some nice stuff we've done.
link |
Like Wikipedia and YouTube.
link |
Do we have it like in a satellite orbiting Earth
link |
with a hard drive?
link |
Like how do we say,
link |
how do we back up human civilization for the good parts
link |
or all of it is good parts
link |
so that it can be preserved longer than our bodies can?
link |
That's kind of a, it's a difficult question.
link |
It also requires the difficult acceptance
link |
of the fact that we may die.
link |
And if we die, we may die suddenly as a civilization.
link |
So let's see, I think it kind of depends on the cataclysm.
link |
We have observed in other parts of the universe
link |
that birds of gamma rays,
link |
these are high energy rays of light
link |
that actually can apparently kill entire galaxy.
link |
So there might be actually nothing event to,
link |
nothing to protect us from it.
link |
I'm also, and I'm looking actually at the past civilizations.
link |
So it's like Aztecs or so,
link |
they disappear from the surface of the Earth
link |
and one can ask, why is it the case?
link |
And the way I'm thinking about it is,
link |
you know, that definitely they had some problem
link |
that they couldn't solve.
link |
And maybe there was a flood
link |
and all of a sudden they couldn't drink,
link |
there was no potable water and they all died.
link |
And I think that so far,
link |
the best solution to such a problems is, I guess, technology.
link |
So I mean, if they would know that you can just boil water
link |
and then drink it after,
link |
then that would save their civilization.
link |
And even now when we look actually at the current pandemic,
link |
it seems that once again, actually science comes to rescue
link |
and somehow science increases size of the action space.
link |
And I think that's a good thing.
link |
Yeah, but nature has a vastly larger action space.
link |
Yeah, but still it might be a good thing for us
link |
to keep on increasing action space.
link |
Well, okay, looking at past civilizations, yes.
link |
But looking at the destruction of human civilization,
link |
perhaps expanding the action space
link |
will add actions that are easily acted upon,
link |
easily executed and as a result destroy us.
link |
I was pondering why actually even
link |
we have negative impact on the globe.
link |
Because if you ask every single individual,
link |
they would like to have clean air.
link |
They would like healthy planet,
link |
but somehow it's not the case that as a collective,
link |
we are not going in this direction.
link |
I think that there exists very powerful system
link |
to describe what we value, that's capitalism.
link |
It assigns actually monetary values to various activities.
link |
At the moment, the problem in the current system
link |
is that there are some things which we value.
link |
There is no cost assigned to it.
link |
So even though we value clean air,
link |
or maybe we also value lack of distraction
link |
on the internet or so,
link |
at the moment these quantities,
link |
companies, corporations can pollute them for free.
link |
So in some sense, I wished,
link |
or like that's I guess purpose of politics
link |
to align the incentive systems.
link |
And we are kind of maybe even moving in this direction.
link |
The first issue is even to be able to measure
link |
the things that we value,
link |
then we can actually assign the monetary value to them.
link |
Yeah, and that's so it's getting the data
link |
and also probably through technology,
link |
enabling people to vote and to move money around
link |
in a way that is aligned with their values.
link |
And that's very much a technology question.
link |
So like having one president and Congress
link |
and voting that happens every four years
link |
or something like that,
link |
that's a very outdated idea.
link |
There could be some technological improvements
link |
to that kind of idea.
link |
So I'm thinking from time to time about these topics,
link |
but it also feels to me that it's a little bit like a,
link |
it's hard for me to actually make correct predictions
link |
what is the appropriate thing to do.
link |
I extremely trust Sam Altman,
link |
our CEO on these topics here.
link |
Okay, I'm more on the side of being,
link |
I guess, naĂŻve hippie that, yeah.
link |
That's your life philosophy.
link |
Well, I think self doubt and I think hippie implies optimism.
link |
Those two things are pretty good way to operate.
link |
I mean, still it is hard for me to actually
link |
understand how the politics works
link |
or exactly how the things would play out.
link |
And Sam is really excellent with it.
link |
What do you think is rarest in the universe?
link |
You said we might be alone.
link |
What's hardest to build is another engineering way
link |
to ask that, life, intelligence or consciousness.
link |
So like you said that we might be alone,
link |
which is the thing that's hardest to get to?
link |
Is it just the origin of life?
link |
Is it the origin of intelligence?
link |
Is it the origin of consciousness?
link |
So let me at first explain in my kind of mental model
link |
what I think is needed for life to appear.
link |
So I imagine that at some point there was this primordial
link |
soup of amino acids and maybe some proteins in the ocean.
link |
And some proteins were turning into some other proteins
link |
And you can almost think about this cycle
link |
of what turns into what,
link |
as there is a graph essentially describing
link |
which substance turns into some other substance.
link |
And essentially life means that all the sudden
link |
in the graph has been created that cycle,
link |
such that the same thing keeps on happening
link |
over and over again.
link |
That's what is needed for life to happen.
link |
And in some sense you can think almost
link |
that you have this gigantic graph
link |
and it needs like a sufficient number of edges
link |
for the cycle to appear.
link |
Then from perspective of intelligence and consciousness,
link |
my current intuition is that they might be quite intertwined.
link |
First of all, it might not be that it's like a binary thing
link |
that you have intelligence or consciousness.
link |
It seems to be a more continuous component.
link |
Let's see, if we look for instance
link |
on the event networks recognizing images,
link |
people are able to show that the activations
link |
of these networks correlate very strongly
link |
with activations in visual cortex of some monkeys.
link |
The same seems to be true about language models.
link |
Also, if you for instance look,
link |
if you train agent in 3D world,
link |
at first it barely recognizes what is going on.
link |
Over the time it recognizes foreground from background.
link |
Over the time it knows where there is a foot
link |
and it just follows it.
link |
Over the time it actually starts having a 3D perception.
link |
So it is possible for instance to look
link |
inside of the head of an agent
link |
and ask what would it see if it looks to the right.
link |
And the crazy thing is initially
link |
when the agents are barely trained
link |
that these predictions are pretty bad,
link |
over the time they become better and better,
link |
you can still see that if you ask what happens
link |
when the head is turned by 360 degrees,
link |
for some time they think that the different thing appears
link |
and then at some stage they understand actually
link |
that the same thing supposed to appear.
link |
So they get like an understanding of 3D structure.
link |
It's also very likely that they have inside some level
link |
of like a symbolic reasoning,
link |
like they're particularly symbols for other agents.
link |
So when you look at DOTA agents,
link |
they collaborate together
link |
and they have some anticipation of
link |
if they would win battle,
link |
they have some expectations with respect to other agents.
link |
I might be too much anthropomorphizing
link |
how the things look for me.
link |
But then the fact that they have a symbol for other agents
link |
makes me believe that at some stage,
link |
as they are optimizing for skills,
link |
they would have also symbol to describe themselves.
link |
This is like a very useful symbol to have.
link |
And this particularity I would call it
link |
like a self consciousness or self awareness.
link |
And still it might be different from the consciousness.
link |
So I guess the way how I'm understanding
link |
the word consciousness,
link |
I'd say the experience of drinking a coffee
link |
or let's say experience of being a bat.
link |
That's the meaning of the word consciousness.
link |
It doesn't mean to be awake.
link |
it might be also somewhat related to memory
link |
and recurrent connections.
link |
So it's kind of like,
link |
if you look at anesthetic drugs,
link |
they might be like essentially,
link |
they disturb brain waves
link |
such that maybe memory is not formed.
link |
And so there's a lessening of consciousness
link |
And so that's the one way to intuit what is consciousness.
link |
There's also kind of another element here.
link |
It could be that it's this kind of self awareness module
link |
that you described.
link |
Plus the actual subjective experience
link |
is a storytelling module
link |
that tells us a story about what we're experiencing.
link |
So let's say, I mean, in meditation,
link |
they teach people not to speak story inside of the head.
link |
And there is also some fraction of population
link |
who doesn't have actually narrator.
link |
I know people who don't have a narrator
link |
and they have to use external people
link |
in order to kind of solve tasks
link |
that require internal narrator.
link |
So it seems that it's possible to have the experience
link |
What are we talking about
link |
when we talk about the internal narrator?
link |
So is that the voice when you like read the book?
link |
Yeah, I thought that that's what you are referring to.
link |
Well, I was referring more on it,
link |
like not an actual voice.
link |
I meant like there's some kind of like subjective experience
link |
feels like it's fundamentally about storytelling
link |
It feels like the feeling is a story
link |
that is much simpler abstraction
link |
than the raw sensory information.
link |
So it feels like it's a very high level abstraction
link |
that is useful for me to feel like entity in this world.
link |
Most useful aspect of it is that because I'm conscious,
link |
I think there's an intricate connection
link |
to me not wanting to die.
link |
So like it's a useful hack to really prioritize not dying.
link |
Like those seem to be somehow connected.
link |
So I'm telling the story of like it's richly feels
link |
like something to be me and the fact that me exists
link |
in this world, I want to preserve me.
link |
And so that makes it a useful agent hack.
link |
So I will just refer maybe to the first part,
link |
as you said about that kind of story of describing who you are.
link |
I was thinking about it even.
link |
So, you know, obviously I like thinking about it
link |
thinking about consciousness.
link |
I like thinking about the AI as well.
link |
And I'm trying to see analogies of these things in AI,
link |
what would it correspond to?
link |
So, you know, OpenAI trained a model called GPT
link |
which can generate pretty amusing text on arbitrary topic.
link |
And one way to control GPT
link |
is by putting into prefix at the beginning of the text
link |
some information what would be the story about.
link |
You can have even chat with GPT by saying
link |
that the chat is with Lex or Elon Musk or so.
link |
And GPT would just pretend to be you or Elon Musk or so.
link |
And it almost feels that this story
link |
that we give ourselves to describe our life,
link |
it's almost like things that you put into context of GPT.
link |
Yeah, and it generates the,
link |
but the context we provide to GPT is multimodal.
link |
It's more, so GPT itself is multimodal.
link |
GPT itself hasn't learned actually
link |
from experience of single human,
link |
but from the experience of humanity.
link |
You can turn it into anything.
link |
And in some sense, by providing context,
link |
it behaves as the thing that you wanted it to be.
link |
It's interesting that people have a stories of who they are.
link |
And as I said, these stories,
link |
they help them to operate in the world.
link |
But it's also interesting,
link |
I guess various people find it out through meditation or so
link |
that there might be some patterns
link |
that you have learned when you were a kid
link |
that actually are not serving you anymore.
link |
And you also might be thinking that that's who you are
link |
and that's actually just a story.
link |
Yeah, so it's a useful hack,
link |
but sometimes it gets us into trouble.
link |
It's a local optima.
link |
It's a local optima.
link |
You wrote that Stephen Hawking, he tweeted,
link |
Stephen Hawking asked what breathes fire into equations,
link |
which meant what makes given mathematical equations
link |
realize the physics of a universe.
link |
Similarly, I wonder what breathes fire into computation.
link |
What makes given computation conscious?
link |
Okay, so how do we engineer consciousness?
link |
How do you breathe fire and magic into the machine?
link |
So it seems clear to me
link |
that not every computation is conscious.
link |
I mean, you can, let's say,
link |
just keep on multiplying one matrix over and over again.
link |
And it might be gigantic matrix.
link |
You can put a lot of computation.
link |
I don't think it would be conscious.
link |
So in some sense, the question is,
link |
what are the computations which could be conscious?
link |
I mean, so one assumption is
link |
that it has to do purely with computation
link |
that you can abstract away matter.
link |
Other possibilities, it's very important
link |
was the realization of computation
link |
that it has to do with some force fields or so
link |
and they bring consciousness.
link |
At the moment, my intuition is that it can be
link |
fully abstracted away.
link |
So in case of computation, you can ask yourself,
link |
what are the mathematical objects
link |
or so that could bring such a properties?
link |
So for instance, if we think about the models,
link |
AI models, what they truly try to do,
link |
or like a models like GPT is,
link |
they try to predict the next word or so.
link |
And this turns out to be equivalent to compressing text.
link |
And because in some sense, compression means
link |
that you learn the model of reality
link |
and you have just to remember what are your mistakes.
link |
The better you are in predicting the,
link |
and in some sense, when we look at our experience,
link |
also when you look for instance at the car driving,
link |
you know in which direction it will go.
link |
You are good like in prediction.
link |
And it might be the case that the consciousness
link |
is intertwined with compression.
link |
It might be also the case that self consciousness
link |
has to do with compressor trying to compress itself.
link |
So I was just wondering what are the objects
link |
in mathematics or computer science,
link |
which are mysterious that could have to do
link |
with consciousness.
link |
And then I thought, you know, you see in mathematics
link |
there is something called Gadot theorem,
link |
which means again, if you have sufficiently
link |
complicated mathematical system,
link |
it is possible to point the mathematical system
link |
In computer science, there is something called
link |
helping problem, it's somewhat similar construction.
link |
So I thought that, you know, if we believe
link |
that under assumption that consciousness
link |
has to do with compression,
link |
then you could imagine that as you are
link |
keeping on compressing things,
link |
then at some point it actually makes sense
link |
for the compressor to compress itself.
link |
Metacompression, consciousness is metacompression.
link |
That's an idea, and in some sense, you know, the crazy.
link |
So, but do you think if we think of a touring machine,
link |
a universal touring machine,
link |
can that achieve consciousness?
link |
So is there something beyond our traditional definition
link |
of computation that's required?
link |
So it's a specific computation.
link |
And I said this computation has to do with compression.
link |
And the compression itself, maybe other way of putting it
link |
is like you are internally creating the model of reality.
link |
In order, it's like you try insight to simplify reality
link |
in order to predict what's gonna happen.
link |
And that also feels somewhat similar to how I think
link |
actually about my own conscious experience.
link |
So clearly I don't have access to reality.
link |
The only access to reality is through, you know,
link |
cable going to my brain.
link |
And my brain is creating a simulation of reality.
link |
And I have access to the simulation of reality.
link |
Are you by any chance aware of the Hutter Prize,
link |
He made this prize for compression of Wikipedia pages.
link |
And there's a few qualities to it.
link |
One, I think has to be perfect compression,
link |
which makes, I think that little quirk
link |
makes it much less applicable to the general task
link |
of intelligence, because it feels like intelligence
link |
is always going to be messy.
link |
Like perfect compression feels like it's not the right goal,
link |
but it's nevertheless a very interesting goal.
link |
So for him, intelligence equals compression.
link |
And so the smaller you make the file,
link |
given a large Wikipedia page,
link |
the more intelligent the system has to be.
link |
Yeah, that makes sense.
link |
So you can make perfect compression if you store errors.
link |
And I think that actually what he meant
link |
is you have algorithm plus errors.
link |
By the way, Hutter is a,
link |
he was a PhD advisor of Shanleck,
link |
who is a deep mind cofounder.
link |
Yeah, yeah, so there's an interesting,
link |
and now he's a deep mind.
link |
There's an interesting network of people.
