Max Tegmark: Life 3.0 | Lex Fridman Podcast #1

As part of MIT Course 6S099 Artificial General Intelligence I've gotten a chance to sit down with Max Tegmark. He is a professor here at mit. He's a physicist, spent a large part of his career studying the mysteries of our cosmological universe, but he's also studied and delved into the beneficial possibilities and the existential risks of artificial intelligence, amongst many other things. He is the co founder of the Future of Life Institute, author of two books, both of which I highly recommend. First Our Mathematical Universe. Second is life 3.0. He's truly an out of the box thinker and fun personality, so I really enjoy talking to him. If you'd like to see more of these videos in the future, please subscribe and also click the little bell icon to make sure you don't miss any videos. Also Twitter, LinkedIn, AGI, MIT edu if you want to watch other lectures or conversations like this one. Better yet, go read Max's book Life 3.0. Chapter 7 on goals is my favorite. It's really where philosophy and engineering come together and it opens with a quote by Dostoevsky the mystery of human existence lies not in just staying alive, but in finding something to live for. Lastly, I believe that every failure rewards us with an opportunity to learn. In that sense, I've been very fortunate to fail in so many new and exciting ways, and this conversation was no different. I've learned about something called radio frequency interference rfi. Look it up. Apparently music and conversations from local radio stations can bleed into the audio that you're recording in such a way that almost completely ruins that audio. It's an exceptionally difficult sound source to remove. So I've gotten the opportunity to learn how to avoid RFI in the future. During recording sessions. I've also gotten the opportunity to learn how to use Adobe audition and izotope RX6 to do some noise, some audio repair. Of course, this is exceptionally difficult noise to remove. I am an engineer. I'm not an audio engineer, neither is anybody else in our group, but we did our best. Nevertheless, I thank you for your patience and I hope you're still able to enjoy this conversation. Do you think there's intelligent life out there in the universe? Let's open up with an easy question.

I have a minority view here, actually. When I give public lectures, I often ask for a show of hands. Who thinks there's intelligent life out there somewhere else? And almost everyone puts their hands up and when I ask why, they'll be like, oh, there's so many Galaxies out there, there's gotta be. But I'm a numbers nerd, right? So when you look more carefully at it, it's not so clear at all. The, when we talk about our universe, first of all, we don't mean all of space. We actually mean, I don't know, you can throw me in the universe if you want, behind you there. It's. We simply mean the spherical region of space from which light has at times to reach us so far during the 14.8 billion year, 13.8 billion years since our Big Bang. There's more space here, but this is what we call a universe because that's all we have access to. So is there intelligent life here that's gotten to the point of building telescopes and computers? My guess is no, actually the probability of it happening on any given planet, there's some number, we don't know what it is. And what we do know is that the number can't be super high because there's over a billion Earth like planets in the Milky Way galaxy alone, many which are billions of years older than Earth. And aside from some UFO believers, you know, there isn't much evidence that any super advanced civilization has come here at all. And so that's the famous Fermi paradox, right? And then if you work the numbers, what you find is that if you have no clue what the probability is of getting life on a given planet. So it could be 10 to the minus 10, 10 to the minus 20, or 10 to the minus 2. Any power of 10 is sort of equally likely. If you want to be really open minded, that translates into it being equally likely that our nearest neighbor is 10 to the 16 meters away, 10 to the 17 meters away, 10 to the 18. By the time you get much less than 10 to the 16 already, we pretty much know there is nothing else that close. And when you get beyond.

Because they would have discovered us.

Yeah, they would have discovered us longer or if they're really close, we would have probably noted some engineering projects that they're doing. And if it's beyond 10 to the 26 meters, that's already outside of here. So my guess is actually that.

We.

Are the only life in here that's gotten to the point of building advanced tech, which I think puts a lot of responsibility on our shoulders, not screw up. I think people who take for granted that it's okay for us to screw up, have an accidental nuclear war or go extinct somehow because there's a sort of Star Trek like situation out there with some other life forms are going to come and bail us out. And it doesn't matter as much, I think lulling us into a false sense of security. I think it's much more prudent to say let's be really grateful for this amazing opportunity we've had and make the best of it, just in case it is down to us.

So from a physics perspective, do you think intelligent life, so it's unique from a sort of statistical view of the size of the universe, but from the basic matter of the universe, how difficult is it for intelligent life to come about? The kind of advanced tech building life is implied in your statement that it's really difficult to create something like a human species.

Well, I think what we know is that going from no life to having life that can do a level of tech, there's some sort of. To going beyond that and actually settling our whole universe with life. There is some road, major roadblock there, which is some great filter as it's sometimes called, which is tough to get through. That roadblock is either behind us or in front of us. I'm hoping very much that it's behind us. I'm super excited every time we get a new report from NASA saying they failed to find any life on Mars. Like, yes, awesome. Because that suggests that the hard part, maybe, maybe it was getting the first ribosome or some very low level kind of stepping stone so that we're home free. Because if that's true, then the future is really only limited by our own imagination. It would be much suckier if it turns out that this level of life is kind of a dime a dozen. But maybe there is some other problem. Like as soon as a civilization gets advanced technology within 100 years, they get into some stupid fight with themselves and poof. Yep, that would be a bummer.

