back to indexMax Tegmark: Life 3.0 | Lex Fridman Podcast #1
link |
As part of MIT course 6S099, Artificial General Intelligence,
link |
I've gotten the chance to sit down with Max Tegmark.
link |
He is a professor here at MIT.
link |
He's a physicist, spent a large part of his career
link |
studying the mysteries of our cosmological universe.
link |
But he's also studied and delved into the beneficial
link |
possibilities and the existential risks
link |
of artificial intelligence.
link |
Amongst many other things, he is the cofounder
link |
of the Future of Life Institute, author of two books,
link |
both of which I highly recommend.
link |
First, Our Mathematical Universe.
link |
Second is Life 3.0.
link |
He's truly an out of the box thinker and a fun personality,
link |
so I really enjoy talking to him.
link |
If you'd like to see more of these videos in the future,
link |
please subscribe and also click the little bell icon
link |
to make sure you don't miss any videos.
link |
Also, Twitter, LinkedIn, agi.mit.edu
link |
if you wanna watch other lectures
link |
or conversations like this one.
link |
Better yet, go read Max's book, Life 3.0.
link |
Chapter seven on goals is my favorite.
link |
It's really where philosophy and engineering come together
link |
and it opens with a quote by Dostoevsky.
link |
The mystery of human existence lies not in just staying alive
link |
but in finding something to live for.
link |
Lastly, I believe that every failure rewards us
link |
with an opportunity to learn
link |
and in that sense, I've been very fortunate
link |
to fail in so many new and exciting ways
link |
and this conversation was no different.
link |
I've learned about something called
link |
radio frequency interference, RFI, look it up.
link |
Apparently, music and conversations
link |
from local radio stations can bleed into the audio
link |
that you're recording in such a way
link |
that it almost completely ruins that audio.
link |
It's an exceptionally difficult sound source to remove.
link |
So, I've gotten the opportunity to learn
link |
how to avoid RFI in the future during recording sessions.
link |
I've also gotten the opportunity to learn
link |
how to use Adobe Audition and iZotope RX 6
link |
to do some noise, some audio repair.
link |
Of course, this is an exceptionally difficult noise
link |
I'm not an audio engineer.
link |
Neither is anybody else in our group
link |
but we did our best.
link |
Nevertheless, I thank you for your patience
link |
and I hope you're still able to enjoy this conversation.
link |
Do you think there's intelligent life
link |
out there in the universe?
link |
Let's open up with an easy question.
link |
I have a minority view here actually.
link |
When I give public lectures, I often ask for a show of hands
link |
who thinks there's intelligent life out there somewhere else
link |
and almost everyone put their hands up
link |
and when I ask why, they'll be like,
link |
oh, there's so many galaxies out there, there's gotta be.
link |
But I'm a numbers nerd, right?
link |
So when you look more carefully at it,
link |
it's not so clear at all.
link |
When we talk about our universe, first of all,
link |
we don't mean all of space.
link |
We actually mean, I don't know,
link |
you can throw me the universe if you want,
link |
it's behind you there.
link |
It's, we simply mean the spherical region of space
link |
from which light has a time to reach us so far
link |
during the 14.8 billion year,
link |
13.8 billion years since our Big Bang.
link |
There's more space here but this is what we call a universe
link |
because that's all we have access to.
link |
So is there intelligent life here
link |
that's gotten to the point of building telescopes
link |
My guess is no, actually.
link |
The probability of it happening on any given planet
link |
is some number we don't know what it is.
link |
And what we do know is that the number can't be super high
link |
because there's over a billion Earth like planets
link |
in the Milky Way galaxy alone,
link |
many of which are billions of years older than Earth.
link |
And aside from some UFO believers,
link |
there isn't much evidence
link |
that any superduran civilization has come here at all.
link |
And so that's the famous Fermi paradox, right?
link |
And then if you work the numbers,
link |
what you find is that if you have no clue
link |
what the probability is of getting life on a given planet,
link |
so it could be 10 to the minus 10, 10 to the minus 20,
link |
or 10 to the minus two, or any power of 10
link |
is sort of equally likely
link |
if you wanna be really open minded,
link |
that translates into it being equally likely
link |
that our nearest neighbor is 10 to the 16 meters away,
link |
10 to the 17 meters away, 10 to the 18.
link |
By the time you get much less than 10 to the 16 already,
link |
we pretty much know there is nothing else that close.
link |
And when you get beyond 10.
link |
Because they would have discovered us.
link |
Yeah, they would have been discovered as long ago,
link |
or if they're really close,
link |
we would have probably noted some engineering projects
link |
that they're doing.
link |
And if it's beyond 10 to the 26 meters,
link |
that's already outside of here.
link |
So my guess is actually that we are the only life in here
link |
that's gotten the point of building advanced tech,
link |
which I think is very,
link |
puts a lot of responsibility on our shoulders, not screw up.
link |
I think people who take for granted
link |
that it's okay for us to screw up,
link |
have an accidental nuclear war or go extinct somehow
link |
because there's a sort of Star Trek like situation out there
link |
where some other life forms are gonna come and bail us out
link |
and it doesn't matter as much.
link |
I think they're leveling us into a false sense of security.
link |
I think it's much more prudent to say,
link |
let's be really grateful
link |
for this amazing opportunity we've had
link |
and make the best of it just in case it is down to us.
link |
So from a physics perspective,
link |
do you think intelligent life,
link |
so it's unique from a sort of statistical view
link |
of the size of the universe,
link |
but from the basic matter of the universe,
link |
how difficult is it for intelligent life to come about?
link |
The kind of advanced tech building life
link |
is implied in your statement that it's really difficult
link |
to create something like a human species.
link |
Well, I think what we know is that going from no life
link |
to having life that can do a level of tech,
link |
there's some sort of two going beyond that
link |
than actually settling our whole universe with life.
link |
There's some major roadblock there,
link |
which is some great filter as it's sometimes called,
link |
which is tough to get through.
link |
It's either that roadblock is either behind us
link |
or in front of us.
link |
I'm hoping very much that it's behind us.
link |
I'm super excited every time we get a new report from NASA
link |
saying they failed to find any life on Mars.
link |
I'm like, yes, awesome.
link |
Because that suggests that the hard part,
link |
maybe it was getting the first ribosome
link |
or some very low level kind of stepping stone
link |
so that we're home free.
link |
Because if that's true,
link |
then the future is really only limited
link |
by our own imagination.
link |
It would be much suckier if it turns out
link |
that this level of life is kind of a dime a dozen,
link |
but maybe there's some other problem.
link |
Like as soon as a civilization gets advanced technology,
link |
within a hundred years,
link |
they get into some stupid fight with themselves and poof.
link |
That would be a bummer.
link |
Yeah, so you've explored the mysteries of the universe,
link |
the cosmological universe, the one that's sitting
link |
I think you've also begun to explore the other universe,
link |
which is sort of the mystery,
link |
the mysterious universe of the mind of intelligence,
link |
of intelligent life.
link |
So is there a common thread between your interest
link |
or the way you think about space and intelligence?
link |
Oh yeah, when I was a teenager,
link |
I was already very fascinated by the biggest questions.
link |
And I felt that the two biggest mysteries of all in science
link |
were our universe out there and our universe in here.
link |
So it's quite natural after having spent
link |
a quarter of a century on my career,
link |
thinking a lot about this one,
link |
that I'm now indulging in the luxury
link |
of doing research on this one.
link |
It's just so cool.
link |
I feel the time is ripe now
link |
for you trans greatly deepening our understanding of this.
link |
Just start exploring this one.
link |
Yeah, because I think a lot of people view intelligence
link |
as something mysterious that can only exist
link |
in biological organisms like us,
link |
and therefore dismiss all talk
link |
about artificial general intelligence as science fiction.
link |
But from my perspective as a physicist,
link |
I am a blob of quarks and electrons
link |
moving around in a certain pattern
link |
and processing information in certain ways.
link |
And this is also a blob of quarks and electrons.
link |
I'm not smarter than the water bottle
link |
because I'm made of different kinds of quarks.
link |
I'm made of up quarks and down quarks,
link |
exact same kind as this.
link |
There's no secret sauce, I think, in me.
link |
It's all about the pattern of the information processing.
link |
And this means that there's no law of physics
link |
saying that we can't create technology,
link |
which can help us by being incredibly intelligent
link |
and help us crack mysteries that we couldn't.
link |
In other words, I think we've really only seen
link |
the tip of the intelligence iceberg so far.
link |
Yeah, so the perceptronium.
link |
So you coined this amazing term.
link |
It's a hypothetical state of matter,
link |
sort of thinking from a physics perspective,
link |
what is the kind of matter that can help,
link |
as you're saying, subjective experience emerge,
link |
consciousness emerge.
link |
So how do you think about consciousness
link |
from this physics perspective?