link |
And he's one of the people that I think
link |
seriously took on the task of
link |
what would an AGI system look like?
link |
I think for the longest time,
link |
the question of AGI was not taken seriously,
link |
or rather rigorously.
link |
And he did just that.
link |
Like mathematically speaking,
link |
what would the model look like?
link |
If you remove the constraints of it having to be,
link |
having to have a reasonable amount of memory,
link |
reasonable amount of running time complexity,
link |
computation time, what would it look like?
link |
And essentially it's a half math,
link |
half philosophical discussion of how would it,
link |
like a reinforcement learning type of framework
link |
look like for an AGI?
link |
Yeah, so he developed the framework
link |
even to describe what's optimal
link |
with respect to reinforcement learning.
link |
Like there is a theoretical framework,
link |
which is as you said, under assumption,
link |
there is infinite amount of memory and compute.
link |
There was actually one person
link |
before his name is Solomonov,
link |
who extended Solomonov work to reinforcement learning,
link |
but there exists a theoretical algorithm,
link |
which is optimal algorithm to build intelligence.
link |
And I can actually explain you the algorithm.
link |
Yes, let's go, let's go, let's go.
link |
So the task itself, you can...
link |
Can I just pause how absurd it is
link |
for a brain in a skull
link |
trying to explain the algorithm for intelligence?
link |
It is pretty crazy.
link |
It is pretty crazy that the brain itself
link |
is actually so small and it can ponder.
link |
How to design algorithms
link |
that optimally solve the problem of intelligence?
link |
Okay, all right, so what's the algorithm?
link |
So first of all, the task itself is describe us,
link |
you have infinite sequence of zeros and ones, okay?
link |
You read n bits and you are about to predict
link |
So that's the task.
link |
And you could imagine that every task
link |
could be casted as such a task.
link |
So if, for instance, you have images and labels,
link |
you can just turn every image
link |
into a sequence of zeros and ones,
link |
then label, you concatenate labels,
link |
and that's actually...
link |
You could start by having training data first,
link |
and then afterwards you have test data.
link |
So theoretically, any problem could be casted
link |
as a problem of predicting zeros and ones
link |
on this infinite tape.
link |
So let's say you read already n bits
link |
and you want to predict n plus one bit.
link |
And I will ask you to write every possible program
link |
that generates these n bits, okay?
link |
So, and you can have, you choose programming language.
link |
It can be Python or C++.
link |
And the difference between programming languages might be,
link |
there is a difference by constant, asymptotically,
link |
your predictions will be equivalent.
link |
So you read n bits, you enumerate all the programs
link |
that produce these n bits in their output.
link |
And then in order to predict n plus one bit,
link |
you actually weight the programs according to their length.
link |
And there is some specific formula how you weight them.
link |
And then the n plus one bit prediction
link |
is the prediction from each of these program
link |
according to their weight.
link |
Like statistically, you pick.
link |
Statistically, yeah.
link |
So the smaller the program,
link |
the more likely you are to pick its output.
link |
So that algorithm is grounded in the hope
link |
or the intuition that the simple answer is the right one.
link |
It's a formalization of it.
link |
It also means like if you would ask the question
link |
after how many years would, you know, stand explode,
link |
you can say, hmm, it's more likely the answer
link |
is two to some power because they're shorter program.
link |
Well, I don't have a good intuition
link |
about how different the space of short programs
link |
are from the space of large programs.
link |
Like what is the universe for short programs?
link |
So, as I said, the things have to agree with n bits.
link |
So even if you have, you need to start,
link |
okay, if you have very short program
link |
and they're like a still some has,
link |
if it's not perfect with prediction of n bits,
link |
you have to start errors.
link |
What are the errors?
link |
And that gives you the full program
link |
that agrees on n bits.
link |
Oh, so you don't agree perfectly with the n bits
link |
and you store errors.
link |
That's like a longer, a longer programs,
link |
slightly longer program
link |
because it can't take these extra bits of errors.
link |
That's fascinating.
link |
What's your intuition about the programs
link |
that are able to do cool stuff
link |
like intelligence and consciousness?
link |
Are they perfectly, like is it,
link |
is there if then statements in them?
link |
So like, is there a lot of exceptions that they're storing?
link |
So you could imagine if there would be tremendous
link |
amount of each statements,
link |
then they wouldn't be that short.
link |
In case of neural networks,
link |
you could imagine that what happens is
link |
when you start with an uninitialized neural network,
link |
it stores internally many possibilities
link |
how the problem can be solved.
link |
And HDD is kind of magnifying some paths
link |
which are slightly similar to the correct answer.
link |
So it's kind of magnifying correct programs.
link |
And in some sense, HDD is a search algorithm
link |
in the program space.
link |
And the program space is represented
link |
by kind of the wiring inside of the neural network.
link |
And there's like an insane number of ways
link |
how the features can be computed.
link |
Let me ask you the high level basic question.
link |
That's not so basic.
link |
What is deep learning?
link |
Is there a way you'd like to think of it
link |
that is different than like a generic textbook definition?
link |
The thing that I hinted just a second ago
link |
is maybe the closest to how I'm thinking these days
link |
about deep learning.
link |
So now the statement is
link |
neural networks can represent some programs.
link |
Seems that various modules
link |
that we are actually adding up to,
link |
like we want networks to be deep
link |
because we want multiple steps of the computation.
link |
And deep learning provides the way
link |
to represent space of programs which is searchable.
link |
And it's searchable with stochastic gradient descent.
link |
So we have an algorithm to search over
link |
humongous number of programs.
link |
And gradient descent kind of bubbles up
link |
the things that tend to give correct answers.
link |
So a neural network with fixed weights,
link |
Do you think of that as a single program?
link |
So there is a work by Christopher Oleg
link |
where he, so he works on interpretability
link |
of neural networks.
link |
And he was able to identify inside of the neural network
link |
for instance, a detector of a wheel for a car
link |
or the detector of a mask for a car.
link |
And then he was able to separate them out
link |
and assemble them together
link |
using a simple program for the detector,
link |
for a car detector.
link |
That's like, if you think of traditionally defined programs,
link |
that's like a function within a program
link |
that this particular neural network was able to find.
link |
And you can tear that out,
link |
just like you can copy and paste from Stack Overflow.
link |
So any program is a composition of smaller programs.
link |
Yeah, I mean, the nice thing about the neural networks
link |
is that it allows the things to be more fuzzy
link |
than in case of programs.
link |
In case of programs, you have this like a branch
link |
in this way or that way.
link |
And the neural networks, they have an easier way
link |
to be somewhere in between or to share things.
link |
What to use the most beautiful, surprising idea
link |
in deep learning and the utilization
link |
of these neural networks,
link |
which by the way, for people who are not familiar,
link |
neural networks is a bunch of, what would you say?
link |
It's inspired by the human brain.
link |
There's neurons, there's connection between those neurons.
link |
There's inputs and there's outputs
link |
and there's millions or billions of those neurons.
link |
And the learning happens by adjusting the weights
link |
on the edges that connect these neurons.
link |
Thank you for giving the definition.
link |
I supposed to do it, but I guess you have enough empathy
link |
to listen there to actually know that that might be useful.
link |
No, that's like, so I'm asking Plato
link |
of like, what is the meaning of life?
link |
He's not gonna answer.
link |
You're being philosophical and deep and quite profound
link |
talking about the space of programs,
link |
which is very interesting,
link |
but also for people who are just not familiar
link |
with the hell we're talking about
link |
when we talk about deep learning.
link |
Anyway, sorry, what is the most beautiful
link |
or surprising idea to you in all the time
link |
you've worked at deep learning?
link |
And you worked on a lot of fascinating projects,
link |
the applications of neural networks.
link |
It doesn't have to be big and profound.
link |
It can be a cool trick.
link |
Yeah, I mean, I'm thinking about the trick,
link |
but like it's still amusing to me that it works at all.
link |
That let's say that the extremely simple
link |
algorithm stochastic gradient descent,
link |
which is something that I would be able to derive
link |
on the piece of paper to high school student
link |
when put at the scale of thousands of machines
link |
actually can create the behaviors,
link |
which we call kind of human like behaviors.
link |
So in general, any applications stochastic gradient descent
link |
to neural networks is amazing to you.
link |
So, or is there a particular application
link |
in natural language, reinforcement learning?
link |
And also, would you attribute that success to,
link |
What profound insight can we take from the fact
link |
that the thing works for gigantic sets of variables?
link |
I mean, the interesting thing is these algorithms,
link |
they were invented decades ago
link |
and people actually gave up on the idea.
link |
And back then they thought that we need
link |
profoundly different algorithms
link |
and they spent a lot of cycles
link |
on very different algorithms.
link |
And I believe that we have seen that various innovations
link |
that say like transformer or dropout
link |
or so they can vastly help.
link |
But it's also remarkable to me
link |
that this algorithm from sixties or so,
link |
or I mean, you can even say that the gradient descent
link |
was invented by Leipniz in I guess 18th century or so.
link |
That actually is the core of learning.
link |
In the past people are, it's almost like a,
link |
out of the maybe an ego,
link |
people are saying that it cannot be the case
link |
that such a simple algorithm is the,
link |
could solve complicated problems.
link |
So they were in search for the other algorithms.
link |
And as I'm saying, like I believe that actually
link |
we are in the game where there is,
link |
there are actually frankly three levers.
link |
There is compute, there are algorithms and there is data.
link |
And if we want to build intelligent systems,
link |
we have to pull all three levers
link |
and they are actually multiplicative.
link |
It's also interesting.
link |
So you ask, is it only compute?
link |
People internally, they did the studies
link |
to determine how much gains they were coming
link |
from different levers.
link |
And so far we have seen that more gains came
link |
from compute than algorithms.
link |
But also we are in the world that in case of compute,
link |
there is a kind of exponential increase in funding.
link |
And at some point it's impossible to invest more.
link |
It's impossible to invest 10 trillion dollars.
link |
We are speaking about the, let's say all taxes in US.
link |
But you're talking about money,
link |
there could be innovation in the compute.
link |
That's true as well.
link |
So I mean, there are like a few pieces.
link |
So one piece is human brain is an incredible supercomputer.
link |
And they're like a,
link |
it has a hundred trillion parameters.
link |
Or like if you try to count the various quantities
link |
in the brain, they're like a neuron, synapses.
link |
They're small number of neurons.
link |
There is a lot of synapses.
link |
It's unclear even how to map synapses
link |
to parameters of neural networks.
link |
But it's clear that there are many more.
link |
So it might be the case that our networks
link |
are still somewhat small.
link |
It also might be the case that they are more efficient
link |
in brain or less efficient by some huge factor.
link |
I also believe that there will be,
link |
like at the moment we are at the stage
link |
that these neural networks,
link |
they require 1000X or like a huge factor
link |
of more data than humans do.
link |
And it will be a matter of,
link |
there will be algorithms that vastly
link |
decrease sample complexity, I believe so.
link |
But the place where we are heading today is
link |
there are domains which contains million X more data.
link |
And even though computers might be 1000 times slower
link |
than humans in learning, that's not the problem.
link |
Like for instance, I believe that it should be possible
link |
to create super human therapists by a,
link |
and they're like even simple steps of doing it.
link |
And the core reason is there is just,
link |
machine will be able to read way more transcripts
link |
of therapists and then it should be able to speak
link |
simultaneously with many more people.
link |
And it should be possible to optimize it all in parallel.
link |
Well, there's now you're touching on something
link |
I deeply care about and think is way harder than we imagine.
link |
What's the goal of a therapist?
link |
What's the goal of therapist?
link |
So, okay, so one goal, now this is terrifying to me,
link |
but there's a lot of people that contemplate suicide,
link |
suffer from depression,
link |
and they could significantly be helped with therapy.
link |
And the idea that an AI algorithm might be in charge of that,
link |
it's like a life and death task.
link |
It's the stakes are high.
link |
So one goal for a therapist, whether human or AI,
link |
is to prevent suicide ideation, to prevent suicide.
link |
How do you achieve that?
link |
So, to be clear, I don't think that the current models
link |
are good enough for such a task
link |
because it requires insane amount of understanding empathy
link |
and the models are far from this place, but it's.
link |
But do you think that understanding empathy,
link |
that signal is in the data?
link |
I think there is some signal in the data, yes.
link |
I mean, there are plenty of transcripts of conversations
link |
and it is possible from it to understand
link |
personalities, it is possible from it to understand
link |
if conversation is friendly, amicable, antagonistic.
link |
It is, I believe that given the fact that the models
link |
that we train now, they can have,
link |
they are chameleons that they can have any personality.
link |
They might turn out to be better in understanding
link |
personality of other people than anyone else.
link |
And they should. Be empathetic.
link |
Yeah, interesting.
link |
But I wonder if there's some level
link |
of multiple modalities required
link |
to be able to be empathetic of the human experience,
link |
whether language is not enough, to understand death,
link |
to understand fear, to understand childhood trauma,
link |
to understand wit and humor required
link |
when you're dancing with a person
link |
who might be depressed or suffering,
link |
both humor and hope and love and all those kinds of things.
link |
So there's another underlying question
link |
which is self supervised versus supervised.
link |
So can you get that from the data
link |
by just reading a huge number of transcripts?
link |
I actually, so I think that reading
link |
a huge number of transcripts is a step one.
link |
It's like the same way as you cannot learn to dance
link |
if just from YouTube.
link |
By watching it, you have to actually try it out yourself.
link |
So I think that here that's a similar situation.
link |
I also wouldn't deploy the system
link |
in the high stakes situations right away,
link |
but kind of see gradually where it goes.
link |
And obviously initially it would have to go hand in hand
link |
But at the moment we are in the situation
link |
that actually there is many more people
link |
who actually would like to have a therapy
link |
or speak with someone than there are therapies out there.
link |
Like I was so fundamentally I was thinking
link |
what are the things that can vastly
link |
increase people well being?
link |
Therapy is one of them.
link |
I think meditation is other one.
link |
I guess maybe human connection is a third one.
link |
And I guess pharmacologically it's also possible.
link |
Maybe direct brain stimulation or something like that.
link |
But these are pretty much options out there.
link |
Then let's say the way I'm thinking about the AGI endeavor
link |
is by default that's an endeavor to increase amount of wealth.
link |
And I believe that we can vastly increase
link |
amount of wealth for everyone.
link |
And simultaneously, so I mean,
link |
these are like two endeavors that make sense to me.
link |
One is like essentially increase amount of wealth.
link |
And second one is increase overall human well being.
link |
And those are coupled together.
link |
And they can, I would say these are different topics.
link |
One can help another.
link |
And you know, therapist is a funny word
link |
because I see friendship and love as therapy.
link |
I mean, so therapist broadly defined
link |
as just friendship as a friend.
link |
So like therapist has a very kind of clinical sense to it.
link |
But what is human connection?
link |
You're like, not to get all Camus and Dostoevsky on you,
link |
but you know, life is suffering and we draw,
link |
we seek connection with other humans
link |
as we desperately try to make sense of this world
link |
in a deep overwhelming loneliness that we feel inside.
