The following is a conversation with Max Tagmark, his second time in the podcast.
In fact, the previous conversation was episode number one of this very podcast.
He is a physicist and artificial intelligence researcher at MIT, co-founder
of the Future of Life Institute and author of Life 3.0, being human
in the age of artificial intelligence.
He's also the head of a bunch of other huge fascinating projects and has
written a lot of different things that you should definitely check out.
He has been one of the key humans who has been outspoken about long-term
existential risks of AI and also its exciting possibilities and solutions
to real world problems.
Most recently at the intersection of AI and physics and also in re-engineering
the algorithms that divide us by controlling the information we see and
thereby creating bubbles and all other kinds of complex social phenomena
that we see today.
In general, he's one of the most passionate and brilliant people I have
the fortune of knowing.
I hope to talk to him many more times on this podcast in the future.
Quick mention of our sponsors, the Jordan Harbinger Show, Four
Sigmatic Mushroom Coffee, BetterHelp Online Therapy and ExpressVPN.
So the choice is Wisdom, Caffeine, Sanity or Privacy.
Choose wisely, my friends, and if you wish, click the sponsor links below
to get a discount at the support this podcast.
As a side note, let me say that much of the researchers in the machine
learning and artificial intelligence communities do not spend much time
thinking deeply about existential risks of AI because our current algorithms
are seen as useful but dumb.
It's difficult to imagine how they may become destructive to the fabric
of human civilization in the foreseeable future.
I understand this mindset, but it's very troublesome.
To me, this is both a dangerous and uninspiring perspective, reminiscent
of a lobster sitting in a pot of lukewarm water that a minute ago was cold.
I feel a kinship with this lobster.
I believe that already the algorithms that drive our interaction on social
media have an intelligence and power that far outstrip the intelligence
and power of any one human being.
Now really is the time to think about this, to define the trajectory of the
interplay of technology and human beings in our society.
I think that the future of human civilization very well may be at stake
over this very question of the role of artificial intelligence in our society.
If you enjoy this thing, subscribe on YouTube, review it on Apple Podcasts,
follow on Spotify, support on Patreon, or connect with me on Twitter at Lex Freedman.
And now, here's my conversation with Max Tagmark.
So, people might not know this, but you were actually episode number one of this
podcast just a couple of years ago, and now we're back.
And it so happens that a lot of exciting things happened in both physics
and artificial intelligence, both fields that you're super passionate about.
Can we try to catch up to some of the exciting things happening in
artificial intelligence, especially in the context of the way it's cracking,
the way it's cracking, open the different problems of the sciences?
Yeah, I'd love to, especially now as we start 2021 here, it's a really fun time to think about
what were the biggest breakthroughs in AI.
Not the ones necessarily the media wrote about, but they really matter.
And what does that mean for our ability to do better science?
What does it mean for our ability to help people around the world?
And what does it mean for new problems that they could cause if we're not smart enough to avoid them?
So, you know, what do we learn basically from this?
Yes, absolutely.
So, one of the amazing things you're part of is the AI Institute for Artificial Intelligence
and Fundamental Interactions.
What's up with this institute?
What are you working on?
What are you thinking about?
The idea is something I'm very on fire with, which is basically AI meets physics.
And, you know, it's been almost five years now since I shifted my own MIT research from physics
to machine learning.
And in the beginning, I noticed a lot of my colleagues, even though they were polite about it,
well, I kind of, what is Max doing?
What is this weird stuff?
He's lost his mind.
But then gradually, I, together with some colleagues, were able to persuade more and more
of the other professors in our physics department to get interested in this.
And now we got this amazing NSF center, so 20 million bucks for the next five years, MIT,
and a bunch of neighboring universities here also.
And I noticed now those colleagues who were looking at me, it'd be funny, I've stopped asking
what the point is of this, because it's becoming more clear.
And I really believe that, of course, AI can help physics,
it can help physics a lot to do better physics, but physics can also help AI a lot,
both by building better hardware.
My colleague, Martin Solzacic, for example, is working on an optical chip for much faster
machine learning, where the computation is done, not by moving electrons around,
but by moving photons around, dramatically less energy use, faster, better.
We can also help AI a lot, I think, by having a different set of tools and a different,
maybe more audacious attitude.
AI has, to a significant extent, been an engineering discipline, where you're just
trying to make things that work, and being less interested in maybe selling them,
then figuring out exactly how they work, and proving theorems about that they will always work.
Contrast that with physics, when Elon Musk sends a rocket to the International Space Station,
they didn't just train with machine learning, all it's fired a little bit left,
more to the left, a bit more to the right, so that also missed.
Let's try here.
No, we figured out Newton's laws of gravitation, and other things,
and got a really deep fundamental understanding, and that's what gives us such confidence in
rockets, and my vision is that in the future, all machine learning systems that actually
have impact on people's lives will be understood at a really, really deep level,
so we trust them not because some sales rep told us to, but because they've earned our trust,
and really safety-critical things even prove that they will always do what we expect them to do.
That's very much the physics mindset, so it's interesting if you look at big breakthroughs
that have happened in machine learning this year, from dancing robots,
it's pretty fantastic, not just because it's cool, but if you just think about
not that many years ago, this YouTube video at this DARPA challenge, where the MIT robot
comes out of the car and face plants, how far we've come in just a few years.
Similarly, Alpha Fold 2, crushing the protein folding problem, we can talk more about
implications for medical research and stuff, but hey, that's huge progress.
You can look at the GPT-3, they can spout off English texts, which sometimes really,
really blows you away. You can look at the DeepMinds Mu Zero, which doesn't just kick
our butt and go and chest and shogi, but also in all these Atari games,
and you don't even have to teach it the rules now. What all of those have in common is,
besides being powerful, is we don't fully understand how they work.
That's fine if it's just some dancing robots, and the worst thing that can happen is that
they face plant, or if they're playing Go, and the worst thing that can happen is that
they make a bad move and lose the game. It's less fine if that's what's controlling your
self-driving car or your nuclear power plant. We've seen already that even though Hollywood
had all these movies where they try to make us worry about the wrong things,
like machines turning evil, the actual bad things that have happened with automation
have not been machines turning evil. They've been caused by over-trust in things we
didn't understand as well as we thought we did. Even very simple automated systems like
what Boeing put into the 737 MAX killed a lot of people. Was it that that little
simple system was evil? Of course not, but we didn't understand it as well as we should have.
And we trusted without understanding. Exactly. We didn't even understand that we didn't
understand. The humility is really at the core of being a scientist. I think step one,
if you want to be a scientist, is don't ever fool yourself into thinking you understand
things when you actually don't. That's probably good advice for humans in general.
I think humility in general can do us good. But in science, it's so spectacular. Why
did we have the wrong theory of gravity ever from Aristotle onward and close until Galileo's
time? Why would we believe something so dumb as that if I throw this water bottle, it's going to
go up with constant speed until it realizes that its natural motion is down? It changes its mind.
People just kind of assumed Aristotle was right. He's an authority. We understand that.
Why did we believe things like that the sun is going around the earth?
Why did we believe that time flows at the same rate for everyone until Einstein?
Same exact mistake over and over again. We just weren't humble enough to acknowledge that we
actually didn't know for sure. We assumed we knew. So we didn't discover the truth because we assumed
there was nothing there to be discovered, right? There was something to be discovered about the
737 max. And if you had been a bit more suspicious and tested it better, we would have found it.
And it's the same thing with most harm that's been done by automation so far, I would say.
Did you hear of a company called Knight Capital?
So good. That means you didn't invest in them earlier. They deployed this automated rating system.
All nice and shiny. They didn't understand it as well as they thought,
and it went about losing 10 million bucks per minute for 44 minutes straight
until someone presumably was like, oh no, shut this up. It wasn't evil? No. It was, again,
misplaced trust, something they didn't fully understand, right? And there have been so many,
even when people have been killed by robots, it's just quite rare still. But in factory accidents,
it's in every single case been not malice, just that the robot didn't understand that a human
is different from an auto part or whatever. And so this is where I think there's so much
opportunity for a physics approach where you just aim for a higher level of understanding.
And if you look at all these systems that we talked about from reinforcement learning systems
and dancing robots to all these neural networks that power GPT-3 and go playing software stuff,
they're all basically black boxes, much like not so different from if you teach a human something,
you have no idea how their brain works, right? Except the human brain at least has been error
corrected during many, many centuries of evolution in a way that some of these systems have not,
right? And my MIT research is entirely focused on demystifying this black box,
intelligible intelligence is my slogan. That's a good line, intelligible intelligence.
Yeah, that we shouldn't settle for something that seems intelligent, but it should be intelligible
so that we actually trust it because we understand it, right? Like, again, Elon trusts his rockets
because he understands Newton's laws and thrust and how everything works.
And let me tell you, can I tell you why I'm optimistic about this?
Yes. I think there is, we've made a bit of a mistake where we,
some people still think that somehow we're never going to understand neural networks.
And we're just going to have to learn to live with this. It's this very powerful black box.
Basically, for those who haven't spent time building their own, it's super simple what
happens inside. You send in a long list of numbers, and then you do a bunch of operations on them,
multiply by matrices, et cetera, et cetera, and some other numbers come out that's output of it.
And then there are a bunch of knobs you can tune. And when you change them, you know, it affects
the computation, the input output relation. And then you just give the computer some definition
of good, and it keeps optimizing these knobs until it performs as good as possible. And often,
you go like, wow, that's really good. This robot can dance. Or this machine is beating me a chest
now. And then in the end, you have something, which even though you can look inside it, you have
very little idea of how it works. You know, you can print out tables of all the millions of
parameters in there. Is it crystal clear now how it's working? And of course not, right?
Many of my colleagues seem willing to settle for that. And I'm like, no, that's like the halfway
point. Now, some have even gone as far as sort of guessing that the mister, the inscrutability of
this is where the some of the power comes from and sort of some sort of mysticism. I think that's
total nonsense. I think the real power of neural networks comes not from inscrutability, but from
differentiability. And what I mean by that is simply that the output depends changes only smoothly
if you tweak your knobs. And then you can use all these powerful methods we have for optimization
in science, we can just tweak them a little bit and see, did that get better or worse?
That's the fundamental idea of machine learning, that the machine itself can keep optimizing
until it gets better. Suppose you wrote this algorithm instead in Python or some other programming
language. And then what what the knobs did was they just changed random letters in your in your
code. Now we'll just epically fail, right? You change one thing. And instead of saying print,
it says SYNTH syntax error, you don't even know, was that for the better or for the worse, right?
This to me is this is what I believe is the fundamental power of neural networks.
And just to clarify the changing the different letters in a program would not be a differentiable
process. It would make it an invalid program, typically. And then you wouldn't even know if
you changed more letters, if it would make it work again, right? So that's the magic of neural
networks, the inscrubility, the differentiability that every every setting of the parameters is a
program. And you can tell is it better or worse, right? And so, so you don't like the poetry of
the mystery of neural networks as the source of its power? I generally like poetry, but not in this
case. It's so misleading. And above all, it's short changes. It's failed. It makes us underestimate
what we can the good things we can accomplish. Because so what we've been doing in my group
is basically step one, train the mysterious neural network to do something well. And then step two,
do some additional AI techniques to see if we can not transform this black box into something
equally intelligent that you can actually understand. So for example, I'll give you one example, this
AI Feynman project that we just published, right? So we took the 100 most famous or complicated
equations from one of my favorite physics textbooks, in fact, the one that got me into physics in the
first place, Feynman lectures on physics. And so you have a formula, you know, maybe it has
what goes into the formula is six different variables. And then what comes out as one. So then
you can make like a giant Excel spreadsheet with seven columns, you put in just random numbers
for the six columns for those six input variables. And then you calculate with the formula of the
seventh column, the output. So maybe it's like the force equals in the last column, some function
of the other. And now the task is, okay, if I don't tell you what the formula was,
can you figure that out from looking at my spreadsheet I gave you this problem is called
symbolic regression. If I tell you that the formula is what we call a linear formula. So it's just
that the output is some sum of all the things input the time some constants, that's the famous
easy problem we can solve. We do it all the time in science and engineering. But the general one,
if it's more complicated functions with logarithms or cosines or other math,
it's a very, very hard one. And probably impossible to do fast in general, just because
the number of formulas with n symbols, not just grows exponentially, just like the number of
passwords you can make grow dramatically with length. But we had this idea that if you first
have a neural network that can actually approximate the formula, you just trained it,
even if you don't understand how it works, that can be the first step towards actually
understanding how it works. So that's what we do first. And then we study that neural network now
and put in all sorts of other data that wasn't in the original training data and
use that to discover simplifying properties of the formula. And that lets us break it apart
often into many simpler pieces in a kind of divide and conquer approach. So we were able
to solve all of those hundred formulas, discover them automatically, plus a whole bunch of other
ones. And it's actually kind of humbling to see that this code, which anyone who wants now is
listening to this can type pip install AI Feynman on the computer and run it. It can actually do
what Johannes Kepler spent four years doing when he stared at Mars data. And he was like,
finally, Eureka, this is an ellipse. This will do it automatically for you in one hour, right?
Or Max Planck, he was looking at how much radiation comes out from the different wavelengths from a
hot object and discovered the famous black body formula. This discovers it automatically.
I'm actually excited about seeing if we can discover not just old formulas again,
but new formulas that no one has seen before.
And do you like this process of using kind of a neural network to find some basic insights and
then dissecting the neural network to then gain the final so that that's in that way you've
forcing the explainability issue, you know, really trying to analyze a neural network
for the things it knows in order to come up with the final beautiful simple theory
underlying the whole the initial system that you were looking at.
I love that. And the reason I'm so optimistic that it can be generalized to so much more is because
that's exactly what we do as human scientists. Think of Galileo whom we mentioned, right?
I bet when he was a little kid, if his dad threw him an apple, he would catch it.
Why? Because he had a neural network in his brain that he had trained to predict the parabolic
orbit of apples that are thrown under gravity. If you throw a tennis ball to a dog, it also
has this same ability of deep learning to figure out how the ball is going to move and catch it.
But Galileo went one step further when he got older. He went back and was like, wait a minute.
I can write down a formula for this. Y equals X squared, a parabola, you know, and he helped
revolutionize physics as we know it, right? So there was a basic neural network in there from
childhood that captured like the experiences of observing different kinds of trajectories.
And then he was able to go back in with another extra little neural network and analyze all
those experiences and be like, wait a minute, there's a deeper rule here. Exactly. He was able
to distill out in symbolic form what that complicated black box neural network was doing, right?
Not only did he, the formula he got ultimately become more accurate, you know, and similarly,
this is how Newton got Newton's laws, which is why Elon can send rockets to the space station now,
right? So it's not only more accurate, but it's also simpler, much simpler, and it's so simple
that we can actually describe it to our friends and each other, right? We've talked about it just
in the context of physics now. But hey, you know, isn't this what we're doing when we're talking to
each other? Also, we go around with our neural networks, just like dogs and cats and chipmunks
and blue jays, and we experience things in the world. But then we humans do this additional
step on top of that, where we then distill out certain high level knowledge that we've extracted
from this in a way that we can communicate it to each other in a symbolic form in English,
in this case, right? So if we can do it, and we believe that we are information processing
entities, then we should be able to make machine learning that does it also.