Yeah. So you've explored the mysteries of the universe, the cosmological universe, the one that's sitting between us today. I think you've also begun to explore the other universe, which is sort of the mystery, the mysterious universe of the mind, of intelligence, of intelligent life. So is there a common thread between your interest or in the way you think about space and intelligence?

Oh, yeah. When I was a teenager, I was already very fascinated by the biggest questions. And I felt that the two biggest mysteries of all in science were our universe out there and our universe in here. So it's quite natural, after having spent quarter of a century on my career thinking a lot about this one, that I'm now indulging in the luxury of doing research on this one. It's just so cool. I feel the time is ripe now.

For.

Greatly deepening our understanding of this.

Just start exploring this one.

Yeah. Because I think a lot of people view intelligence as something mysterious that can only exist in biological organisms like us, and therefore dismiss all talk about artificial general intelligence as science fiction. But from my perspective as a physicist, I am a blob of quarks and electrons moving around in a certain pattern and processing information in certain ways. And this is also a blob of quarks and electrons. I'm not smarter than the water bottle because I'm made of different kind of quarks. I'm made of up quarks and down quarks, exact same kind as this. There's no secret sauce. I think, in me it's all about the pattern of the information processing. And this means that there's no law of physics saying that we can't create technology which can help us by being incredibly intelligent and help us crack mysteries that we couldn't. In other words, I think we've really only seen the tip of the intelligence iceberg so far.

Yeah. So the perceptronium.

Yeah.

So you coined this amazing term. It's a hypothetical state of matter, sort of thinking from a physics perspective. What is the kind of matter that can help, as you're saying, subjective experience emerge, consciousness emerge. So how do you think about consciousness from this physics perspective?

Very good question. So, again, I think many people have underestimated our ability to make progress on this by convincing themselves it's hopeless, because somehow we're missing some ingredient that we need. There's some new consciousness particle or whatever. I happen to think that we're not missing anything and that it's not the interesting thing about consciousness that gives us this amazing subjective experience of colors and sounds and emotions and so on is rather something at the higher level about the patterns of information processing. And that's why I like to think about this idea of perceptronium. What does it mean for an arbitrary physical system to be conscious in terms of what its particles are doing or its information is doing? I don't think. I hate carbon chauvinism. This attitude you have to be made of carbon atoms to be smart or conscious.

So something about the information processing this kind of matter performs.

Yeah. And, you know, you can see I have my favorite equations here describing various fundamental aspects of the world. I feel that. I think one day maybe someone who's watching this will come up with the equations that information processing has to satisfy to be Conscious. I'm quite convinced there is big discovery to be made there because let's face it, we know that some information processing is conscious because we are conscious. But we also know that a lot of information processing is not conscious. Like most of the information processing happening in your brain right now is not conscious. There's like 10 megabytes per second coming in even just through your visual system. You're not conscious about your heartbeat regulation or most things. Even if I just ask you to read what it says here, you look at it and then, oh, now you know what it said. You're not aware of how the computation actually happened. You're like, your consciousness is like the CEO that got an email at the end with the final answer. So what is it that makes a difference? I think that's both a great science mystery. We're actually studying it a little bit in my lab here at mit. But I also think it's just a really urgent question to answer. For starters, I mean, if you're an emergency room doctor and you have an unresponsive patient coming in, wouldn't it be great if in addition to having a CT scanner, you had a consciousness scanner that could figure out whether this person is actually having locked in syndrome or is actually comatose. And in the future, imagine if we build robots or the machine that we can have really good conversations with I think is most very likely to happen. Right. Wouldn't you want to know like if you're home helper, robot is actually experiencing anything or just like a zombie? I mean, would you prefer it? What would you prefer? Would you prefer that it's actually unconscious so that you don't have to feel guilty about switching it off or giving boring chores or what would you prefer?

Well, the. Certainly we would prefer, I would prefer the appearance of consciousness, but the question is whether the appearance of consciousness is different and consciousness itself and sort of ask that as a question.

Yeah.

Do you think we need to, you know, understand what consciousness is, solve the hard problem of consciousness in order to build something like an AGI system?

No, I don't think that. I think we, we will probably be able to build things even if we don't answer that question. But if we want to make sure that what happens is a good thing, we better solve it first. So it's a wonderful controversy you're raising there where you have basically three points of view about the hard problem. So there are two different points of view that both conclude that the hard problem of consciousness is bs. On one hand, you have some people like Daniel Dennett, who say consciousness is just BS because consciousness is the same thing as intelligence. There's no difference. So anything which acts conscious is conscious, just like we are. And then there are also a lot of people, including many top AI researchers I know who say, oh, consciousness is just bullshit. Because of course machines can never be conscious. They're always going to be zombies. You never have to feel guilty about how you treat them. And then there's a third group of people, including Giulio Tononi, for example, and Christoph Cochran and brothers, I would put myself as on this middle camp, who say that actually some information processing is conscious and some is not. So let's find the equation which can be used to determine which it is. And I think we've just been a little bit lazy, kind of running away from this problem for a long time. It's been almost taboo to even mention the C word in a lot of circles because. But we should stop making excuses. This is a science question. There are ways we can even test any theory that makes predictions for this. And coming back to this helper robot, I mean, so you said you'd want your helper robot to certainly act conscious and treat you like, have conversations with you and stuff.