link |
So I think connection has to do with understanding.
link |
And I think that almost like a lack of understanding
link |
causes suffering if you speak with someone
link |
and do you feel ignored that actually causes pain
link |
if you are feeling deeply understood that actually
link |
they might not even tell you what to do in life,
link |
but like a pure understanding.
link |
Or just being heard.
link |
Understanding is a kind of, that's a lot, you know,
link |
just being heard, feel like you're being heard.
link |
Like somehow that's a alleviation temporarily
link |
of the loneliness that if somebody knows you're here
link |
with their body language, with the way they are,
link |
with the way they look at you, with the way they talk,
link |
you feel less alone for a brief moment.
link |
Yeah, very much agree.
link |
So I thought in the past about somewhat similar question
link |
to yours, which is what is love?
link |
Rather what is connection?
link |
And obviously I think about these things
link |
from AI perspective, what would it mean?
link |
So I said that the, you know,
link |
intelligence has to do with some compression,
link |
which is more or less like you can say
link |
almost understanding of what is going around.
link |
It seems to me that other aspect
link |
is there seem to be reward functions.
link |
And you can have, you know, reward for food,
link |
for maybe human connection, for let's say warmth, sex,
link |
And it turns out that the various people
link |
might be optimizing slightly different reward functions.
link |
They essentially might care about different things.
link |
And in case of love, at least the love between two people,
link |
you can say that the, you know,
link |
boundary between people dissolves to such extent
link |
that they end up optimizing each other reward functions.
link |
Oh, that's interesting.
link |
Celebrate the success of each other.
link |
Or in some sense, I would say love means helping others
link |
to optimize their reward functions,
link |
not your reward functions,
link |
not the things that you think are important,
link |
but the things that the person cares about,
link |
you try to help them to optimize it.
link |
So love is, if you think of two reward functions,
link |
you just, it's a condition.
link |
You combine them together.
link |
Yeah, pretty much.
link |
Maybe like with a weight,
link |
and it depends like the dynamic of the relationship.
link |
Yeah, I mean, you could imagine that if you are fully
link |
optimizing someone's reward function without yours,
link |
then maybe you are creating codependency
link |
or something like that.
link |
I'm not sure what's the appropriate weight.
link |
But the interesting thing is,
link |
I even think that the individual person,
link |
we ourselves, we are actually less of a unified insight.
link |
So for instance, if you look at the donut,
link |
on the one level, you might think,
link |
oh, this is like a look space that I would like to eat it.
link |
On another level, you might tell yourself,
link |
I shouldn't be doing it because I want to gain muscles.
link |
So, and you know, you might do it regardless,
link |
kind of against yourself.
link |
So it seems that even within ourselves,
link |
they're almost like a kind of intertwined personas.
link |
And I believe that the self love
link |
means that the love between all these personas,
link |
which also means being able to love yourself
link |
when we are angry or stressed or so.
link |
Combining all those reward functions
link |
of the different selves you have.
link |
Yeah, and accepting that they are there.
link |
Like, you know, often people,
link |
they have a negative self talk,
link |
or they say, I don't like when I'm angry.
link |
And like I try to imagine,
link |
if there would be like a small baby Lex,
link |
like a five years old, angry,
link |
and then you're like, you shouldn't be angry.
link |
Like, stop being angry.
link |
But like, instead, actually you want Lex to come over,
link |
give him a hug, and this, like I say, it's fine.
link |
Okay, you can be angry as long as you want.
link |
And then he would stop, or maybe not.
link |
Or maybe not, but you cannot expect it even.
link |
Yeah, but still, that doesn't explain it.
link |
Still, that doesn't explain the why of love.
link |
Like, why is love part of the human condition?
link |
Why is it useful to combine the reward functions?
link |
It seems like that doesn't, I mean,
link |
I don't think reinforcement learning frameworks
link |
can give us answers to why.
link |
Even the Hutter framework has an objective function
link |
So we came to existence as a consequence
link |
of evolutionary process.
link |
And in some sense, the purpose of evolution is survival.
link |
And then this complicated optimization objective
link |
baked into us, let's say compression,
link |
which might help us operate in the reward,
link |
and it baked into us various reward functions.
link |
Then to be clear, at the moment,
link |
we are operating in the regime,
link |
which is somewhat out of distribution
link |
where the even evolution optimized us.
link |
It's almost like love is a consequence
link |
of cooperation that we've discovered is useful.
link |
Correct. In some way, it's even the case if you...
link |
I just love the idea that love is out of distribution.
link |
Or it's not out of distribution.
link |
It's like, as you said, it evolved for cooperation.
link |
And I believe that in some sense,
link |
cooperation ends up helping each of us individually.
link |
So it makes sense evolutionary.
link |
And there is, in some sense, and love means
link |
there is this dissolution of boundaries
link |
that you have shared the reward function.
link |
And we evolved to actually identify ourselves
link |
with larger groups.
link |
So we can identify ourselves with a family.
link |
We can identify ourselves with a country
link |
to such extent that people are willing
link |
to give away their life for country.
link |
So we are wired, actually, even for love.
link |
And at the moment, I guess, maybe it
link |
would be somewhat more beneficial if we would identify
link |
ourselves with all the humanity as a whole.
link |
So you can clearly see, when people travel around the world,
link |
when they run into a person from the same country,
link |
they say, oh, which city you are.
link |
And all of a sudden, they find all these similarities.
link |
They befriend those folks earlier than others.
link |
So there is some sense of the belonging.
link |
And I would say, I think it would be overall good thing
link |
to the world for people to move towards,
link |
I think it's even called open individualism,
link |
move toward the mindset of a larger and larger groups.
link |
So the challenge there, that's a beautiful vision,
link |
and I share it, to expand that circle of empathy,
link |
that circle of love towards the entirety of humanity.
link |
But then you start to ask, well, where do you draw the line?
link |
Because why not expand it to other conscious beings?
link |
And then finally, for our discussion,
link |
something I think about is why not expand it to AI systems?
link |
Like, we start respecting each other
link |
when the entity on the other side has the capacity to suffer,
link |
because then we develop a capacity to empathize.
link |
And so I could see AI systems that
link |
are interacting with humans more and more
link |
having conscious like displays.
link |
So they display consciousness through language
link |
and through other means.
link |
And so then the question is, well, is that consciousness?
link |
Because they're acting conscious.
link |
And so the reason we don't like torturing animals
link |
is because they look like they're suffering
link |
when they're tortured.
link |
And if AI looks like it's suffering when it's tortured,
link |
how is that not requiring of the same kind of empathy from us
link |
and respect and rights that animals do and other humans do?
link |
I think it requires empathy as well.
link |
I mean, I would like, I guess, us or humanity
link |
or so make a progressing understanding what consciousness
link |
is, because I don't want just to be speaking
link |
about the philosophy, but rather actually make a scientific.
link |
There was a time that people thought
link |
that there is a force of life and the things that
link |
have this force, they are alive.
link |
And I think that there is actually
link |
a path to understand exactly what consciousness is.
link |
And in some sense, it might require essentially putting
link |
probes inside of a human brain.
link |
What Neuralink does.
link |
So the goal there, I mean, there's
link |
several things with consciousness that
link |
make it a real discipline, which is one,
link |
is rigorous measurement of consciousness.
link |
And then the other is the engineering
link |
of consciousness, which may or may not be related.
link |
I mean, you could also run into trouble,
link |
like, for example, in the United States,
link |
the Department of DOT, Department of Transportation,
link |
and a lot of different places put a value on human life.
link |
I think DOT's value is $9 million per person.
link |
So in that same way, you can get into trouble
link |
if you put a number on how conscious a being is,
link |
because then you can start making policy.
link |
If a cow is 0.1 or like 10% as conscious as a human,
link |
then you can start making calculations
link |
and it might get you into trouble.
link |
But then again, that might be a very good way to do it.
link |
I would like to move to that place that actually we
link |
have scientific understanding what consciousness is.
link |
And then we'll be able to actually assign value.
link |
And I believe that there is even the path
link |
for the experimentation in it.
link |
So we said that you could put the probes inside of the brain.
link |
There is actually a few other things
link |
that you could do with devices like Neuralink.
link |
So you could imagine that the way even
link |
to measure if AI system is conscious
link |
is by literally just plugging into the brain.
link |
I mean, that assumes that it's kind of easy.
link |
But plugging into the brain and asking
link |
a person if they feel that their consciousness expanded.
link |
This direction, of course, has some issues.
link |
You can say, if someone takes a psychedelic drug,
link |
they might feel that their consciousness expanded,
link |
even though that drug itself is not conscious.
link |
So you can't fully trust the self report of a person
link |
saying their consciousness is expanded or not.
link |
Let me ask you a little bit about psychedelics.
link |
There have been a lot of excellent research
link |
on different psychedelics, psilocybin, MDMA, even DMT,
link |
drugs in general, marijuana too.
link |
What do you think psychedelics do to the human mind?
link |
It seems they take the human mind to some interesting places.
link |
Is that just a little hack, a visual hack,
link |
or is there some profound expansion of the mind?
link |
I don't believe in magic.
link |
I believe in science, in causality.
link |
And then as I said, I think that our subjective experience
link |
of reality is we live in the simulation run by our brain.
link |
And the simulation that our brain runs,
link |
they can be very pleasant or very hellish.
link |
Drugs, they are changing some hyperparameters
link |
of the simulation.
link |
It is possible thanks to change of these hyperparameters
link |
to actually look back on your experience
link |
and even see that the given things that we took for granted,
link |
they are changeable.
link |
So they allow to have an amazing perspective.
link |
There is also, for instance, the fact
link |
that after DMT people can see the full movie inside
link |
of their head, gives me further belief
link |
that the brain can generate the full movie,
link |
that the brain is actually learning
link |
the model of reality to such extent
link |
that it tries to predict what's going to happen next.
link |
Yeah, very high resolution.
link |
So it can replay realities.
link |
Extremely high resolution.
link |
Yeah, and it's also kind of interesting to me
link |
that somehow there seems to be some similarity
link |
between these drugs and meditation itself.
link |
And I actually started even these days
link |
to think about meditation as a psychedelic.
link |
Do you practice meditation?
link |
I practice meditation.
link |
I mean, I went a few times on the retreats
link |
and it feels like after second or third day of meditation,
link |
there is almost like a sense of tripping.
link |
What is a meditation retreat entail?
link |
So you wake up early in the morning
link |
and you meditate for an extended period of time.
link |
Just to say, Jake's been trying.
link |
Yeah, so it's optimized, even though there are other people,
link |
it's optimized for isolation.
link |
So you don't speak with anyone.
link |
You don't actually look into other people's eyes.
link |
And you sit on the chair.
link |
So vipassana meditation tells you to focus on the breath.
link |
So you try to put all the attention
link |
into breathing in and breathing out.
link |
And the crazy thing is that as you focus attention like that,
link |
after some time, there starts coming back
link |
like some memories that you completely forgotten.
link |
It almost feels like you have a mailbox
link |
and then you are just archiving email one by one.
link |
And at some point, there is like an amazing feeling
link |
of getting to mailbox zero, zero emails.
link |
And it's very pleasant.
link |
It's kind of, it's crazy to me that once you resolve
link |
these inner stories or inner traumas,
link |
then once there is nothing left,
link |
the default state of human mind is extremely peaceful and happy.
link |
Like some sense, it feels that the,
link |
it feels at least to me in the way how when I was a child
link |
that I can look at any object and it's very beautiful.
link |
I have a lot of curiosity about the simple things.
link |
And that's where the usual in meditation takes me.
link |
Are you, what are you experiencing?
link |
Are you just taking in simple sensory information
link |
and are just enjoying the rawness of that sensory information?
link |
So there's no, there's no memories or all that kind of stuff.
link |
You're just enjoying being.
link |
Yeah, pretty much.
link |
I mean, still there is a, that it's,
link |
it's thoughts are slowing down sometimes they pop up,
link |
but it's also somehow the extended meditation
link |
takes you to the space that they are way more friendly.
link |
You're way more positive.
link |
There is also this, this thing that we've extended.
link |
It almost feels that the,
link |
it almost feels that we are constantly getting a little bit of a reward function
link |
and we are just spreading this reward function on various activities.
link |
But if you stay still for extended period of time,
link |
it kind of accumulates, accumulates, accumulates.
link |
And there is a, there is a sense,
link |
there is a sense that some point it passes some threshold
link |
and it feels as drop is falling into kind of ocean of love and bliss.
link |
And that's like a, this is like a very pleasant.
link |
And that's what I'm saying, like a,
link |
that corresponds to the subjective experience.
link |
Some people, I guess, in spiritual community,
link |
they describe it that that's the reality.
link |
And I would say I believe that they're like all sorts of subjective experience
link |
that one can have.
link |
And I believe that, for instance, meditation might take you to the subjective experiences,
link |
which are very pleasant, collaborative.
link |
And I would like a word to move toward a more collaborative place.
link |
Yeah, I would say that's very pleasant and I enjoy doing stuff like that.
link |
I wonder how that maps to your mathematical model of love
link |
with the reward function combining a bunch of things.
link |
It seems like our life then is we're just,
link |
we have this reward function and we're accumulating a bunch of stuff in it with weights.
link |
It's like a, like multi objective.
link |
And what meditation is, is you just remove them, remove them
link |
until the weight on one or just a few is very high.
link |
And that's where the pleasure comes from.
link |
Yeah, so something similar to how I'm thinking about it.
link |
So I told you that there is this like, there is a story of who you are.
link |
And I think almost about it as a, you know, text prepended to GPT.
link |
And some people refer to it as ego.
link |
Okay, it's like a story who, who, who you are.
link |
So ego is the prompt for GPT three or GPT.
link |
And that's the description of you.
link |
And then with meditation, you can get to the point that actually you
link |
experience things without the prompt.
link |
And you experience things like as they are, you are not biased
link |
over the description, how they supposed to be.
link |
That's very pleasant.
link |
And then we've respected the reward function.
link |
It's possible to get to the point that there is the solution of self.
link |
And therefore you can say that you're, you're having a, you're,
link |
or like a your brain attempts to simulate the reward function of everyone else or
link |
like everything that's, that there is this like a love which feels like a
link |
oneness with everything.
link |
And that's also, you know, very beautiful, very pleasant.
link |
At some point you might have a lot of altruistic thoughts during that
link |
moment and then the self always comes back.
link |
How would you recommend if somebody is interested in meditation, like a big
link |
thing to take on as a project?
link |
Would you recommend a meditation retreat?
link |
How many days, what kind of thing would you recommend?
link |
I think that actually retreat is the way to go.
link |
It almost feels that, as I said, like a meditation is a psychedelic,
link |
but when you take it in the small dose, you might barely feel it.
link |
Once you get the high dose, actually you're going to feel it.