Well, do you think the entire thing could be learning? Because there, this dissection process,
like for AI Feynman, the secondary stage feels like something like reasoning. And the initial
step feels like more like the more basic kind of differentiable learning. Do you think the whole
thing could be differentiable learning? Do you think the whole thing could be basically neural
networks on top of each other? It's like turtles all the way down? Can it be neural networks all
the way down? I mean, that's a really interesting question. We know that in your case, it is neural
networks all the way down because that's all you have in your skull as a bunch of neurons
doing their thing. But if you ask the question more generally, what algorithms are being used
in your brain, I think it's super interesting to compare. I think we've gotten a little bit
backwards historically because we humans first discovered good old fashioned AI, the logic based
AI that we often call Go-Fi for good old fashioned AI. And then more recently, we did machine learning
because it required bigger computers. So we had to discover it later. So we think of machine
learning with neural networks as the modern thing and the logic based AI as the old fashioned thing.
But if you look at evolution on Earth, it's actually been the other way around. I would say
that, for example, an eagle has a better vision system than I have using and dogs are just as
good at casting tennis balls as I am. All this stuff which is done by training in neural network
and not interpreting it in words is something so many of our animal friends can do at least as
well as us. What is it that we humans can do that the chipmunks and the eagles cannot?
It's more to do with this logic based stuff where we can extract out information in symbols
in language and now even with equations if you're a scientist. So basically what happened was first
we built these computers that could multiply numbers real fast and manipulate symbols and we
felt they were pretty dumb. Then we made neural networks that can see as well as a cat can and
do a lot of this inscrutable black box neural networks. What we humans can do also is put the
two together in a useful way. In our own brain. Yes, in our own brain. So if we ever want to get
artificial general intelligence that can do all jobs as well as humans can, then that's what's
going to be required to be able to combine the neural networks with symbolic, combine the old
AI with a new AI in a good way. We do it in our brains and there seems to be basically two strategies
I see in industry now. One scares the heebie-jeebies out of me and the other one I find much more
encouraging. Okay, can we break them apart? Which other two? The one that scares the heebie-jeebies
out of me is this attitude that we're just going to make ever bigger systems that we still don't
understand until they can be as smart as humans. What could possibly go wrong? I think it's just
such a reckless thing to do and unfortunately, and if we actually succeed as a species to build
artificial general intelligence, then we still have no clue how it works. I think at least 50%
chance we're going to be extinct before too long. It's just going to be an utter epic own goal.
So that 44-minute losing money problem or the paperclip problem where we don't understand
how it works and it's just in a matter of seconds runs away in some kind of direction that's going
to be very problematic. Even long before you have to worry about the machines themselves somehow
deciding to do things and to us that we have to worry about people using machines that are short
of AI, AGI and power to do bad things. I mean, just take a moment and if anyone who's not worried
particularly about advanced AI, just take 10 seconds and just think about your least favorite
leader on the planet right now. Don't tell me who it is. You want to keep this apolitical,
but just see the face in front of you, that person for 10 seconds. Now imagine that that
person has this incredibly powerful AI under their control and can use it to impose their
will on the whole planet. How does that make you feel? Yeah. Can we break that apart just
briefly? For the 50% chance that we'll run into trouble with this approach, do you see the bigger
worry in that leader or humans using the system to do damage or are you more worried? And I think
I'm in this camp more worried about like accidental unintentional destruction of everything. So like
humans trying to do good and in a way where everyone agrees it's kind of good, it's just
that they're trying to do good without understanding. Because I think every evil leader in history
to some degree thought they're trying to do good. Oh yeah, I'm sure Hitler thought he was doing
good. I've been reading a lot about Stalin. I'm sure Stalin legitimately thought that communism
was good for the world and that he was doing good. I think Mao Zedong thought what he was doing
with a great leap forward was good too. Yeah. I'm actually concerned about both of those.
Before I promised to answer this in detail, but before we do that, let me finish answering the
first question because I told you that there were two different routes we could get to artificial
general intelligence and one scares the FPGVs out of me, which is this one where we build something,
we just say bigger neural networks, ever more hardware and it's just trained like more data and
poof. Now it's very powerful. That I think is the most unsafe and reckless approach. The alternative
to that is the intelligible intelligence approach instead, where we say neural networks is just a
tool for the first step to get the intuition. But then we're going to spend also serious
resources on other AI techniques for demystifying this black box and figuring out what it's actually
doing so we can convert it into something that's equally intelligent, but that we actually understand
what it's doing. Maybe we can even prove theorems about it that this car here will never be hacked
when it's driving because here's the proof. There is a whole science of this. It doesn't work for
neural networks. There are big black boxes, but it works well and certain other kinds of codes.
That approach I think is much more promising. That's exactly why I'm working on it, frankly,
not just because I think it's cool for science, but because I think the more we understand these
systems, the better the chances that we can make them do the things that are good for us that are
actually intended, not unintended. So you think it's possible to prove things about something
as complicated as a neural network? That's the hope? Well, ideally, there's no reason
there has to be a neural network in the end either. We discovered that Newton's laws of gravity
with neural network in Newton's head, but that's not the way it's programmed into the
navigation system of Elon Musk's rocket anymore. It's written in C++ or I don't know what language
he uses exactly. Then there are software tools called symbolic verification. DARPA in the U.S.
Military has done a lot of really great research on this because they really want to understand
that when they build weapon systems, they don't just go fire at random or malfunction.
There's even a whole operating system called cell three that's been developed by a DARPA grant
where you can actually mathematically prove that this thing can never be hacked. Well, one day,
I hope that will be something you can say about the OS that's running on our laptops too, as you
know, but we're not there. But I think we should be ambitious, frankly. And if we can use machine
learning to help do the proofs and so on as well, then it's much easier to verify that a proof is
correct than to come up with a proof in the first place. That's really the core idea here.
If someone comes on your podcast and says they proved the Riemann hypothesis or some sensational
new theorem, it's much easier for someone else to take some math grad students and check,
oh, there's an error here on equation five, or this really checks out than it was to discover
the proof. Yeah, although some of those proofs are pretty complicated, but yes, it's still
nevertheless much easier to verify the proof. I love the optimism. Even with the security of
systems, there's a cynicism that pervades people who think about this, which is like, oh, it's
hopeless. In the same sense, exactly like you're saying when you own networks, oh, it's hopeless
to understand what's happening. With security, people are just like, well, there's always going
to be attack vectors and ways to attack the system, but you're right. We're just very new
with these computational systems. We're new with these intelligent systems, and it's not out of the
realm of possibilities. Just like people that understand the movement of the stars and the
planets and so on. It's entirely possible that within hopefully soon, but it could be within
100 years, we start to have an obvious laws of gravity about intelligence. God forbid about
consciousness too, that one. Agreed. Of course, if you're selling computers that get hacked a
lot that's in your interest as a company that people think it's impossible to make it safe,
so nobody's going to get the idea of suing you, but I want to really inject optimism here.
It's absolutely possible to do much better than we're doing now. Your laptop does so much stuff.
You don't need the music player to be super safe in your future self-driving car. If someone hacks
it and starts playing music, you don't like the world on end. But what you can do is you
can break out and say that your drive computer that controls your safety must be completely
physically decoupled entirely from the entertainment system, and it must physically be such that it
can't take on over-the-air updates while you're driving. It can have ultimately some operating
system on it, which is symbolically verified and proven. It's always going to do what it's
supposed to do. And companies should take that attitude. They should look at everything they do
and say, what are the few systems in our company that threaten the whole life of the company
if they get hacked and have the highest standards for them, and then they can save money by going
for the El Chippo poorly understood stuff for the rest. This is very feasible, I think.
Coming back to the bigger question that you worried about, that there'll be unintentional
failures, I think, there are two quite separate risks here. We talked a lot about one of them,
which is that the goals are noble of the human. The human says, I want this airplane to not crash
because this is not Muhammad Atta and now flying the airplane. And now there's this technical
challenge of making sure that the autopilot is actually going to behave as the pilot wants.
If you set that aside, there's also the separate question. How do you make sure
that the goals of the pilot are actually aligned with the goals of the passenger? How do you
make sure very much more broadly that if we can all agree as a species that we would like things
to kind of go well for humanity as a whole, that the goals are aligned here, the alignment problem.
And yeah, there's been a lot of progress in the sense that there's suddenly
huge amounts of research going on about it. I'm very grateful to Elon Musk for giving us that
money five years ago so we could launch the first research program on technical AI safety
and alignment. There's a lot of stuff happening. But I think we need to do more than just make
sure little machines do always what their owners do. That wouldn't have prevented September 11.
If Mohammed Atta said, okay, autopilot, please fly into World Trade Center. And it's like, okay.
That even happened. In a different situation, there was this depressed pilot named Andreas
Lubitz who told his German wings passenger jet to fly into the Alps. He just told the computer to
change the altitude to 100 meters or something like that. And you know what the computer said?
Right. Okay. And it had the freaking topographical map of the Alps in there. It had GPS, everything.
No one had bothered teaching it even the basic kindergarten ethics of like, no, we never want
airplanes to fly into mountains under any circumstances. And so we have to think beyond
just the technical issues and think about how do we align in general incentives on this planet
for the greater good. So starting with simple stuff like that, every airplane that has a computer in
it should be taught whatever kindergarten ethics it's smart enough to understand. Like, no, don't
fly into fixed objects. If the pilot tells you to do so, then go on autopilot mode, send an email to
the cops and land at the latest airport, nearest airport, you know, any car with a forward facing
camera should just be programmed by the by the manufacturers that it will never accelerate into
a human ever. That would avoid things like the niece attack and many horrible terrorist vehicle
attacks where they deliberately did that, right? This was not some sort of thing. Oh, you know,
US and China, different views on no, there was not a single car manufacturer on the world in
the world who wanted the cars to do this. They just hadn't thought to do the alignment. And if
you look at more broadly problems that happen on this planet, the vast majority have to do a poor
alignment. I mean, think about this go back really big, because I know this is you're so good at
that. Yeah, in the very so long ago in evolution, we had these genes. And they wanted to make copies
of themselves. That's really all they cared about. So they some genes say, Hey, I'm going to
build a brain on this body I'm in so that I can get better at making copies myself. And then they
decided for their benefit to get copied more to align your brain's incentives with their
incentives. So it didn't want you to starve to death. So it gave you an incentive to eat.
And it wanted you to make copies of the genes. So it gave you an incentive to fall in love and
do all sorts of naughty things to make copies of itself, right? So that was successful value
alignment done on the genes. They created something more intelligent than themselves,
but they made sure to try to align the values. But then something went a little bit
against the idea of what the genes wanted, because a lot of humans discovered, Hey, you know,
yeah, we really like this business about sex that the genes have made us enjoy.
But we don't want to have babies right now. So we're going to hack the genes and use birth control.
And I really feel like drinking a Coca-Cola right now, but I don't want to get a potbelly. So I'm
going to drink Diet Coke. We have all these things we've figured out, because we're smarter than the
genes, how we can actually subvert their intentions. So it's not surprising that we humans now,
when we're in the role of these genes, creating other non-human entities with a lot of power,
have to face the same exact challenge. How do we make other powerful entities,
have incentives that are aligned with ours, and that so they won't hack them? Corporations,
for example, right? We humans decided to create corporations because it can benefit us greatly.
Now all of a sudden there's a supermarket. I can go buy food there. I don't have to hunt. Awesome.
And then to make sure that this corporation would do things that were good for us and not
bad for us, we created institutions to keep them in check. Like if the local supermarket sells
poisonous food, then the owners of the supermarket have to spend some years reflecting behind bars,
right? So we created incentives to get to align them. But of course, just like we were able to
see through this thing, well, birth control, if you're a powerful corporation, you also have an
incentive to try to hack the institutions that are supposed to govern you. Because you ultimately,
as a corporation, have an incentive to maximize your profits. Just like you have an incentive
to maximize the enjoyment your brain has, not for your genes. So if they can figure out a way of
bribing regulators, then they're going to do that. In the US, we kind of caught on to that and
made laws against corruption and bribery. Then in the late 1800s, Teddy Roosevelt realized that,
no, we were still being kind of hacked because the Massachusetts railroad companies had like a
bigger budget than the state of Massachusetts. And they were doing a lot of very corrupt stuff.
So he did the whole trust busting thing to try to align these other non human entities,
the companies, again, more with the incentives of Americans as a whole. But it's not surprising,
though, that this is a battle you have to keep fighting. Now, we have even larger companies
than we ever had before. And of course, they're going to try to, again, support the institutions,
not because I think people make a mistake of getting all too black thinking about things in
terms of good and evil, like arguing about whether corporations are good or evil or whether
robots are good or evil. A robot isn't good or evil. It's tool. And you can use it for great
things like robotic surgery or for bad things. And a corporation also is a tool, of course.
And if you have good incentives to the corporation, it'll do great things like start a
hospital or a grocery store. If you have really bad incentives, then it's going to start maybe
with marketing addictive drugs to people, and you'll have an opioid epidemic, right?
It's all about, we should not make a mistake of getting into some sort of fairy tale good,
evil thing about corporations or robots. We should focus on putting the right incentives
in place. My optimistic vision is that if we can do that, then we can really get good things.
We're not doing so great with that right now, either on AI, I think, or on other
intelligent, non-human entities like big companies. We just have a new
secretary of defense. There's going to start up now in the Biden administration who is,
was an active member of the board of Raytheon. I have nothing against Raytheon.
I'm not a pacifist, but there's an obvious conflict of interest if someone
is in the job where they decide who's they're going to contract with.
Maybe we need another Teddy Roosevelt to come along again and say, hey, we want what's good
for all Americans, and we need to go do some serious realigning again of the incentives
that we're giving to these big companies, and then we're going to be better off.
It seems that naturally with human beings, just like you beautifully described the history of
this whole thing, it all started with the genes and they're probably pretty upset by all the
unintended consequences that happened since. It seems that it kind of works out. It's in this
collective intelligence that emerges at the different levels. It seems to find, sometimes
last minute, a way to realign the values or keep the values aligned. It finds a way.
Like different leaders, different humans pop up all over the place that reset the system.
Do you have an explanation why that is, or is that just survivor bias, and also is that
different, somehow fundamentally different than with the AI systems where you're no longer dealing
with something that was a direct, maybe companies are the same, a direct byproduct of the evolutionary
process? I think there is one thing which has changed. That's why I'm not all optimistic.
That's why I think there's about a 50% chance if we take the dumb route with artificial
intelligence that humanity will be extinct in this century. First, just the big picture.
Yeah, companies need to have incentives. Even governments. We used to have governments. Usually
there were just some king who was the king because his dad was the king. Then there were some benefits
of having this powerful kingdom or empire of any sort because then it could prevent a lot of local
squabbles. So at least everybody in that region would stop warring against each other and their
incentives of different cities in the kingdom became more aligned. That was the whole selling
point. Harari has a beautiful piece on how empires were collaboration enablers. Harari
says invented money for that reason so we could have better alignment and we could trade even
with people we didn't know. This stuff has been playing out since time immemorial. What's changed
is that it happens on ever larger scales. The technology keeps getting better because science
gets better. Now we can communicate over larger distances, transport things faster over larger
distances. The entities get ever bigger but our planet is not getting bigger anymore. In the past,
you could have one experiment that just totally screwed up like Easter Island where they actually
managed to have such poor alignment that when they went extinct, people there, there was no one
else to come back and replace them. If Elon Musk doesn't get us to Mars and then we go extinct
on a global scale and we're not coming back, that's the fundamental difference.