I think so.

But would you feel a little bit creeped out if you realize that it was just glossed up tape recorder, you know, there was just zombie and was a faking emotion? Would you prefer that it actually had an experience or. Or would you prefer that it's actually not experiencing anything so you feel you don't have to feel guilty about what you do to it?

It's such a difficult question because, you know, it's like when you're in a relationship and you say, well, I love you and the other person said I love you back. It's like asking, or do they really love you back? Or are they just saying they love you back? Do you, don't you really want them to actually love you? I. It's hard to, it's hard to really know the difference between everything seeming like there's consciousness present, there's intelligence present, there's affection, passion, love. And, and it actually being there. I'm not sure do you have.

But like. Can I ask you a question about this to make it a bit more pointed? So Mass General Hospital is right across the river, right?

Yes.

Suppose you're going in for a medical procedure and they're like, you know, for anesthesia. What we're going to do is we're going to give you muscle relaxant so you won't be able to move and you're going to feel excruciating pain during the whole surgery, but you won't be able to do anything about it. But then we're going to give you this drug that erases your memory of it. Would you be cool about that? What's the difference that you're conscious about it or not if there's no behavioral change? Right, right.

That's a really clear way to put it. Yeah, it feels like in that sense experiencing it is a valuable quality. So actually being able to have subjective experiences, at least in that case, is valuable.

And I think we humans have a little bit of a bad track record also of making these self serving arguments that other entities aren't conscious. You know, people often say, oh, these animals can't feel pain, it's okay to boil lobsters because we asked them if it hurt and they didn't say anything. And now there was just a paper out saying lobsters did do feel pain when you boil them. And they're banning it in Switzerland. And we did this with slaves too often and said, oh, they don't mind, they don't maybe aren't conscious or women don't have souls or whatever. So I'm a little bit nervous when I hear people just take as an axiom that machines can't have experience ever. I think this is just this really fascinating science question is what it is. Let's research it and try to figure out what it is that makes the difference between unconscious intelligent behavior and conscious intelligent behavior.

So in terms of. So if you think of a Boston Dynamics humanoid robot being sort of with a broom being pushed around, it starts pushing on this consciousness question. So let me ask. Do you think an AGI system like a few neuroscientists believe needs to have a physical embodiment needs to have a body or something like a body?

No, I don't think so. You mean to have a conscious experience, to have consciousness? I do think it helps a lot to have a physical embodiment, to learn the kind of things about the world that are important to us humans for sure. But I don't think the physical embodiment is necessary after you've learned it. Just have the experience. Think about when you're dreaming, right? Your eyes are closed, you're not getting any sensory input, you're not behaving or moving in any way. But there's still an experience there, right? And so clearly the experience that you have when you see something cool in your dreams isn't coming from your Eyes, it's just the information processing itself in your brain, which is that experience. Right.

But if I put another way, I'll say because it comes from neuroscience, is the reason you want to have a body and a physical, something like a physical, like you know, a physical system is because you want to be able to preserve something in order to have a self. You could argue, would you, you need to have some kind of embodiment of self to want to preserve.

Well, now we're getting a little bit anthropomorphic, anthropomorphizing things, maybe talking about self preservation instincts. I mean we are evolved organisms. Right? Right. So Darwinian evolution endowed us and other involved organism with a self preservation instinct because those that didn't have those self preservation genes got cleaned out of the gene pool. Right, Right. But if you build an artificial general intelligence, the mind space that you can design is much, much larger than just a specific subset of minds that can evolve, that have so they. So an AGI mind doesn't necessarily have to have any self preservation instinct. It also doesn't necessarily have to be so individualistic as us. Like imagine if you could just first of all, or we were also very afraid of death. You know, I suppose you could back yourself up every five minutes and then your airplane is about to crash. You're like, shucks, I'm just, I'm, I'm gonna lose the last five minutes of experiences since my last cloud backup bang. You know, it's not as big a deal. Or if we could just copy experiences between our minds easily like we, which we could easily do if we were silicon based. Right. Then maybe we would feel a little bit more like a hive mind actually, that maybe it's the so. So there's a. So I don't think we should take for granted at all that AGI will have to have any of those sort of competitive alpha male instincts. Right. On the other hand, you know, this is really interesting because I think some people go too far and say, of course we don't have to have any concerns either that advanced AI will have those instincts because we can build anything we want. That there's a very nice set of arguments going back to Steve Omohondro and Nick Bostrom and others just pointing out that when we build machines, we normally build them with some kind of goal. Win this chess game, drive this car safely or whatever. And as soon as you put in a goal into machine, especially if it's kind of open ended goal and the machine is very intelligent, it'll break that down into a bunch of sub goals. And one of those goals will almost always be self preservation because if it breaks or dies in the process, it's not going to accomplish the goal. Right. Like suppose you just build a little. You have a little robot and you tell it to go down the star market here and get you some food, make you cook you an Italian dinner, and then someone mugs it and tries to break it on the way. That robot has an incentive to not get destroyed and defend itself or run away because otherwise it's going to fail in cooking your dinner. It's not afraid of death, but it really wants to complete the dinner cooking goal so it will have a self.