link |
Very good question.
link |
So again, I think many people have underestimated
link |
our ability to make progress on this
link |
by convincing themselves it's hopeless
link |
because somehow we're missing some ingredient that we need.
link |
There's some new consciousness particle or whatever.
link |
I happen to think that we're not missing anything
link |
and that it's not the interesting thing
link |
about consciousness that gives us
link |
this amazing subjective experience of colors
link |
and sounds and emotions.
link |
It's rather something at the higher level
link |
about the patterns of information processing.
link |
And that's why I like to think about this idea
link |
What does it mean for an arbitrary physical system
link |
to be conscious in terms of what its particles are doing
link |
or its information is doing?
link |
I don't think, I hate carbon chauvinism,
link |
this attitude you have to be made of carbon atoms
link |
to be smart or conscious.
link |
There's something about the information processing
link |
that this kind of matter performs.
link |
Yeah, and you can see I have my favorite equations here
link |
describing various fundamental aspects of the world.
link |
I feel that I think one day,
link |
maybe someone who's watching this will come up
link |
with the equations that information processing
link |
has to satisfy to be conscious.
link |
I'm quite convinced there is big discovery
link |
to be made there because let's face it,
link |
we know that so many things are made up of information.
link |
We know that some information processing is conscious
link |
because we are conscious.
link |
But we also know that a lot of information processing
link |
Like most of the information processing happening
link |
in your brain right now is not conscious.
link |
There are like 10 megabytes per second coming in
link |
even just through your visual system.
link |
You're not conscious about your heartbeat regulation
link |
Even if I just ask you to like read what it says here,
link |
you look at it and then, oh, now you know what it said.
link |
But you're not aware of how the computation actually happened.
link |
Your consciousness is like the CEO
link |
that got an email at the end with the final answer.
link |
So what is it that makes a difference?
link |
I think that's both a great science mystery.
link |
We're actually studying it a little bit in my lab here
link |
at MIT, but I also think it's just a really urgent question
link |
For starters, I mean, if you're an emergency room doctor
link |
and you have an unresponsive patient coming in,
link |
wouldn't it be great if in addition to having
link |
a CT scanner, you had a consciousness scanner
link |
that could figure out whether this person
link |
is actually having locked in syndrome
link |
or is actually comatose.
link |
And in the future, imagine if we build robots
link |
or the machine that we can have really good conversations
link |
with, which I think is very likely to happen.
link |
Wouldn't you want to know if your home helper robot
link |
is actually experiencing anything or just like a zombie,
link |
I mean, would you prefer it?
link |
What would you prefer?
link |
Would you prefer that it's actually unconscious
link |
so that you don't have to feel guilty about switching it off
link |
or giving boring chores or what would you prefer?
link |
Well, certainly we would prefer,
link |
I would prefer the appearance of consciousness.
link |
But the question is whether the appearance of consciousness
link |
is different than consciousness itself.
link |
And sort of to ask that as a question,
link |
do you think we need to understand what consciousness is,
link |
solve the hard problem of consciousness
link |
in order to build something like an AGI system?
link |
No, I don't think that.
link |
And I think we will probably be able to build things
link |
even if we don't answer that question.
link |
But if we want to make sure that what happens
link |
is a good thing, we better solve it first.
link |
So it's a wonderful controversy you're raising there
link |
where you have basically three points of view
link |
about the hard problem.
link |
So there are two different points of view.
link |
They both conclude that the hard problem of consciousness
link |
On one hand, you have some people like Daniel Dennett
link |
who say that consciousness is just BS
link |
because consciousness is the same thing as intelligence.
link |
There's no difference.
link |
So anything which acts conscious is conscious,
link |
And then there are also a lot of people,
link |
including many top AI researchers I know,
link |
who say, oh, consciousness is just bullshit
link |
because, of course, machines can never be conscious.
link |
They're always going to be zombies.
link |
You never have to feel guilty about how you treat them.
link |
And then there's a third group of people,
link |
including Giulio Tononi, for example,
link |
and Krzysztof Koch and a number of others.
link |
I would put myself also in this middle camp
link |
who say that actually some information processing
link |
is conscious and some is not.
link |
So let's find the equation which can be used
link |
to determine which it is.
link |
And I think we've just been a little bit lazy,
link |
kind of running away from this problem for a long time.
link |
It's been almost taboo to even mention the C word
link |
in a lot of circles because,
link |
but we should stop making excuses.
link |
This is a science question and there are ways
link |
we can even test any theory that makes predictions for this.
link |
And coming back to this helper robot,
link |
I mean, so you said you'd want your helper robot
link |
to certainly act conscious and treat you,
link |
like have conversations with you and stuff.
link |
But wouldn't you, would you feel,
link |
would you feel a little bit creeped out
link |
if you realized that it was just a glossed up tape recorder,
link |
you know, that was just zombie and was a faking emotion?
link |
Would you prefer that it actually had an experience
link |
or would you prefer that it's actually
link |
not experiencing anything so you feel,
link |
you don't have to feel guilty about what you do to it?
link |
It's such a difficult question because, you know,
link |
it's like when you're in a relationship and you say,
link |
And the other person said, I love you back.
link |
It's like asking, well, do they really love you back
link |
or are they just saying they love you back?
link |
Don't you really want them to actually love you?
link |
It's hard to, it's hard to really know the difference
link |
between everything seeming like there's consciousness
link |
present, there's intelligence present,
link |
there's affection, passion, love,
link |
and it actually being there.
link |
I'm not sure, do you have?
link |
But like, can I ask you a question about this?
link |
Like to make it a bit more pointed.
link |
So Mass General Hospital is right across the river, right?
link |
Suppose you're going in for a medical procedure
link |
and they're like, you know, for anesthesia,
link |
what we're going to do is we're going to give you
link |
muscle relaxants so you won't be able to move
link |
and you're going to feel excruciating pain
link |
during the whole surgery,
link |
but you won't be able to do anything about it.
link |
But then we're going to give you this drug
link |
that erases your memory of it.
link |
Would you be cool about that?
link |
What's the difference that you're conscious about it
link |
or not if there's no behavioral change, right?
link |
Right, that's a really, that's a really clear way to put it.
link |
That's, yeah, it feels like in that sense,
link |
experiencing it is a valuable quality.
link |
So actually being able to have subjective experiences,
link |
at least in that case, is valuable.
link |
And I think we humans have a little bit
link |
of a bad track record also of making
link |
these self serving arguments
link |
that other entities aren't conscious.
link |
You know, people often say,
link |
oh, these animals can't feel pain.
link |
It's okay to boil lobsters because we ask them
link |
if it hurt and they didn't say anything.
link |
And now there was just a paper out saying,
link |
lobsters do feel pain when you boil them
link |
and they're banning it in Switzerland.
link |
And we did this with slaves too often and said,
link |
oh, they don't mind.
link |
They don't maybe aren't conscious
link |
or women don't have souls or whatever.
link |
So I'm a little bit nervous when I hear people
link |
just take as an axiom that machines
link |
can't have experience ever.
link |
I think this is just a really fascinating science question
link |
Let's research it and try to figure out
link |
what it is that makes the difference
link |
between unconscious intelligent behavior
link |
and conscious intelligent behavior.
link |
So in terms of, so if you think of a Boston Dynamics
link |
human or robot being sort of with a broom
link |
being pushed around, it starts pushing
link |
on a consciousness question.
link |
So let me ask, do you think an AGI system
link |
like a few neuroscientists believe
link |
needs to have a physical embodiment?
link |
Needs to have a body or something like a body?
link |
No, I don't think so.
link |
You mean to have a conscious experience?
link |
To have consciousness.
link |
I do think it helps a lot to have a physical embodiment
link |
to learn the kind of things about the world
link |
that are important to us humans, for sure.
link |
But I don't think the physical embodiment
link |
is necessary after you've learned it
link |
to just have the experience.
link |
Think about when you're dreaming, right?
link |
Your eyes are closed.
link |
You're not getting any sensory input.
link |
You're not behaving or moving in any way
link |
but there's still an experience there, right?
link |
And so clearly the experience that you have
link |
when you see something cool in your dreams
link |
isn't coming from your eyes.
link |
It's just the information processing itself in your brain
link |
which is that experience, right?
link |
But if I put it another way, I'll say
link |
because it comes from neuroscience
link |
is the reason you want to have a body and a physical
link |
something like a physical, you know, a physical system
link |
is because you want to be able to preserve something.
link |
In order to have a self, you could argue,
link |
would you need to have some kind of embodiment of self
link |
to want to preserve?
link |
Well, now we're getting a little bit anthropomorphic
link |
into anthropomorphizing things.
link |
Maybe talking about self preservation instincts.
link |
I mean, we are evolved organisms, right?