And that's a mistake you would rather we don't make for that reason. In the past, of course,
history is full of fiascos but it was never the whole planet. And then okay, now there's this
nice uninhabited land here, some other people could move in and organize things better. This is
different. The second thing which is also different is that technology gives us so much more
empowerment both to do good things and also to screw up. In the Stone Age, even if you had someone
whose goals were really poorly aligned, maybe he was really pissed off because his Stone Age
girlfriend dumped him and he just wanted to kill as many people as he could. How many could he
really take out with a rock and a stick before he was overpowered? Just handful, right? Now,
with today's technology, if we have an accidental nuclear war between Russia and the US, which we
almost have about a dozen times and then we have a nuclear winter, it could take out seven billion
people or six billion people or we don't know. So the scale of damage is bigger than we can do. There's
obviously no law of physics that says that technology will never get powerful enough that
we could wipe out our species entirely. That would just be fantasy to think that science is
somehow doomed not to get more powerful than that. And it's not at all unfeasible in our lifetime
that someone could design a designer pandemic, which spreads as easily as COVID, but just basically
kills everybody. We already had smallpox. They killed one third of everybody who got it.
What do you think of the, here's an intuition, maybe it's completely naive and this optimistic
intuition I have, which it seems, and maybe it's a biased experience that I have, but it seems like
the most brilliant people I've met in my life, all are really like fundamentally good human beings.
And not like naive, good, like they really want to do good for the world in a way that
well maybe is aligned to my sense of what good means. And so I have a sense that the
people that will be defining the very cutting edge of technology, there will be much more
of the ones that are doing good versus the ones that are doing evil. So the race
optimistic on the us always like last minute coming up with a solution. So if there's an
engineered pandemic that has the capability to destroy most of the human civilization,
it feels like to me either leading up to that before or as it's going on, there will be,
we're able to rally the collective genius of the human species. I could tell by your smile
that you're at least some percentage doubtful, but could that be a fundamental law of human
nature? That evolution only creates karma is beneficial, good is beneficial, and therefore
will be all right. I hope you're right. I would really love it if you're right. If there's some
sort of law of nature that says that we always get lucky in the last second because of karma, but
I prefer not playing it so close and gambling on that. And I think, in fact, I think it can
be dangerous to have too strong faith in that because it makes us complacent. Like if someone
tells you, you never have to worry about your house burning down, then you're not going to put
in a smoke detector because why would you need to? Even if it's sometimes very simple precautions,
we don't take them. If you're like, oh, the government is going to take care of everything
for us, I can always trust my politicians. We advocate our own responsibility. I think it's
a healthier attitude to say, yeah, maybe things will work out, but maybe I'm actually going to
have to myself step up and take responsibility. And the stakes are so huge. I mean, if we do this
right, we can develop all this ever more powerful technology and cure all diseases and create a
future where humanity is healthy and wealthy for not just the next election cycle, but like
billions of years throughout our universe. That's really worth working hard for and not just
sitting and hoping for some sort of fairytale karma.
Well, I just mean, so you're absolutely right. From the perspective of the individual,
for me, the primary thing should be to take responsibility and to build the solutions that
your skill set allows to build. Yeah, which is a lot. I think we underestimate often very much
how much good we can do. If you or anyone listening to this is completely confident
that our government would do a perfect job on handling any future crisis with engineered pandemics
or future AI. The one or two people out there. On what actually happened in 2020.
Do you feel that the government by and large around the world is handled flawlessly?
That's a really sad and disappointing reality that hopefully is a wake up call for everybody.
For the scientists, for the engineers, for the researchers in AI especially,
it was disappointing to see how inefficient we were at collecting the right amount of data
in a privacy preserving way and spreading that data and utilizing that data to make decisions,
all that kind of stuff. Yeah, I think when something bad happens to me, I made myself
a promise many years ago that I would not be a whiner. So when something bad happens to me,
of course, it's just a process of disappointment. But then I try to focus on what did I learn from
this that can make me a better person in the future. And there's usually something to be
learned when I fail. And I think we should all ask ourselves, what can we learn from
the pandemic about how we can do better in the future? And you mentioned there's
a really good lesson. You know, we were not as resilient as we thought we were.
And we were not as prepared, maybe as we wish we were. You can even see very stark contrast
around the planet. South Korea, they have over 50 million people. Do you know how many deaths
they have from COVID last time I checked? No, it's about 500. Why is that? Well,
the short answer is that they had prepared. They were incredibly quick, incredibly quick to get on
it with very rapid testing and contact tracing and so on, which is why they never had more cases
than they could contract trace effectively, right? They never even have to have the kind
of big lockdowns we had in the West. But the deeper answer to it's not just Koreans are just
somehow better people. The reason I think they were better prepared was because they
had already had a pretty bad hit from the SARS pandemic, which never became a pandemic,
something like 17 years ago, I think. So it's kind of fresh memory that, you know,
we need to be prepared for pandemics. So they were, right? And so maybe this is a lesson here
for all of us to draw from COVID, that rather than just wait for the next pandemic or the next
problem with AI getting out of control or anything else, maybe we should just actually
set aside a tiny fraction of our GDP to have people very systematically do some horizon
scanning and say, okay, what are the things that could go wrong? And let's do get out and see which
are the more likely ones and which are the ones that are actually actionable and then be prepared.
So one of the observations is one little ant slash human that I am of disappointment is the
political division over information that has been observed that I observed this year,
that it seemed the discussion was less about sort of what happened and understanding what
happened deeply and more about there's different truths out there. And it's like an argument,
my truth is better than your truth. And it's like red versus blue or different, like it was like
this ridiculous discourse that doesn't seem to get at any kind of notion of the truth. It's not
like some kind of scientific process. Even science got politicized in ways that's very
hard breaking to me. You have an exciting project on the AI front of trying to rethink
one of the, you mentioned corporations, there's one of the other collective intelligence systems
that have emerged through all of this is social networks and just to spread the internet is
the spread of information on the internet, our ability to share that information. There's all
different kinds of news sources and so on. And so you said like, that's from first principles,
let's rethink how we think about the news, how we think about information. Can you talk about this
amazing effort that you're undertaking? Oh, I'd love to. This has been my big COVID projects,
it's been nights and weekends on ever since the lockdown. The segue into this actually,
let me come back to what you said earlier that you had this hope that in your experience people
who you felt were very talented, often idealistic and wanted to do good. Frankly, I feel the same
about all people by and large. There are always exceptions, but I think the vast majority of
everybody regardless of education and whatnot really are fundamentally good. So how can it be
that people still do so much nasty stuff? I think it has everything to do with this,
with the information that we're given. If you go into Sweden 500 years ago and you start telling
all the farmers that those Danes in Denmark, they're so terrible people and we have to
invade them because they've done all these terrible things that you can't fact check
yourself. A lot of people in Sweden did that. And we've seen so much of this today in the world,
both geopolitically, where we are told that China is bad and Russia is bad and Venezuela is bad
and people in those countries are often told that we are bad. And we also see it at a micro level
where people are told that, oh, those who voted for the other party are bad people.
It's not just an intellectual disagreement, but they're bad people and we're getting ever
more divided. And so how do you reconcile this with intrinsic goodness in people? I think it's
pretty obvious that it has again to do with this, with the information that we're fed and given.
We evolved to live in small groups where you might know 30 people in total. So you
then had a system that was quite good for assessing who you could trust and who you
could not. And if someone told you that Joe there is a jerk, but you had interacted with him
yourself and seen him in action, you would quickly realize maybe that that's actually
not quite accurate. But now that the most people on the planet are people we've never met,
it's very important that we have a way of trusting the information we're given.
So okay, so where does the news project come in? Well, throughout history, you can go read
Machiavelli from the 1400s and you'll see how already then there were busy manipulating people
with propaganda and stuff. Propaganda is not new at all. And the incentives to manipulate people
is just not new at all. What is it that's new? What's new is machine learning meets propaganda.
That's what's new. That's why this has gotten so much worse. Some people like to blame certain
individuals like in my liberal university bubble, many people blame Donald Trump and say it was his
fault. I see it differently. I think Donald Trump just had this extreme skill at playing this game
in the machine learning algorithm age. A game he couldn't have played 10 years ago. So what's
changed? What's changed is, well, Facebook and Google and other companies. And I don't want,
I'm not bad man, I think them, I have a lot of friends who work for these companies,
the good people, they deployed machine learning algorithms just to increase their profit a little
bit to just maximize the time people spent watching ads. And they had totally underestimated how
effective they were going to be. This was again, the black box, non intelligible intelligence.
They just noticed, oh, we're getting more ad revenue. Great. It took a long time
until even realized why and how damaging this was for society. Because of course,
what the machine learning figured out was that the by far most effective way of gluing you to
your little rectangle was to show you things that triggered strong emotions, anger, et cetera,
resentment. And if it was true or not, didn't really matter. It was also easier to find stories
that weren't true. If you weren't limited, that's just the limitation to show people.
That's a very limiting fact.
Before long, we've got these amazing filter bubbles on a scale we had never seen before.
A couple of this to the fact that also the online news media were so effective that they
killed a lot of print journalism. There's less than half as many journalists now in America,
I believe, as there was a generation ago. He just couldn't compete with the online
advertising. So all of a sudden, most people are not getting even reading newspapers. They get
their news from social media. And most people only get news in their little bubble. So
along comes now some people like Donald Trump who figured out among the first successful
politicians to figure out how to really play this new game and become very, very influential.
But I think Donald Trump took advantage of it. The fundamental conditions were created by machine
learning taking over the news media. So this is what motivated my little COVID project here.
I said before, machine learning and tech in general is not evil, but it's also not good.
It's just a tool that you can use for good things or bad things. And as it happens,
machine learning and news was mainly used by the big players, big tech, to manipulate people
and to watch as many ads as possible, which had this unintended consequence of really screwing
up our democracy and fragmenting it into filter bubbles. So I thought, well, machine learning
algorithms are basically free. They can run on your smartphone for free also if someone gives
them away to you, right? There's no reason why they only have to help the big guy to manipulate
the little guy. They can just as well help the little guy to see through all the manipulation
attempts from the big guy. So did this project, you can go to improvethenews.org. The first thing
we've built is this little news aggregator. Looks a bit like Google News, except it has
these sliders on it to help you break out of your filter bubble. So if you're reading,
you can click, click and go to your favorite topic. And then if you just slide the left,
right slider away all the way over to the left. There's two sliders, right? Yeah. There's the
one, the most obvious one is the one that has left, right labeled on us. You go to left,
you get one set of articles, you go to the right, you see a very different truth appearing.
Well, that's literally left and right on the political spectrum. Yeah. So if you're reading
about immigration, for example, it's very, very noticeable. And I think step one always,
if you want to not get manipulated is just to be able to recognize the techniques people use. So
it's very helpful to just see how they spin things on the two sides. I think many people
are under the misconception that the main problem is fake news. It's not. I had an amazing team of
MIT students where we did an academic project to use machine learning to detect the main kinds of
bias over the summer. And yes, of course, sometimes there's fake news where someone just claims
something that's false, right? Like, oh, Hillary Clinton just got divorced or something. Yes.
But what we see much more of is actually just omissions. If you go to, there's some stories
which just won't be mentioned by the left or the right because it doesn't suit their agenda.
And then they also mentioned other ones very, very, very much. So, for example, we've had
a number of stories about the Trump family's financial dealings. And then there's been
a bunch of stories about the Biden families, Hunter Biden's financial dealings, right?
Surprise, surprise. They don't get equal coverage on the left and the right.
One side loves to cover the Biden, Hunter Biden's stuff. And one side loves to cover the Trump.
You can never guess which is which, right? But the great news is if you're a normal American
citizen and you dislike corruption in all its forms, then slide, slide, you can just look
at both sides and you'll see all those political corruption stories. It's really liberating to
just take in the both sides, the spin on both sides. It somehow unlocks your mind to like
think on your own, to realize that, I don't know, it's the same thing that was useful, right,
in the Soviet Union times for when everybody was much more aware that they're surrounded by
propaganda, right? That is so interesting what you're saying, actually. So, Noam Chomsky used to
be our MIT colleague once said that propaganda is to democracy, what violence is to totalitarianism.
And what he means by that is if you have a really totalitarian government, you don't need
propaganda. People will do what you want them to do anyway, out of fear, right? But otherwise,
you need propaganda. So, I would say actually that the propaganda is much higher quality
in democracies, much more believable. And it's brilliant. It's really striking when I talk to
colleagues, science colleagues like from Russia and China and so on, I noticed they are actually
much more aware of the propaganda in their own media than many of my American colleagues are
about the propaganda in Western media. That's brilliant. That means the propaganda in the
Western media is just better. That's so brilliant. But once you realize that, you realize there's
also something very optimistic there that you can do about it, right? Because, first of all, omissions,
as long as there's no outright censorship, you can just look at both sides and pretty quickly
these pieces together, a much more accurate idea of what's actually going on, right?
And develop a natural skepticism, too. Yeah.
Analytical scientific mind about what you're taking information.
Yeah. And I think I have to say, sometimes I feel that some of us in the academic bubble are too
arrogant about this and somehow think, oh, it's just people who aren't as educated. When we are
often just as gullible also, because we read only our media and don't see through things,
anyone who looks at both sides like this in comparison will immediately start noticing
the shenanigans being pulled. And I think what I tried to do with this app is that the
big tech has, to some extent, tried to blame the individual for being manipulated much like
big tobacco tried to blame the individuals entirely for smoking. And then later on,
you know, our government stepped up and say, actually, you can't just blame little kids
for starting to smoke. We have to have more responsible advertising and this and that.
I think it's a bit the same here. It's very convenient for a big tech to blame.
So it's just people who are so dumb and get fooled. The blame usually comes in saying,
oh, it's just human psychology. People just want to hear what they already believe.
But Professor David Rand at MIT actually partly debunked that with a really nice study
showing that people tend to be interested in hearing things that go against what they believe
if it's presented in a respectful way. Like suppose, for example, that you have a company
and you're just about to launch this project and you're convinced it's going to work and someone
says, you know, Lex, I hate to tell you this, but this is going to fail. And here's why.
Would you be like, shut up, I don't want to hear it. La, la, la, la, la, la, la, la.
Would you? You would be interested, right? And also, if you're on an airplane back in the
pre-COVID times, you know, and the guy next to you is clearly from the opposite side of the
political spectrum, but is very respectful and polite to you. Wouldn't you be kind of interested
to hear a bit about how he or she thinks about things? Of course. But it's not so easy to find
out respectful disagreement now because like, for example, if you are a Democrat and you're like,
oh, I want to see something on the other side. So you just go brightbar.com. And then after the
first 10 seconds, you feel deeply insulted by something and they, it's not going to work. Or
if you take someone who votes Republican and they go to something on the left and they just get very
offended very quickly by them having put a deliberately ugly picture of Donald Trump on
the front page or something. It doesn't really work. So this news aggregator also has a nuanced
slider, which you can pull to the right. And then to make it easier to get exposed to actually more
sort of academic style or more respectful portrayals of different views. And finally,
the one kind of bias I think people are mostly aware of is the left, right, because it's so
obvious because both left and right are very powerful here, right? Both of them have well
funded TV stations and newspapers. And it's kind of hard to miss. But there's another one,
the establishment slider, which is it's also really fun. I love to play with it. And that's
more about corruption. Yes, because if you have a society that where almost all the powerful
entities want you to believe a certain thing, that's what you're going to read. And both the
big media mainstream media on the left and on the right, of course, and powerful companies
can push back very hard, like tobacco companies pushed back very hard back in the day when some
newspapers started writing articles about tobacco being dangerous. So it was hard to get a lot of
coverage about it initially. And also if you look geopolitically, right, of course, in any country
when you read their media, you're mainly going to be reading a lot about articles about how our
country is the good guy, and the other countries are the bad guys, right? So if you want to have a
really more nuanced understanding, you know, like the Germans used to be told that the British used
to be told that the French were the bad guys and the French used to be told that British were the
bad guys. Now they visit each other's countries a lot and have a much more nuanced understanding.