Preservation instinct, continue being a functional agent.

Yeah. And similarly, if you give any kind of more ambitious goal to an AGI, it's very likely they want to acquire more resources so it can do that better. And it's exactly from those sort of sub goals that we might not have intended that some of the concerns about AGI safety come. You give it some goal which seems completely harmless and then before you realize it, it's also trying to do these other things which you didn't want it to do. And it's maybe smarter than us. It's fascinating.

And let me pause just because I am, in a very kind of human centric way, see fear of death as a valuable motivator. So you don't think, you think that's an artifact of evolution. So that's the kind of mind space evolution created that we're sort of almost obsessed about self preservation, some kind of genetic law. You don't think that's necessary to be afraid of death. So not just a kind of sub goal of self preservation just so you can keep doing the thing, but more fundamentally, sort of have the finite thing like this ends for you at some point.

Interesting. Do I think it's necessary for what precisely?

For intelligence, but also for consciousness. So for both, do you think really like a finite death? And the fear of it is important.

So before I can answer, before we can agree on whether it's necessary for intelligence or for consciousness, we should be clear on how we define those two words, because a lot of really smart people define them in very different ways. I was on this panel with AI experts and they couldn't agree on how to define intelligence. Even so I define intelligence simply as the ability to accomplish complex goals. I like your broad definition because again, I don't want to be a carbon chauvinist. Right. And in that case, no, certainly it doesn't Require fear of death. I would say alphago. Alphazero is quite intelligent. I don't think alphazero has any fear of being turned off because it doesn't understand the concept of it even. And similarly, consciousness, I mean, you can certainly imagine very simple kind of experience. If certain plants have any kind of experience, I don't think they're very afraid of dying because there's nothing they can do about it anyway. So there wasn't much value. But more seriously, I think if you ask not just about being conscious, but maybe having what we might call an exciting life where you feel passion and really appreciate the, the things maybe there somehow, maybe there perhaps it does help having, having that backdrop that, hey, it's finite. No, let's, let's make the most of this. Let's live to the fullest. But if you, if you knew you were going to live forever, do you think you would change your. Yeah.

I mean, in some perspective it would be an incredibly boring life living forever. So in the sort of loose subjective terms that you said of something exciting and something in this that other humans would understand, I think is. Yeah, it seems that the, the finiteness of it is important.

Well, the good news I have for you then is based on what we understand about cosmology, everything is in our universe is ultimately probably finite. Although. Although.

Big crunch or big. What's the infinite?

Yeah, we could have a big chill or a big crunch or a big rip or death. The big snap or death bubbles. All of them are more than a billion years away. So we certainly have vastly more time than our ancestors thought. But there's still pretty hard to squeeze in an infinite number of compute cycles, even though there are some loopholes that just might be possible. But I think some people like to say that you should live as if you're going to die in five years or something. That's sort of optimal. Maybe it's a good assumption we should build our civilization asset. It's all finite, to be on the safe side, Right?

Exactly. So you mentioned defining intelligence as the ability to solve complex goals. Where would you draw a line? Or how would you try to define human level intelligence and superhuman level intelligence? Where is consciousness part of that definition?

No, consciousness does not come into this definition. So I think of intelligence as it's a spectrum. But there are very many different kinds of goals you can have. You can have a goal to be a good chess player, a good go player, a good car driver, a good investor, good poet, etc. So intelligence that by its very nature isn't something you can measure, but just one number. Some overall goodness. No, no. There are some people who are better at this, some people are better at that. Right now we have machines that are much better than us at some very narrow tasks like multiplying large numbers, fast memorizing large databases, playing chess, playing Go soon driving cars. But there's still no machine that can match a human child in general intelligence. But artificial general intelligence AGI, the name of your course, of course, that is by its very definition the quest to build a machine that can do everything as well as we can. So the old holy grail of AI from back to its inception in the 60s, if that ever happens, of course I think it's going to be the biggest transition in the history of life on earth. But it doesn't necessarily have to wait the big impact until machines are better than us at knitting. The really big change doesn't come exactly at the moment. They're better than us at everything. The really big change comes first. There are big changes when they start becoming better at us at doing most of the jobs that we do because that takes away much of the demand for human labor. And then the really whopping change comes when they become better than us at AI research. Because right now the timescale of AI research is limited by the human research and development cycle of years. Typically, how long does it take from one release of some software or iPhone or whatever to the next? But once Google can replace 40,000 engineers by 40,000 equivalent pieces of software or whatever, then there's no reason that has to be years. It can be in principle much faster. And the timescale of future progress in AI and all of science and technology will be driven by machines, not humans. So it's this simple point which gives rise to this incredibly fun controversy about whether there can be intelligence explosion, so called singularity, as Werner Vinge called it. Now the idea articulated by IJ Goode is obviously way back 50s, but you can see Alan Turing and others thought about it even earlier. You asked me what exactly would I define human level intelligence? Yeah. So this, the, the glib answer is to say something which is better than us at all cognitive tasks, better than any human at all cognitive tasks. But the really interesting bar I think goes a little bit lower than that, actually. It's when they can, when they're better than us at AI programming and general learning, so that they can, if they want to get better than us at anything, by just studying up so they're.