link |
So Darwinian evolution endowed us
link |
and other evolved organism with a self preservation instinct
link |
because those that didn't have those self preservation genes
link |
got cleaned out of the gene pool, right?
link |
But if you build an artificial general intelligence
link |
the mind space that you can design is much, much larger
link |
than just a specific subset of minds that can evolve.
link |
So an AGI mind doesn't necessarily have
link |
to have any self preservation instinct.
link |
It also doesn't necessarily have to be
link |
so individualistic as us.
link |
Like, imagine if you could just, first of all,
link |
or we are also very afraid of death.
link |
You know, I suppose you could back yourself up
link |
every five minutes and then your airplane
link |
is about to crash.
link |
You're like, shucks, I'm gonna lose the last five minutes
link |
of experiences since my last cloud backup, dang.
link |
You know, it's not as big a deal.
link |
Or if we could just copy experiences between our minds
link |
easily like we, which we could easily do
link |
if we were silicon based, right?
link |
Then maybe we would feel a little bit more
link |
like a hive mind actually, that maybe it's the,
link |
so I don't think we should take for granted at all
link |
that AGI will have to have any of those sort of
link |
competitive as alpha male instincts.
link |
On the other hand, you know, this is really interesting
link |
because I think some people go too far and say,
link |
of course we don't have to have any concerns either
link |
that advanced AI will have those instincts
link |
because we can build anything we want.
link |
That there's a very nice set of arguments going back
link |
to Steve Omohundro and Nick Bostrom and others
link |
just pointing out that when we build machines,
link |
we normally build them with some kind of goal, you know,
link |
win this chess game, drive this car safely or whatever.
link |
And as soon as you put in a goal into machine,
link |
especially if it's kind of open ended goal
link |
and the machine is very intelligent,
link |
it'll break that down into a bunch of sub goals.
link |
And one of those goals will almost always
link |
be self preservation because if it breaks or dies
link |
in the process, it's not gonna accomplish the goal, right?
link |
Like suppose you just build a little,
link |
you have a little robot and you tell it to go down
link |
the store market here and get you some food,
link |
make you cook an Italian dinner, you know,
link |
and then someone mugs it and tries to break it
link |
That robot has an incentive to not get destroyed
link |
and defend itself or run away,
link |
because otherwise it's gonna fail in cooking your dinner.
link |
It's not afraid of death,
link |
but it really wants to complete the dinner cooking goal.
link |
So it will have a self preservation instinct.
link |
Continue being a functional agent somehow.
link |
And similarly, if you give any kind of more ambitious goal
link |
to an AGI, it's very likely they wanna acquire
link |
more resources so it can do that better.
link |
And it's exactly from those sort of sub goals
link |
that we might not have intended
link |
that some of the concerns about AGI safety come.
link |
You give it some goal that seems completely harmless.
link |
And then before you realize it,
link |
it's also trying to do these other things
link |
which you didn't want it to do.
link |
And it's maybe smarter than us.
link |
So it's fascinating.
link |
And let me pause just because I am in a very kind
link |
of human centric way, see fear of death
link |
as a valuable motivator.
link |
So you don't think, you think that's an artifact
link |
of evolution, so that's the kind of mind space
link |
evolution created that we're sort of almost obsessed
link |
about self preservation, some kind of genetic flow.
link |
You don't think that's necessary to be afraid of death.
link |
So not just a kind of sub goal of self preservation
link |
just so you can keep doing the thing,
link |
but more fundamentally sort of have the finite thing
link |
like this ends for you at some point.
link |
Do I think it's necessary for what precisely?
link |
For intelligence, but also for consciousness.
link |
So for those, for both, do you think really
link |
like a finite death and the fear of it is important?
link |
So before I can answer, before we can agree
link |
on whether it's necessary for intelligence
link |
or for consciousness, we should be clear
link |
on how we define those two words.
link |
Cause a lot of really smart people define them
link |
in very different ways.
link |
I was on this panel with AI experts
link |
and they couldn't agree on how to define intelligence even.
link |
So I define intelligence simply
link |
as the ability to accomplish complex goals.
link |
I like your broad definition, because again
link |
I don't want to be a carbon chauvinist.
link |
And in that case, no, certainly
link |
it doesn't require fear of death.
link |
I would say alpha go, alpha zero is quite intelligent.
link |
I don't think alpha zero has any fear of being turned off
link |
because it doesn't understand the concept of it even.
link |
And similarly consciousness.
link |
I mean, you could certainly imagine very simple
link |
kind of experience.
link |
If certain plants have any kind of experience
link |
I don't think they're very afraid of dying
link |
or there's nothing they can do about it anyway much.
link |
So there wasn't that much value in, but more seriously
link |
I think if you ask, not just about being conscious
link |
but maybe having what you would, we might call
link |
an exciting life where you feel passion
link |
and really appreciate the things.
link |
Maybe there somehow, maybe there perhaps it does help
link |
having a backdrop that, Hey, it's finite.
link |
No, let's make the most of this, let's live to the fullest.
link |
So if you knew you were going to live forever
link |
do you think you would change your?
link |
Yeah, I mean, in some perspective
link |
it would be an incredibly boring life living forever.
link |
So in the sort of loose subjective terms that you said
link |
of something exciting and something in this
link |
that other humans would understand, I think is, yeah
link |
it seems that the finiteness of it is important.
link |
Well, the good news I have for you then is
link |
based on what we understand about cosmology
link |
everything is in our universe is probably
link |
ultimately probably finite, although.
link |
Big crunch or big, what's the, the infinite expansion.
link |
Yeah, we could have a big chill or a big crunch
link |
or a big rip or that's the big snap or death bubbles.
link |
All of them are more than a billion years away.
link |
So we should, we certainly have vastly more time
link |
than our ancestors thought, but there is still
link |
it's still pretty hard to squeeze in an infinite number
link |
of compute cycles, even though there are some loopholes
link |
that just might be possible.
link |
But I think, you know, some people like to say
link |
that you should live as if you're about to
link |
you're going to die in five years or so.
link |
And that's sort of optimal.
link |
Maybe it's a good assumption.
link |
We should build our civilization as if it's all finite
link |
to be on the safe side.
link |
So you mentioned defining intelligence
link |
as the ability to solve complex goals.
link |
Where would you draw a line or how would you try
link |
to define human level intelligence
link |
and superhuman level intelligence?
link |
Where is consciousness part of that definition?
link |
No, consciousness does not come into this definition.
link |
So, so I think of intelligence as it's a spectrum
link |
but there are very many different kinds of goals
link |
You can have a goal to be a good chess player
link |
a good goal player, a good car driver, a good investor
link |
good poet, et cetera.
link |
So intelligence that by its very nature
link |
isn't something you can measure by this one number
link |
or some overall goodness.
link |
There are some people who are more better at this.
link |
Some people are better than that.
link |
Right now we have machines that are much better than us
link |
at some very narrow tasks like multiplying large numbers
link |
fast, memorizing large databases, playing chess
link |
playing go and soon driving cars.
link |
But there's still no machine that can match
link |
a human child in general intelligence
link |
but artificial general intelligence, AGI
link |
the name of your course, of course
link |
that is by its very definition, the quest
link |
to build a machine that can do everything
link |
as well as we can.
link |
So the old Holy grail of AI from back to its inception
link |
in the sixties, if that ever happens, of course
link |
I think it's going to be the biggest transition
link |
in the history of life on earth
link |
but it doesn't necessarily have to wait the big impact
link |
until machines are better than us at knitting
link |
that the really big change doesn't come exactly
link |
at the moment they're better than us at everything.
link |
The really big change comes first
link |
there are big changes when they start becoming better
link |
at us at doing most of the jobs that we do
link |
because that takes away much of the demand
link |
And then the really whopping change comes
link |
when they become better than us at AI research, right?
link |
Because right now the timescale of AI research
link |
is limited by the human research and development cycle
link |
of years typically, you know
link |
how long does it take from one release of some software
link |
or iPhone or whatever to the next?
link |
But once Google can replace 40,000 engineers
link |
by 40,000 equivalent pieces of software or whatever
link |
but then there's no reason that has to be years
link |
it can be in principle much faster
link |
and the timescale of future progress in AI
link |
and all of science and technology will be driven
link |
by machines, not humans.
link |
So it's this simple point which gives right
link |
this incredibly fun controversy
link |
about whether there can be intelligence explosion
link |
so called singularity as Werner Vinge called it.
link |
Now the idea is articulated by I.J. Good
link |
is obviously way back fifties
link |
but you can see Alan Turing
link |
and others thought about it even earlier.
link |
So you asked me what exactly would I define
link |
human level intelligence, yeah.
link |
So the glib answer is to say something
link |
which is better than us at all cognitive tasks
link |
with a better than any human at all cognitive tasks
link |
but the really interesting bar
link |
I think goes a little bit lower than that actually.