I don't think there's going to be any more wars between France and Germany. But on the geopolitical
scale, it's just as much as ever, you know, big Cold War now, US, China, and so on. And
if you want to get a more nuanced understanding of what's happening geopolitically,
then it's really fun to look at this establishment slider because it turns out there are tons of
little newspapers, both on the left and on the right, who sometimes challenge establishment and
say, you know, maybe we shouldn't actually invade Iraq right now. Maybe this weapons and mass
destruction thing is BS. If you look at journalism research afterwards, you can actually see that
quickly. Both CNN and Fox were very pro. Let's get rid of Saddam. There are weapons of mass
destruction. Then there were a lot of smaller newspapers. They were like, wait a minute,
this evidence seems a bit sketchy. And maybe we, but of course, they were so hard to find. Most
people didn't even know they existed, right? Yet it would have been better for American
national security if those voices had also come up. I think it harmed America's national security
actually that we invaded Iraq. And arguably there's a lot more interest in that kind of thinking too
from those small sources. So like the, when you say big, it's more about kind of the reach of the
broadcast. But it's not big in terms of the interest. I think there's a lot of interest in that kind of
anti-establishment or like skepticism towards out of the box thinking. There's a lot of interest in
that kind of thing. Do you see this news project or something like it being basically taken over
the world as the main way we consume information? Like how do we get there? The idea is brilliant.
It's a beauty. You're calling it your little project in 2020, but how does that become the
new way we consume information? I hope first of all, just to plan the little seed there because
normally the big barrier of doing anything in media is you need a ton of money, but this costs
no money at all. I've just been paying myself. You pay a tiny amount of money each month to
Amazon to run the thing in their cloud. There will never be any ads. The point is not to make
any money off of it. We just train machine learning algorithms to classify the articles and stuff,
so it just kind of runs by itself. If it actually gets good enough at some point that it starts
catching on, it could scale. If other people carbon copy it and make other versions that are
better, that's the more the merrier. I think there's a real opportunity for machine learning to
empower the individual against the powerful players. As I said in the beginning here,
it's been mostly the other way around so far that the big players have the AI and then they
tell people, this is the truth. This is how it is, but it can just as well go the other way
around. When the internet was born, actually, a lot of people had this hope that maybe this will
be a great thing for democracy, make it easier to find out about things. Maybe machine learning
and things like this can actually help again. I have to say, I think it's more important than
ever now because this is very linked also to the whole future of life as we discussed earlier.
We're getting this ever more powerful tack. It's pretty clear, if you look on the one or two
generation, three generation timescale, that there are only two ways this can end geopolitically.
Yeah. Either it ends great for all humanity or ends terribly for all of us. There's really
no in between. Technology knows no borders. You can't have people fighting when the weapons
just keep getting ever more powerful indefinitely. Eventually, the luck runs out.
Right now, we have, I love America, but the fact of the matter is what's good for America is not
opposites in the long term to what's good for other countries. It would be if this was some
sort of zero-sum game like it was thousands of years ago when the only way one country could
get more resources was to take land from other countries because that was basically the resource.
Right? Look at the map of Europe. Some countries kept getting bigger and smaller, endless wars,
but then since 1945, there hasn't been any war in Western Europe and they all got way richer
because of tech. The optimistic outcome is that the big winner in this century is going to be
America and China and Russia and everybody else because technology just makes us all
healthier and wealthier and we just find some way of keeping the peace on this planet.
But I think, unfortunately, there are some pretty powerful forces right now that are pushing in
exactly the opposite direction and trying to demonize other countries, which just makes it
more likely that this ever more powerful tech we're building is going to be in disastrous ways.
Yeah, for aggression versus cooperation, that kind of thing.
Yeah, even look at just military AI now. It was so awesome to see these dancing robots.
I loved it, right? But one of the biggest growth areas in robotics now is, of course,
autonomous weapons. And 2020 was like the best marketing year ever for autonomous weapons because
in both Libya, civil war, and in Nagorno-Karabakh, they made the decisive difference, right?
And everybody else is like watching this. Oh, yeah, we want to build autonomous weapons too.
And in Libya, you had, on one hand, our ally, the United Arab Emirates,
that were flying their autonomous weapons that they bought from China,
bombing Libyans. And on the other side, you had our other ally, Turkey, flying their drones.
They had no skin in the game, any of these other countries. And of course,
it was the Libyans who really got screwed. In Nagorno-Karabakh, you had actually, again,
just now Turkey is sending drones built by this company that was actually
founded by a guy who went to MIT AeroAstrode. Do you know that?
No.
Bakratiar. Yeah. So MIT has a direct responsibility for ultimately this. And a lot of civilians
were killed there. And so because it was militarily so effective, now suddenly there's
like a huge push. Oh, yeah, yeah, let's go build ever more autonomy into these weapons.
And it's going to be great. And I think actually, people who are obsessed about some sort of future
terminers, NATO scenario right now, should start focusing on the fact that we have
too much more urgent threats happening for machine learning. One of them is the whole
destruction of democracy that we've talked about now, where our flow of information is
being manipulated by machine learning. And the other one is that right now, this is the year
when the big arms race and out of control arms race in at least autonomous weapons is going to
start or it's going to stop. So you have a sense that there is like 2020 was a instrumental catalyst
for the race of for the autonomous weapons race. Yeah, because it was the first year when they
proved decisive in the battlefield. And these ones are still not fully autonomous, mostly
they're remote controlled, right? But we could very quickly make things about the size and cost
of a smartphone, which you just put in the GPS coordinates or the face of the one you want to
kill a skin color or whatever and it flies away and you know, does it and the real good reason
why the US and all the other superpowers should put the kibosh on this is the same reason we decided
to put the kibosh on bio weapons. So, you know, we gave the future of life award that we can talk
more about later, Matthew Messelsen from Harvard before for convincing Nixon to ban bio weapons.
And I asked him, how did you do it? And he was like, well, I just said, look, we don't want
there to be a $500 weapon of mass destruction that even all our enemies can afford, even non state
actors. And Nixon was like, good point. You know, it's in America's interest that the powerful
weapons are all really expensive. So only we can afford them or maybe some more stable adversaries,
right? Nuclear weapons are like that. But bio weapons were not like that. That's why we banned
them. And that's why you never hear about them now. That's why we love biology.
So you have a sense that it's possible for the big powerhouses in terms of the big nations in
the world to agree that autonomous weapons is not a race we want to be on. That doesn't end well.
Yeah, because we know it's just going to end in mass proliferation and every terrorist
everywhere is going to have these super cheap weapons that they will use against us.
And our politicians have to constantly worry about being assassinated every time they go outdoors
by some anonymous little mini drone, you know, we don't want that. And even if the US and China
and everyone else could just agree that you can only build these weapons if they cost at least
10 million bucks. That would be a huge win for the superpowers. And frankly, for everybody,
people often push back and say, well, it's so hard to prevent cheating. But hey,
you can say the same about bio weapons, you know, take any of your RMIT colleagues in biology.
Of course, they could build some nasty bio weapon if they really wanted to.
But first of all, they don't want to because they think it's disgusting because of the stigma.
And second, even if there's some sort of nutcase and want to, it's very likely that
some of their grad students or someone would rat them out because everyone else thinks it's
so disgusting. And in fact, we now know there was even a fair bit of cheating on the bio weapons
ban. But none, no countries use them because it was so stigmatized that it just wasn't worth
revealing that they had cheated. You talk about drones, but you kind of think
that drones is the remote operation. Which they are mostly. Yes. But you're not taking the next
intellectual step of like, where does this go? You're kind of saying the problem with drones
is that you're removing yourself from direct violence. Therefore, you're not able to sort
of maintain the common humanity required to make the proper decisions strategically.
But that's the criticism as opposed to like, if this is automated, and just exactly as you said,
if you automate it, and there's a race, then you just go into the technology and get better
and better and better, which means getting cheaper and cheaper and cheaper. And unlike perhaps
nuclear weapons, which is connected to resources in a way like it's hard to get the, it's hard to
engineer. Yeah. It feels like it's, there's too much overlap between the tech industry
and autonomous weapons to where you could have smartphone type of
cheapness. If you look at drones, it's a, for $1,000, you can have an incredible system that's
able to maintain flight autonomously for you and take pictures and stuff. You could see that going
into the autonomous weapon space that's, but like, why is that not thought about or discussed enough
in the public? Do you think you see those dancing Boston Dynamics robots and everybody has this kind
of, like as if this is like a far future. Yeah. They have this like fear, like, oh, this will be
Terminator in like some, I don't know, unspecified 20, 30, 40 years. And they don't think about, well,
this is like some much less dramatic version of that is actually happening now. It's not going to
have, it's not going to be legged. It's not going to be dancing, but it's already has the capability
to use artificial intelligence to kill humans. Yeah. The Boston Dynamics leg robots, I think the
reason we imagine them holding guns is just because you've all seen Arnold Schwarzenegger, right?
That's our reference point. That's pretty useless. That's not going to be the main military
use of them. They might be useful in law enforcement in the future. And then there's a whole debate
about you want robots showing up at your house with guns telling you who'll be perfectly obedient
to whatever dictator controls them. But let's leave that aside for a moment and look at what's
actually relevant now. So there's a spectrum of things you can do with AI in the military. And
again, to put my card on the table, I'm not the pacifist. I think we should have good defense.
So, for example, a predator drone is a room basically a fancy little remote control airplane.
Right. There's a human piloting it and the decision ultimately about whether to kill
somebody with it is made by a human still. And this is a line I think we should never cross.
There's a current DOD policy. Again, you have to have a human in the loop. I think algorithms
should never make life or death decisions. They should be left to humans. Now, why might we cross
that line? Well, first of all, these are expensive, right? So for example, when, when, when, when
Azerbaijan had all these drones and Armenia didn't have any, they start trying to jerry rig little
cheap things fly around. And but then of course, the meanings would jam them, or the other areas
would jam them. And remote control things can be jammed. That makes them inferior. Also,
there's a bit of a time delay between, you know, if we're piloting something far away,
speed of light, and the human has a reaction time as well, it would be nice to eliminate
that jamming possibility in the time delay by having it fully autonomous. But now you might be,
so then if you do, but now you might be crossing that exact line, you might program it to just,
oh yeah, dear drone, go hover over this country for a while. And whenever you find someone who is
a bad guy, you know, kill them. Now the machine is making these sort of decisions. And you and
some people who defend this still say, well, that's morally fine, because we are the good guys,
and we will tell it, the definition of bad guy that we think is moral. But now,
it would be very naive to think that if ISIS buys that same drone, that they're going to use our
definition of bad guy. Maybe for them, bad guy is someone wearing a US Army uniform. Or maybe
there will be some weird ethnic group who decides that someone of another ethnic group,
they are the bad guys, right? The thing is, human soldiers, with all our faults, right,
we still have some basic wiring in us. Like, no, it's not okay to kill kids and civilians.
And Thomas Reppin has none of that. It's just going to do whatever is programmed. It's like the
perfect Adolf Eichmann on steroids. Like, they told him, Adolf Eichmann, you know,
you want you to do this and this and this to make the Holocaust more efficient. And he was like,
YAHOOL! And off he went and did it, right? Do we really want to make machines that are like that?
Like, completely amoral and will take the user's definition of who's the bad guy? And do we then
want to make them so cheap that all our adversaries can have them? Like, what could possibly go wrong?
That's the, that's, I think, the big argument for why we want to, this year, really put the
kibosh on this. And I think you can tell there's a lot of very active debate even going on within
the U.S. military and undoubtedly in other militaries around the world also about whether we
should have some sort of international agreement to at least require that these weapons have to be
above a certain size and cost, you know, so that things just don't totally spiral out of control.
And finally, just for your question, but is it possible to stop it? Because some people tell me,
oh, just give up, you know. But again, so, so Matthew Messelsen again from Harvard, right,
who the bio weapons hero, he had exactly this criticism mostly with bio weapons. People were
like, how can you check for sure that the Russians aren't cheating? And he told me this, I think,
really ingenious insight. He said, you know, Max, some people think you have to have inspections
and things, and you have to make sure that people you can catch the cheaters with 100% chance.
You don't need 100%. He said, 1% is usually enough. Because if it's just an enemy, if it's
another big state, like suppose China and the US have signed the treaty,
drawing a certain line and saying, yeah, these kind of drones are okay, but these fully autonomous
ones are not. Now, suppose you are China, and you have cheated and secretly developed some
clandestine little thing, or you're thinking about doing it, you know, what's your calculation that
you do? Well, you're like, okay, what's the probability that we're going to get caught?
If the probability is 100%, of course, we're not going to do it.
But if the probability is 5% that we're going to get caught, then it's going to be like a huge
embarrassment for us. And we still have our nuclear weapons anyway, so it doesn't really
make any enormous difference in terms of deterring the US.
And that feeds the stigma that you kind of establish, like this fabric, this universal
stigma over the thing. Exactly. So it's very reasonable for them to say, well,
you know, we probably get away with it. But if we don't, then the US will know we cheated,
and then they're going to go full tilt with their program and say, look, the Chinese are
cheaters, and that's good. Now we have all these weapons against us, and that's bad. So
the stigma alone is very, very powerful. And again, look what happened with bio weapons,
right? It's been 50 years now. When was the last time you read about a bioterrorism attack?
The only deaths I really know about with bio weapons that have happened,
when we Americans managed to kill some of our own with anthrax, you know, the idiot who sent
them to Tom Daschle and others and letters, right? And similarly, in Svetlovsk in the Soviet Union,
they had some anthrax and some lab there. Maybe they were cheating or who knows,
and it leaked out and killed a bunch of Russians. I'd say that's a pretty good success, right?
50 years, just two own goals by the superpowers, and then nothing. And that's why,
whenever I ask anyone what they think about biology, they think it's great.
They associate it with new cures, new diseases, maybe a good vaccine. This is how I want to
think about AI in the future. And I want others to think about AI too, as a source of all these
great solutions to our problems, not as, oh, AI. Oh, yeah. That's the reason I feel scared going
outside these days. Yeah, it's kind of brilliant that the bio weapons and nuclear weapons, we've
figured out, I mean, of course, there's still a huge source of danger, but we figured out some
way of creating rules and social stigma over these weapons that then creates a stability
to whatever that game theoretic stability. Exactly. And we don't have that with AI,
and you're kind of screaming from the top of the mountain about this, that we need to find that,
because just like, it's very possible with the future of life, as you point out, Institute
Awards pointed out that with nuclear weapons, we could have destroyed ourselves quite a few
times. And it's a learning experience that doesn't, it's very costly.
We gave this future life award, we gave it the first time to this guy, Vasily Arkhipov.