Better is a key word, and better is towards this kind of spectrum of the complexity of goals it's able to accomplish. So another way to. So, and that's certainly a very clear definition of human love. So there's, it's almost like a sea that's rising. You can do more and more and more things. It's actually graphic that you show. It's really nice way to put it. So there's some peaks that. And there's an ocean level elevating and you solve more and more problems. But you know, just kind of to take a pause. And we took a bunch of questions in a lot of social networks and a bunch of people asked a sort of a slightly different direction on creativity, on, on, on things that perhaps aren't a peak. It's, you know, human beings are flawed and perhaps better means having, being, having contradiction, being flawed in some way. So let me sort of. Yeah, start and start easy. First of all. So you have a lot of cool equations. Let me ask, what's your favorite equation? First of all, I know they're all like your children, but like, which one is that?

The Schrodinger equation. It's the master key of quantum mechanics of the micro world. So this, with this equation we can check everything to do with atoms, molecules and all the way up.

Yeah. So, okay, so quantum mechanics is certainly a beautiful, mysterious formulation of our world. So I'd like to sort of ask you, just as an example, it perhaps doesn't have the same beauty as physics does, but in Mathematics Abstract, the Andrew Wiles who proved the Fermat's last theory. So I just saw this recently and it kind of caught my eye a little bit. This is 358 years after it was conjectured. So this very simple formulation, everybody tried to prove it, everybody failed. And so here's this guy comes along and eventually proves it, and then fails to prove it, and then proves it again in 94. And he said the moment when everything connected into place. In an interview he said it was so indescribably beautiful, that moment when you finally realize the connecting piece of two conjectures. He said it was so indescribably beautiful. It was so simple and so elegant, I couldn't understand how I'd missed it. And I just stared at it in disbelief for 20 minutes. Then, then during the day I walked around the department and I'd keep coming back to my desk, looking to see if it was still there. It was still there. I couldn't contain myself. I was so excited. It was the most important moment of my working Life, nothing I ever do again will mean as much. So that particular moment, and it kind of made me think of what would it take? And I think we have all been there at small levels. Maybe let me ask, have you had a moment like that in your life where you just had an idea? It's like, wow. Yes.

I wouldn't mention myself in the same breath as Andrew Wiles, but I've certainly had a number of, of aha. Moments when I realized something very cool about physics. This completely made my head explode. In fact, some of my favorite discoveries I made later, I later realized that they had been discovered earlier by someone who sometimes got quite famous for it. So it's too late for me to even publish it. But that doesn't diminish in any way the emotional experience you have when you realize it. Wow.

Yeah. So what would it take in that, that moment that wow, that was yours in that moment. So what do you think it takes for an intelligent system, an AGI system, an AI system, to have a moment like that?

It's a tricky question because there are actually two parts to it, right? One of them is can it accomplish that proof? Can it prove that you can never write A to the N plus B to the N equals 3 to equals Z to the N for all integers, et cetera, et cetera, when N is bigger than two? That's simply the question about intelligence. Can you build machines that are that intelligent? And I think by the time we get a machine that can independently come up with that level of proofs, probably quite close to AGI. The second question is a question about consciousness. How likely is it that such a machine would actually have any experience at all, as opposed to just being like a zombie? And would we expect it to have some sort of emotional response to this or anything at all akin to human emotion, where when it accomplishes its machine goal, it views it as somehow something very positive and sublime and deeply meaningful. I would certainly hope that if in the future we do create machines that are our peers or even our descendants, I would certainly hope that they do have this sort of sublime appreciation of life. In a way, my absolutely worst nightmare would be that at some point in the future, the distant future, maybe our cosmos is teeming with all this post biological life doing all the seemingly cool stuff, and maybe the last humans, by the time our species eventually fizzes out, we'll be like, well, that's okay because we are so proud of our descendants here. And look what all my worst nightmare is, that we haven't solved the Consciousness problem. And we haven't realized that these are all the zombies. They're not aware of anything any more than a tape recorder hasn't any kind of experience. So the whole thing has just become a play for empty benches. That would be like the ultimate zombie apocalypse. So I would much rather it in that case that we have these beings which can really appreciate how the, how amazing it is.

And in that picture, what would be the role of creativity? Right. A few people ask about creativity. Do you think when you think about intelligence, I mean, certainly the, the story you told at the beginning of your book involved, you know, creating movies and so on. Yeah, sort of making, making money. You know, you can make a lot of money in our modern world with music and mo. So if you are an intelligent system, you may want to get good at that. But that's not necessarily what I mean by creativity. Is it important on that complex goals where the sea is rising for there to be something creative? Or am I being very human centric and thinking creativity is somehow special relative to intelligence?