link |
It's when they can, when they're better than us
link |
at AI programming and general learning
link |
so that they can if they want to get better
link |
than us at anything by just studying.
link |
So they're better is a key word and better is towards
link |
this kind of spectrum of the complexity of goals
link |
it's able to accomplish.
link |
So another way to, and that's certainly
link |
a very clear definition of human love.
link |
So there's, it's almost like a sea that's rising
link |
you can do more and more and more things
link |
it's a geographic that you show
link |
it's really nice way to put it.
link |
So there's some peaks that
link |
and there's an ocean level elevating
link |
and you solve more and more problems
link |
but just kind of to take a pause
link |
and we took a bunch of questions
link |
and a lot of social networks
link |
and a bunch of people asked
link |
a sort of a slightly different direction
link |
on creativity and things that perhaps aren't a peak.
link |
Human beings are flawed
link |
and perhaps better means having contradiction
link |
being flawed in some way.
link |
So let me sort of start easy, first of all.
link |
So you have a lot of cool equations.
link |
Let me ask, what's your favorite equation, first of all?
link |
I know they're all like your children, but like
link |
which one is that?
link |
This is the shirt in your equation.
link |
It's the master key of quantum mechanics
link |
of the micro world.
link |
So this equation will protect everything
link |
to do with atoms, molecules and all the way up.
link |
So quantum mechanics is certainly a beautiful
link |
mysterious formulation of our world.
link |
So I'd like to sort of ask you, just as an example
link |
it perhaps doesn't have the same beauty as physics does
link |
but in mathematics abstract, the Andrew Wiles
link |
who proved the Fermat's last theorem.
link |
So he just saw this recently
link |
and it kind of caught my eye a little bit.
link |
This is 358 years after it was conjectured.
link |
So this is very simple formulation.
link |
Everybody tried to prove it, everybody failed.
link |
And so here's this guy comes along
link |
and eventually proves it and then fails to prove it
link |
and then proves it again in 94.
link |
And he said like the moment when everything connected
link |
into place in an interview said
link |
it was so indescribably beautiful.
link |
That moment when you finally realize the connecting piece
link |
of two conjectures.
link |
He said, it was so indescribably beautiful.
link |
It was so simple and so elegant.
link |
I couldn't understand how I'd missed it.
link |
And I just stared at it in disbelief for 20 minutes.
link |
Then during the day, I walked around the department
link |
and I keep coming back to my desk
link |
looking to see if it was still there.
link |
It was still there.
link |
I couldn't contain myself.
link |
It was the most important moment on my working life.
link |
Nothing I ever do again will mean as much.
link |
So that particular moment.
link |
And it kind of made me think of what would it take?
link |
And I think we have all been there at small levels.
link |
Maybe let me ask, have you had a moment like that
link |
in your life where you just had an idea?
link |
It's like, wow, yes.
link |
I wouldn't mention myself in the same breath
link |
as Andrew Wiles, but I've certainly had a number
link |
of aha moments when I realized something very cool
link |
about physics, which has completely made my head explode.
link |
In fact, some of my favorite discoveries I made later,
link |
I later realized that they had been discovered earlier
link |
by someone who sometimes got quite famous for it.
link |
So it's too late for me to even publish it,
link |
but that doesn't diminish in any way.
link |
The emotional experience you have when you realize it,
link |
Yeah, so what would it take in that moment, that wow,
link |
that was yours in that moment?
link |
So what do you think it takes for an intelligence system,
link |
an AGI system, an AI system to have a moment like that?
link |
That's a tricky question
link |
because there are actually two parts to it, right?
link |
One of them is, can it accomplish that proof?
link |
Can it prove that you can never write A to the N
link |
plus B to the N equals three to that equal Z to the N
link |
for all integers, et cetera, et cetera,
link |
when N is bigger than two?
link |
That's simply a question about intelligence.
link |
Can you build machines that are that intelligent?
link |
And I think by the time we get a machine
link |
that can independently come up with that level of proofs,
link |
probably quite close to AGI.
link |
The second question is a question about consciousness.
link |
When will we, how likely is it that such a machine
link |
will actually have any experience at all,
link |
as opposed to just being like a zombie?
link |
And would we expect it to have some sort of emotional response
link |
to this or anything at all akin to human emotion
link |
where when it accomplishes its machine goal,
link |
it views it as somehow something very positive
link |
and sublime and deeply meaningful?
link |
I would certainly hope that if in the future
link |
we do create machines that are our peers
link |
or even our descendants, that I would certainly
link |
hope that they do have this sublime appreciation of life.
link |
In a way, my absolutely worst nightmare
link |
would be that at some point in the future,
link |
the distant future, maybe our cosmos
link |
is teeming with all this post biological life doing
link |
all the seemingly cool stuff.
link |
And maybe the last humans, by the time
link |
our species eventually fizzles out,
link |
will be like, well, that's OK because we're
link |
so proud of our descendants here.
link |
And look what all the, my worst nightmare
link |
is that we haven't solved the consciousness problem.
link |
And we haven't realized that these are all the zombies.
link |
They're not aware of anything any more than a tape recorder
link |
has any kind of experience.
link |
So the whole thing has just become
link |
a play for empty benches.
link |
That would be the ultimate zombie apocalypse.
link |
So I would much rather, in that case,
link |
that we have these beings which can really
link |
appreciate how amazing it is.
link |
And in that picture, what would be the role of creativity?
link |
A few people ask about creativity.
link |
When you think about intelligence,
link |
certainly the story you told at the beginning of your book
link |
involved creating movies and so on, making money.
link |
You can make a lot of money in our modern world
link |
with music and movies.
link |
So if you are an intelligent system,
link |
you may want to get good at that.
link |
But that's not necessarily what I mean by creativity.
link |
Is it important on that complex goals
link |
where the sea is rising for there
link |
to be something creative?
link |
Or am I being very human centric and thinking creativity
link |
somehow special relative to intelligence?
link |
My hunch is that we should think of creativity simply
link |
as an aspect of intelligence.
link |
And we have to be very careful with human vanity.
link |
We have this tendency to very often want
link |
to say, as soon as machines can do something,
link |
we try to diminish it and say, oh, but that's
link |
not real intelligence.
link |
Isn't it creative or this or that?
link |
The other thing, if we ask ourselves
link |
to write down a definition of what we actually mean
link |
by being creative, what we mean by Andrew Wiles, what he did
link |
there, for example, don't we often mean that someone takes
link |
a very unexpected leap?
link |
It's not like taking 573 and multiplying it
link |
by 224 by just a step of straightforward cookbook
link |
like rules, right?
link |
You can maybe make a connection between two things
link |
that people had never thought was connected or something
link |
I think this is an aspect of intelligence.
link |
And this is actually one of the most important aspects of it.
link |
Maybe the reason we humans tend to be better at it
link |
than traditional computers is because it's
link |
something that comes more naturally if you're
link |
a neural network than if you're a traditional logic gate
link |
based computer machine.
link |
We physically have all these connections.
link |
And you activate here, activate here, activate here.
link |
My hunch is that if we ever build a machine where you could
link |
just give it the task, hey, you say, hey, I just realized
link |
I want to travel around the world instead this month.
link |
Can you teach my AGI course for me?
link |
And it's like, OK, I'll do it.
link |
And it does everything that you would have done
link |
and improvises and stuff.
link |
That would, in my mind, involve a lot of creativity.
link |
Yeah, so it's actually a beautiful way to put it.
link |
I think we do try to grasp at the definition of intelligence
link |
is everything we don't understand how to build.
link |
So we as humans try to find things
link |
that we have and machines don't have.
link |
And maybe creativity is just one of the things, one
link |
of the words we use to describe that.
link |
That's a really interesting way to put it.
link |
I don't think we need to be that defensive.
link |
I don't think anything good comes out of saying,
link |
well, we're somehow special, you know?
link |
Contrary wise, there are many examples in history
link |
of where trying to pretend that we're somehow superior
link |
to all other intelligent beings has led to pretty bad results,
link |
Nazi Germany, they said that they were somehow superior
link |
Today, we still do a lot of cruelty to animals
link |
by saying that we're so superior somehow,
link |
and they can't feel pain.
link |
Slavery was justified by the same kind
link |
of just really weak arguments.
link |
And I don't think if we actually go ahead and build
link |
artificial general intelligence, it
link |
can do things better than us, I don't
link |
think we should try to found our self worth on some sort
link |
of bogus claims of superiority in terms
link |
of our intelligence.
link |
I think we should instead find our calling
link |
and the meaning of life from the experiences that we have.
link |
I can have very meaningful experiences
link |
even if there are other people who are smarter than me.
link |
When I go to a faculty meeting here,
link |
and we talk about something, and then I certainly realize,
link |
oh, boy, he has an old prize, he has an old prize,
link |
he has an old prize, I don't have one.