link |
So even cold turkey, if you haven't really seriously meditated for a long
link |
period of time, just go to retreat.
link |
Yeah, how many days?
link |
How many days? Start weekend one.
link |
Weekend, so like two, three days.
link |
And it's like, it's interesting that first or second day, it's hard and
link |
at some point it becomes easy.
link |
There's a lot of seconds in a day.
link |
How hard is the meditation retreat, just sitting there in a chair?
link |
So the thing is actually, it literally just depends on your, on your own framing.
link |
Like if you are in the mindset that you are waiting for it to be over or
link |
you are waiting for Nirvana to happen, it will be very unpleasant.
link |
And in some sense, even the difficulty, it's not even in the lack
link |
of being able to speak with others.
link |
Like you're sitting there, your legs will hurt from sitting.
link |
In terms of like the practical things, do you experience kind of discomfort,
link |
like physical discomfort of just sitting, like your, your butt being numb,
link |
your legs being sore, all that kind of stuff?
link |
Yes, you experience it and then the, the, they teach you to observe it rather.
link |
And it's like a, the crazy thing is, you at first might have a feeling
link |
toward trying to escape it.
link |
And that becomes very apparent that that's extremely unpleasant.
link |
And then you just, just observe it.
link |
And at some point it, it just becomes, it just is.
link |
It's like a, I remember that we've, Ilya told me some time ago that, you know,
link |
he takes a cold shower and his mindset of taking a cold, cold shower was
link |
to embrace suffering.
link |
This is your style?
link |
Yeah, it's my style.
link |
So my style is actually, I also sometimes take cold showers.
link |
It is purely observing how the water goes through my body, like a purely
link |
being present, not trying to escape from there.
link |
And I would say then it actually becomes pleasant.
link |
It's not like, well, that, that's interesting.
link |
I, I'm also, that means that's, that's the way to deal with anything really
link |
difficult, especially in the physical space is to observe it.
link |
To say it's pleasant.
link |
It's a, I would use a different word.
link |
You're, you're accepting of the full beauty of reality, I would say,
link |
because it's a pleasant, but yeah, in some sense, it is pleasant.
link |
That's the only way to deal with a cold shower is to become an observer
link |
and to find joy in it.
link |
Same with like really difficult physical exercise or like running for a
link |
really long time, endurance events, just anytime you're exhausted, any kind of
link |
pain, I think the only way to survive it is not to resist it.
link |
It's to observe it.
link |
You mentioned Ilya, Ilya Satskever.
link |
It's very, he's our chief scientist, but also he's very close friend of mine.
link |
You co founded Open AI with you.
link |
I've spoken with him a few times.
link |
I really enjoy talking to him.
link |
His mind, just like yours, works in fascinating ways.
link |
Both of you are not able to define deep learning simply.
link |
What's it like having him as somebody you have technical discussions with
link |
on in space and machine learning, deep learning AI, but also life?
link |
What's it like when these two agents get into a self play situation in a room?
link |
What's it like collaborating with him?
link |
So I believe that we have extreme respect to each other.
link |
So I love Ilya's insight, both like I guess about consciousness, life, AI.
link |
But in terms of the, it's interesting to me because you're a brilliant thinker in
link |
the space and machine learning, like intuition, like digging deep in what works, what doesn't,
link |
why it works, why it doesn't.
link |
I'm wondering if there's interesting deep discussions you've had with him in the past
link |
or disagreements that were very productive.
link |
So I can say I also understood over the time where are my strengths.
link |
So obviously we have plenty of discussions.
link |
And I myself have plenty of ideas, but I consider Ilya one of the most prolific AI scientists
link |
in the entire world.
link |
And I think that I realize that maybe my super skill is being able to bring people to collaborate
link |
together, that I have some level of empathy that is unique in AI world.
link |
And that might come from either meditation, psychedelics, or let's say I read just hundreds
link |
of books on this topic.
link |
And I also went through a journey of, I develop all sorts of algorithms.
link |
So I think that maybe I can, that's my super human skill.
link |
Ilya is one of the best AI scientists, but then I'm pretty good in assembling teams.
link |
And I'm also not holding to people like I'm growing people.
link |
And then people become managers at OpenAI, that's room any of them like a research managers.
link |
So you find places where you're excellent and he finds like his deep scientific insights
link |
And you find ways you can, the puzzle pieces fit together.
link |
Like, you know, ultimately, for instance, let's say Ilya, he doesn't manage people.
link |
That's not what he likes or so.
link |
I like hanging out with people.
link |
By default, I'm an extrovert and I care about people.
link |
So that fits perfectly together.
link |
But I mean, I also just like your intuition about various problems in machine learning.
link |
He's definitely one I really enjoy.
link |
I remember talking to him about something I was struggling with, which is coming up with
link |
a good model for pedestrians for human beings that cross the street in the context of autonomous
link |
And he immediately started to like formulate a framework within which you can evolve a
link |
model for pedestrians, like through self play, all that kind of mechanisms.
link |
The depth of thought on a particular problem, especially problems he doesn't know anything
link |
about is fascinating to watch.
link |
And it makes you realize like, yeah, the limits that the human intellect might be limitless.
link |
Or it's just impressive to see a descendant of ape come up with clever ideas.
link |
I mean, so even in the space of deep learning, when you look at various people, there are
link |
people now who invented some breakthroughs once, but there are very few people who did
link |
it multiple times.
link |
And you can think if someone invented it once, that might be just a sure luck.
link |
And if someone invented it multiple times, you know, if the probability of inventing
link |
it once is one over a million, then probability of inventing it twice or three times would
link |
be one over a million square or to the power of three, which would be just impossible.
link |
So it literally means that it's given that it's not the luck.
link |
And Ilya is one of these few people who have a lot of these inventions in his arsenal.
link |
It also feels that, you know, for instance, if you think about folks like Gauss or Euler,
link |
you know, at first they read a lot of books, and then they did thinking, and then they
link |
And that's how it feels with Ilya.
link |
You know, at first he read stuff, and then he spent his thinking cycles.
link |
And that's a really good way to put it.
link |
When I talk to him, I see thinking.
link |
He's actually thinking.
link |
Like he makes me realize that there's like deep thinking that the human mind can do.
link |
Like most of us are not thinking deeply.
link |
Like you really have to put in a lot of effort to think deeply.
link |
Like I have to really put myself in a place where I think deeply about a problem.
link |
It takes a lot of effort.
link |
It's like an airplane taking off or something.
link |
You have to achieve deep focus.
link |
He's just, his brain is like a vertical takeoff in terms of airplane analogy.
link |
So it's interesting.
link |
But I mean, Cal Newport talks about this, his ideas of deep work.
link |
It's, you know, most of us don't work much at all in terms of like deeply think about
link |
particular problems, whether it's in math, engineering, all that kind of stuff.
link |
You want to go to that place often.
link |
And that's real hard work.
link |
And some of us are better than others at that.
link |
So I think that the big piece has to do with actually even engineering our environment
link |
that such that it's conducive to that.
link |
So see both Ilya and I on the frequent basis, we kind of disconnect ourselves from the world
link |
in order to be able to do extensive amount of thinking.
link |
So Ilya usually, he just leaves iPad at hand.
link |
He loves his iPad.
link |
And for me, I'm even sometimes, you know, just going for a few days to different location
link |
So I'm turning off my phone and there is no access to me.
link |
And that's extremely important for me to be able to actually just formulate new thoughts
link |
to do deep work rather than to be reactive.
link |
And the older I am, the more of this like a random tasks are at hand.
link |
Before I go on to that thread, let me return to our friend GPT.
link |
Let me ask you another ridiculously big question.
link |
Can you give an overview of what GPT three is?
link |
Or like you say in your Twitter bio, GPT and plus one.
link |
How it works and why it works.
link |
So GPT three is a humongous neural network.
link |
Let's assume that we know what is neural network by definition.
link |
And it is trained on the entire internet just to predict next word.
link |
So let's say it sees part of the article and the only task that it has at hand, it is to
link |
say what would be the next word?
link |
What would be the next word?
link |
And it becomes really exceptional at the task of figuring out what's the next word.
link |
So you might ask, why would this be an important task?
link |
Why would it be important to predict what's the next word?
link |
And it turns out that a lot of problems can be formulated as a text completion problem.
link |
So GPT is purely learning to complete the text.
link |
As you could imagine, for instance, if you are asking a question, who is president of
link |
United States, then GPT can give you an answer to it.
link |
It turns out that many more things can be formulated this way.
link |
You can format text in the way that you have sent us in English.
link |
You make it even look like some content of a website elsewhere, which would be teaching
link |
people how to translate things between languages.
link |
So it would be EN colon, text in English, FR colon, and then you ask model to continue.
link |
And it turns out that such a model is predicting translation from English to French.
link |
The crazy thing is that this model can be used for way more sophisticated tasks.
link |
So you can format text such that it looks like a conversation between two people.
link |
And that might be a conversation between you and Elon Musk.
link |
And because the model read all the text about Elon Musk, it will be able to predict Elon
link |
Musk words as it would be Elon Musk.
link |
It will speak about colonization of Mars, about sustainable future and so on.
link |
And it's also possible to even give arbitrary personality to the model.
link |
You can say, here is a conversation that we've a friendly AI bot.
link |
And the model will complete the text as a friendly AI bot.
link |
So I mean, how do I express how amazing this is?
link |
So just to clarify a conversation, generating a conversation between me and Elon Musk, it
link |
wouldn't just generate good examples of what Elon would say.
link |
It would get the syntax all correct.
link |
So like interview style, it would say like Elon Cohen and Lex Cohen.
link |
It's not just like inklings of semantic correctness.
link |
It's like the whole thing, grammatical, syntactic, semantic.
link |
It's just really, really impressive generalization.
link |
Yeah, I mean, I also want to provide some caveat so it can generate few paragraphs of
link |
But as you go to longer pieces, it actually goes off the rails.
link |
If you would try to write a book, it won't work out this way.
link |
What way does it go off the rails, by the way?
link |
Is there interesting ways in which it goes off the rails?
link |
What falls apart first?
link |
So the model is trained on all the existing data that is out there, which means that it
link |
is not trained on its own mistakes.
link |
So for instance, if it would make a mistake, then I kept so to give you an example.
link |
So let's say I have a conversation with a model pretending that is Elon Musk, and then
link |
I start putting some, I'm start actually making up things which are not factual.
link |
Sounds like Twitter.
link |
But I got you, sorry.
link |
I would say that Elon is my wife, and the model will just keep on carrying it on.
link |
And in some sense, if you would have a normal conversation with Elon, he would be, what
link |
Yeah, there would be some feedback.
link |
So the model is trained on things that humans have written, but through the generation process,
link |
there's no human in the loop feedback.
link |
That's fascinating.
link |
So it's magnified.
link |
It's like the errors get magnified and magnified.
link |
And it's also interesting, I mean, first of all, humans have the same problem.
link |
It's just that we will make fewer errors and magnify the errors slower.
link |
I think that actually what happens with humans is if you have a wrong belief about the world
link |
as a kid, then very quickly you will learn that it's not correct because they are grounded
link |
in reality and they are learning from your new experience.
link |
But do you think the model can correct itself too?
link |
Won't it, through the power of the representation, and so the absence of Elon Musk being your
link |
wife, information on the internet, won't it correct itself?
link |
There won't be examples like that.
link |
So the errors will be subtle at first.
link |
And in some sense, you can also say that the data that is not out there is the data which
link |
would represent how the human learns.
link |
And maybe the model would be trained on such a data, then it would be better off.
link |
How intelligent is GPT3, do you think?
link |
When you think about the nature of intelligence, it seems exceptionally impressive.
link |
But then if you think about the big AGI problem, is this footsteps along the way to AGI?
link |
So it seems that intelligence itself is a multiple axis of it.
link |
And I would expect that the systems that we are building, they may end up being superhuman
link |
on some axis and subhuman on some other axis.
link |
It would be surprising to me on all axis simultaneously, they would become superhuman.
link |
Of course, people ask this question, is GPT a spaceship that would take us to Moon?
link |
Or are we building a ladder to Heaven that we are just building bigger and bigger ladder?
link |
And we don't know in some sense which one of these two works better.
link |
Which one is better?
link |
I like Stairway to Heaven, it's a good song, so I'm not exactly sure which one is better.
link |
But you're saying the spaceship to the Moon is actually effective?
link |
And people who criticize GPT, they say, you are just building a ladder and it will never
link |
And at the moment, I would say the way I'm thinking is like a scientific question.
link |
And I'm also in heart, I'm a builder creator.
link |
And I'm thinking, let's try out, let's see how far it goes.
link |
And so far, we see constantly that there is a progress.
link |
So do you think GPT 4, GPT 5, GPT N plus 1, there will be a phase shift, like a transition
link |
to a place where we'll be truly surprised.
link |
And again, GPT 3 is already very truly surprising.
link |
The people that criticize GPT 3 as a Stairway, as a ladder to Heaven, I think two people
link |
quickly get accustomed to how impressive it is that the prediction of the next word can
link |
achieve such depth of semantics, accuracy of syntax, grammar, and semantics.
link |
Do you think GPT 4 and 5 and 6 will continue to surprise us?
link |
I mean, definitely, there will be more impressive models.
link |
There is a question, of course, if there will be a phase shift.
link |
And also even the way I'm thinking about these models is that when we build these models,
link |
we see some level of the capabilities, but we don't even fully understand everything
link |
that the model can do.
link |
And actually one of the best things to do is to allow other people to probe the model
link |
to even see what is possible.
link |
Since using GPT as an API and opening it up to the world.
link |
I mean, so when I'm thinking from perspective of obviously various people that have concerns
link |
about AGI, including myself, and then when I'm thinking from perspective what's the strategy
link |
even to deploy these things to the world, then the one strategy that I have seen many
link |
times working is the iterative deployment that you deploy slightly better versions
link |
and you allow other people to criticize you.
link |
So you actually or try it out, you see where are their fundamental issues.
link |
And it's almost you don't want to be in that situation that you are holding into powerful
link |
system and there's like a huge overhang, then you deploy it and it might have a random chaotic
link |
impact on the world.
link |
So you actually want to be in the situation that you are gradually deploying systems.
link |
I asked this question of Ilya, let me ask you this question.
link |
I've been reading a lot about Stalin and power.
link |
If you're in possession of a system that's like AGI, that's exceptionally powerful.
link |
Do you think your character and integrity might become corrupted?
link |
Like famously power corrupts and absolutely power corrupts, absolutely.
link |
So I believe that you want at some point to work toward distributing the power.
link |
I think that you want to be in the situation that actually AGI is not controlled by a small
link |
number of people, but essentially by a larger collective.
link |
So the thing is that requires a George Washington style move in the ascent to power.
link |
There's always a moment when somebody gets a lot of power and they have to have the integrity
link |
and the moral compass to give away that power.
link |
That humans have been good and bad throughout history at this particular step.