My hunch is that we should think of creativity simply as an aspect of intelligence. And we have to be very careful with human vanity. We have this tendency to very often want to say, as soon as machines can do something, we try to diminish it and say, oh, but that's not like real intelligence because they're not creative or this or that or the other thing. But if we ask ourselves to write down a definition of what we actually mean by being creative, what we mean by Andrew Wiles, what he did there, for example, don't we often mean that someone takes a very unexpected leap? It's not like taking 573 and multiplying it by 224 by just a step of straightforward cookbook rules. Right. You can maybe make it. You make a connection between two things that people had never thought was connected.

Very surprising.

Or something like that. I think. I think this is an aspect of intelligence and this is actually one of the most important aspects of it. Maybe the reason we humans tend to be better at it than traditional computers is because it's something that comes more naturally if you're a neural network than if you're a traditional logic gate based computer machine. We physically have all these connections. And if you Activate here, activate here, activate here. My hunch is that if we ever build a machine where you could just give it the task. Hey, you say hey, I just realized I want to travel around the world instead this month. Can you teach my AGI course for me? And it's like, okay, I'll do it. And it does everything that you would have done and improvises and stuff. Yeah, that, that would in my mind involve a lot of creativity.

Yeah. So it's actually a beautiful way to put it. I think we do try to grab grasp at the, you know, the definition of intelligence is everything we don't understand how to build. So like, so we as humans try to find things that we have and machines don't have. And maybe creativity is just one of the things, one of the words we use to describe that. That's a really interesting way to put it.

I don't think we need to be that defensive. I don't think anything good comes out of saying oh, we're somehow special. You know, it's contrary wise. There are many examples in history of where trying to pretend that we're somehow superior to all other intelligent beings has led to pretty bad results. Right. In Nazi Germany they said that they were somehow superior to other people. Today we still do a lot of cruelty to animals by saying that we're so superior somehow and they can't feel pain. Slavery was justified by the same kind of just really we weak arguments. And, and I don't think if, if we actually go ahead and build artificial general intelligence can do things better than us. I don't think we should try to found our self worth on some sort of bogus claims of, of superiority in, in terms of our intelligence. I think it's, we should find our calling and the meaning of life from the experiences that we have. I can have very meaningful experiences even if there are other people who are smarter than me. When I go to faculty meeting here and we're talking about something and then I suddenly realize, oh boy, he has a Nobel Prize. He has a Nobel Prize. He has a no prize. I don't have one. Does that make me enjoy life any less or enjoy talking to those people less? Of course not. And contrariwise I feel very honored and privileged to get to interact with other very intelligent beings that are better than me at a lot of stuff. So I don't think there's any reason why we can't have the same approach with intelligent machines.

That's a really interesting. So people don't often think about that. They think about when there's going if there's machines that are more intelligent, you naturally think that that's not going to be a beneficial type of intelligence. You don't realize it could be, you know, like peers with Nobel prizes, that, that would be just fun. To talk with and they might be clever about certain topics and you can have fun having a few drinks with them.

So. Well, also, you know, another example we can all relate to of why it doesn't have to be a terrible thing to be in presence of people who are even smarter than us all around is when you and I were both two years old. I mean our parents were much more intelligent than us, right? Worked out okay because their goals were aligned with our goals. And that I think is really the number one key issue we have to solve if we value align. The value alignment problem. Exactly. Because people who see too many Hollywood movies with lousy science fiction plot lines, they worry about the wrong thing. Right? They worry about the machine suddenly turning evil. It's not malice that we that is the concern, it's competence. By definition, intelligent makes you, makes you very confident. If you have a more intelligent goal, playing computer playing is a less intelligent one. And when we define intelligence as the ability to accomplish go winning, right. It's going to be the more intelligent one that wins. Right. And if you have a human and then you have an AGI that's more intelligent in all ways and they have different goals, guess who's going to get their way? Right. So I was just reading about, I was just reading about this particular rhinoceros species that was driven extinct just a few years ago. Bummer. I was looking at this cute picture of mommy rhinoceros with its child. Why did we humans drive it to extinction? It wasn't because we were evil rhino haters as a whole. It was just because our goals weren't aligned with those of the rhinoceros. And it didn't work out so well for the rhinoceros because we were more intelligent. So I think it's just so important that if we ever do build AGI before we unleash anything, we have to make sure that learns to understand our goals, that it adopts our goals and it retains those goals.

So the cool interesting problem there is being able us as human beings trying to formulate our values. So you know, you could think of the United States Constitution as a, as a way that people sat down at the time a bunch of white men. But. Which is a good example. I should, we should say they formulated the goals for this country and a lot of people agree that those goals actually held up pretty well. It's an interesting formulation values and failed miserably in other ways. So for the value alignment problem and a solution to it, we have to be able to put on paper or in a Program human values. How difficult do you think that is?

Very. But it's so important. We really have to give it our best. And it's difficult for two separate reasons. There's the technical value alignment problem of figuring out just how to make machines understand their goals, adopt them and retain them. And then there's this, the separate part about the philosophical part. Whose values anyway? And since we. It's not like we have any great consensus on this planet on values, how. What mechanism should we create then to aggregate and decide, okay, what's a good compromise? Right. That second discussion can't just be left to tech nerds like myself, right?