He was on, most people haven't even heard of him. Yeah, can you say who he is? Vasily Arkhipov, he
has, in my opinion, made the greatest positive contribution to humanity of any human in modern
history. And maybe it sounds like hyperbole here, like I'm just over the top, but let me tell you
the story, and I think maybe you'll agree. So during the Cuban Missile Crisis,
we Americans first didn't know that the Russians had sent four submarines, but we caught two of
them. And we didn't know that, so we dropped practice depth charges on the one that he was on,
trying to force it to the surface. But we didn't know that this nuclear submarine actually was
a nuclear submarine with a nuclear torpedo. We also didn't know that they had an authorization
to launch it without clearance from Moscow. And we also didn't know that they were running out of
electricity. Their batteries were almost dead. They were running out of oxygen. Sailors were
fainting left and right. The temperature was about 110, 120 Fahrenheit on board. It was really
hellish conditions, really just a kind of doomsday. And at that point, these giant explosions start
happening from Americans dropping these. The captain thought World War III had begun. They
decided that they were going to launch the nuclear torpedo. And one of them shouted,
you know, we're all going to die, but we're not going to disgrace our Navy. We don't know what
would have happened if there had been a giant mushroom cloud all of a sudden against Americans.
But since everybody had their hands on the triggers, you don't have to be too creative to
think that it could have led to an all out nuclear war, in which case we wouldn't be having this
conversation now, right? What actually took place was they needed three people to approve this.
The captain had said, yes, there was the Communist Party political officer. He also said, yes,
let's do it. And the third man was this guy, Vasily Arkhipov, who said, and yet,
yeah, he for some reason, he was just more chilled than the others. And he was the right man at the
right time. I don't want us as a species rely on the right person being there at the right time.
You know, we tracked down his family living in relative poverty outside Moscow.
Maybe he flew his daughter. He had passed away and flew them to London. They had never been to
the West even. It was incredibly moving to get to honor them for this. The next year,
we gave this Future Life Award to Stanislav Petrov. Have you heard of him?
Yes.
So he was in charge of the Soviet early warning station, which was built with Soviet technology
and honestly not that reliable. It said that there were five US missiles coming in again.
If they had launched at that point, we probably wouldn't be having this conversation.
He decided based on just mainly gut instinct to just not escalate this. And I'm very glad he
wasn't replaced by an AI that was just automatically following orders. And then we gave the third one
to Matthew Messelsen. Last year, we gave this award to these guys who actually used technology
for good, not avoiding something bad, but for something good. The guys who eliminated this
disease, which is way worse than COVID, that had killed half a billion people in its violent
century, smallpox. So we mentioned it earlier. COVID, on average, kills less than 1% of people
who get it. Smallpox, about 30%. And they just, ultimately, Viktor Zhdanov and Bill Fagy,
most of my colleagues have never heard of either of them, one American, one Russian,
they did this amazing effort. Not only was Zhdanov able to get the US and the Soviet
Union to team up against smallpox during the Cold War, but Bill Fagy came up with this ingenious
strategy for making it actually go all the way to defeat the disease without funding for
vaccinating everyone. And as a result, we went from 15 million deaths the year I was born in
smallpox. So what do we have in COVID now? A little bit short of 2 million, right?
To zero deaths, of course, this year. And forever. There have been 200 million people,
they estimate, who would have died since then by smallpox had it not been for this. So
isn't science awesome when you use it for good? And the reason we want to celebrate these sort
of people is to remind them of this. Science is so awesome when you use it for good.
And those awards, actually, the variety there, it's a very interesting picture. So the first
two are looking at, it's kind of exciting to think that these average humans, in some sense,
that there are products of billions of other humans that came before them, evolution,
and some little, you said gut, but there's something in there that stopped the annihilation
of the human race. And that's a magical thing, but that's like this deeply human thing. And then
there's the other aspect where that's also very human, which is to build solution to the
existential crises that we're facing, to build it, to take responsibility, to come up with
different technologies and so on. And both of those are deeply human. The gut and the mind,
whatever that is. The best is when they work together. Arkhipov, I wish I could have met him,
of course, but he had passed away. He was really a fantastic military officer, combining all the
best traits that we in America admire in our military. Because first of all, he was very loyal,
of course, he never even told anyone about this during his whole life, even though you think he
had some bragging rights, right? But he just was like, this is just business, just doing my job.
It only came out later after his death. And second, the reason he did the right thing was not
because he was some sort of liberal, or some sort of, not because he was just, oh, you know,
peace and love. It was partly because he had been the captain on another submarine that had
a nuclear reactor meltdown. And it was his heroism that helped contain this. That's why he died of
cancer later also. But he's seen many of his crew members die. And I think for him, that gave him
this gut feeling that, you know, if there's a nuclear war between the US and the Soviet Union,
the whole world is going to go through what I saw my dear crew members suffer through. It wasn't
just an abstract thing for him. I think it was real. And second, though, not just a gut, the mind,
right? He was, for some reason, very level headed personality and very smart guy,
which is exactly what we want our best fighter pilots to be also. I never forget Neil Armstrong
when he's landing on the moon and almost running out of gas. And he doesn't even change,
really say 30 seconds, doesn't even change the tone of voice just keeps going. Archipelago,
I think was just like that. So when the explosions start going off and his captain is screaming,
and we should nuke them and all, he's like, I don't think the Americans are trying to sink us.
I think they're trying to send us a message. That's pretty badass. Yes, coolness. Because he said,
if they wanted to sink us. And he said, listen, listen, it's alternating one loud explosion on
the left, one on the right, one on the left, one on the right. He was the only one to notice this
pattern. And he's like, I think this is them trying to send us a signal that they want us to
surface. And they're not going to sink us. And somehow, this is how he then managed to ultimately
with his combination of gut and also just cool analytical thinking was able to deescalate the
whole thing. And yeah, so this is some of the best in humanity. I guess coming back to what we
talked about earlier is the combination of the neural network, the instinctive, you know,
with, I'm getting tearing up here, getting emotional, but he is one of my superheroes
having both the heart, you know, and the mind combined. And especially in that time,
there's something about the, I mean, this is a very in America, people are used to this kind
of idea of being the individual of like on your own thinking. Yeah. I think under in the Soviet
Union under communism, it's actually much harder to do that. Oh, yeah, he didn't even, he even
he didn't get any accolades either when he came back for this, right? They just want to
hush the whole thing up. Yeah, there's echoes of that with Chernobyl, there's all kinds of
that that's one, that's, that's a really hopeful thing that amidst big centralized powers,
whether it's companies or states, there's still the power of the individual to think on their
own to act. But I think we need to think of people like this, not as a panacea we can always count
on, but rather as a wake up call, you know, so because of them, because of Archipelago, we are
alive to learn from this lesson, to learn from the fact that we shouldn't keep playing Russian
roulette and almost have a nuclear war by mistake now and then, because relying on luck is not a
good long term strategy. If you keep playing Russian roulette over and over again, the probability
of surviving just drops exponentially with time. Yeah. And if you have some probability of having
an accidental nuclear war every year, the probability of not having one also drops exponentially. I
think we can do better than that. So I think the message is very clear. There are once in a while
shit happens. And there's a lot of very concrete things we can do to reduce the risk of things
like that happening in the first place. On the AI front, if we just link on that for a second.
And so you're friends with, you often talk with Elon Musk throughout history,
did a lot of interesting things together. He has a set of fears about the future of artificial
intelligence, AGI. Do you have a sense, we've already talked about the things we should be
worried about with AI. Do you have a sense of the shape of his fears in particular about AI,
of the, which subset of what we've talked about, whether it's creating, you know,
it's that direction of creating sort of these giant computational systems that are not explainable.
They're not intelligible intelligence. Or is it the, and then like as a branch of that,
is it the manipulation by big corporations of that or individual evil people to use that for
destruction or the unintentional consequences? Do you have a sense of where his thinking is on this?
From my many conversations with Elon, I certainly have a model of how he thinks. It's
actually very much like the way I think also, I'll elaborate on it a bit. I just want to
push back on when you said evil people. I don't think it's a very helpful concept,
evil people. Sometimes people do very, very bad things, but they usually do it because they
think it's a good thing. Because somehow other people had told them that that was a good thing
or given them incorrect information or whatever, right? I believe in the fundamental goodness
of humanity that if we educate people well and they find out how things really are,
people generally want to do good and be good.
But hence the value alignment. It's about information, about knowledge. And then once
we have that, we'll likely be able to do good in the way that's aligned with everybody else who
thinks it's good. And it's not just the individual people we have to align. So we don't just want
people to be educated to know the way things actually are and to treat each other well.
But we also would need to align other non-human entities. We talked about corporations or has
to be institutions so that what they do is actually good for the country they're in and
we should align, make sure that what countries do is actually good for the species as a whole,
et cetera. Coming back to Elon, my understanding of how Elon sees this is really quite similar
to my own, which is one of the reasons I like him so much and enjoy talking with him so much.
I think he's quite different from most people in that he
thinks much more than most people about their really big picture,
not just what's going to happen in the next election cycle, but in millennia,
millions and billions of years from now. And when you look in this more cosmic
perspective, it's so obvious that we're gazing out into this universe that as far as we can tell
is mostly dead with life being an almost imperceptibly tiny perturbation. And he sees
this enormous opportunity for our universe to come alive, first to become an interplanetary
species. Mars is obviously just first stop on this cosmic journey. And precisely because he
thinks more long term, it's much more clear to him than to most people that what we do with
this Russian roulette thing we keep playing with our nukes is a really poor strategy,
really reckless strategy. And also that we're just building these ever more powerful AI systems
that we don't understand is also a really reckless strategy. I feel Elon is a humanist
in the sense that he wants an awesome future for humanity. He wants it to be us that control the
machines rather than the machines that control us. And why shouldn't we insist on that? We're
building them after all, right? Why should we build things that just make us into some little cog
in the machinery that has no further say in the matter? It's not my idea of an inspiring future
either. Yeah, if you think on the cosmic scale in terms of both time and space,
so much is put into perspective. Yeah. Whenever I have a bad day, that's what I think about.
It immediately makes me feel better. It makes me sad that for us individual humans, at least for
now, the ride ends too quickly. We don't get to experience the cosmic scale. Yeah. I mean,
I think of our universe sometimes as an organism that has only begun to wake up a tiny bit. Just
like the very first little glimmers of consciousness you have in the morning when you start coming
around. Before the coffee. Before the coffee. Even before you get out of bed, before you even open
your eyes, you start to wake up a little bit. There's something here. That's very much how I
think of what we are. All those galaxies out there, I think they're really beautiful,
but why are they beautiful? They're beautiful because conscious entities are actually observing
them and experiencing them through our telescopes. I define consciousness as subjective experience,
whether it be colors or emotions or sounds. So beauty is an experience. Meaning is an experience.
Purpose is an experience. If there was no conscious experience observing these galaxies,
they wouldn't be beautiful. If we do something dumb with advanced AI in the future here and
Earth originating, life goes extinct. And that was it for this. If there is nothing else
with telescopes in our universe, then it's kind of game over for beauty and meaning and
purpose in our whole universe. And I think that would be just such an opportunity lost, frankly.
And I think when Elon points this out, he gets very unfairly maligned in the media for all the
dumb media bias reasons we talked about. They want to print precisely the things about Elon out of
context that are really clickbaity. He has gotten so much flak for this summoning the demon statement.
I happen to know exactly the context because I was in the front row when he gave that talk. It
was at MIT, you'll be pleased to know. It was the AeroAstro anniversary. They had Buzz Aldrin
there from the moon landing, the whole house, a Kresge auditorium packed with MIT students.
And he had this amazing Q&A. It might have gone for an hour and they talked about rockets and
Mars and everything. At the very end, this one student was actually in my class asked him,
what about AI? Elon makes this one comment, and they take this out of context, print it,
goes viral. Is it like with AI, we're summoning the demons or something like that?
And try to cast him as some sort of doom and gloom dude. You know Elon. He's not the doom
and gloom dude. He is such a positive visionary. And the whole reason he warns about this is because
he realizes more than most what the opportunity cost is of screwing up, that there is so much
awesomeness in the future that we can and our descendants can enjoy if we don't screw up.
Right? I get so pissed off when people try to cast him as some sort of
technophobic Luddite. And at this point, it's kind of ludicrous when I hear people say that
people who worry about artificial general intelligence are Luddites, because of course,
if you look more closely, you have some of the most outspoken
people making warnings are people like Professor Stuart Russell from Berkeley,
who's written the best selling AI textbook, you know, claiming that he is a Luddite who
doesn't understand AI is the joke is really on the people who said it. But I think more broadly,
this message is really not sunk in at all. What it is that people worry about, they think
that Elon and Stuart Russell and others are worried about the dancing robots
picking up an AR 15 and going on a rampage, right? They think they're worried about robots turning
evil. They're not. I'm not. The risk is not
malice. It's competence. The risk is just that we build some systems that are incredibly
competent, which means they're always going to get their goals accomplished,
even if they clash with our goals. That's the risk. Why did we humans drive the West African
black rhino extinct? Is it because we're malicious, evil rhinoceros haters? No,
it's just because our goals didn't align with the goals of those rhinos and tough luck for the rhinos.
So the point is just we don't want to put ourselves in the position of those rhinos,
creating something more powerful than us if we haven't first figured out how to align the goals.
I'm optimistic. I think we could do it if we worked really hard on it because I spent a lot
of time around intelligent entities that were more intelligent than me. My mom and my dad,
and I was little, and that was fine because their goals were actually aligned with mine quite well.
But we've seen today many examples of where the goals of our powerful systems are not so aligned.
So those click-through optimization algorithms that are polarized social media, they were actually
pretty poorly aligned with what was good for democracy, it turned out. And again,
almost all problems we've had in machine learning, again, came so far, not from Alice,
but from poor alignment. And that's exactly why that's why we should be concerned about in the
future. Do you think it's possible that with systems like Neuralink and brain-computer interfaces,
you know, again, thinking of the cosmic scale, Elon's talked about this, but others have as
well throughout history of figuring out how the exact mechanism of how to achieve that kind of
alignment. So one of them is having a symbiosis with AI, which is like coming up with clever ways
where we're like stuck together in this weird relationship, whether it's biological or in
some kind of other way. Do you think that's a possibility of having that kind of symbiosis,
or do we want to instead kind of focus on this distinct entities of us humans talking to these
intelligible, self-doubting AIs, maybe like Stuart Russell thinks about it. We're self-doubting and
full of uncertainty, and then have our AI systems that are full of uncertainty, we communicate back
and forth, and in that way achieve symbiosis. I honestly don't know. I would say that because
we don't know for sure what, if any of our, which of any of our ideas will work, but we do know that
if we don't, I'm pretty convinced that if we don't get any of these things to work and just
barge ahead, then our species is probably going to go extinct this century. This century,
you think we're facing this crisis is a 21st century crisis. This century will be remembered
on a hard drive. On a hard drive somewhere or maybe by future generations is like
there will be future future life-ositude awards for people that have done something about AI.
It could also end even worse, whether we're not superseded by leaving any AI behind either,
where we just totally wipe out on Easter Island. Our century is long. There are still
79 years left of it. Think about how far we've come just in the last 30 years.
We can talk more about what might go wrong, but you asked me this really good question about
what's the best strategy? Is it Neuralink or Russell's approach or whatever? I think
when we did the Manhattan Project, we didn't know if any of our four ideas for enriching uranium
and getting out the uranium-235 were going to work, but we felt this was really important
to get it before Hitler did. You know what we did? We tried all four of them.