link |
And I wonder, I wonder we like blind ourselves in, for example, between nations, a race towards
link |
the AI race between nations, we might blind ourselves and justify to ourselves the development
link |
of AI without distributing the power because we want to defend ourselves against China,
link |
against Russia, that kind of logic.
link |
And I wonder how we design governance mechanisms that prevent us from becoming power hungry
link |
and in the process destroying ourselves.
link |
I have been thinking about this topic quite a bit, but I also want to admit that once
link |
again I actually want to rely way more on some outman, on a hero and excellent block
link |
on how even to distribute wealth.
link |
And he proposed in his blog to tax equity of the companies rather than profit and to
link |
And this is an example of Washington move.
link |
I guess I personally have insane trust in some.
link |
He already spent plenty of money running universal basic income project that gives me, I guess,
link |
maybe some level of trust to him, but I also, I guess, love him as a friend.
link |
I wonder, because we're sort of summoning a new set of technologies, I wonder if we'll
link |
be cognizant, like you're describing the process of open AI, but it could also be at other
link |
places like in the US government, right?
link |
Both China and the US are now full steam ahead on autonomous weapons systems development.
link |
And that's really worrying to me because in the framework of something being a national
link |
security danger or a military danger, you can do a lot of pretty dark things that blind
link |
our moral compass.
link |
And I think AI will be one of those things.
link |
In some sense, the mission and the work you're doing at open AI is like the counterbalance
link |
So you want to have more open AI and less autonomous weapons systems.
link |
I like these statements.
link |
To be clear, this is interesting, and I'm thinking about it myself, but this is a place
link |
that I put my trust actually in some sense, because it's extremely hard for me to reason
link |
I mean, one important statement to make is it's good to think about this.
link |
No question about it.
link |
So even like low level, quote unquote, engineer.
link |
Like there's such a, I remember I programmed a car, our RC car.
link |
They went really fast, like 30, 40 miles an hour.
link |
And I remember I was like sleep deprived, so I programmed it pretty crappily, and like
link |
So it's doing some basic computer vision and it's going around on track, but it's going
link |
And there was a bug in the code that the car just went, it didn't turn, it went straight,
link |
full speed and smash into the wall.
link |
And I remember thinking the seriousness with which you need to approach the design of artificial
link |
intelligence systems and the programming of artificial intelligence systems is high because
link |
the consequences are high.
link |
That little car smashing into the wall, for some reason, I immediately thought of like
link |
an algorithm that controls nuclear weapons, having the same kind of bug.
link |
And so like the lowest level engineer and the CEO of a company all need to have the seriousness
link |
in approaching this problem and thinking about the worst case consequences.
link |
So I think that is true.
link |
I mean, what I also recognize in myself and others even asking this question is that it
link |
evokes a lot of fear and the fear itself ends up being actually quite deabilitating.
link |
The place where I arrived at the moment might sound cheesy or so, but it's almost to build
link |
things out of love rather than fear.
link |
I can focus on how I can maximize the value, how the systems that I'm building might be
link |
I'm not saying that the fear doesn't exist out there and it totally makes sense to minimize
link |
But I don't want to be working because I'm scared.
link |
I want to be working out of passion, out of curiosity, out of looking forward for the
link |
With the definition of love arising from rigorous practice of empathy, so not just like your
link |
own conception of what is good for the world, but always listening to others.
link |
Like at the love where I'm considering reward functions of others.
link |
To limit infinity is like one to N where N is 7 billion or whatever it is.
link |
Not projecting my reward functions on others.
link |
Can we just take a step back to something else super cool, which is Open AI Codex?
link |
Can you give an overview of what Open AI Codex and GitHub copilot is, how it works and why
link |
the hell it works so well?
link |
So with GPT3, we noticed that the system train on all the language out there started having
link |
some rudimentary coding capabilities.
link |
So we're able to implement addition functions between two numbers and indeed it can write
link |
Python or JavaScript code for that.
link |
And then we thought we might as well just go full steam ahead and try to create a system
link |
that is actually good at what we are doing every day ourselves, which is programming.
link |
We optimize models for proficiency in coding.
link |
We actually even created models that both have a comprehension of language and code.
link |
And Codex is API for these models.
link |
So it's first pre trained on language and then, I don't know if you can say fine tuned
link |
because there's a lot of code, but it's language and code.
link |
It's language and code.
link |
It's also optimized for various things like let's say low latency and so on.
link |
Codex is the API that's similar to GPT3.
link |
We expect that there will be proliferation of the potential products that can use coding
link |
capabilities and I can speak about it in a second.
link |
Copilot is a first product developed by GitHub.
link |
So as we're building models, we wanted to make sure that these models are useful and
link |
we work together with GitHub on building the first product.
link |
Copilot is actually, as you code, it suggests you code completions.
link |
And we have seen in the past, there are like various tools that can suggest how to like
link |
a few characters of the code or the line of code.
link |
The thing about Copilot is it can generate 10 lines of code.
link |
It's often the way how it works is you often write in the comment what you want to happen
link |
because people in comments, they describe what happens next.
link |
So these days when I code, instead of going to Google to search for the appropriate code
link |
to solve my problem, I say, oh, for this array, could you smooth it?
link |
And then it imports some appropriate libraries and say it uses NumPy convolution or so that
link |
I was not even aware that exists.
link |
Many does the appropriate thing.
link |
So you write a comment, maybe the header of a function and it completes the function.
link |
Of course, you don't know what is the space of all the possible small programs it can
link |
What are the failure cases?
link |
How many edge cases?
link |
How many subtle errors there are?
link |
How many big errors there are?
link |
It's hard to know, but the fact that it works at all in a large number of cases is incredible.
link |
It's like a kind of search engine into code that's been written on the Internet.
link |
So for instance, when you search things online, then usually you get to some particular case
link |
like if you go to Stack Overflow, people describe that one particular situation and then they
link |
seek for a solution.
link |
But in case of Copilot, it's aware of your entire context and in context is, oh, these
link |
are the libraries that they are using.
link |
That's the set of the variables that is initialized.
link |
And on the spot, it can actually tell you what to do.
link |
So the interesting thing is, and we think that the Copilot is one possible product
link |
using Codex, but there is a place for many more.
link |
So internally, we tried out to create other fun products.
link |
So it turns out that a lot of tools out there, let's say Google Calendar or Microsoft Word
link |
or so, they all have internal API to build plugins around them.
link |
So there is a way in the sophisticated way to control Calendar or Microsoft Word.
link |
Today if you want more complicated behaviors from these programs, you have to add a new
link |
button for every behavior.
link |
But it is possible to use Codex and tell, for instance, to Calendar, could you schedule
link |
an appointment with Lex next week after 2 p.m. and either write corresponding piece
link |
And that's the thing that actually you want.
link |
So what you figure out is there's a lot of programs with which you can interact through
link |
And so there, you can generate that code from natural language.
link |
That's fascinating.
link |
That's somewhat like also closest to what was the promise of Siri or Alexa.
link |
So previously, all these behaviors, they were hand hard coded.
link |
And it seems that Codex on the fly can pick up the API of, let's say, given software.
link |
And then it can turn language into use of this API.
link |
So without hard coding, it can translate to machine language.
link |
So for example, this would be really exciting for me, like for Adobe products like Photoshop,
link |
which I think ActionScript, I think there's a scripting language that communicates with
link |
Same with Premiere.
link |
And do you could imagine that that allows even to do coding by voice on your phone?
link |
So for instance, in the past, as of today, I'm not editing Word documents on my phone
link |
because it's just the keyboard is too small.
link |
But if I would be able to tell to my phone, you know, make the header large, then move
link |
the paragraphs around, and it does actually what I want.
link |
So I can tell you one more cool thing, or even how I'm thinking about Codex.
link |
So if you look actually at the evolution of computers, we started with very primitive
link |
interfaces, which is a punch card.
link |
And punch card, essentially, you make holes in the plastic card to indicate zeros and ones.
link |
And during that time, there was a small number of specialists who were able to use computers.
link |
And by the way, people even suspected that there is no need for many more people to use
link |
But then we moved from punch cards to, at first, assembly and C. And these programming
link |
languages, they were slightly higher level.
link |
They allowed many more people to code.
link |
And they also led to more of a proliferation of technology.
link |
And further on, there was a jump to, say, from C++ to Java and Python.
link |
And every time it has happened, more people are able to code, and we build more technology.
link |
And it's even hard to imagine now if someone will tell you that you should write code in
link |
assembly instead of, let's say, Python or Java or JavaScript.
link |
And Codex is yet another step toward kind of bringing computers closer to humans, such
link |
that you communicate with a computer with your own language, rather than with a specialized
link |
And I think that it will lead to increase of number of people who can code.
link |
And the kind of technologies that those people will create, it's innumerable, it could be
link |
a huge number of technologies we're not predicting at all, because that's less and less requirement
link |
of having a technical mind, a programming mind.
link |
You're not opening it to the world of other kinds of minds, creative minds, artistic minds,
link |
all that kind of stuff.
link |
I would like, for instance, biologists who work on DNA to be able to program and not
link |
to need to spend a lot of time learning it.
link |
And I believe that's a good thing to the world.
link |
And I would actually add that.
link |
So at the moment, I'm a managing Codex team and also language team.
link |
And I believe that there is like a plenty of brilliant people out there, and they should
link |
So what's the language in the Codex?
link |
So those are kind of their overlapping teams, it's like GPT, the raw language.
link |
And then the Codex is like applied to programming.
link |
And they are quite intertwined.
link |
There are many more teams involved making these models extremely efficient and deployable.
link |
For instance, there are people who are working to make our data centers amazing, or there
link |
are people who work on putting these models into production, or even pushing it at the
link |
very limit of the scale.
link |
So all aspects from the infrastructure to the actual machine learning?
link |
So I'm just saying there are multiple teams, while the team working on Codex and language,
link |
I guess, I'm directly managing them.
link |
I would love to hire.
link |
If you're interested in machine learning, this is probably one of the most exciting problems
link |
and like systems to be working on, because it's actually, it's pretty cool.
link |
Like what the program said, this is like generating a program, it's a very interesting, very interesting
link |
problem that has echoes of reasoning and intelligence in it.
link |
And I think there's a lot of fundamental questions that you might be able to sneak up to by generating
link |
One more exciting thing about the programs is that, so I said that the, you know, the
link |
in case of language, that one of the troubles is even evaluating language.
link |
So when the things are made up, you need somehow either a human to say that this doesn't make
link |
Or so in case of program, there is one extra lever that we can actually execute programs
link |
and see what they evaluate to.
link |
So the process might be somewhat more automated in order to improve the qualities of generations.
link |
Oh, that's fascinating.
link |
So like the, wow, that's really interesting.
link |
So for language, the, you know, the simulation to actually execute it as a human mind.
link |
For programs, there is a, there is a computer on which you can evaluate it.
link |
That's a brilliant little insight that the thing compiles and runs.
link |
And second, you can evaluate on a do automated unit testing.
link |
And in some sense, it seems to mean that we'll be able to make a tremendous progress.
link |
You know, we are in the paradigm that there is way more data and there is like a transcription
link |
of millions of, of software engineers.
link |
So, I mean, you just mean because I was going to ask you about reliability, the thing about
link |
programs is you don't know if they're going to like a program that's controlling a nuclear
link |
power plant has to be very reliable.
link |
So I wouldn't start with controlling nuclear power plant, maybe one day, but that's not
link |
actually, that's not on the current roadmap.
link |
That's not the step one, you know, it's the Russian thing.
link |
You just want to go to the most powerful destructive thing right away run by JavaScript.
link |
So it's a lower impact.
link |
But nevertheless, what you make me realize it is possible to achieve some levels of reliability
link |
You could imagine that, you know, maybe there are ways for a model to write even code for
link |
testing itself and so on.
link |
And there exist ways to create the feedback loops that the model could keep on improving.
link |
By writing programs that generate tests, for instance, for instance.
link |
And that's how we get consciousness because it's meta compression.
link |
That's what you're going to write.
link |
That's the comment.
link |
That's the prompt that generates consciousness, compressor of compressors.
link |
You just write that.
link |
Do you think the code that generates consciousness would be simple?
link |
So let's see, I mean, ultimately, the core idea behind will be simple, but there will
link |
be also decent amount of engineering involved.
link |
In some sense, it seems that spreading these models on many machines, it's not that trivial.
link |
And we find all sorts of innovations that make our models more efficient.
link |
I believe that first models that I guess are conscious are truly intelligent.
link |
They will have all sorts of tricks.
link |
But then again, there's a certain argument that maybe the tricks are temporary things.
link |
Yeah, they might be temporary things.
link |
And in some sense, it's also even important to know that even the cost of a trick.
link |
So sometimes, people are eager to put the trick while forgetting that there is a cost
link |
Or like a long term cost.
link |
Long term cost or maintenance or maybe even flexibility of code to actually implement
link |
So even if you have something that gives you 2X, but it requires 1,000 lines of code,
link |
I'm not sure if it's actually worth it.
link |
So in some sense, if it's 5 lines of code and 2X, I would take it.
link |
And we see many of this, but also that requires some level of, I guess, lack of attachment
link |
to code that we are willing to remove it.
link |
So you led the OpenAI robotics team.
link |
Can you give an overview of the cool things you're able to accomplish, what are you most
link |
So when we started robotics, we knew that actually reinforcement learning works.
link |
And it is possible to solve fairly complicated problems.
link |
Like for instance, AlphaGo is an evidence that it is possible to build superhuman and
link |
Dota2 is an evidence that it's possible to build superhuman agents playing Dota.
link |
So I asked myself a question, you know, what about robots out there?
link |
Would we train machines to solve arbitrary tasks in the physical world?
link |
Our approach was, I guess, let's pick a complicated problem that if we would solve it, that means
link |
that we made some significant progress in the domain, and then we went after the problem.
link |
So we noticed that actually the robots out there, they are kind of at the moment optimized
link |
So you can have a robot that if you have a robot opening a battle, it's very likely
link |
that the end factor is a battle opener.
link |
And in some sense, that's a hack to be able to solve a task, which makes any task easier.
link |
And I asked myself, so what would be a robot that can actually solve many tasks?
link |
And we concluded that human hands have such a quality that indeed they are, you know,
link |
you have five kind of tiny arms attached individually, they can manipulate pretty broad spectrum
link |
So we went after a single hand, like a trying to solve Rubik's Cube single handed, we picked
link |
this task because we thought that there is no way to hard code it.
link |
And also we picked a robot on which it would be hard to hard code it.
link |
And we went after the solution such that it could generalize to other problems.
link |
And just to clarify, it's one robotic hand solving the Rubik's Cube.
link |
The hard part is in the solution to the Rubik's Cube is the manipulation of the, of like having
link |
it not fall out of the hand, having it use the five baby arms to what is it like rotate
link |
different parts of the Rubik's Cube to achieve the solution.
link |
So what, what was the hardest part about that?
link |
What was the approach taken there?