link |
Does that make me enjoy life any less
link |
or enjoy talking to those people less?
link |
And the contrary, I feel very honored and privileged
link |
to get to interact with other very intelligent beings that
link |
are better than me at a lot of stuff.
link |
So I don't think there's any reason why
link |
we can't have the same approach with intelligent machines.
link |
That's a really interesting.
link |
So people don't often think about that.
link |
They think about when there's going,
link |
if there's machines that are more intelligent,
link |
you naturally think that that's not
link |
going to be a beneficial type of intelligence.
link |
You don't realize it could be like peers with Nobel prizes
link |
that would be just fun to talk with,
link |
and they might be clever about certain topics,
link |
and you can have fun having a few drinks with them.
link |
Well, also, another example we can all
link |
relate to of why it doesn't have to be a terrible thing
link |
to be in the presence of people who are even smarter than us
link |
all around is when you and I were both two years old,
link |
I mean, our parents were much more intelligent than us,
link |
Worked out OK, because their goals
link |
were aligned with our goals.
link |
And that, I think, is really the number one key issue
link |
we have to solve if we value align the value alignment
link |
Because people who see too many Hollywood movies
link |
with lousy science fiction plot lines,
link |
they worry about the wrong thing, right?
link |
They worry about some machine suddenly turning evil.
link |
It's not malice that is the concern.
link |
By definition, intelligent makes you very competent.
link |
If you have a more intelligent goal playing,
link |
computer playing is a less intelligent one.
link |
And when we define intelligence as the ability
link |
to accomplish goal winning, it's going
link |
to be the more intelligent one that wins.
link |
And if you have a human and then you
link |
have an AGI that's more intelligent in all ways
link |
and they have different goals, guess who's
link |
going to get their way, right?
link |
So I was just reading about this particular rhinoceros species
link |
that was driven extinct just a few years ago.
link |
Ellen Bummer is looking at this cute picture of a mommy
link |
rhinoceros with its child.
link |
And why did we humans drive it to extinction?
link |
It wasn't because we were evil rhino haters as a whole.
link |
It was just because our goals weren't aligned
link |
with those of the rhinoceros.
link |
And it didn't work out so well for the rhinoceros
link |
because we were more intelligent, right?
link |
So I think it's just so important
link |
that if we ever do build AGI, before we unleash anything,
link |
we have to make sure that it learns
link |
to understand our goals, that it adopts our goals,
link |
and that it retains those goals.
link |
So the cool, interesting problem there
link |
is us as human beings trying to formulate our values.
link |
So you could think of the United States Constitution as a way
link |
that people sat down, at the time a bunch of white men,
link |
which is a good example, I should say.
link |
They formulated the goals for this country.
link |
And a lot of people agree that those goals actually
link |
held up pretty well.
link |
That's an interesting formulation of values
link |
and failed miserably in other ways.
link |
So for the value alignment problem and the solution to it,
link |
we have to be able to put on paper or in a program
link |
How difficult do you think that is?
link |
But it's so important.
link |
We really have to give it our best.
link |
And it's difficult for two separate reasons.
link |
There's the technical value alignment problem
link |
of figuring out just how to make machines understand our goals,
link |
adopt them, and retain them.
link |
And then there's the separate part of it,
link |
the philosophical part.
link |
Whose values anyway?
link |
And since it's not like we have any great consensus
link |
on this planet on values, what mechanism should we
link |
create then to aggregate and decide, OK,
link |
what's a good compromise?
link |
That second discussion can't just
link |
be left to tech nerds like myself.
link |
And if we refuse to talk about it and then AGI gets built,
link |
who's going to be actually making
link |
the decision about whose values?
link |
It's going to be a bunch of dudes in some tech company.
link |
And are they necessarily so representative of all
link |
of humankind that we want to just entrust it to them?
link |
Are they even uniquely qualified to speak
link |
to future human happiness just because they're
link |
good at programming AI?
link |
I'd much rather have this be a really inclusive conversation.
link |
But do you think it's possible?
link |
So you create a beautiful vision that includes the diversity,
link |
cultural diversity, and various perspectives on discussing
link |
rights, freedoms, human dignity.
link |
But how hard is it to come to that consensus?
link |
Do you think it's certainly a really important thing
link |
that we should all try to do?
link |
But do you think it's feasible?
link |
I think there's no better way to guarantee failure than to
link |
refuse to talk about it or refuse to try.
link |
And I also think it's a really bad strategy
link |
to say, OK, let's first have a discussion for a long time.
link |
And then once we reach complete consensus,
link |
then we'll try to load it into some machine.
link |
No, we shouldn't let perfect be the enemy of good.
link |
Instead, we should start with the kindergarten ethics
link |
that pretty much everybody agrees on
link |
and put that into machines now.
link |
We're not doing that even.
link |
Look at anyone who builds this passenger aircraft,
link |
wants it to never under any circumstances
link |
fly into a building or a mountain.
link |
Yet the September 11 hijackers were able to do that.
link |
And even more embarrassingly, Andreas Lubitz,
link |
this depressed Germanwings pilot,
link |
when he flew his passenger jet into the Alps killing over 100
link |
people, he just told the autopilot to do it.
link |
He told the freaking computer to change the altitude
link |
And even though it had the GPS maps, everything,
link |
the computer was like, OK.
link |
So we should take those very basic values,
link |
where the problem is not that we don't agree.
link |
The problem is just we've been too lazy
link |
to try to put it into our machines
link |
and make sure that from now on, airplanes will just,
link |
which all have computers in them,
link |
but will just refuse to do something like that.
link |
Go into safe mode, maybe lock the cockpit door,
link |
go over to the nearest airport.
link |
And there's so much other technology in our world
link |
as well now, where it's really becoming quite timely
link |
to put in some sort of very basic values like this.
link |
Even in cars, we've had enough vehicle terrorism attacks
link |
by now, where people have driven trucks and vans
link |
into pedestrians, that it's not at all a crazy idea
link |
to just have that hardwired into the car.
link |
Because yeah, there are a lot of,
link |
there's always going to be people who for some reason
link |
want to harm others, but most of those people
link |
don't have the technical expertise to figure out
link |
how to work around something like that.
link |
So if the car just won't do it, it helps.
link |
So let's start there.
link |
So there's a lot of, that's a great point.
link |
So not chasing perfect.
link |
There's a lot of things that most of the world agrees on.
link |
Yeah, let's start there.
link |
Let's start there.
link |
And then once we start there,
link |
we'll also get into the habit of having
link |
these kind of conversations about, okay,
link |
what else should we put in here and have these discussions?
link |
This should be a gradual process then.
link |
Great, so, but that also means describing these things
link |
and describing it to a machine.
link |
So one thing, we had a few conversations
link |
with Stephen Wolfram.
link |
I'm not sure if you're familiar with Stephen.
link |
Oh yeah, I know him quite well.
link |
So he is, he works with a bunch of things,
link |
but cellular automata, these simple computable things,
link |
these computation systems.
link |
And he kind of mentioned that,
link |
we probably have already within these systems
link |
already something that's AGI,
link |
meaning like we just don't know it
link |
because we can't talk to it.
link |
So if you give me this chance to try to at least
link |
form a question out of this is,
link |
I think it's an interesting idea to think
link |
that we can have intelligent systems,
link |
but we don't know how to describe something to them
link |
and they can't communicate with us.
link |
I know you're doing a little bit of work in explainable AI,
link |
trying to get AI to explain itself.
link |
So what are your thoughts of natural language processing
link |
or some kind of other communication?
link |
How does the AI explain something to us?
link |
How do we explain something to it, to machines?
link |
Or you think of it differently?
link |
So there are two separate parts to your question there.
link |
One of them has to do with communication,
link |
which is super interesting, I'll get to that in a sec.
link |
The other is whether we already have AGI
link |
but we just haven't noticed it there.
link |
There I beg to differ.
link |
I don't think there's anything in any cellular automaton
link |
or anything or the internet itself or whatever
link |
that has artificial general intelligence
link |
and that it can really do exactly everything
link |
we humans can do better.
link |
I think the day that happens, when that happens,
link |
we will very soon notice, we'll probably notice even before
link |
because in a very, very big way.
link |
But for the second part, though.
link |
Wait, can I ask, sorry.
link |
So, because you have this beautiful way
link |
to formulating consciousness as information processing,
link |
and you can think of intelligence
link |
as information processing,
link |
and you can think of the entire universe
link |
as these particles and these systems roaming around
link |
that have this information processing power.
link |
You don't think there is something with the power
link |
to process information in the way that we human beings do
link |
that's out there that needs to be sort of connected to.
link |
It seems a little bit philosophical, perhaps,
link |
but there's something compelling to the idea
link |
that the power is already there,
link |
which the focus should be more on being able
link |
to communicate with it.