Here, I think it's analogous where the greatest threat that's ever faced
are species and, of course, US national security by implication. We don't have any method that's
guaranteed to work, but we have a lot of ideas. We should invest pretty heavily in pursuing all
of them with an open mind and hope that one of them at least works. The good news is the century
is long and it might take decades until we have artificial general intelligence. We have some
time, hopefully, but it takes a long time to solve these very, very difficult problems.
It's going to actually be the most difficult problem we were ever trying to solve as a species.
We have to start now so we don't want to have, rather than begin thinking about it the night
before some people who've had too much Red Bull switch it on. We have to, coming back to your
question, we have to pursue all of these different avenues and see.
If you're my investment advisor and I was trying to invest in the future,
how do you think the human species is most likely to destroy itself in this century?
If the crises, many of the crises we're facing are really before us within the next 100 years,
how do we make explicit, make known the unknowns and solve those problems to avoid the biggest
starting with the biggest existential crisis?
So as your investment advisor, how are you planning to make money on us going,
destroying ourselves? I have to ask.
I don't know. It might be the Russian origins that somehow is involved.
At the micro level of detailed strategies, of course, these are unsolved problems.
For AI alignment, we can break it into three sub-problems that are all unsolved.
I think you want first to make machines understand our goals, then adopt our goals,
and then retain our goals. So to hit on all three real quickly.
The problem when Andreas Lubitz told his autopilot to fly into the Alps,
was that the computer didn't even understand anything about his goals, right? It was too dumb.
It could have understood, actually, but you would have had to put some effort in as a
system designer to don't fly into mountains. So that's the first challenge.
How do you program into computers human values, human goals?
We could start rather than saying, oh, it's so hard, we should start with the simple stuff,
as I said, self-driving cars, airplanes, just put in all the goals that we all agree on already,
and then have a habit of whenever machine gets smarter so they can understand one level
higher goals, put them into. The second challenge is getting them to adopt the goals.
It's easy for situations like that where you just program it in, but when you have
self-learning systems like children, any parent knows that there's a difference between getting
our kids to understand what we want them to do and to actually adopt our goals. With humans,
with children, fortunately, they go through this phrase, first they're too dumb to understand
what we want our goals are, and then they have this period of some years when they're both
smart enough to understand them and malleable enough that we have a chance to raise them well,
and then they become teenagers kind of too late. But we have this window with machines,
the challenges, the intelligence might grow so fast that that window is pretty short.
So that's a research problem. The third one is how do you make sure they keep the goals
if they keep learning more and getting smarter? Many sci-fi movies are about how you have something
which initially was aligned, but then things kind of go off keel. And my kids were very,
very excited about their Legos when they were little. Now they're just gathering dust in
the basement. If we create machines that are really on board with the goal of taking care
of humanity, we don't want them to get as bored with us as my kids got with Legos.
So this is another research challenge. How can you make some sort of recursively self-improving
system retain certain basic goals? That said, a lot of adult people still play with Legos. So
maybe we succeeded with the Legos. I like your optimism. So not all AI systems have to maintain
the goals, right? Just some fraction. Yeah. So there's a lot of talented AI researchers now
who have heard of this and want to work on it. Not so much funding for it yet.
Of the billions that go into building AI more powerful, it's only a minuscule fraction. So for
going into the safety research, my attitude is generally we should not try to slow down the
technology, but we should greatly accelerate the investment in this sort of safety research
and also make sure. This was very embarrassing last year, but the NSF decided to give out
six of these big institutes. We got one of them for AI and Science, you asked me about.
Another one was supposed to be for AI safety research, and they gave it to people studying
oceans and climate and stuff. So I'm all for studying oceans and climates, but we need to
actually have some money that actually goes into AI safety research also and doesn't just get grabbed
by whatever. That's a fantastic investment. And then at the higher level, you asked this question,
okay, what can we do? What are the biggest risks? I think we cannot just consider this to be only
a technical problem. Again, because if you solve only the technical problem, can I play with your
robot? Get our machines to just blindly obey the orders we give them. So we can always trust
that it will do what we want. That might be great for the owner of the robot. That might not be so
great for the rest of humanity if that person is that least favorite world leader or whatever you
imagine. So we have to also take a look at the apply alignment, not just to machines,
but to all the other powerful structures. That's why it's so important to strengthen our democracy
again. As I said, to have institutions make sure that the playing field is not rigged so that
corporations are given the right incentives to do the things that both make profit and are good for
people to make sure that countries have incentives to do things that are both good for their people
and don't screw up the rest of the world. And this is not just something for AI nerds to geek
out on. This is an interesting challenge for political scientists, economists, and so many
other thinkers. So one of the magical things that perhaps makes this earth quite unique
is that it's home to conscious beings. So you mentioned consciousness. Perhaps as a small aside,
because we didn't really get specific to what, how we might do the alignment. Like you said,
is there just a really important research problem? But do you think engineering consciousness into
AI systems is a possibility? Is something that we might one day do? Or is there something fundamental
to consciousness that is, is there something about consciousness that is fundamental to humans and
humans only? I think it's possible. I think both consciousness and intelligence are information
processing, certain types of information processing. And that fundamentally, it doesn't matter whether
the information is processed by carbon atoms in neurons and brains or by silicon atoms and
so on in our technology. Some people disagree. This is what I think as a physicist that I,
and the consciousness is the same kind of, you said consciousness is information processing.
So meaning, you know, I think you had a quote of something like it's information
knowing itself, that kind of thing. I think consciousness, yeah, is the way information
feels when it's being processed in certain complex ways. We don't know exactly what those
complex ways are. It's clear that most of the information processing in our brains does not
create an experience. We're not even aware of it, right? Like, for example, you're not aware of your
heartbeat regulation right now, even though it's clearly being done by your body, right? It's just
kind of doing its own thing. When you go jogging, there's a lot of complicated stuff about how you
put your foot down. And we know it's hard. That's why robots used to fall over so much. But you're
mostly unaware about it. Your brain, your CEO consciousness module just sends an email, hey,
you know, I want to keep jogging along this path. The rest is on autopilot, right? So most of it is
not conscious. But somehow, there are some of the information processing, which is, we don't know
what exactly. I think this is a science problem that I hope one day will have some equation for
or something. So we can be able to build a consciousness detector and say, yeah, here,
there is some consciousness. Here, there's not. Oh, don't boil that lobster because it's feeling
pain, or it's okay, because it's not feeling pain. Right now, we treat this as sort of just
metaphysics. But it would be very useful in emergency rooms to know if a patient has locked
in syndrome and is conscious, or if they are actually just out. And in the future, if you build a
very, very intelligent helper robot to take care of you, you know, I think you'd like to know
if you should feel guilty about shutting it down. Or if it's just like a zombie going through
emotions like a fancy tape recorder, right? And once we can make progress on the science of
consciousness and figure out what is conscious and what isn't, then we, assuming we want to create
positive experiences and not suffering, we'll probably choose to build some machines that are
deliberately unconscious, that do incredibly boring, repetitive jobs in an iron mine somewhere
or whatever. And maybe we'll choose to create helper robots for the elderly that are conscious,
so that people don't just feel creeped out that the robot is just faking it
when it acts like it's sad or happy. Like I said, elderly, I think everybody gets pretty
deeply lonely in this world. And so there's a place I think for everybody to have a connection
with conscious beings, whether they're human or otherwise. But I know for sure that I would,
if I had a robot, if I was going to develop any kind of personal emotional connection with it,
I would be very creeped out if I knew it in intellectual level that the whole thing was
just a fraud. You know, today you can buy a little talking doll for a kid, which will say
things and the little child will often think that this is actually conscious and even real
secrets to it that then go on the internet and with all sorts of creepy repercussions.
I would not want to be just hacked and tricked like this. If I was going to be developing real
emotional connections with the robot, I would want to know that this is actually real. It's
acting conscious, acting happy because it actually feels it. And I think this is not sci-fi. I think
it's possible to measure, to come up with tools. After we understand the science of
consciousness, you're saying we'll be able to come up with tools that can measure consciousness
and definitively say like this thing is experiencing the things it says it's experiencing.
Kind of by definition, if it is a physical phenomena, information processing,
and we know that some information processing is conscious and some isn't, well, then there is
something there to be discovered with the methods of science. Giulio Tononi has stuck his neck out
the farthest and written down some equations for a theory. Maybe that's right. Maybe it's wrong.
We certainly don't know. But I applaud that kind of efforts to take this, say this is not just
something that philosophers can have beer and muse about, but something we can measure and study.
And coming, being that back to us, I think what we would probably choose to do, as I said, is if
we cannot figure this out, choose to be quite mindful about what sort of consciousness,
if any, we put in different machines that we have. And certainly, we wouldn't want to make,
we should not be making you much machines that suffer without us even knowing it, right?
And if at any point, someone decides to upload themselves like Ray Kurzweil wants to do,
I don't know if you've had him on your show. We agree, but then COVID happens, so we're waiting
it out a little bit. You know, suppose he uploads himself into this robo-ray and it talks like him
and acts like him and laughs like him. And before he powers off his biological body,
he would probably be pretty disturbed if he realized that there's no one home. This robot
is not having any subjective experience, right? If humanity gets replaced by machine descendants,
do all these cool things and build spaceships and go to intergalactic rock concerts,
and it turns out that they are all unconscious, just going through the motions. Wouldn't that
be like the ultimate zombie apocalypse, right? Just a play for empty benches?
Yeah, I have a sense that there's some kind of, once we understand consciousness better,
we'll understand that there's some kind of continuum and it would be a greater appreciation.
And we'll probably understand, just like you said, it'd be unfortunate if it's a trick.
We'll probably definitely understand that love is indeed a trick that will play on each other,
that we humans are, we convince ourselves we're conscious, but we're really,
you know, awesome trees and dolphins are all the same kind of consciousness.
Can I try to cheer you up a little bit with the philosophical thought here about the love part?
Yes, let's do it.
You might say, okay, love is just a collaboration enabler,
and then maybe you can go and get depressed about that. But I think that would be the wrong
conclusion, actually. You know, I know that the only reason I enjoy food is because my
genes hacked me and they don't want me to starve to death, not because they care about me consciously
enjoying succulent delights of pistachio ice cream, but they just want me to make copies of
them. The whole thing. So in a sense, the whole, the whole enjoyment of food is also a scam
like this. But does that mean I shouldn't take pleasure in this pistachio ice cream?
I love pistachio ice cream. And I can tell you, I have, I have, I know this is an experimental
fact. I enjoy pistachio ice cream every bit as much, even though I scientifically know exactly
why, what kind of scam this was. Your genes really appreciate that you like the pistachio ice cream.
Well, but I, my mind appreciates it too, you know, and I have a conscious experience right
now. Ultimately, all of my brain is also just something the genes built to copy themselves.
But so what, you know, I'm grateful that, yeah, thanks genes for doing this. But, you know,
now it's my brain that's in charge here, and I'm going to enjoy my conscious experience.
Thank you very much. And not just the pistachio ice cream, but also the love I feel for my amazing
wife and all the other delights of being conscious. I don't actually, Richard Feynman,
I think said this so well. He is also the guy, you know, really got me into physics.
Some art friend said that science kind of just is the party pooper. It's kind of ruins the fun,
right? When like, you have a beautiful flower, says the artist, and then the scientist is going
to deconstruct that into just a blob of quarks and electrons. And Feynman pushed back on that in
such a beautiful way, which I think also can be used to push back and make you not feel guilty
about falling in love. So here's what Feynman basically said. He said to his friend, you know,
yeah, I can also as a scientist see that this is a beautiful flower. Thank you very much. Maybe I
can't draw as good a painting as you because I'm not as talented an artist. But yeah,
I can really see the beauty in it. And it just, it also looks beautiful to me.
But in addition to that, Feynman said, as a scientist, I see even more beauty
that the artist did not see, right? Suppose this is a flower on a blossoming apple tree.
He could say this tree has more beauty in it than just the colors and the fragrance. This tree
is made of air, Feynman wrote. This is one of my favorite Feynman quotes ever. And it took
the carbon out of the air and bound it in using the flaming heat of the sun, you know,
to turn the air into tree. And when you burn logs in your fireplace, it's really beautiful to think
that this is being reversed. Now the tree is going, the wood is going back into air. And in
this flaming, beautiful dance of the fire that the artist can see is the flaming light of the sun
that was bound in to turn the air into tree. And then the ashes is the little residue that
didn't come from the air that the tree sucked out of the ground, you know, Feynman said,
these are beautiful things. And science just adds, it doesn't subtract. And I feel exactly that way
about love and about pistachio ice cream also. I can understand that there is even more nuance
to the whole thing, right? At this very visceral level, you can fall in love just as much as
someone who knows nothing about neuroscience. But you can also appreciate this even greater beauty
in it. Isn't it remarkable that it came about from from this completely lifeless universe,
just a bunch of hot blob of plasma expanding? And then over the eons, you know, gradually,
first the strong nuclear force decided to combine quarks together into nuclei and then the electric
force bound in electrons and made atoms and then they clustered it from gravity and you got planets
and stars and this and that. And then natural selection came along and the genes had their
little thing and you started getting what went from seeming like a completely pointless universe
that we're just trying to increase entropy and approach heat depth into something that looked
more goal oriented. Isn't that kind of beautiful? And then this goal orientedness through revolution
got ever more sophisticated where you got ever more. And then you started getting this thing,
which is kind of like DeepMind's mu zero and steroids self, the ultimate self play is not
what what DeepMind's AI does against itself to get better at the go. It's what all these little
cork blobs did against each other in the game of survival of the fittest. You know, when you had
really dumb bacteria living in a simple environment, there wasn't much incentive to get an intelligent,
but then the life made environment more complex. And then there was more incentive to get even
smarter. And, and that gave the other organisms more incentive to also get smarter. And then
here we are now, just like, like mu zero learn to become world master at the go and chess from
playing as itself, by just playing against itself, all the quirks here on our planet, the electrons
have created giraffes and elephants and humans and love. I just find that really beautiful.
And I mean, that just adds to the enjoyment of love. It doesn't subtract anything. Do you feel
a little more? I feel way better. That was, that was incredible. So this self play of quirks,
taking back to the beginning of our conversation a little bit, you've, there's so many exciting
possibilities about artificial intelligence, understanding the basic laws of physics.
Do you think AI will help us unlock? There's been a quite a bit of excitement throughout
the history of physics of coming up with more and more general simple laws that explain the
nature of our reality. And then the ultimate of that would be a theory of everything that combines
everything together. Do you think it's possible that, well, one, we humans, but perhaps AI systems
will figure out a theory of physics that unifies all the laws of physics? Yeah, I think it's
absolutely, absolutely possible. I think it's very clear that we're going to see a great boost to
science. We're already seeing a boost actually from machine learning, helping science. Alpha
fold was an example, you know, decades old protein folding problem. And gradually, yeah,
unless we go extinct by doing something dumb, like we discussed, I think it's
very likely that our understanding of physics will become so good that our technology will
no longer be limited by human intelligence, but instead be limited by the laws of physics.