link |
What are you most proud of?
link |
Obviously we have like a strong belief in reinforcement learning.
link |
And you know, one path, it is to do reinforcement learning, the real world.
link |
Other path is to the simulation, in some sense, the tricky part about the real world is at
link |
the moment our models, they require a lot of data, there is essentially no data.
link |
And I think we decided to go through the path of the simulation and in simulation, you can
link |
have infinite amount of data.
link |
The tricky part is the fidelity of the simulation.
link |
And also can you in simulation represent everything that you represent otherwise in the real world?
link |
And you know, it turned out that, that, you know, because there is lack of fidelity, it
link |
is possible to what we, what we arrived at is training a model that doesn't solve one
link |
simulation, but it actually solves the entire range of simulations, which vary in terms of
link |
like what's exactly the friction of the cube or the weight or so.
link |
And the single AI that can solve all of them ends up working well with the reality.
link |
How do you generate the different simulations?
link |
So you know, there's plenty of parameters out there, we just pick them randomly.
link |
And in simulation, model just goes for thousands of years and keeps on solving Rubik's Cube
link |
And the thing is the neural network that we used.
link |
And as it presses, for instance, the side of the, of the cube, it can sense, oh, that's
link |
actually this side was difficult to press, I should press it stronger.
link |
And throughout this process, kind of learns even how to, how to solve this particular
link |
instance of the Rubik's Cube back even mass, it's kind of like a, you know, sometimes when
link |
you go to a gym and after, after bench press, you try to lift the glass and you kind of
link |
forgot and, and, and your head goes like right away because kind of you got used to maybe
link |
different weight and it takes a second to adjust.
link |
And this kind of a, of a memory that model gained through the process of interacting
link |
with the cube in the simulation.
link |
I appreciate you speaking to the audience with a bench press, all the bros in the audience
link |
probably working out right now.
link |
There's probably somebody listening to this actually doing bench press.
link |
So maybe put the bar down and pick up the water bottle and you'll know exactly what,
link |
what Jack is talking about.
link |
So what, what was the hardest part of getting the whole thing to work?
link |
You know, the hardest part is at the moment when it comes to physical work, when it comes
link |
to robots, they require maintenance.
link |
It's hard to replicate a million times.
link |
It's, it's also, it's hard to replay things exactly.
link |
I remember this situation that one guy at our company, he had like a model that performs
link |
way better than other models in solving Rubik's Cube.
link |
And you know, we kind of didn't know what's going on, why it's that.
link |
And it turned out that, you know, he was running it from his laptop that had better
link |
CPU or better, maybe local GPU as well.
link |
And because of that, there was less of a latency and the model was the same.
link |
And that actually made solving Rubik's Cube more reliable.
link |
So in some sense, there might be some subtle bugs like that when it comes to running things
link |
in the real world, even hinting on that.
link |
You could imagine that the initial models, you would like to have models, which are insanely
link |
huge neural networks, and you would like to give them even more time for thinking.
link |
And when you have these real time systems, then you might be constrained actually by the
link |
amount of latency.
link |
And ultimately, I would like to build a system that it is worth for you to wait five minutes
link |
because it gives you the answer that you're willing to wait for five minutes.
link |
So latency is a very unpleasant constraint under which to operate.
link |
And also, there is actually one more thing which is tricky about robots.
link |
There is actually no, not much data.
link |
So the data that I'm speaking about would be a data of first person experience from the
link |
robot, and like gigabytes of data like that, if we would have gigabytes of data like that
link |
of robot solving parties problems, it would be very easy to make a progress on robotics.
link |
And you can see that in case of text or code, there is a lot of data, like a first person
link |
perspective data on writing code.
link |
Yeah, so you had this, you mentioned this really interesting idea that if you were to
link |
build like a successful robotics company, so OpenAS mission is much bigger than robotics.
link |
This is one of the, one of the things you've worked on.
link |
But if it was a robotics company, that you wouldn't so quickly dismiss supervised learning.
link |
Correct, that you would build a robot that was perhaps like an empty shell like dumb
link |
and they would operate under teleoperation.
link |
So you would invest, that's just one way to do it, invest in human, like direct human
link |
control of the robots as it's learning.
link |
And over time, add more and more automation.
link |
So let's say that's how I would build a robotics company today.
link |
If I would be building robotics company, which is spend $10 million or so, recording human
link |
trajectories, controlling a robot.
link |
After you find a thing that the robot should be doing that there's a market fit for, like
link |
that you can make a lot of money with that product.
link |
So I would record data and then I would essentially train supervised learning model on it.
link |
That might be the path today.
link |
Look term, I think that actually what is needed is to train powerful models over video.
link |
So you have seen maybe a models that can generate images like Dali.
link |
And people are looking into models generating videos.
link |
They're like, are these algorithmic questions, even how to do it?
link |
And it's unclear if there is enough compute for this purpose.
link |
But I suspect that the models that which would have a level of understanding of video, same
link |
as GPT has a level of understanding of text, could be used to train robots to solve tasks.
link |
They would have a lot of common sense.
link |
If one day, I'm pretty sure one day, there will be a robotics company.
link |
I mean, the primary source of income is from robots that is worth over $1 trillion.
link |
What do you think that company would do?
link |
I think sell driving cars, no.
link |
It's interesting because my mind went to personal robotics, robots in the home.
link |
It seems like there's much more market opportunity there.
link |
I think it's very difficult to achieve.
link |
I mean, this might speak to something important, which is I understand self driving much better
link |
than I understand robotics in the home.
link |
So I understand how difficult it is to actually solve self driving to a level, not just the
link |
actual computer vision and the control problem and just the basic problem of self driving,
link |
but creating a product that would undeniably be that would cost less money, like it would
link |
save you a lot of money.
link |
It would order the magnitude less money that could replace Uber drivers, for example.
link |
So car sharing is autonomous.
link |
It creates a similar or better experience in terms of how quickly you get from A to B
link |
or just whatever, the pleasantness of the experience, the efficiency of the experience,
link |
the value of the experience, and at the same time, the car itself costs cheaper.
link |
I think that's very difficult to achieve.
link |
I think there's a lot more low hanging fruit in the home.
link |
I also want to give you a perspective on how challenging it would be at home or maybe kind
link |
of depends on the exact problem that you'd be solving.
link |
If we're speaking about these robotic arms, and hence, these things, they cost tens of
link |
thousands of dollars or maybe 100K.
link |
And maybe, obviously, maybe there would be economy of scale, these things would be cheaper.
link |
But actually, for any household to buy it, the price would have to go down to maybe 1,000
link |
I personally think that so self driving car provides a clear service.
link |
I don't think robots in the home, there'll be a trillion dollar company will just be
link |
all about service, meaning it will not necessarily be about like a robotic arm.
link |
That helps you, I don't know, open a bottle or wash dishes or any of that kind of stuff.
link |
It has to be able to take care of that whole, the therapist thing you mentioned.
link |
I think that's, of course, there's a line between what is a robot and what is not.
link |
Like, does it really need a body?
link |
But AI system with some embodiment, I think.
link |
So the tricky part when you think actually what's the difficult part is when the robot
link |
has, like when there is a diversity of the environment with which the robot has to interact,
link |
that becomes hard.
link |
So, you know, on one spectrum, you have industrial robots, as they are doing over and over the
link |
same thing, it is possible to some extent to prescribe the movements.
link |
And with very small amount of intelligence, the movement can be repeated millions of times.
link |
And there are also, you know, various pieces of industrial robots where it becomes harder
link |
Like, for instance, in case of Tesla, maybe a matter of putting a rack inside of a car.
link |
And you know, because the rack kind of moves their own, it's not that easy, it's not exactly
link |
the same every time.
link |
It ends up being the case that you need actually humans to do it.
link |
While, you know, welding cars together, it's a very repetitive process.
link |
And in case of self driving itself, the difficulty has to do with the diversity of the environment.
link |
But still the car itself, the problem that you are solving is you try to avoid even interacting
link |
You are not touching anything around, because touching itself is hard.
link |
And then if you would have in the home robot that, you know, has to touch things and like
link |
if these things, they change the shape, if there is a huge variety of things to be touched,
link |
then that's difficult.
link |
If you are speaking about the robot, which there is, you know, head, that it's smiling
link |
in some way with cameras, that it doesn't, you know, touch things, that's relatively
link |
So to both agree and to push back.
link |
So you're referring to touch like soft robotics, like the actual touch.
link |
But I would argue that you could formulate just basic interaction between like non contact
link |
interaction is also a kind of touch.
link |
And that might be very difficult to solve.
link |
That's the basic this, not disagreement, but that's the basic open question to me with
link |
self driving cars and disagreement with Elon, which is how much interaction is required
link |
to solve self driving cars, how much touch is required.
link |
You said that in your intuition, touch is not required in my intuition to create a product
link |
that's compelling to use, you're going to have to interact with pedestrians, not just
link |
avoid pedestrians, but interact with them.
link |
When we drive around in major cities, we're constantly threatening everybody's life with
link |
And that's how they respect us.
link |
There's a game theory going on with pedestrians.
link |
And I'm afraid you can't just formulate autonomous driving as a collision avoidance problem.
link |
So I think it goes beyond like a collision avoidance is the first order approximation.
link |
But then at least in case of Tesla, they are gathering data from people driving their cars.
link |
And I believe that's an example of supervised learning data that they can train their models
link |
on, and they are doing it, which can give a model this like another level of behavior
link |
that is needed to actually interact with the real world.
link |
Yeah, it's interesting how much data is required to achieve that.
link |
What do you think of the whole Tesla autopilot approach, the computer vision based approach
link |
with multiple cameras and as a data engine, it's a multi task, multi headed neural network,
link |
and it's this fascinating process of similar to what you're talking about with the robotics
link |
approach, which is you deploy a neural network and then there's humans that use it and then
link |
it runs into trouble in a bunch of places and that stuff is sent back.
link |
So the deployment discovers a bunch of edge cases and those edge cases are sent back for
link |
supervised annotation, thereby improving the neural network and that's deployed again.
link |
It goes over and over until the network becomes really good at the task of driving, becomes
link |
What do you think of that kind of approach to robotics?
link |
I believe that's the way to go.
link |
So in some sense, even when I was speaking about collecting trajectories from humans,
link |
that's like a first step and then you deploy the system and then you have humans revising
link |
all the issues and in some sense, like at this approach converges to system that doesn't
link |
make mistakes because for the cases where there are mistakes, you gather data how to
link |
fix them and the system will keep on improving.
link |
So there's a very, to me, difficult question of how long that converging takes, how hard
link |
The other aspect of autonomous vehicles probably applies to certain robotics applications
link |
As the quality of the system converges, so one, there's a human factors perspective of
link |
psychology of humans being able to supervise those, even with teleoperation, those robots.
link |
And the other is society willing to accept robots.
link |
Currently society is much harsher on self driving cars than it is on human driven cars
link |
in terms of the expectation of safety.
link |
So the bar is set much higher than for humans.
link |
So if there's a death in an autonomous vehicle that's seen as much more dramatic than a death
link |
in a human driven vehicle.
link |
Part of the success of deployment of robots is figuring out how to make robots part of
link |
society, both on the just the human side, on the media journalist side and also on the
link |
policy government side.
link |
And that seems to be, maybe you can put that into the objective function to optimize.
link |
And that is definitely a tricky one.
link |
And I wonder if that is actually the trickiest part for self driving cars or any system that's
link |
It's not the algorithm, it's the society accepting it.
link |
Yeah, I would say, I believe that the part of the process of deployment is actually showing
link |
people that the given things can be trusted.
link |
And trust is also like a glass that is actually really easy to crack it and damage it.
link |
And I think that's actually very common with innovation, that there is some resistance
link |
And it's just a natural progression.
link |
So in some sense, people will have to keep on proving that indeed these systems are worth
link |
And I would say, I also found out that often the best way to convince people is by letting
link |
them experience it.
link |
That's the case with Tesla autopilot, for example.
link |
That's the case with, yeah, with basically robots in general.
link |
It's kind of funny to hear people talk about robots like there's a lot of fear, even like
link |
But when they actually interact with them, there's joy.
link |
I love interacting with them.
link |
And the same with the car.
link |
With a robot, if it starts being useful, I think people immediately understand.
link |
And if the product is designed well, they fall in love.
link |
It's actually even similar when I'm thinking about Copilot, the GitHub Copilot.
link |
There was a spectrum of responses that people had.
link |
And ultimately, the important piece was to let people try it out.
link |
And then many people just loved it.
link |
Yeah, programmers.
link |
But like some of them, you know, they came with a fear.
link |
But then you try it out and you think, actually, that's cool.
link |
And you know, you can try to resist the same way as, you know, you could resist moving
link |
from punch cards to, let's say, C++ or so.
link |
And it's a little bit futile.
link |
So we talked about generation of program, generation of language, even self supervised
link |
learning in the visual space for robotics and then reinforcement learning.
link |
What do you and like this whole beautiful spectrum of AI, do you think is a good benchmark,
link |
a good test to strive for, to achieve intelligence?
link |
That's a strong test of intelligence.
link |
You know, it started with Alan Turing and the Turing test.
link |
Maybe you think natural language conversation is a good test.
link |
So you know, it would be nice if for instance, machine would be able to solve Riemann hypothesis
link |
That would be, I think that would be very impressive.
link |
So theorem proving, is that to you, proving theorems is a good, oh, like one thing that
link |
the machine did, you would say, damn.
link |
That would be quite, quite impressive.
link |
I mean, the tricky part about the benchmarks is, you know, as we are getting closer with
link |
them, we have to invent new benchmarks.
link |
There is actually no ultimate benchmark out there.
link |
See, my thought with the Riemann hypothesis would be the moment the machine proves it,
link |
would say, okay, well, then the problem was easy.
link |
That's what happens.
link |
And I mean, in some sense, that's actually what happens over there in AI that like we
link |
get used to things very quickly.
link |
You know something I talked to Rodney Brooks, I don't know if you know that is, he called
link |
AlphaZero homework problem, because he was saying like, there's nothing special about
link |
It's not a big leap.
link |
And I didn't, well, he's coming from one of the aspects that we referred to is he was
link |
part of the founding of iRobot, which deployed now tens of millions of robot in the home.
link |
So if you see robots that are actually in the homes of people as the legitimate instantiation
link |
of artificial intelligence, then yes, maybe an AI that plays a silly game like go and
link |
chess is not a real accomplishment.
link |
But to me, it's a fundamental leap.
link |
But I think we as humans then say, okay, well, then that that game of chess or go wasn't
link |
that difficult compared to the thing that's currently unsolved.
link |
So my intuition is that from perspective of the evolution of these AI systems, we'll
link |
at first see the tremendous progress in digital space, and the main thing about digital space
link |
is also that you can, everything is, there is a lot of recorded data, plus you can very
link |
rapidly deploy things to billions of people.
link |
While in case of physical space, the deployment part takes multiple years.