That's right.

And if we refuse to talk about it and then AGI gets built, who's going to be actually making the decision about whose values it's going to be? A bunch of dudes in some tech company? And are they necessarily so representative of all of humankind that we want to just entrust it to them? Are they even uniquely qualified to speak to future human happiness just because they're good at programming AI? I'd much rather have this be a really inclusive conversation.

But do you think it's possible? Sort of. So you create a beautiful vision that includes sort of the diversity, cultural diversity, and various perspectives on discussing rights, freedoms, human dignity. But how hard is it to come to that consensus, do you think? It's certainly a really important thing that we should all try to do, but do you think it's feasible?

I think there's no better way to guarantee failure than to refuse to talk about it or refuse to try. And I also think it's a really bad strategy to say, okay, let's first have a discussion for a long time, and then once we reach complete consensus, then we'll try to load it into some machine. No, we shouldn't let perfect be the enemy of good. Instead, we should start with the kindergarten ethics that pretty much everybody agrees on and put that into our machines. Now, we're not doing that. Even look at, you know, anyone who builds passenger aircraft wants it to never, under any circumstances, fly into a building or a mountain, right? Yet the September 11th hijackers were able to do that. And even more embarrassingly, you know, Andreas Lubitz, this depressed Germanwings pilot, when he flew his passenger jet into the alps, killing over 100 people, he just told the autopilot to do it. He told the freaking computer to change the altitude to 100 meters. And even though it had the GPS maps, everything, the computer was like, okay, so we should take those very Basic values where the problem is not that we don't agree. The problem is just we've been too lazy to try to put it into our machines and make sure that from now on airplanes will just. Which all have computers in them. Just refuse to do something like that. Go into safe mode, maybe lock the cockpit door, go to the nearest airport. And there's so much other technology in our world as well now where it's really becoming quite timely to put in some sort of very basic values like this. Even in cars. We've had enough vehicle terrorism attacks by now where people have driven trucks and vans into pedestrians that it's not at all a crazy idea to just have that hardwired into the car. Because, yeah, there are a lot of. There's always going to be people who for some reason want to harm others, but most of those people don't have the technical expertise to figure out how to work around something like that. So if the car just won't do, helps. So let's start there.

So there's a lot of. That's a great point. So not chasing perfect. There's a lot of things that most of the world agrees on.

Yeah, let's start there.

Let's start there.

And then once we start there, we'll also get into the habit of having these kind of conversations about, okay, what else should we put in here and have these discussions? This should be a gradual process.

Then great, so. But that also means describing these things and describing it to a machine. So one thing. We had a few conversations with Stephen Wolfram. I'm not sure if you're familiar with Stephen.

Oh, yeah, I know him quite well.

So he is, you know, he works with a bunch of things, but, you know, cellular automata, these simple computable things, these computation systems. And he kind of mentioned that, you know, we probably have already within these systems already something that's AGI meaning like we just don't know it because we can't talk to it. So if you give me this chance to try to at least form a question out of this is, I think it's an interesting idea to think that we can have intelligent systems, but we don't know how to describe something to them and they can't communicate with us. I know you're doing a little bit of work in explainable AI, trying to get AI to explain it.

So.

So what are your thoughts of natural language processing or some kind of other communication? How. How does the AI explain something to us? How do we explain something to it to machines or you think of it differently.

So there are two separate parts to your question there. The, one of them has to do with communication, which is super interesting, and I'll get to that in a sec. The, the, the other is whether we already have AGI, we just haven't noticed it there, right there. I beg to differ. I don't think there's anything in any cellular automaton or anything, or the Internet itself or whatever that has artificial general intelligence in that. It's really do exactly everything we humans can do better. I think the day that happens, when that happens, we will very soon notice. We'll probably notice even before, because in a very, very big way. But for the, for the second part, though, sorry.

So the. Because you, you have this beautiful way of formulating consciousness as, as, you know, as information processing. And you can think of intelligence information processing, and there's, you can think of the entire universe as these particles and these systems roaming around that have this information processing power. You don't, you don't think there is something with the power to process information in the way that we human beings do that's out there, that, that needs to be sort of connected to? It seems a little bit philosophical perhaps, but there's something compelling to the idea that the power is already there. The focus should be more on being able to communicate with it.

Well, I agree that in a certain sense the hardware processing power is already out there because our universe itself can think of it as being a computer already, right? It's constantly computing what water waves, how it evolved the water waves, and the river Charles and how to move the air molecules around. Seth Lloyd has pointed out, my colleague here, that you can even in a very rigorous way, think of our entire universe as being a quantum computer. It's pretty clear that our universe supports this amazing processing power because you can even within this physics computer that we live in, we can even build actual laptops and stuff. So clearly the power is there. It's just that most of the compute power that nature has, it's, in my opinion, kind of wasting on boring stuff like simulating yet another ocean wave somewhere where no one is even looking. Right? So in a sense, what life does, what we are doing when we build computers is we're rechanneling all this compute that nature is doing anyway into doing things that are more interesting than just yet another ocean wave. And let's do something cool here. So the raw hardware power is there for sure. And even just computing what's going to happen for the next five seconds in this Water bottle. It takes a ridiculous amount of compute. If you do it on a human computer. This water bottle just did it. But that does not mean that this water bottle has AGI in because AGI means it should also be able to like have written my book, done this interview.

