Max Tegmark is a physicist and AI researcher at MIT. Please support this podcast by checking out our sponsors:
- The Jordan Harbinger Show: https://www.jordanharbinger.com/lex/
- Four Sigmatic: https://foursigmatic.com/lex and use code LexPod to get up to 60% off
- BetterHelp: https://betterhelp.com/lex to get 10% off
- ExpressVPN: https://expressvpn.com/lexpod and use code LexPod to get 3 months free

EPISODE LINKS:
News Project Explainer Video: https://www.youtube.com/watch?v=PRLF17Pb6vo
News Project Website: https://www.improvethenews.org/
Max's Twitter: https://twitter.com/tegmark
Max's Website: https://space.mit.edu/home/tegmark/
Future of Life Institute: https://futureoflife.org/
Lex Fridman Podcast #1: https://www.youtube.com/watch?v=Gi8LUnhP5yU

PODCAST INFO:
Podcast website: https://lexfridman.com/podcast
Apple Podcasts: https://apple.co/2lwqZIr
Spotify: https://spoti.fi/2nEwCF8
RSS: https://lexfridman.com/feed/podcast/
Full episodes playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4
Clips playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOeciFP3CBCIEElOJeitOr41

OUTLINE:
0:00 - Introduction
2:49 - AI and physics
16:07 - Can AI discover new laws of physics?
24:57 - AI safety
42:33 - Extinction of human species
53:31 - How to fix fake news and misinformation
1:15:05 - Autonomous weapons
1:30:28 - The man who prevented nuclear war
1:40:36 - Elon Musk and AI
1:54:14 - AI alignment
2:00:16 - Consciousness
2:09:20 - Richard Feynman
2:13:30 - Machine learning and computational physics
2:24:28 - AI and creativity
2:35:42 - Aliens
2:51:25 - Mortality

CONNECT:
- Subscribe to this YouTube channel
- Twitter: https://twitter.com/lexfridman
- LinkedIn: https://www.linkedin.com/in/lexfridman
- Facebook: https://www.facebook.com/LexFridmanPage
- Instagram: https://www.instagram.com/lexfridman
- Medium: https://medium.com/@lexfridman
- Support on Patreon: https://www.patreon.com/lexfridman

The following is a conversation with Max Tegmark, his second time on the podcast.

In fact, the previous conversation was episode number 1 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 1 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 1 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, 4 Sigmatic Mushroom Coffee, Better Help online therapy, and ExpressVPN.

So the choice is wisdom, caffeine, sanity, or privacy.

And if you wish, click the sponsor links below to get a discount to 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 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 1 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, Alex Friedman.

And now, here's my conversation with Max Tagmark.

So people might not know this, but you were actually episode number 1 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 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 that media wrote about, but that 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?

So, 1 of the amazing things you're part of is the AI Institute for Artificial Intelligence and Fundamental interactions.

The idea is something I'm very on fire with, which is basically AI meets physics.

And It's been almost 5 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, were like,

But then gradually, together with some colleagues, we're 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 5 years, MIT and a bunch of neighboring universities here also.

And I noticed now those colleagues who were looking at me funny have stopped asking why I'm, what the point is of this because it's becoming more clear.

And I really believe that, of course, AI can help physics a lot to do better physics, but physics can also help AI a lot, both by building better hardware.

My colleague, Marin Solzhachich, for example, is working on 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.

You know, AI has, to a significant extent, been an engineering discipline where you're just trying to make things that work and being less, more interested in maybe selling them than in figuring out exactly how they work and proving theorems about that they will always work, right?

You know, when Elon Musk sends a rocket to the International Space Station, they didn't just train with machine learning, oh, let's fire it a little bit more to the left, a bit more to the right, 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, right?

So we trust them not because some sales rep told us to, but because they've earned our trust.

We can, and really safety critical things even prove that they will always do what we expect them to do.

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.

Similarly, alpha fold too, 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 GPT-3 that can spout off English texts, which sometimes really, really blows you away.

You can look at the Google, at DeepMind's Mew 0, which doesn't just kick our butt in Go and Chess 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.

And that's fine if it's just some dancing robots and the worst thing that can happen is 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, right?

It's less fine if that's what's controlling your self-driving car or your nuclear power plant.

And 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 overtrust 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, right?

Was it that that little simple system was evil?

Of course not, but we didn't understand it as well as we should have, right?

We didn't even understand that we didn't understand.

The Humility is really at the core of being a scientist.

I think step 1, 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.

Like, 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 gonna go up with constant speed until it realizes that its natural motion is down.

Because 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?

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.

So I don't know if you, did you hear of a company called Knight Capital?

So good, that means you didn't invest in them earlier.

They deployed this automated trading system.

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 off.

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, this is 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.

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 and stuff.