link |
You have to manufacture things and, you know, delivering it to actual people is very hard.
link |
So I'm expecting that the first and the prices in digital space of goods they would go down
link |
to, let's say, marginal costs are two zero.
link |
And also the question is how much of our life will be in digital because it seems like we're
link |
heading towards more and more of our lives being in the digital space.
link |
So like innovation in the physical space might become less and less significant.
link |
Like why do you need to drive anywhere if most of your life is spent in virtual reality?
link |
I still would like, you know, to, at least at the moment, my impression is that I would
link |
like to have a physical contact with other people and that's very important to me.
link |
We don't have a way to replicate it in the computer.
link |
It might be the case that over the time it will change.
link |
Like in 10 years from now, why not have like an arbitrary infinite number of people you
link |
can interact with, some of them are real, some are not, with arbitrary characteristics
link |
that you can define based on your own preferences.
link |
I think that's maybe where we are heading and maybe I'm resisting the future.
link |
I'm telling you, if I got to choose, if I could live in Elder Scroll Skyrim versus the
link |
real world, I'm not so sure I would stay with the real world.
link |
Yeah, I mean, the question is, will VR be sufficient to get us there or do you need
link |
to, you know, plug electrodes in the brain and it would be nice if these electrodes
link |
wouldn't be invasive?
link |
Or at least like provably nondestructive.
link |
But in a digital space, do you think we'll be able to solve the Turing test, the spirit
link |
of the Turing test, which is do you think we'll be able to achieve compelling natural
link |
language conversation between people, like have friends that are AI systems on the Internet?
link |
I totally think it's doable.
link |
Do you think the current approach to GPT will take us there?
link |
So there is, you know, the part of at first learning all the content out there and I think
link |
that Steel System should keep on learning as it speaks with you.
link |
And I think that should work.
link |
The question is how exactly to do it and, you know, obviously we have people at the
link |
open air asking these questions and kind of at first pre training on all existing content
link |
is like a backbone and is a decent backbone.
link |
Do you think AI needs a body connecting to our robotics question to truly connect with
link |
humans or can most of the connection be in the digital space?
link |
We know that there are people who met each other online and they felt in love.
link |
So it seems that it's conceivable to establish connection, which is purely through Internet.
link |
Of course, it might be more compelling, the more modalities you add.
link |
So it would be like you're proposing like a Tinder, but for AI, are you like swipe right
link |
left and half the systems are AI and the other is humans and you don't know which is which?
link |
That would be our formulation of Turing test.
link |
The moment AI is able to achieve more swipe right or left, whatever, the moment it's able
link |
to be more attractive than other humans, it passes the Turing test.
link |
Then you would pass the Turing test in attractiveness.
link |
Well, no, like attractiveness just to clarify.
link |
There will be conversation.
link |
Not just visual, right?
link |
It's also attractiveness with wit and humor and whatever makes conversation pleasant for
link |
So you're saying it's possible to achieve in the digital space.
link |
In some sense, I would almost ask the question, why wouldn't that be possible?
link |
Well, I have this argument with my dad all the time.
link |
He thinks that touch and smell are really important.
link |
So they can be very important.
link |
And I'm saying the initial systems, they won't have it.
link |
Still I wouldn't, like there are people being born without these senses.
link |
And you know, I believe that they can still fall in love and have meaningful life.
link |
I wonder if it's possible to go close to all the way by just training on transcripts
link |
Like, I wonder how far that takes us.
link |
So I think that actually still you want images, I would like.
link |
So I don't have kids, but I could imagine having AI tutor, it has to see kids drawing
link |
some pictures on the paper.
link |
And also facial expressions, all that kind of stuff.
link |
We use dogs and humans use their eyes to communicate with each other.
link |
I think that's a really powerful mechanism of communication.
link |
Body language too, that words are much lower bandwidth.
link |
And for body language, we still, you know, we can have a system that displays an image
link |
of its artificial expression on the computer.
link |
It doesn't have to move, you know, mechanical pieces or so.
link |
So I think that, you know, that there is like kind of a progression.
link |
You can imagine that text might be the simplest to tackle, but this is not a complete human
link |
experience at all.
link |
You expand it to let's say images both for input and output.
link |
And what you describe is actually the final, I guess, frontier.
link |
What makes us human, the fact that we can touch each other or smell or so.
link |
And it's the hardest from perspective of data and deployment.
link |
And I believe that these things might happen gradually.
link |
Are you excited by that possibility?
link |
This particular application of human to AI friendship and interaction.
link |
So let's see, like would you, do you look forward to a world you said you're living
link |
with a few folks and you're very close friends with them?
link |
Do you look forward to a day where one or two of those friends are AI systems?
link |
So if the system would be truly wishing me well, rather than being in the situation that
link |
it optimizes for my time to interact with the system.
link |
The line between those is, it's a gray, it's a gray area.
link |
I think that's the distinction between love and possession.
link |
And these things, they might be often correlated for humans, but you might find that there
link |
are some friends with whom you haven't spoke for months.
link |
And then, you know, you pick up the phone, it's as the time hasn't passed.
link |
They are not holding to you.
link |
And I wouldn't like to have AI system that, you know, it's trying to convince me to spend
link |
I would like the system to optimize for what I care about and help me in achieving my own
link |
But there's some, I mean, I don't know, there's some manipulation, there's some possessiveness,
link |
there's some insecurities, there's fragility, all those things are necessary to form a close
link |
friendship over time, to go through some dark shit together, some bliss and happiness together.
link |
I feel like there's a lot of greedy self center behavior within that process.
link |
My intuition, but I might be wrong, is that human computer interaction doesn't have to
link |
go through computer being greedy, possessive and so on.
link |
It is possible to train systems, maybe, that they actually, you know, they are, I guess,
link |
prompted or fine tuned or so to truly optimize for what you care about.
link |
And you could imagine that, you know, the way how the process would look like is at
link |
some point, we as humans, we look at the transcript of the conversation or like an entire interaction
link |
and we say, actually, here there was more loving way to go about it.
link |
Maybe supervise system toward being more loving or maybe we train the system such that it
link |
has a reward function toward being more loving.
link |
Or maybe the possibility of the system being an asshole and manipulative and possessive
link |
every once in a while is a feature, not a bug.
link |
Because some of the happiness that we experience when two souls meet each other, when two humans
link |
meet each other, is a kind of break from the assholes in the world.
link |
And so you need assholes in AI as well, because like, it'll be like a breath of fresh air
link |
to discover in AI that the three previous AI's you had are too friendly or cruel or whatever.
link |
It's like some kind of mix and then this one is just right, but you need to experience
link |
the full spectrum.
link |
Like, I think you need to be able to engineer assholes.
link |
Because there's some level to us being appreciated, to appreciate the human experience.
link |
We need the dark and the light.
link |
So that kind of reminds me.
link |
I met a while ago at the meditation retreat, one woman and a beautiful, beautiful woman
link |
and she had a crush, she had the trouble of walking on one leg.
link |
I asked her what has happened and she said that five years ago she was in Maui, Hawaii
link |
and she was eating a salad and some snail fell into the salad and apparently there
link |
are neurotoxic snails over there and she got into coma for a year.
link |
And apparently there is a high chance of even just dying, but she was in the coma.
link |
At some point she regained partially consciousness.
link |
She was able to hear people in the room.
link |
People behave as she wouldn't be there.
link |
At some point she started being able to speak, but she was mumbling like barely able to express
link |
Then at some point she got into wheelchair.
link |
And at some point she actually noticed that she can move her toe and then she knew that
link |
she will be able to walk.
link |
And then that's where she was five years after and she said that since then she appreciates
link |
the fact that she can move her toe.
link |
And I was thinking, do I need to go through such experience to appreciate that I can move
link |
Wow, that's a really good story, a really deep example, yeah.
link |
And in some sense it might be the case that we don't see light if we haven't went through
link |
the darkness, but I wouldn't say that we shouldn't.
link |
We shouldn't assume that that's the case, we may be able to engineer shortcuts.
link |
Yeah, Ilya had this belief that maybe one has to go for a week or six months to some
link |
challenging camp to just experience a lot of difficulties and then comes back.
link |
And actually everything is bright, everything is beautiful.
link |
I'm with Ilya, it must be a Russian thing, where are you from originally?
link |
I'm tempted to say that explains a lot, but yeah, there's something about the Russian,
link |
the necessity of suffering.
link |
I believe suffering or rather struggle is necessary.
link |
I believe that struggle is necessary, I mean in some sense you even look at the story of
link |
any superhero in the movie, it's not that it was like I ever forgot it goes easy, easy,
link |
I like how that's your ground truth, is the story of superheroes, okay.
link |
You mentioned that you used to do research at night and go to bed at like 6am or 7am,
link |
I still do that often.
link |
What sleep schedules have you tried to make for a productive and happy life?
link |
Is there some interesting wild sleeping patterns that you engaged that you found that works
link |
really well for you?
link |
I tried at some point decreasing number of hours of sleep, like a gradually, like a half
link |
an hour every few days to this, I was hoping to just save time.
link |
That clearly didn't work for me, at some point there's like a face shift and I felt tired
link |
There was a time that I used to work during the nights, the nice thing about the nights
link |
is that no one disturbs you and even I remember when I was meeting for the first time with
link |
Greg Brockman, his CTO and chairman of OpenAI, our meeting was scheduled to 5pm and I overstepped
link |
Over slept for the meeting at 5pm, yeah.
link |
Now you sound like me, that's hilarious, okay, yeah.
link |
At the moment in some sense, my sleeping schedule also has to do with the fact that I'm interacting
link |
with people, I sleep without an alarm.
link |
So yeah, the team thing, you mentioned the extrovert thing because most humans operate
link |
during a certain set of hours, you're forced to then operate at the same set of hours.
link |
But I'm not quite there yet, I found a lot of joy just like you said, working through
link |
the night because it's quiet, because the world doesn't disturb you and there's some
link |
aspect counter to everything you're saying, there's some joyful aspect to sleeping through
link |
the mess of the day because people are having meetings and sending emails and there's drama,
link |
I can sleep through all the meetings.
link |
You know, I have meetings every day and they prevent me from having sufficient amount of
link |
time for focus work.
link |
And then I modified my calendar and I said that I'm out of office Wednesday, Thursday
link |
and Friday every day and I'm having meetings only Monday and Tuesday and that was the positively
link |
influence my mood that I have literally like three days for fully focused work.
link |
So there's better solutions to this problem than staying awake all night.
link |
Okay, you've been part of development of some of the greatest ideas in artificial intelligence.
link |
What would you say is your process for developing good novel ideas?
link |
You have to be aware that clearly there are many other brilliant people around.
link |
So you have to ask yourself a question, why they give an idea, let's say wasn't tried
link |
by someone else and in some sense it has to do with, you know, kind of simple, it might
link |
sound simple, but like a thinking outside of the box and what do I mean here?
link |
So for instance, for a while, people in academia, they assumed that you have a fixed data set
link |
and then you optimize the algorithms in order to get the best performance.
link |
And that was so in great assumption that no one thought about training models on anti
link |
internet or like that, maybe some people thought about it, but it felt too many as unfair.
link |
And in some sense, that's almost like a, it's not my idea or so, but that's an example
link |
of breaking a typical assumption.
link |
So you want to be in the paradigm that you're breaking a typical assumption.
link |
In the context of the AI community, getting to pick your data set is cheating.
link |
And in some sense, so that was a, that was assumption that many people had out there.
link |
And then if you free yourself from assumptions, then they are likely to achieve something
link |
that others cannot do.
link |
And in some sense, if you are trying to do exactly the same things as others, it's very
link |
likely that you're going to have the same results.
link |
I, but there's also that kind of tension, which is asking yourself the question, why
link |
haven't others done this?
link |
Because I mean, I get a lot of good ideas, but I think probably most of them suck when
link |
they meet reality.
link |
So actually, I think the other big piece is getting into habit of generating ideas, training
link |
your brain towards generating ideas and not even suspending judgment of the ideas.
link |
So in some sense, I noticed myself that even if I'm in the process of generating ideas,
link |
if I tell myself, oh, that was a bad idea, then that actually interrupts the process
link |
and I cannot generate more ideas because I'm actually focused on the negative part, why
link |
But I created also environment in the way that it's very easy for me to store new ideas.
link |
So for instance, next to my bed, I have a voice recorder.
link |
And it happens to me often, like I wake up during the night and I have some idea.
link |
In the past, I was writing them down on my phone, but that means turning on the screen
link |
and that wakes me up or like pulling a paper, which requires turning on the light.
link |
These days, I just start recording it.
link |
What do you think?
link |
I don't know if you know who Jim Keller is.
link |
I know Jim Keller.
link |
He's a big proponent of thinking hard on a problem right before sleep so that he can
link |
sleep through it and solve it in his sleep or like come up with radical stuff in his
link |
He was trying to get me to do this.
link |
So it happened from my experience perspective.
link |
It happened to me many times during the high school days when I was doing mathematics that
link |
I had the solution to my problem as I woke up.
link |
At the moment, regarding thinking hard about the given problem is I'm trying to actually
link |
devote substantial amount of time to think about important problems, not just before
link |
I'm organizing the huge chunks of time such that I'm not constantly working on the urgent
link |
problems, but I actually have time to think about the important one.
link |
So you do it naturally, but his idea is that you prime your brain to make sure that that's
link |
Oftentimes, people have other worries in their life that's not fundamentally deep problems
link |
like I don't know, just stupid drama in your life and even at work, all that kind of stuff.
link |
He wants to pick the most important problem that you're thinking about and go to bed
link |
I think that's why I mean, the other thing that comes to my mind is also I feel the most
link |
fresh in the morning.
link |
So during the morning, I try to work on the most important things rather than just being
link |
pulled by urgent things or checking email or so.
link |
What do you do with the because I've been doing the voice recorder thing too, but I end up
link |
recording so many messages is hard to organize.
link |
I have the same problem.
link |
Now I have heard that Google Pixel is really good in transcribing text and I might get
link |
Google Pixel just for the sake of transcribing text.
link |
People listening to this, if you have a good voice recorder suggestion that transcribes,
link |
please let me know.
link |
Some of it is this has to do with the open AI codex too.
link |
Some of it is simply like the friction.
link |
I need apps that remove that friction between voice and the organization of the resulting
link |
transcripts and all that kind of stuff.
link |
But yes, you're right.
link |
Like during for me is walking, sleep too, but walking and running, especially running
link |
get a lot of thoughts during running and there's no good mechanism for recording thoughts.
link |
So one more thing that I do, I have a separate phone which has no apps.
link |
Maybe it's like audible or let's say Kindle.
link |
No one has this phone number, this kind of my meditation phone.
link |
And I try to expand the amount of time that that's the phone that I'm having.
link |
It has also Google Maps if I need to go somewhere.
link |
And I also use this phone to write down ideas.
link |
That's a really good idea.