So our tech today is limited by what we've been able to invent, right? I think as AI progresses,
it'll just be limited by the speed of light and other physical limits, which would mean it's going
to be just dramatically beyond where we are now. Do you think it's a fundamentally mathematical
pursuit of trying to understand the laws of this that govern our universe from a mathematical
perspective? It's almost like if it's AI, it's exploring the space of theorems and those kinds
of things. Or is there some other, is there some other more computational ideas, more sort of
empirical ideas? They're both, I would say. It's really interesting to look out at the landscape
of everything we call science today. So here you come now with this big new hammer that says
machine learning on it and ask, you know, where are there some nails that you can help with here
that you can hammer? Ultimately, if machine learning gets the point that it can do everything
better than us, it will be able to help across the whole space of science. But maybe we can anchor it
by starting a little bit right now near term and see how we kind of move forward. So right now,
first of all, you have a lot of big data science where, for example, with telescopes,
we are able to collect way more data every hour than a grad student can just pour over like in
the old times, right? And machine learning is already being used very effectively, even at MIT,
like to find planets around other stars, to detect exciting new signatures of new particle
physics in the sky, and to detect the ripples in the fabric of space time that we call gravitational
waves caused by enormous black holes crashing into each other halfway across the observable
universe. Machine learning is running and taking it right now, you know, doing all these things,
and it's really helping all these experimental fields. There is a separate front of physics,
computational physics, which is getting an enormous boost also. So we had to do all our
computations by hand, right? People would have these giant books with tables of logarithms,
and oh my God, it pains me to even think how long time it would have taken to do simple stuff.
Then we started to get calculators and computers that could do some basic math for us.
Now, what we're starting to see is kind of a shift from go-fi computational physics
to neural network computational physics. What I mean by that is most computational physics
would be done by humans programming in the intelligence of how to do the computation into
the computer. Just as when Gary Kasparov got his posterior kicked by IBM's Deep Blue in chess,
humans had programmed in exactly how to play chess. Intelligence came from the humans. It
wasn't learned, right? Mu0 can be not only Kasparov in chess, but also stock fish,
which is the best go-fi chess program by learning. We're seeing more of that now,
that shift beginning to happen in physics. Let me give you an example. Lattice QCD is an area
of physics whose goal is basically to take the periodic table and just compute the whole thing
from first principles. This is not the search for theory of everything. We already know the theory
that's supposed to produce as output the periodic table, which atoms are stable, how heavy they are,
all that good stuff. They're spectral lines. It's a theory, Lattice QCD, you can put it on
your t-shirt. Our colleague Frank Wilczek got the Nobel Prize for working on it,
but the math is just too hard for us to solve. We have not been able to start with these equations
and solve them to the extent that we can predict, oh yeah, and then there is carbon, and this is
what the spectrum of the carbon atom looks like. But the awesome people are building these super
computer simulations where you just put in these equations and you make a big cubic lattice of space,
or actually it's a very small lattice because you're going down to the subatomic scale,
and you try to solve it. But it's just so computationally expensive that we still haven't
been able to calculate things as accurately as we measure them in many cases. And now machine
learning is really revolutionizing this. So my colleague Fiola Shanahan at MIT, for example,
she's been using this really cool machine learning technique called normalizing flows,
where she's realized she can actually speed up the calculation dramatically
by having the AI learn how to do things faster. Another area like this where we suck up an
enormous amount of super computer time to do physics is black hole collisions. So now that
we've done the sexy stuff of detecting a bunch of this, LIGO and other experiments, we want to be
able to know what we're seeing. And so it's a very simple conceptual problem. It's the two-body
problem. Newton solved it for classical gravity hundreds of years ago, but the two-body problem
is still not fully solved. For black holes. Yes, a nice thing is gravity, because they won't
just orbit each other forever anymore, two things. They give off gravitational waves,
and eventually they crash into each other. And the game, what you want to do is you want to figure
out, okay, what kind of wave comes out as a function of the masses of the two black holes,
as a function of how they're spinning, relative to each other, etc. And that is so hard. It can
take months of super computer time on massive numbers of cores to do it, you know. Wouldn't
it be great if you can use machine learning to greatly speed that up, right? Now you can use
the expensive old GOFI calculation as the truth, and then see if machine learning can figure out
smarter, faster way of getting the right answer. Yet another area, like computational physics,
these are probably the big three that suck up the most computer time, lattice QCD, black hole
collisions, and cosmological simulations, where you take not a subatomic thing and try to figure
out the mass of the proton, but you take something that's enormous and try to look at how all the
galaxies get formed in there. There again, there are a lot of very cool ideas right now about how
you can use machine learning to do this sort of stuff better. The difference between this and the
big data is you kind of make the data yourself, right? And then finally, we're looking over the
physical landscape and seeing what can we hammer with machine learning. So we talked about experimental
data, big data, discovering cool stuff that we humans then look more closely at. Then we talked
about taking the expensive computations we're doing now and figuring out how to do the much faster
and better with AI. And finally, let's go really theoretical. So things like discovering equations,
having deep fundamental insights, this comes, this is something closest to what I've been
doing in my group. We talked earlier about the whole AI Feynman project where if you just have
some data, how do you automatically discover equations that seem to describe this well that
you can then go back as a human and work with and test and explore. And you asked a really good
question also about if this is sort of a search problem in some sense, that's very deep actually
what you said because it is. Suppose I ask you to prove some mathematical theorem. What is a
proof in math? It's just a long string of steps, logical steps that you can write out with symbols.
And once you find it, it's very easy to write a program to check whether it's a valid proof or not.
So why is it so hard to prove it then? Well, because there are ridiculously many possible
candidate proofs you could write down, right? If the proof contains 10,000 symbols, even if
there are only 10 options for what each symbol could be, that's 10 to the power of 1000 possible
proofs, which is way more than there are atoms in our universe, right? So you could say it's
trivial to prove these things. You just write a computer, generate all strings, and then check,
is this a valid proof? No. Is this a valid proof? No. And then you just keep doing this forever.
It is fundamentally a search problem. You just want to search the space of all strings
of symbols to find the one that is the proof, right? And there's a whole area of machine
learning called search. How do you search with some giant space to find the needle in the haystack?
It's easier in cases where there's a clear measure of good, like you're not just right or
wrong, but this is better and this is worse. You can maybe get some hints as to which direction to
go in. That's why we talked about neural networks work so well. I mean, that's such a human thing
of that moment of genius of figuring out the intuition of good, essentially. I mean, we thought
that that was... Or is it? Maybe it's not right. We thought that about chess, right? Exactly.
That the ability to see like 10, 15, sometimes 20 steps ahead was not a calculation that humans
were performing. It was some kind of weird intuition about different patterns, about board
positions, about the relative positions, that somehow stitching stuff together. And a lot of
it is just like intuition. But then you have like alpha, like zero be the first one that did
the self-play. It just came up with this. It was able to learn through self-play mechanism,
this kind of intuition. Exactly. But just like you said, it's so fascinating to think
whether in the space of totally new ideas, can that be done in developing theorems?
We know it can be done by neural networks because we did it with the neural networks
in the craniums of the great mathematicians of humanity, right? And I'm so glad you brought
up alpha zero because that's the counter example. It turned out we were flattering ourselves when we
said intuition is something different. It's only humans can do it. It's not the information
processing. It used to be that way. Again, it's really instructive, I think, to compare
the chess computer deep blue that beat Kasparov with alpha zero that beat Lisadol at the go.
Because for deep blue, there was no intuition. Humans had programmed in some intuition.
After humans had played a lot of games, they told the computer, you know,
count the pawn as one point, the bishop is three points, the rook is five points, and so on. You
add it all up, and then you add some extra points for past pawns and subtract if the opponent has
it and blah, blah, blah, blah. And then what deep blue did was just search. Just very brute force
tried many, many moves ahead, all these combinations and approved research, and it could think much
faster than Kasparov, and it won, right? And that, I think, inflated our egos in a way it
shouldn't have, because people started to say, yeah, yeah, it's just brute force search, but
has no intuition. Alpha zero really popped our bubble there, because what alpha zero does,
yes, it does also do some of that research, but it also has this intuition module,
which in GeekSpeak is called a value function, where it just looks at the board and comes up
with a number for how good is that position. The difference was no human told it how good
the position is. It just learned it. And mu zero is the coolest or scariest of all,
depending on your mood, because the same basic AI system will learn what the good board position
is, regardless of whether it's chess, or Go, or Shogi, or Pac-Man, or Lady Pac-Man, or Breakout,
or Space Invaders, or any number, a bunch of other games. You don't tell it anything, and it gets
this intuition after a while for what's good. So this is very hopeful for science, I think,
because if it can get intuition for what's a good position there, maybe it can also get intuition
for what are some good directions to go if you're trying to prove something. I often,
one of the more most fun things in my science career is when I've been able to prove some
theorem about something, and it's very heavily intuition guided, of course. I don't sit and
try all random strings. I have a hunch that this reminds me a little bit about this other proof
I've seen for this thing. So maybe I, first, what if I try this? No, that didn't work out.
But this reminds me, actually, the way this failed reminds me of that. So combining the intuition
with all these brute force capabilities, I think it's going to be able to help physics, too.
Do you think there will be a day when an AI system, being the primary contributor,
let's say 90% plus wins the Nobel Prize in physics?
Obviously, they'll give it to the humans, because we humans don't like to give prizes to machines.
It'll give it to the humans behind the system. You could argue that AI has already been involved
in some Nobel Prizes, probably maybe something with black holes and stuff like that.
Yeah, we don't like giving prizes to other life forms. If someone wins a horse racing contested,
they don't give the prize to a horse either. That's true. But do you think that we might be
able to see something like that in our lifetimes when AI. So the first system, I would say,
that makes us think about a Nobel Prize seriously is Alpha Fold. It's making us think about in
medicine physiology, a Nobel Prize, perhaps discoveries that are a direct result of something
that's discovered by Alpha Fold. Do you think in physics we might be able to see that in our
lifetimes? I think what's probably going to happen is more of a blurring of the distinctions.
Today, if somebody uses a computer to do a computation that gives them the normal prize,
nobody's going to dream of giving the prize to the computer. They're going to be like,
that was just a tool. I think for these things also, people are just going to,
for a long time, view the computer as a tool. But what's going to change is the ubiquity
of machine learning. I think at some point in my lifetime, finding a human physicist who
knows nothing about machine learning is going to be about almost as hard as it is today finding
a human physicist who doesn't says, oh, I don't know anything about computers or I don't use math.
It would just be a ridiculous concept. But the thing is, there is a magic moment though,
like with Alpha Zero, when the system surprises us in a way where the best people in the world
truly learn something from the system in a way where you feel like it's another entity.
The way people, the way Magnus Carlson, the way certain people are looking at the work of Alpha
Zero, it truly is no longer a tool in the sense that it doesn't feel like a tool. It feels like
some other entity. There is a magic difference where you're like, if an AI system is able to
come up with an insight that surprises everybody in some major way that's a phase shift in our
understanding of some particular science or some particular aspect of physics, I feel like that
is no longer a tool. Then you can start to say that it perhaps deserves the prize.
For sure, the more important, the more fundamental transformation of the 21st century
science is exactly what you're saying, which is probably everybody will be doing machine
learning to some degree. If you want to be successful at unlocking the mysteries of science,
you should be doing machine learning. But it's just exciting to think about whether there'll be
one that comes along that's super surprising and they'll make us question who the real inventors
are in this world. I think the question of, isn't if it's going to happen, but when?
But it's important, in my mind, the time when that happens is also more or less the same time
when we get artificial general intelligence. Then we have a lot bigger things to worry about
than whether we should get the Nobel Prize or not. Because when you have machines that can outperform
our best scientists at science, they can probably outperform us at a lot of other stuff as well,
which can at a minimum make them incredibly powerful agents in the world. I think
it's a mistake to think we only have to start worrying about loss of control when machines
get to AGI across the board, when they can do all our jobs. Long before that, they'll be hugely
influential. We talked at length about how the hacking of our minds with algorithms trying to
get us glued to our screens has already had a big impact on society. There was an incredibly
dumb algorithm in the grand scheme of things, the supervised machine learning, yet it had huge
impact. I don't want us to be lulled into false sense of security and think there won't be any
societal impact until things reach human level because it's happening already. I was just
thinking the other week, when I see some scaremonger going, oh, the robots are coming. The implication
is always that they're coming to kill us. Maybe you should have worried about that if you were
in Nagorno-Karabakh during the recent war there. But more seriously, the robots are coming right
now, but they're mainly not coming to kill us. They're coming to hack us. They're coming to
hack our minds into buying things that maybe we didn't need to vote for people who may not have
our best interest in mind. And it's kind of humbling, I think, actually, as a human being to
admit that it turns out that our minds are actually much more hackable than we thought.
And the ultimate insult is that we are actually getting hacked by the machine learning algorithms
that are in some objective sense much dumber than us. But maybe we shouldn't be so surprised
because how do you feel about the cute puppies? Love them. You would probably argue that in some
across-the-board measure, you're more intelligent than they are. But boy, our cute puppies good at
hacking us, right? They move into our house, persuade us to feed them and do all these things.
And what do they ever do for us other than being cute and making us feel good, right?
So if puppies can hack us, maybe we shouldn't be so surprised if pretty dumb machine learning
algorithms can hack us too. Not to speak of cats, which is another level. And I think we should,
to counter your previous point about there, let us not think about evil creatures in this world.
We can all agree that cats are as close to objective evil as we can get. But that's just
me saying that. Okay, so you have you seen the cartoon? I think it's maybe the onion
with this incredibly cute kitten. And it just says underneath something that thinks about murder
all day. Exactly. That's accurate. You mentioned offline that there might be a link between
post-biological AGI and SETI. So last time we talked, you've talked about this intuition that
we humans might be quite unique in our galactic neighborhood. Perhaps our galaxy, perhaps the
entirety of the observable universe, we might be the only intelligence civilization here, which is
and you argue pretty well for that thought. So I have a few little questions around this. One,
the scientific question, in which way would you be, if you were wrong in that intuition,
in which way do you think you would be surprised? Why were you wrong? We find out that you ended
up being wrong. In which dimension? So is it because we can't see them? Is it because the
nature of their intelligence or the nature of their life is totally different than we can
possibly imagine? Is it because the, I mean, something about the great filters and surviving
them? Or maybe because we're being protected from signals, all those explanations for
why we haven't heard a big loud like red light that says we're here. Yeah. So there are actually
two separate things there that I could be wrong about the two separate claims that I made, right?
One one of them is I made the claim, I think most civilizations,
when you're going from simple bacteria like things to space, fair space colonizing civilizations,
they spend only a very, very tiny fraction of their other life being where we are.
That I could be wrong about. The other one I could be wrong about is quite different statement
that I think that actually I'm guessing that we are the only civilization in our observable
universe from which light has reached us so far that that's actually gotten far enough to invent
telescopes. So let's talk about maybe both of them in turn because they really are different.
The first one, if you look at the n equals one, the data point we have on this planet,
right? So we spent four and a half billion years futzing around on this planet with life, right?
We got, and most of it was pretty lame stuff from an intelligence perspective, you know,
bacteria and then the dinosaurs spent, then the things gradually accelerated, right? Then the
dinosaurs spent over 100 million years stomping around here without even inventing smartphones.