link |
Well, I agree that in a certain sense,
link |
the hardware processing power is already out there
link |
because our universe itself can think of it
link |
as being a computer already, right?
link |
It's constantly computing what water waves,
link |
how it devolved the water waves in the River Charles
link |
and how to move the air molecules around.
link |
Seth Lloyd has pointed out, my colleague here,
link |
that you can even in a very rigorous way
link |
think of our entire universe as being a quantum computer.
link |
It's pretty clear that our universe
link |
supports this amazing processing power
link |
because you can even,
link |
within this physics computer that we live in, right?
link |
We can even build actual laptops and stuff,
link |
so clearly the power is there.
link |
It's just that most of the compute power that nature has,
link |
it's, in my opinion, kind of wasting on boring stuff
link |
like simulating yet another ocean wave somewhere
link |
where no one is even looking, right?
link |
So in a sense, what life does, what we are doing
link |
when we build computers is we're rechanneling
link |
all this compute that nature is doing anyway
link |
into doing things that are more interesting
link |
than just yet another ocean wave,
link |
and let's do something cool here.
link |
So the raw hardware power is there, for sure,
link |
but then even just computing what's going to happen
link |
for the next five seconds in this water bottle,
link |
takes a ridiculous amount of compute
link |
if you do it on a human computer.
link |
This water bottle just did it.
link |
But that does not mean that this water bottle has AGI
link |
because AGI means it should also be able to,
link |
like I've written my book, done this interview.
link |
And I don't think it's just communication problems.
link |
I don't really think it can do it.
link |
Although Buddhists say when they watch the water
link |
and that there is some beauty,
link |
that there's some depth and beauty in nature
link |
that they can communicate with.
link |
Communication is also very important though
link |
because I mean, look, part of my job is being a teacher.
link |
And I know some very intelligent professors even
link |
who just have a bit of hard time communicating.
link |
They come up with all these brilliant ideas,
link |
but to communicate with somebody else,
link |
you have to also be able to simulate their own mind.
link |
Build well enough and understand model of their mind
link |
that you can say things that they will understand.
link |
And that's quite difficult.
link |
And that's why today it's so frustrating
link |
if you have a computer that makes some cancer diagnosis
link |
and you ask it, well, why are you saying
link |
I should have this surgery?
link |
And if it can only reply,
link |
I was trained on five terabytes of data
link |
and this is my diagnosis, boop, boop, beep, beep.
link |
It doesn't really instill a lot of confidence, right?
link |
So I think we have a lot of work to do
link |
on communication there.
link |
So what kind of, I think you're doing a little bit of work
link |
in explainable AI.
link |
What do you think are the most promising avenues?
link |
Is it mostly about sort of the Alexa problem
link |
of natural language processing of being able
link |
to actually use human interpretable methods
link |
So being able to talk to a system and it talk back to you,
link |
or is there some more fundamental problems to be solved?
link |
I think it's all of the above.
link |
The natural language processing is obviously important,
link |
but there are also more nerdy fundamental problems.
link |
Like if you take, you play chess?
link |
Of course, I'm Russian.
link |
You speak Russian?
link |
Yes, I speak Russian.
link |
Excellent, I didn't know.
link |
When did you learn Russian?
link |
I speak very bad Russian, I'm only an autodidact,
link |
but I bought a book, Teach Yourself Russian,
link |
read a lot, but it was very difficult.
link |
That's why I speak so bad.
link |
How many languages do you know?
link |
Wow, that's really impressive.
link |
I don't know, my wife has some calculation,
link |
but my point was, if you play chess,
link |
have you looked at the AlphaZero games?
link |
The actual games, no.
link |
Check it out, some of them are just mind blowing,
link |
And if you ask, how did it do that?
link |
You go talk to Demis Hassabis,
link |
I know others from DeepMind,
link |
all they'll ultimately be able to give you
link |
is big tables of numbers, matrices,
link |
that define the neural network.
link |
And you can stare at these tables of numbers
link |
till your face turn blue,
link |
and you're not gonna understand much
link |
about why it made that move.
link |
And even if you have natural language processing
link |
that can tell you in human language about,
link |
oh, five, seven, points, two, eight,
link |
still not gonna really help.
link |
So I think there's a whole spectrum of fun challenges
link |
that are involved in taking a computation
link |
that does intelligent things
link |
and transforming it into something equally good,
link |
equally intelligent, but that's more understandable.
link |
And I think that's really valuable
link |
because I think as we put machines in charge
link |
of ever more infrastructure in our world,
link |
the power grid, the trading on the stock market,
link |
weapon systems and so on,
link |
it's absolutely crucial that we can trust
link |
these AIs to do all we want.
link |
And trust really comes from understanding
link |
in a very fundamental way.
link |
And that's why I'm working on this,
link |
because I think the more,
link |
if we're gonna have some hope of ensuring
link |
that machines have adopted our goals
link |
and that they're gonna retain them,
link |
that kind of trust, I think,
link |
needs to be based on things you can actually understand,
link |
preferably even improve theorems on.
link |
Even with a self driving car, right?
link |
If someone just tells you it's been trained
link |
on tons of data and it never crashed,
link |
it's less reassuring than if someone actually has a proof.
link |
Maybe it's a computer verified proof,
link |
but still it says that under no circumstances
link |
is this car just gonna swerve into oncoming traffic.
link |
And that kind of information helps to build trust
link |
and helps build the alignment of goals,
link |
at least awareness that your goals, your values are aligned.
link |
And I think even in the very short term,
link |
if you look at how, you know, today, right?
link |
This absolutely pathetic state of cybersecurity
link |
that we have, where is it?
link |
Three billion Yahoo accounts we can't pack,
link |
almost every American's credit card and so on.
link |
Why is this happening?
link |
It's ultimately happening because we have software
link |
that nobody fully understood how it worked.
link |
That's why the bugs hadn't been found, right?
link |
And I think AI can be used very effectively
link |
for offense, for hacking,
link |
but it can also be used for defense.
link |
Hopefully automating verifiability
link |
and creating systems that are built in different ways
link |
so you can actually prove things about them.
link |
And it's important.
link |
So speaking of software that nobody understands
link |
how it works, of course, a bunch of people ask
link |
about your paper, about your thoughts
link |
of why does deep and cheap learning work so well?
link |
But what are your thoughts on deep learning?
link |
These kind of simplified models of our own brains
link |
have been able to do some successful perception work,
link |
pattern recognition work, and now with AlphaZero and so on,
link |
do some clever things.
link |
What are your thoughts about the promise limitations
link |
Great, I think there are a number of very important insights,
link |
very important lessons we can always draw
link |
from these kinds of successes.
link |
One of them is when you look at the human brain,
link |
you see it's very complicated, 10th of 11 neurons,
link |
and there are all these different kinds of neurons
link |
and yada, yada, and there's been this long debate
link |
about whether the fact that we have dozens
link |
of different kinds is actually necessary for intelligence.
link |
We can now, I think, quite convincingly answer
link |
that question of no, it's enough to have just one kind.
link |
If you look under the hood of AlphaZero,
link |
there's only one kind of neuron
link |
and it's ridiculously simple mathematical thing.
link |
So it's just like in physics,
link |
it's not, if you have a gas with waves in it,
link |
it's not the detailed nature of the molecule that matter,
link |
it's the collective behavior somehow.
link |
Similarly, it's this higher level structure
link |
of the network that matters,
link |
not that you have 20 kinds of neurons.
link |
I think our brain is such a complicated mess
link |
because it wasn't evolved just to be intelligent,
link |
it was involved to also be self assembling
link |
and self repairing, right?
link |
And evolutionarily attainable.
link |
And so on and so on.
link |
So I think it's pretty,
link |
my hunch is that we're going to understand
link |
how to build AGI before we fully understand
link |
how our brains work, just like we understood
link |
how to build flying machines long before
link |
we were able to build a mechanical bird.
link |
Yeah, that's right.
link |
You've given the example exactly of mechanical birds
link |
and airplanes and airplanes do a pretty good job
link |
of flying without really mimicking bird flight.
link |
And even now after 100 years later,
link |
did you see the Ted talk with this German mechanical bird?
link |
I heard you mention it.
link |
Check it out, it's amazing.
link |
But even after that, right,
link |
we still don't fly in mechanical birds
link |
because it turned out the way we came up with was simpler
link |
and it's better for our purposes.
link |
And I think it might be the same there.
link |
That's one lesson.
link |
And another lesson, it's more what our paper was about.
link |
First, as a physicist thought it was fascinating
link |
how there's a very close mathematical relationship
link |
actually between our artificial neural networks
link |
and a lot of things that we've studied for in physics
link |
go by nerdy names like the renormalization group equation
link |
and Hamiltonians and yada, yada, yada.
link |
And when you look a little more closely at this,
link |
at first I was like, well, there's something crazy here
link |
that doesn't make sense.