link |
Often actually what I end up doing is even sending a message from that phone to the other
link |
So that's actually my way of recording messages or I just put them into notes.
link |
What advice would you give to a young person, high school, college, about how to be successful?
link |
You've done a lot of incredible things in the past decade.
link |
So maybe you have some...
link |
There might be something.
link |
There might be something.
link |
I mean, it might sound simplistic or so, but I would say literally just follow your passion
link |
double down on it.
link |
And if you don't know what's your passion, just figure out what could be a passion.
link |
So the step might be an exploration.
link |
When I was in elementary school, it was math and chemistry.
link |
And I remember for some time I gave up on math because my school teacher, she told me
link |
And I guess maybe an advice would be just ignore people if they tell you that you're
link |
You mentioned something offline about chemistry and explosives.
link |
What was that about?
link |
So a story goes like that.
link |
I got into chemistry.
link |
Maybe I was like a second grade of my elementary school, third grade.
link |
I started going to chemistry classes.
link |
I really love building stuff.
link |
And I did all the experiments that they describe in the book, like how to create oxygen with
link |
vinegar and baking soda or so.
link |
So I did all the experiments.
link |
And at some point, I was like, so what's next?
link |
And the explosives, they also you have a clear reward signal if the thing worked or not.
link |
So I remember at first I got interested in producing hydrogen.
link |
That was kind of a funny experiment from school.
link |
You can just burn it.
link |
And then I moved to nitroglycerin.
link |
So that's also relatively easy to synthesize.
link |
I started producing essentially dynamite and detonating it with a friend.
link |
Remember, there was at first like maybe two attempts that I went with a friend to detonate
link |
And it didn't work out.
link |
And like a third time, he was like, ah, it won't work, like let's don't waste time.
link |
And now we were, I was carrying this tube with dynamite, I don't know, pound or so dynamite
link |
in my backpack were like riding on the bike to the edges of the city.
link |
And attempt number three, this was be attempt number three, attempt number three.
link |
And now we we dig a hole to put it inside.
link |
It actually had the, you know, electrical detonator.
link |
We draw a cable behind the tree.
link |
I even, I never had, I haven't ever seen like a explosion before.
link |
So I thought that there will be a lot of sound.
link |
But you know, we're like laying down and I'm holding the cable and the battery at some
link |
point, you know, we kind of like a three to one.
link |
And I just connected it and it felt like at the ground shake, it was like a more like
link |
And then the soil started kind of lifting up and started falling on us.
link |
And then now the friend said, let's, let's make sure next time we have helmets, but it's
link |
also, you know, I'm happy that nothing happened to me.
link |
It could have been the case that I lost the limb or so.
link |
But that's childhood of an engineering mind with a strong reward signal of an explosion.
link |
And my, there's some aspect of a chemist, the chemist, I know like my dad with plasma
link |
chemistry, plasma physics, he was very much into explosives too.
link |
It's a worrying quality of people that work in chemistry that they love.
link |
I think it is that exactly is the, the strong signal that the thing worked.
link |
There is no doubt.
link |
There's some magic.
link |
It's almost like a reminder that physics works, that chemistry works.
link |
It's almost like a little glimpse at nature that you yourself engineer.
link |
That's why I really like artificial intelligence, especially robotics, is you create a little
link |
And in some sense, even for me with explosives, the motivation was creation rather than distraction.
link |
In terms of advice, I forgot to ask about just machine learning and deep learning for
link |
people who are specifically interested in machine learning.
link |
How would you recommend it get into the field?
link |
So I would say re implement everything.
link |
And also there is plenty of courses for like from scratch.
link |
So on different levels of abstraction in some sense, but I would say re implement something
link |
from scratch, re implement something from a paper, re implement something, you know,
link |
from podcasts that you have heard about, I would say that's a powerful way to understand
link |
So it's often the case that you read the description and you think you understand, but you truly
link |
understand once you build it, then you actually know what really mattered in the description.
link |
Is there a particular topics that you find people just fall in love with?
link |
I've seen, I tend to really enjoy reinforcement learning because it's much more, it's much
link |
easier to get to a point where you feel like you created something special, like fun games
link |
Yeah, as opposed to like re implementing from scratch, more like supervised learning kind
link |
So, you know, if someone would optimize for things to be rewarding, then it feels that
link |
the things that are somewhat generative, they have such a property.
link |
So you have, for instance, adversarial networks, or you have just even generative language
link |
And you can even see, internally, we have seen this thing with our releases.
link |
So we have, we released the recent two models.
link |
There is one model called Dali that generates images.
link |
And there is other model called Clip that actually you provide various possibilities,
link |
what could be the answer to what is on the picture.
link |
And it can tell you which one is the most likely.
link |
And in some sense, in case of the first one, Dali, it is very easy for you to understand
link |
that actually there is magic going on.
link |
And in the case of the second one, even though it is insanely powerful, and you know, people
link |
from a vision community, they, as they started probing it inside, they actually understood
link |
It's difficult for a person at first to see how well it works.
link |
And that's the same as you said, that in case of supervised learning models, you might not
link |
kind of see, or it's not that easy for you to understand the strength.
link |
Even though you don't believe in magic, to see the magic, the generative, that's really
link |
So anything that's generative, because then you are at the core of the creation, you get
link |
to experience creation without much effort, unless you have to do it from scratch.
link |
And it feels that, you know, humans are wired, there is some level of reward for creating
link |
Of course, different people have a different weight on this reward.
link |
In the big objective function of a person.
link |
You wrote that beautiful is what you intensely pay attention to.
link |
Even a cockroach is beautiful if you look very closely.
link |
Can you expand on this?
link |
So what I wrote here actually corresponds to my subjective experience that I had through
link |
extended periods of meditation.
link |
It's pretty crazy that at some point, the meditation gets you to the place that you
link |
have really increased focus, increased attention, and then you look at the very simple objects
link |
that were all the time around you, and look at the table or on the pen, or at the nature.
link |
And you notice more and more details, and it becomes very pleasant to look at it.
link |
And once again, it kind of reminds me my childhood, like a just pure joy of being.
link |
It's also, I have seen even the reverse effect that by default, regardless of what we possess,
link |
we very quickly get used to it.
link |
And you know, you can have a very beautiful house.
link |
And if you don't put sufficient effort, you're just going to get used to it.
link |
And it doesn't bring any more joy regardless of what you have.
link |
Well, I actually, I find that material possessions get in the way of that experience of pure joy.
link |
So I've always, I've been very fortunate to just find joy in simple things, just like
link |
you're saying, just like, I don't know, objects in my life, just stupid objects like this
link |
cup, like thing, you know, just objects sounds, okay, I'm not being eloquent, but literally
link |
objects in the world, they're just full of joy, because it's like, I can't believe,
link |
one I can't believe that I'm fortunate enough to be alive to experience these objects.
link |
And then two, I can't believe humans are clever enough to build these objects.
link |
The hierarchy of pleasure that that provides is infinite.
link |
I mean, even if you look at the cup of water, so you know, you see, first like a level of
link |
like a reflection of light, but then you think, no man, there's like a trillions upon trillions
link |
of particles bouncing against each other.
link |
There is also the tension on the surface that, you know, if the back, back could like a stand
link |
on it and move around.
link |
And you think it also has this like a magical property that as you decrease temperature,
link |
it actually expands in volume, which allows for the, you know, legs to freeze on the,
link |
on the surface and then at the bottom to have actually not freeze, which allows for life
link |
like a crazy, you look in detail at some object and you think actually, you know, this table,
link |
that was just the figment of someone's imagination at some point.
link |
And then there was like a thousands of people involved to actually manufacture it and put
link |
And by default, no one cares.
link |
And then you can start thinking about evolution, how it all started from single cell organisms
link |
that led to this table.
link |
And these thoughts, they give me life appreciation and even lack of thoughts, just the pure raw
link |
signal also gives the life appreciation.
link |
See the thing is, and then that's coupled for me with the sadness that the whole ride
link |
ends and perhaps is deeply coupled in that the fact that this experience, this moment
link |
ends gives it gives it an intensity that I'm not sure I would otherwise have.
link |
So in that same way, I tried to meditate on my own death often.
link |
Do you think about your mortality?
link |
Are you afraid of death?
link |
So fear of death is like one of the most fundamental fears that each of us has.
link |
We might be not even aware of it.
link |
It requires to look inside to even recognize that it's out there.
link |
There is still, let's say, this property of nature that if things would last forever,
link |
then they would be also boring to us.
link |
The fact that the things change in some way gives any meaning to them.
link |
I also, you know, found out that it seems to be very healing to people to have these
link |
short experiences, like I guess, psychedelic experiences in which they experience death
link |
of self, in which they let go of this fear and then maybe can even increase the appreciation
link |
It seems that many people, they can easily comprehend the fact that their money is finite
link |
while they don't see that time is finite.
link |
I have this discussion with Ilya from time to time.
link |
He's like, you know, man, life will pass very fast.
link |
At some point, I will be 40, 50, 60, 70, and then it's over.
link |
This is true, which also makes me believe that every single moment, it is so unique
link |
that should be appreciated.
link |
And it also makes me think that I should be acting on my life, because otherwise it will
link |
I also like this framework of thinking from Jeff Bezos on regret minimization, that like
link |
I would like if I will be at the death bed to look back on my life and not regret that
link |
I haven't done something.
link |
It's usually you might regret that you haven't tried.
link |
I'm fine with failing, but I haven't tried.
link |
What's the need to turn over currents, try to live a life that if you had to live it
link |
infinitely many times, that would be the you'd be OK with that kind of life.
link |
So try to live it optimally.
link |
I can say that it's almost like unavailable to me where I am in my life.
link |
I'm extremely grateful for actually people whom I met.
link |
I would say I think that I'm decently smart and so on.
link |
But I think that actually to a great extent where I am has to do with the people who I
link |
Would you be OK if after this conversation you died?
link |
So if I'm dead, then it kind of I don't have a choice anymore.
link |
So in some sense, there's like a plenty of things that I would like to try out in my
link |
I feel that I'm gradually going one by one and I'm just doing them.
link |
I think that the list will be always infinite.
link |
So might as well go today.
link |
I mean, to be clear, I'm not looking forward to dying.
link |
I would say if there is no choice, I would accept it.
link |
But like in some sense, if there would be a choice, if there would be possibility to
link |
live, I would fight for a living.
link |
I find it's more honest and real to think about dying today at the end of the day.
link |
That seems to me, at least to my brain, more honest slap in the face as opposed to I still
link |
have 10 years today, then I'm much more about appreciating the cup and the table and so
link |
on and less about silly, worldly accomplishments and all those kinds of things.
link |
We have in the company, a person who at some point found out that they have cancer.
link |
And that also gives huge perspective with respect to what matters now.
link |
And often people in situations like that, they conclude that actually what matters is
link |
human connection and love.
link |
And people conclude also if you have kids, kids, family, you I think tweeted, we don't
link |
assign the minus infinity reward to our death.
link |
Such a reward would prevent us from taking any risk.
link |
We wouldn't be able to cross the road in fear of being hit by a car.
link |
So in the objective function, you mentioned fear of death might be fundamental to the
link |
So as I said, let's assume that they're like a reward functions in our brain.
link |
And the interesting thing is even realization how different reward functions can play with
link |
As a matter of fact, I wouldn't say that you should assign infinite negative reward to
link |
anything because that messes up the math.
link |
The math doesn't work out.
link |
It doesn't work out.
link |
As you said, even government or some insurance companies, they assign $9 million to human
link |
And I'm just saying it with respect to, that might be a hard statement to ourselves, but
link |
in some sense that there's a finite value of our own life.
link |
I'm trying to put it from perspective of being less or of being more ego less and realizing
link |
fragility of my own life.
link |
And in some sense, the fear of death might prevent you from acting because anything can
link |
And I'm sure actually if you were to put death in the objective function, there's probably
link |
so many aspects to death and fear of death and realization of death and mortality.
link |
There's just whole components of finiteness of not just your life, but every experience
link |
and so on that you're going to have to formalize mathematically.
link |
And also, that might lead to you spending a lot of compute cycles on this like a deliberating
link |
this terrible future instead of experiencing now.
link |
But in some sense, it's also kind of unpleasant simulation to run in your head.
link |
Do you think there's an objective function that describes the entirety of human life?
link |
So usually the way you ask that is what is the meaning of life?
link |
Is there a universal objective function that captures the why of life?
link |
So yeah, I mean, I suspect that they will ask this question, but it's also a question
link |
that I ask myself many, many times.
link |
See, I can tell you a framework that I have these days to think about this question.
link |
So I think that fundamentally meaning of life has to do with some of our reward functions
link |
that we have in brain and they might have to do with, let's say, for instance, curiosity
link |
or human connection, which might mean understanding others.
link |
It's also possible for a person to slightly modify their reward function, usually they
link |
mostly stay fixed, but it's possible to modify reward function and you can pretty much choose.
link |
So in some sense, reward functions, optimizing reward functions, they will give you life satisfaction.
link |
Is there some randomness in the function?
link |
I think when you are born, there is some randomness like you can see that some people, for instance,
link |
they care more about building stuff.
link |
Some people care more about caring for others.
link |
Some people, there are all sorts of default reward functions and then in some sense, you
link |
can ask yourself, was this the satisfying way for you to go after this reward function
link |
and you just go after this reward function and some people also ask, are these reward
link |
I almost think about it as, let's say, if you would have to discover mathematics, in mathematics
link |
you are likely to run into various objects, like complex numbers or differentiation, some
link |
other objects and these are very natural objects that arise and similarly, the reward functions
link |
that we are having in our brain, they are somewhat very natural that there is a reward
link |
function for understanding, like a comprehension, curiosity and so on.
link |
So in some sense, they are in the same way natural as their natural objects in mathematics.
link |
So you know there's the old sort of debate, is mathematics invented or discovered?
link |
You're saying reward functions are discovered.
link |
So nature provides…
link |
So nature provided some, you can still, let's say, expand it throughout the life.
link |
Some of the reward functions, they might be futile, like for instance, there might be
link |
a reward function, maximize amount of wealth.
link |
And this is more like a learning reward function.
link |
But we know also that some reward functions, if you optimize them, you won't be quite
link |
Well, I don't know which part of your reward function resulted in you coming today, but
link |
I am deeply appreciative that you did spend your valuable time with me.
link |
Wojciech is really fun talking to you.
link |
You're brilliant, you're a good human being and it's an honor to meet you and an honor
link |
Thanks for talking to me, brother.
link |
Thank you Lexelad, I appreciated your questions, I had a lot of time being here.
link |
Thanks for listening to this conversation with Wojciech Saremba.
link |
To support this podcast, please check out our sponsors in the description.
link |
And now, let me leave you with some words from Arthur C. Clarke, who is the author of
link |
2001 A Space Odyssey.
link |
It may be that our role on this planet is not to worship God, but to create Him.
link |
Thank you for listening and I hope to see you next time.