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.

And my MIT research is entirely focused on demystifying this black box.

That's a good line, intelligible intelligence.

Yeah, we shouldn't settle for something that seems intelligent, but it should be intelligible so that we actually trust it because we understand it.

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?

I think there is, we've made a bit of a mistake where we, some people still think that somehow we're never gonna understand neural networks.

And we're just gonna have to learn to live with this.

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 the 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 at chess now.

And in the end, you have something which even though you can look inside it, you have very little idea of how it works.

You can print out tables of all the millions of parameters in there, is it crystal clear now how it's working, no of course not.

Many of my colleagues seem willing to settle for that.

And I'm like, no, that's like the halfway point.

Some have even gone as far as sort of guessing that the mister, the inscrutability of this is where some of the power comes from and some sort of mysticism.

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 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 the knobs did was they just changed random letters in your code.

You change 1 thing and instead of saying print, it says sint, 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 of 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.

So that's the magic of neural networks, the inscrutability.

The differentiability, that every setting of the parameters is a program, and you can tell is

So you don't like the poetry of the mystery of neural networks as the source of its power?

It's so misleading, and above all, it shortchanges us, it makes us underestimate what we can, the good things we can accomplish.

So what we've been doing in my group is basically step 1, train the mysterious neural network to do something well.

And then step 2, do some additional AI techniques to see if we can now transform this black box into something equally intelligent that you can actually understand.

So for example, I'll give you 1 example, this AI Feynman project that we just published, right?

So we took the 100 most famous or complicated equations from 1 of my favorite physics textbooks.

In fact, the 1 that got me into physics in the first place, the Feynman Lectures on Physics.

And so you have a formula, you know, maybe it has what goes into the formula is 6 different variables and then what comes out is 1.

So then you can make like a giant Excel spreadsheet with 7 columns.

You put in just random numbers for the 6 columns for those 6 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 times some constants, that's the famous easy problem we can solve.

We do it all the time in science and engineering.

But the general 1, if it's more complicated functions with logarithms or cosines or other math, it's a very, very hard 1 and probably impossible to do fast in general just because the number of formulas with N symbols 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 a first step towards actually understanding how it works.

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 and a kind of divide and conquer approach.

So we were able to solve all of those 100 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, who's listening to this, can type pip install AI Feynman on the computer and run it.

It can actually do what Johannes Kepler spent 4 years doing when he stared at Mars data until he's like, finally, Eureka, this is an ellipse.

This will do it automatically for you in 1 hour, right?

Or Max Planck, he was looking at how much radiation comes out at different wavelengths from a hot object and discovered the famous black body formula.

I'm actually excited about seeing if we can discover not just old formulas again, but new formulas that no 1 has seen before.

I do 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's, in that way, you've, forcing the explainability issue, really trying to analyze the neural network for the things it knows in order to come up with the final beautiful simple theory underlying the initial system that you were looking at.

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.

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 gonna move and catch it.

But Galileo went 1 step further when he got older.

He went back and was like, wait a minute.

And he helped revolutionize physics as we know it, right?

was a basic neural network in there from childhood that captured 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.

Not only did the formula he got ultimately become more accurate, and similarly this is how Newton got Newton's laws, which is why Elon can send rockets to the space station now.

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.

We've talked about it just in the context of physics now, but hey, 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.

Because 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?

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 is a bunch of neurons doing their thing, right?

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 gone a little bit backwards historically because we humans first discovered good old-fashioned AI, the logic-based AI that we often called 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, it's something so many of our animal friends can do, at least as well as us, right?

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, right, where we can extract out information in symbols, in language, and now even with equations if you're a scientist, right?

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.

And 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 2 together in a useful way.

So if we ever wanna get artificial general intelligence that can do all jobs as well as humans can, right, then that's what's gonna be required, to be able to combine the neural networks with symbolic, combine the old AI with the new AI in a good way.

We do it in our brains, and there seems to be basically 2 strategies I see in industry now.

1 scares the heebie-jeebies out of me and the other 1 I find much more encouraging.

The 1 that scares the heebie-jeebies out of me is this attitude that we're just gonna make ever bigger systems that we still don't understand until they can be as smart as humans.

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 gonna be extinct before too long.

It's just gonna be an utter epic own goal.

So it's that 44 minute, or it's losing money problem, or like the paperclip problem, like where we don't understand how it works and it 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 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.

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, so Can we break that apart just briefly?

For the 50% chance that we'll run to 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 like 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 thought they're, to some degree, thought they were trying to do good.

Oh yeah, I'm sure Hitler thought he was doing good.

Yeah, Stalin, I've been reading a lot about Stalin.

See all Lex Fridman transcripts on Youtube

Max Tegmark: AI and Physics | Lex Fridman Podcast #155