link |
Because we know that if you even want to build
link |
a super simple neural network to tell apart cat pictures
link |
and dog pictures, right,
link |
that you can do that very, very well now.
link |
But if you think about it a little bit,
link |
you convince yourself it must be impossible
link |
because if I have one megapixel,
link |
even if each pixel is just black or white,
link |
there's two to the power of 1 million possible images,
link |
which is way more than there are atoms in our universe,
link |
right, so in order to,
link |
and then for each one of those,
link |
I have to assign a number,
link |
which is the probability that it's a dog.
link |
So an arbitrary function of images
link |
is a list of more numbers than there are atoms in our universe.
link |
So clearly I can't store that under the hood of my GPU
link |
or my computer, yet somehow it works.
link |
So what does that mean?
link |
Well, it means that out of all of the problems
link |
that you could try to solve with a neural network,
link |
almost all of them are impossible to solve
link |
with a reasonably sized one.
link |
But then what we showed in our paper
link |
was that the fraction, the kind of problems,
link |
the fraction of all the problems
link |
that you could possibly pose,
link |
that we actually care about given the laws of physics
link |
is also an infinite testimony, tiny little part.
link |
And amazingly, they're basically the same part.
link |
Yeah, it's almost like our world was created for,
link |
I mean, they kind of come together.
link |
Yeah, well, you could say maybe where the world was created
link |
for us, but I have a more modest interpretation,
link |
which is that the world was created for us,
link |
but I have a more modest interpretation,
link |
which is that instead evolution endowed us
link |
with neural networks precisely for that reason.
link |
Because this particular architecture,
link |
as opposed to the one in your laptop,
link |
is very, very well adapted to solving the kind of problems
link |
that nature kept presenting our ancestors with.
link |
So it makes sense that why do we have a brain
link |
in the first place?
link |
It's to be able to make predictions about the future
link |
So if we had a sucky system, which could never solve it,
link |
we wouldn't have a world.
link |
So this is, I think, a very beautiful fact.
link |
We also realize that there's been earlier work
link |
on why deeper networks are good,
link |
but we were able to show an additional cool fact there,
link |
which is that even incredibly simple problems,
link |
like suppose I give you a thousand numbers
link |
and ask you to multiply them together,
link |
and you can write a few lines of code, boom, done, trivial.
link |
If you just try to do that with a neural network
link |
that has only one single hidden layer in it,
link |
but you're going to need two to the power of a thousand
link |
neurons to multiply a thousand numbers,
link |
which is, again, more neurons than there are atoms
link |
That's fascinating.
link |
But if you allow yourself to make it a deep network
link |
with many layers, you only need 4,000 neurons.
link |
It's perfectly feasible.
link |
That's really interesting.
link |
So on another architecture type,
link |
I mean, you mentioned Schrodinger's equation,
link |
and what are your thoughts about quantum computing
link |
and the role of this kind of computational unit
link |
in creating an intelligence system?
link |
In some Hollywood movies that I will not mention by name
link |
because I don't want to spoil them.
link |
The way they get AGI is building a quantum computer.
link |
Because the word quantum sounds cool and so on.
link |
First of all, I think we don't need quantum computers
link |
I suspect your brain is not a quantum computer
link |
in any profound sense.
link |
So you don't even wrote a paper about that
link |
a lot many years ago.
link |
I calculated the so called decoherence time,
link |
how long it takes until the quantum computerness
link |
of what your neurons are doing gets erased
link |
by just random noise from the environment.
link |
And it's about 10 to the minus 21 seconds.
link |
So as cool as it would be to have a quantum computer
link |
in my head, I don't think that fast.
link |
On the other hand,
link |
there are very cool things you could do
link |
with quantum computers.
link |
Or I think we'll be able to do soon
link |
when we get bigger ones.
link |
That might actually help machine learning
link |
do even better than the brain.
link |
one, this is just a moonshot,
link |
but learning is very much same thing as search.
link |
If you're trying to train a neural network
link |
to get really learned to do something really well,
link |
you have some loss function,
link |
you have a bunch of knobs you can turn,
link |
represented by a bunch of numbers,
link |
and you're trying to tweak them
link |
so that it becomes as good as possible at this thing.
link |
So if you think of a landscape with some valley,
link |
where each dimension of the landscape
link |
corresponds to some number you can change,
link |
you're trying to find the minimum.
link |
And it's well known that
link |
if you have a very high dimensional landscape,
link |
complicated things, it's super hard to find the minimum.
link |
Quantum mechanics is amazingly good at this.
link |
Like if I want to know what's the lowest energy state
link |
this water can possibly have,
link |
incredibly hard to compute,
link |
but nature will happily figure this out for you
link |
if you just cool it down, make it very, very cold.
link |
If you put a ball somewhere,
link |
it'll roll down to its minimum.
link |
And this happens metaphorically
link |
at the energy landscape too.
link |
And quantum mechanics even uses some clever tricks,
link |
which today's machine learning systems don't.
link |
Like if you're trying to find the minimum
link |
and you get stuck in the little local minimum here,
link |
in quantum mechanics you can actually tunnel
link |
through the barrier and get unstuck again.
link |
That's really interesting.
link |
Yeah, so it may be, for example,
link |
that we'll one day use quantum computers
link |
that help train neural networks better.
link |
That's really interesting.
link |
Okay, so as a component of kind of the learning process,
link |
Let me ask sort of wrapping up here a little bit,
link |
let me return to the questions of our human nature
link |
and love, as I mentioned.
link |
you mentioned sort of a helper robot,
link |
but you could think of also personal robots.
link |
Do you think the way we human beings fall in love
link |
and get connected to each other
link |
is possible to achieve in an AI system
link |
and human level AI intelligence system?
link |
Do you think we would ever see that kind of connection?
link |
Or, you know, in all this discussion
link |
about solving complex goals,
link |
is this kind of human social connection,
link |
do you think that's one of the goals
link |
on the peaks and valleys with the raising sea levels
link |
that we'll be able to achieve?
link |
Or do you think that's something that's ultimately,
link |
or at least in the short term,
link |
relative to the other goals is not achievable?
link |
I think it's all possible.
link |
And I mean, in recent,
link |
there's a very wide range of guesses, as you know,
link |
among AI researchers, when we're going to get AGI.
link |
Some people, you know, like our friend Rodney Brooks
link |
says it's going to be hundreds of years at least.
link |
And then there are many others
link |
who think it's going to happen much sooner.
link |
maybe half or so of AI researchers
link |
think we're going to get AGI within decades.
link |
So if that happens, of course,
link |
then I think these things are all possible.
link |
But in terms of whether it will happen,
link |
I think we shouldn't spend so much time asking
link |
what do we think will happen in the future?
link |
As if we are just some sort of pathetic,
link |
your passive bystanders, you know,
link |
waiting for the future to happen to us.
link |
Hey, we're the ones creating this future, right?
link |
So we should be proactive about it
link |
and ask ourselves what sort of future
link |
we would like to have happen.
link |
We're going to make it like that.
link |
Well, what I prefer is just some sort of incredibly boring,
link |
zombie like future where there's all these
link |
mechanical things happening and there's no passion,
link |
no emotion, no experience, maybe even.
link |
No, I would of course, much rather prefer it
link |
if all the things that we find that we value the most
link |
about humanity are our subjective experience,
link |
passion, inspiration, love, you know.
link |
If we can create a future where those things do happen,
link |
where those things do exist, you know,
link |
I think ultimately it's not our universe
link |
giving meaning to us, it's us giving meaning to our universe.
link |
And if we build more advanced intelligence,
link |
let's make sure we build it in such a way
link |
that meaning is part of it.
link |
A lot of people that seriously study this problem
link |
and think of it from different angles
link |
have trouble in the majority of cases,
link |
if they think through that happen,
link |
are the ones that are not beneficial to humanity.
link |
And so, yeah, so what are your thoughts?
link |
What's should people, you know,
link |
I really don't like people to be terrified.
link |
What's a way for people to think about it
link |
in a way we can solve it and we can make it better?
link |
No, I don't think panicking is going to help in any way.
link |
It's not going to increase chances
link |
of things going well either.
link |
Even if you are in a situation where there is a real threat,
link |
does it help if everybody just freaks out?
link |
No, of course, of course not.
link |
I think, yeah, there are of course ways
link |
in which things can go horribly wrong.
link |
First of all, it's important when we think about this thing,
link |
about the problems and risks,
link |
to also remember how huge the upsides can be
link |
if we get it right, right?
link |
Everything we love about society and civilization
link |
is a product of intelligence.
link |
So if we can amplify our intelligence
link |
with machine intelligence and not anymore lose our loved one
link |
to what we're told is an incurable disease
link |
and things like this, of course, we should aspire to that.