And then very recently, you know, it's only, we've only spent 400 years going from Newton
to us, right? In terms of technology. And we've looked at what we've done even,
you know, when I was a little kid, there was no internet even. So it's, I think it's pretty
likely for in this case of this planet, right, that we're either going to really get our act
together and start spreading life into space, the century and doing all sorts of great things,
or we're going to wipe out. It's a little hard. If I could be wrong in the sense that maybe what
happened on this earth is very atypical. And for some reason, what's more common on other planets
is that they spend an enormously long time futzing around with the ham radio and things,
but they just never really take it to the next level for reasons I don't have. I haven't understood
and I'm humble and open to that. But I would bet at least 10 to one that our situation is more
typical. Because the whole thing with Moore's law and accelerating technology, it's pretty obvious
why it's happening. Everything that grows exponentially, we call it an explosion, whether
it's a population explosion or a nuclear explosion, it's always caused by the same thing. It's that
the next step triggers a step after that. Today's technology enables tomorrow's technology,
and that enables the next level. Because the technology is always better, of course,
the steps can come faster and faster. On the other question that I might be wrong about,
that's the much more controversial one, I think. But before we close out on this thing about
the first one, if it's true that most civilizations spend only a very short amount of their total
time in the stage, say, between inventing telescopes or mastering electricity and leaving
there and doing space travel, if that's actually generally true, but then that should apply also
elsewhere out there. We should be very, very surprised if we find some random civilization
and we happen to catch them exactly in that very, very short stage. It's much more likely that we
find a planet full of bacteria, or that we find some civilization that's already post-biological
and has done some really cool galactic construction projects in their galaxy.
Would we be able to recognize them, do you think? Is it possible that we just can't? I mean,
this post-biological world, could it be just existing in some other dimension? Could it be
just all a virtual reality game for them or something? I don't know. That it changes completely
where we won't be able to detect. We have to be, honestly, very humble about this. I think I said
earlier the number one principle being scientists is you have to be humble and willing to acknowledge
that everything we think, guess, might be totally wrong. Of course, you could imagine some civilization
where they all decide to become Buddhists and very inward looking and just move into their little
virtual reality and not disturb the flora and fauna around them and we might not notice them.
But this is a numbers game, right? If you have millions of civilizations out there or billions
of them, all it takes is one with a more ambitious mentality that decides, hey, we are going to go
out and settle a bunch of other solar systems in maybe galaxies and then it doesn't matter if
they're a bunch of quiet Buddhists. We're still going to notice that expansionist one, right?
And it seems like quite the stretch to assume that, you know, we know even in our own galaxy that there
are probably a billion or more planets that are pretty Earth-like and many of them were formed
over a billion years before ours, so had a big head start. So if you actually assume also that
life happens kind of automatically on an Earth-like planet, I think it's pretty
quite the stretch to then go and say, okay, so we have billions of and other billion
civilizations out there that also have our level of tech and they all decided to become Buddhists
and not a single one decided to go like go Hitler on the galaxy and say we need to go on and colonize
or not a single one decided for more benevolent reasons to go out and get more resources.
That seems like a bit of a stretch, frankly. And this leads into the second thing you challenge
me to be that I might be wrong about, how rare or common is life, you know? So Francis Drake,
when he wrote down the Drake equation, multiplied together, I used a number of factors and said
we don't know any of them, so we know even less about what you get when you multiply together
the whole product. Yeah. Since then, a lot of those factors have become much better known.
One of his big uncertainties was how common is it that a solar system even has a planet?
Right. Well, now we know it very common.
Earth-like planets, we know it better. There are diamond dozen, there are many,
many others, even in our galaxy. At the same time, you know, we have, thanks to, I'm a big
supporter of the SETTI project and its cousins, and I think we should keep doing this and we've
learned a lot. We've learned that so far, all we have is still unconvincing hints,
nothing more, right? And there are certainly many scenarios where it would be dead obvious.
If there were 100 million other human-like civilizations in our galaxy, it would not
be that hard to notice some of them with today's technology, and we haven't, right? So what we
can, what we can say is, well, okay, we can rule out that there is a human-level civilization on
the moon, and in fact, the many nearby solar systems where we, we cannot rule out, of course,
that there is something like Earth sitting in a galaxy five billion light years away.
But we've ruled out a lot, and that's already kind of shocking, given that there are all these
planets there, you know, so like, where are they? Where are they all? That's the classic Fermi paradox.
So my argument, which might very wrong, it's very simple really, it just goes like this,
okay, we have no clue about this. It could be the probability of getting life on a random planet,
it could be 10 to the minus one a priori or 10 to the minus 10, 10 to the minus 20, 10 to the
minus 30, 10 to the minus 40. Basically, every order of magnitude is about equally likely.
When you then do the math, and ask how close is our nearest neighbor, it's again equally likely
that it's 10 to the 10 meters away, 10 to 20 meters away, 10 to the 30 meters away. We can,
we have some nerdy ways of talking about this with Bayesian statistics and a uniform log prior,
but that's irrelevant. This is the simple basic argument. And now comes the data. So we can say,
okay, how many were, there are all these orders of magnitude 10 to the 26 meters away, there's the
edge of our observable universe. If it's farther than that light hasn't even reached us yet.
If it's less than 10 to the 16 meters away, well, it's within Earth's rate, it's no farther
way than the sun, we can definitely rule that out, you know. So I think about it like this,
a priori, before we looked with telescopes, you know, it could be 10 to 10 meters, 10 to 20, 10
to 30, 10 to 40, 10 to 50, 10 to blah, blah, blah, equally likely anywhere here. And now we've ruled
out like this chunk. Yeah. And most of it is outside. And here is the edge of our global
universe. Yes. Yep. So I'm certainly not saying I don't think there's any life elsewhere in space.
If space is infinite, then you're basically 100% guaranteed that there is. But the probability
that there is life, that the nearest neighbor, it happens to be in this little region between
where we would have seen it already. And when where we will never see it, there's actually
significantly less than one, I think. And I think there's a moral lesson from this, which is really
important, which is to be good stewards of this planet and this shot we've had, you know, it can
be very dangerous to say, Oh, you know, it's fine if we nuke our planet or ruin the climate or mess
it up with unaligned AI, because I know there is this nice Star Trek fleet out there, they're
going to swoop in and take over where we failed, just like it wasn't the big deal that the Easter
Island losers wiped themselves out. That's a dangerous way of loading yourself into false
sense of security. If it's actually the case that it might be up to us, and only us, the whole future
of intelligent life in our observable universe, then I think it's both, it really puts a lot of
responsibility on our shoulders, inspiring. It's a little bit terrifying, but it's also inspiring.
But it's empowering, I think most of all, because because the biggest problem today is I see this
even when I teach, right? So many people feel that it doesn't matter what they do, or we do,
we feel disempowered. Oh, it makes no difference. This is about as far from that as you can come,
but we realize that what we do on our little spinning ball here in our lifetime, you know,
could make the difference for the entire future of life in our universe. You know,
how empowering is that? Yeah, survival of consciousness. I mean, on the other,
a very similar kind of empowering aspect of the Drake equation is, say there is a huge number
of intelligent civilizations that spring up everywhere. But because of the Drake equation,
which is the lifetime of a civilization, maybe many of them hit a wall. And just like you said,
it's clear that that for us, the great filter, the one possible great filter seems to be coming,
you know, in the next 100 years. So it's also empowering to say, okay, well,
we have a chance to not, I mean, the way great filters work, you just get most of them.
Exactly. Nick Bostrom has articulated this really beautifully too. Every time
yet another search for life on Mars comes back negative or something. I'm like, yes!
Our odds for us surviving this is the best. You already made the argument and brought
rush there, right? But just the uncockets, right? The point is, we already know
there is a crap ton of planets out there that are Earth-like. And we also know that most of them
do not seem to have anything like our kind of life on them. So what went wrong? There's
clearly one step along the evolutionary, at least one filter roadblock in going from no life
to spacefaring life. And where is it? Is it in front of us, or is it behind us? Right?
If there's no filter behind us, and we keep finding all sorts of little mice on Mars,
and whatever, right? That's actually very depressing because that makes it much more
likely that the filter is in front of us. And what actually is going on is the ultimate
dark joke that whenever a civilization invents sufficiently powerful tech, it's just set
their clock and then after a little while it goes poof for one reason or other and wipes itself out.
Now wouldn't that be utterly depressing if we're actually doomed? Whereas if it turns out
that there is a really, there is a great filter early on that for whatever reason seems to be
really hard to get to the stage of sexually reproducing organisms or even the first ribosome
or whatever, right? Or maybe you have lots of planets with dinosaurs and cows, but for some
reason they tend to get stuck there and never invent smartphones. All of those are huge
big boosts for our own odds because being there done that, it doesn't matter how hard or unlikely
it was that we got past that roadlock because we already did. And then that makes it likely that
the filter is in our own hands. We're not doomed. So that's why I think the fact that
life is rare in the universe, it's not just something that there is some evidence for,
but also something we should actually hope for.
So that's the end, the mortality, the death of human civilization that we've been discussing
in life, maybe prospering beyond any kind of great filter. Do you think about your own death?
Does it make you sad that you may not witness some of the, you know, you lead a research group
on working some of the biggest questions in the universe, actually, both on the physics and the
AI side? Does it make you sad that you may not be able to see some of these exciting things
come to fruition that we've been talking about? Of course. Of course it sucks, the fact that I'm
going to die. I remember once when I was much younger, my dad made this remark that life is
fundamentally tragic. And I'm like, what are you talking about, daddy? And then many years later,
I felt now I feel I totally understand what he means. You know, we grow up, we're little kids,
and everything is infinite, and it's so cool. And then suddenly we find out that actually, you know,
you got us only this is that you can get game over at some point. So of course it's, it's, it's,
it's something that's sad. Are you afraid?
No, not in the sense that I think anything terrible is going to happen after I die or
anything like that. No, I think it's really going to be a game over. But it's more that
it makes me very acutely aware of I just have what a wonderful gift this is that it gets to
be alive right now. And, and is a steady reminder to just live life to the fullest and really enjoy
it. Because it is finite, you know, and I think actually, and we know we all get the regular
reminders when someone near and dear to us dies, that that one day it's going to be our turn,
adds this kind of focus. I wonder what it would feel like, actually, to be an immortal being,
if they might even enjoy some of the wonderful things of life a little bit less, just because
there isn't that finiteness. Do you think that could be a feature, not a bug, the fact that we
beings are finite? Maybe there's lessons for engineering and artificial intelligence systems
as well that are conscious. Like, do you think it makes, is it possible that the reason the
pistachio ice cream is delicious is the fact that you're going to die one day? And you will not have
all the pistachio ice cream that you could eat because of that fact?
Well, let me say two things. First of all, it's actually quite profound what you're saying. I do
think I appreciate the pistachio ice cream a lot more knowing that I will, there's only a finite
number of times I get to enjoy that. And I can only remember a finite number of times in the past.
And, moreover, my life is not so long that it just starts to feel like things are repeating
themselves in general. It's so new and fresh. I also think, though, that
death is a little bit overrated in the sense that it comes from a sort of outdated view of
physics and what life actually is. Because if you ask, okay, what is it that's going to die exactly?
What am I really? When I say I feel sad about the idea of myself dying, am I really sad that the skin
cell here is going to die? Of course not, because it's going to die next week anyway and I'll grow
a new one, right? And it's not any of my cells that I'm associating really with who I really am,
nor is it any of my atoms or quarks or electrons. In fact, basically all of my atoms
get replaced on a regular basis, right? So what is it that's really me? From a more modern physics
perspective, it's the information in processing Amy. That's where my memories, that's my memories,
that's my values, my dreams, my passion, my love. That's what's really fundamentally me.
And frankly, not all of that will die when my body dies. Like Richard Feynman, for example,
his body died of cancer, you know? But many of his ideas that he felt made him very him actually
live on. This is my own little personal tribute to Richard Feynman, right? I try to keep a little
bit of him alive in myself. I've even quoted him today, right?
Yeah, he almost came alive for a brief moment in this conversation.
Yeah. Yeah. And this honestly gives me some solace. You know, when I work as a teacher,
I feel if I can actually share a bit about myself that my students feel worthy enough
to copy and adopt as a part of things that they know or they believe or aspire to,
now I live on also a little bit in them, right? And so being a teacher is a little bit of what I
that's something also that contributes to making me a little teeny bit less mortal,
right? Because I'm not at least not all going to die all at once, right?
And I find that a beautiful tribute to people we generally respect. If we can, if we can remember
them and carry in us the things that we felt was the most awesome about them, right? Then they live on.
And I'm getting a bit emotional here, but it's a very beautiful idea you bring up there.
I think we should stop this old fashioned materialism and just equate who we are
with our quirks and electrons. There's no scientific basis for that, really. And
it's also very uninspiring. Now, if you look a little bit towards the future, right?
One thing which really sucks about humans dying is that even though some of their
teachings and memories and stories and ethics and so on will be copied by those around them,
hopefully, a lot of it can't be copied and just dies with them with a brain. And that really
sucks. That's the fundamental reason why we find it so tragic when someone goes from having all
this information there to just gone ruin, right? With more post biological intelligence,
that's going to shift a lot, right? The only reason it's so hard to make a backup of your brain
in its entirety is exactly because it wasn't built for that, right? If you have a future
machine intelligence, there's no reason for why it has to die at all if you want to copy it
whatever is it into some other quirk blob, right? You can copy not just some of it, but all of it,
right? And so in that sense, you can get immortality because all the information can be
copied out of any individual entity. And it's not just mortality that will change if we get
more post biological life. It's also with that, I think, very much the whole individualism we
have now, right? The reason that we make such a big difference between me and you is exactly because
we're a little bit limited in how much we can copy. Like, I would just love to go like this
and copy your Russian skills, Russian speaking skills. Wouldn't it be awesome? But I can't.
I have to actually work for years. I want to get better on it. But if we were robots,
just copy and paste freely, then that loses completely. It washes away the sense of what
immortality is. And also individuality a little bit, right? We would start feeling much more,
maybe we would feel much more collaborative with each other if we can just, hey, I'll give you
my, you can give me your Russian and I'll give you whatever. And suddenly you can speak Swedish.
Maybe that's less a bad trade for you, but whatever else you want from my brain, right?
And there have been a lot of sci-fi stories about hive minds and so on where experiences
can be more broadly shared. And I think one, we don't, I don't pretend to know what it would feel
like to be a super intelligent machine, but I'm quite confident that however it feels about
mortality and individuality will be very, very different from how it is for us.
Well, for us, mortality and finiteness seems to be pretty important at this particular moment.
And so all good things must come to an end, just like this conversation Max.
I saw that coming. Sorry, this is the world's worst transition. I could talk to you forever.
It's such a huge honor that you spend time with me.
It's my honor. Thank you so much for getting me essentially to start this podcast by doing
the first conversation, making me realize falling in love with conversation in itself.
And thank you so much for inspiring so many people in the world with your books,
with your research, with your talking, and with other like this ripple effect of friends,
including Elon and everybody else that you inspire. So thank you so much for talking today.
Thank you. I feel so fortunate that you're doing this podcast and getting so many
interesting voices out there into the ether and not just the five second sound bites,
but so many of the interviews are what you do. You really let people go in
into depth in a way which we sorely need in this day and age and that I got to be number one.
But I feel super honored. Yeah, you started it. Thank you so much, Max.
Thanks for listening to this conversation with Max Tegmark. And thank you to our sponsors,
The Jordan Harbinger Show, ForSigmatic Mushroom Coffee, BetterHelp Online Therapy,
and ExpressVPN. So the choice is wisdom, caffeine, sanity, or privacy. Choose wisely,
my friends. And if you wish, click the sponsor links below to get a discount and to support this
podcast. And now let me leave you with some words from Max Tegmark. If consciousness is the way that
information feels when it's processed in certain ways, then it must be substrate independent.
It's only the structure of information processing that matters, not the structure of the matter doing
the information processing. Thank you for listening and hope to see you next time.