link |
So that can be a motivator, I think,
link |
reminding ourselves that the reason we try to solve problems
link |
is not just because we're trying to avoid gloom,
link |
but because we're trying to do something great.
link |
But then in terms of the risks,
link |
I think the really important question is to ask,
link |
what can we do today that will actually help
link |
make the outcome good, right?
link |
And dismissing the risk is not one of them.
link |
I find it quite funny often when I'm in discussion panels
link |
about these things,
link |
how the people who work for companies,
link |
always be like, oh, nothing to worry about,
link |
nothing to worry about, nothing to worry about.
link |
And it's only academics sometimes express concerns.
link |
That's not surprising at all if you think about it.
link |
Upton Sinclair quipped, right,
link |
that it's hard to make a man believe in something
link |
when his income depends on not believing in it.
link |
And frankly, we know a lot of these people in companies
link |
that they're just as concerned as anyone else.
link |
But if you're the CEO of a company,
link |
that's not something you want to go on record saying
link |
when you have silly journalists who are gonna put a picture
link |
of a Terminator robot when they quote you.
link |
So the issues are real.
link |
And the way I think about what the issue is,
link |
is basically the real choice we have is,
link |
first of all, are we gonna just dismiss the risks
link |
and say, well, let's just go ahead and build machines
link |
that can do everything we can do better and cheaper.
link |
Let's just make ourselves obsolete as fast as possible.
link |
What could possibly go wrong?
link |
That's one attitude.
link |
The opposite attitude, I think, is to say,
link |
here's this incredible potential,
link |
let's think about what kind of future
link |
we're really, really excited about.
link |
What are the shared goals that we can really aspire towards?
link |
And then let's think really hard
link |
about how we can actually get there.
link |
So start with, don't start thinking about the risks,
link |
start thinking about the goals.
link |
And then when you do that,
link |
then you can think about the obstacles you want to avoid.
link |
I often get students coming in right here into my office
link |
for career advice.
link |
I always ask them this very question,
link |
where do you want to be in the future?
link |
If all she can say is, oh, maybe I'll have cancer,
link |
maybe I'll get run over by a truck.
link |
Yeah, focus on the obstacles instead of the goals.
link |
She's just going to end up a hypochondriac paranoid.
link |
Whereas if she comes in and fire in her eyes
link |
and is like, I want to be there.
link |
And then we can talk about the obstacles
link |
and see how we can circumvent them.
link |
That's, I think, a much, much healthier attitude.
link |
And I feel it's very challenging to come up with a vision
link |
for the future, which we are unequivocally excited about.
link |
I'm not just talking now in the vague terms,
link |
like, yeah, let's cure cancer, fine.
link |
I'm talking about what kind of society
link |
do we want to create?
link |
What do we want it to mean to be human in the age of AI,
link |
in the age of AGI?
link |
So if we can have this conversation,
link |
broad, inclusive conversation,
link |
and gradually start converging towards some,
link |
some future that with some direction, at least,
link |
that we want to steer towards, right,
link |
then we'll be much more motivated
link |
to constructively take on the obstacles.
link |
And I think if I had, if I had to,
link |
if I try to wrap this up in a more succinct way,
link |
I think we can all agree already now
link |
that we should aspire to build AGI
link |
that doesn't overpower us, but that empowers us.
link |
And think of the many various ways that can do that,
link |
whether that's from my side of the world
link |
of autonomous vehicles.
link |
I'm personally actually from the camp
link |
that believes this human level intelligence
link |
is required to achieve something like vehicles
link |
that would actually be something we would enjoy using
link |
and being part of.
link |
So that's one example, and certainly there's a lot
link |
of other types of robots and medicine and so on.
link |
So focusing on those and then coming up with the obstacles,
link |
coming up with the ways that that can go wrong
link |
and solving those one at a time.
link |
And just because you can build an autonomous vehicle,
link |
even if you could build one
link |
that would drive just fine without you,
link |
maybe there are some things in life
link |
that we would actually want to do ourselves.
link |
Right, like, for example,
link |
if you think of our society as a whole,
link |
there are some things that we find very meaningful to do.
link |
And that doesn't mean we have to stop doing them
link |
just because machines can do them better.
link |
I'm not gonna stop playing tennis
link |
just the day someone builds a tennis robot and beat me.
link |
People are still playing chess and even go.
link |
Yeah, and in the very near term even,
link |
some people are advocating basic income, replace jobs.
link |
But if the government is gonna be willing
link |
to just hand out cash to people for doing nothing,
link |
then one should also seriously consider
link |
whether the government should also hire
link |
a lot more teachers and nurses
link |
and the kind of jobs which people often
link |
find great fulfillment in doing, right?
link |
We get very tired of hearing politicians saying,
link |
oh, we can't afford hiring more teachers,
link |
but we're gonna maybe have basic income.
link |
If we can have more serious research and thought
link |
into what gives meaning to our lives,
link |
the jobs give so much more than income, right?
link |
And then think about in the future,
link |
what are the roles that we wanna have people
link |
continually feeling empowered by machines?
link |
And I think sort of, I come from Russia,
link |
from the Soviet Union.
link |
And I think for a lot of people in the 20th century,
link |
going to the moon, going to space was an inspiring thing.
link |
I feel like the universe of the mind,
link |
so AI, understanding, creating intelligence
link |
is that for the 21st century.
link |
So it's really surprising.
link |
And I've heard you mention this.
link |
It's really surprising to me,
link |
both on the research funding side,
link |
that it's not funded as greatly as it could be,
link |
but most importantly, on the politician side,
link |
that it's not part of the public discourse
link |
except in the killer bots terminator kind of view,
link |
that people are not yet, I think, perhaps excited
link |
by the possible positive future
link |
that we can build together.
link |
So we should be, because politicians usually just focus
link |
on the next election cycle, right?
link |
The single most important thing I feel we humans have learned
link |
in the entire history of science
link |
is they were the masters of underestimation.
link |
We underestimated the size of our cosmos again and again,
link |
realizing that everything we thought existed
link |
was just a small part of something grander, right?
link |
Planet, solar system, the galaxy, clusters of galaxies.
link |
And we now know that the future has just
link |
so much more potential
link |
than our ancestors could ever have dreamt of.
link |
This cosmos, imagine if all of Earth
link |
was completely devoid of life,
link |
except for Cambridge, Massachusetts.
link |
Wouldn't it be kind of lame if all we ever aspired to
link |
was to stay in Cambridge, Massachusetts forever
link |
and then go extinct in one week,
link |
even though Earth was gonna continue on for longer?
link |
That sort of attitude I think we have now
link |
on the cosmic scale, life can flourish on Earth,
link |
not for four years, but for billions of years.
link |
I can even tell you about how to move it out of harm's way
link |
when the sun gets too hot.
link |
And then we have so much more resources out here,
link |
which today, maybe there are a lot of other planets
link |
with bacteria or cow like life on them,
link |
but most of this, all this opportunity seems,
link |
as far as we can tell, to be largely dead,
link |
like the Sahara Desert.
link |
And yet we have the opportunity to help life flourish
link |
around this for billions of years.
link |
So let's quit squabbling about
link |
whether some little border should be drawn
link |
one mile to the left or right,
link |
and look up into the skies and realize,
link |
hey, we can do such incredible things.
link |
Yeah, and that's, I think, why it's really exciting
link |
that you and others are connected
link |
with some of the work Elon Musk is doing,
link |
because he's literally going out into that space,
link |
really exploring our universe, and it's wonderful.
link |
That is exactly why Elon Musk is so misunderstood, right?
link |
Misconstrued him as some kind of pessimistic doomsayer.
link |
The reason he cares so much about AI safety
link |
is because he more than almost anyone else appreciates
link |
these amazing opportunities that we'll squander
link |
if we wipe out here on Earth.
link |
We're not just going to wipe out the next generation,
link |
all generations, and this incredible opportunity
link |
that's out there, and that would really be a waste.
link |
And AI, for people who think that it would be better
link |
to do without technology, let me just mention that
link |
if we don't improve our technology,
link |
the question isn't whether humanity is going to go extinct.
link |
The question is just whether we're going to get taken out
link |
by the next big asteroid or the next super volcano
link |
or something else dumb that we could easily prevent
link |
with more tech, right?
link |
And if we want life to flourish throughout the cosmos,
link |
AI is the key to it.
link |
As I mentioned in a lot of detail in my book right there,
link |
even many of the most inspired sci fi writers,
link |
I feel have totally underestimated the opportunities
link |
for space travel, especially at the other galaxies,
link |
because they weren't thinking about the possibility of AGI,
link |
which just makes it so much easier.
link |
So that goes to your view of AGI that enables our progress,
link |
that enables a better life.
link |
So that's a beautiful way to put it
link |
and then something to strive for.
link |
So Max, thank you so much.
link |
Thank you for your time today.
link |
It's been awesome.
link |
Thank you so much.