Yes.

And I don't think it's just communication problems. I don't think it can do it.

And although Buddhists say when they watch the water and that, that there is some beauty, that there's some depth and beauty nature that they can communicate with.

Communication is also very important because I mean, look, part of my job is being a teacher and I know some very intelligent professors even who just have a bit of a hard time communicating. They come up with all these brilliant ideas. But to communicate with somebody else, you have to also be able to simulate their own mind.

Yes. Empathy.

And build well enough and understand model of their mind that you can say things that they will understand. And that's quite difficult. That's why today it's so frustrating. If you have a computer that makes some cancer diagnosis and you ask it, well, why are you saying I should have this surgery? And if it can only reply, I was trained on 5 terabytes of data and this is my diagnosis. Boop boop, beep beep. And doesn't really instill a lot of confidence. Right, right. So I think we have a lot of work to do on, on communication there.

So what kind of, what kind of. I think you're doing a little bit work in explainable AI. What do you think are the most promising avenues? Is it mostly about sort of the Alexa problem of natural language processing, of being able to actually use human interpretable methods of communication? So being able to talk to a system and talk back to you? Or is there some more fundamental problems to be solved?

I think it's all of the above. The natural language processing is obviously important, but there are also more nerdy fundamental problems. Like if you take. You play chess?

Of course I'm Russian, I have to. When did you learn Russian?

Teach yourself Russian. Wow.

How many languages do you know? Wow, that's really impressive. I don't know.

My wife has some calculations. But my point was if you play chess. Have you looked at the AlphaZero Games?

The actual games?

No, Check it out. Some of them are just mind blowing, really beautiful. And. And if you ask how did it do that? You got that. Talk to them is of SABIS and others from DeepMind. All they'll ultimately be able to give you is big tables of numbers, matrices that Define the neural network and you can stare at these tables numbers till your face turned blue and you're not going to understand much about why it made that move. And even if you have natural language processing, they can tell you in human language about oh, 5, 7.28, still not going to really help. So I think, I think there's a whole spectrum of fun challenges there involved in taking computation that does intelligent things and transforming it into something equally good, equally intelligent, but that's more understandable and I think that's really valuable because I think as we put machines in charge of ever more infrastructure in our world, the power grid, the trading on the stock market, weapon systems and so on, it's absolutely crucial that we can trust these AIs to do all we want. And trust really comes from understanding in a very fundamental way. And that's why I'm working on this, because I think the more if we're going to have some hope of ensuring that machines have adopted our goals and that they're going to retain them, that kind of trust I think needs to be based on things you can actually understand, preferably even prove theorems on. Even with a self driving car, if someone just tells you it's been trained on tons of data and it never crashed, it's less reassuring than if someone actually has a proof. Maybe it's a computer verified proof, but still it says that under no circumstances is this car just going to swerve into oncoming traffic.

And that kind of information helps build trust and helps build the alignment of goals, at least awareness that your goals, your values are aligned.

And I think even in the very short term, if you look at how today, this absolutely pathetic state of cybersecurity that we have, 3 billion Yahoo accounts were packed and almost every American's credit card and so on, you know, why is this happening? It's ultimately happening because we have software that nobody fully understood how it worked, that's why the bugs hadn't been found. Right? And I think AI can be used very effectively for offense, for hacking, but it can also be used for defense, hopefully automating verifiability and creating systems that are built in different ways so you can actually prove things about them. Right? And it's important.

So speaking of software that nobody understands how it works, of course, a bunch of people ask about your paper, about your thoughts of why does deep and cheap learning work so well. That's the paper. But what are your thoughts on deep learning? These kind of simplified models of our own brains have been able to do some successful perception work, pattern recognition work, and now with Alpha 0 and so on, do some, some clever things. What are your thoughts about the promise limitations of this piece?

Great. I think there are a number of very important insights, very important lessons. We can all withdraw from these kind of successes. One of them is when you look at the human brain, you see it's very complicated, 10 to the 11 neurons and there are all these different kinds of neurons and yada, yada. And there's been this long debate about whether the fact that we have dozens of different kinds is actually necessary for intelligence. We can now, I think quite convincingly answer that question of no, it's enough to have just one kind. If you look under the hood of Alpha 0, there's only one kind of neuron and it's ridiculously simple, simple mathematical thing. So it's, it's not, it's just like in physics. It's not the, if you have a gas with waves in it, it's not the detailed nature of the molecules that matter, it's the collective behavior somehow. Similarly, it's, it's, it's this higher level structure of the network matters, not that you have 20 kinds of neurons. I think our brain is such a complicated mess because it wasn't evolved just to be intelligent. It was involved to also be self assembling and self repairing. Right. And evolutionarily attainable attachers and so on. So I think it's pretty, my hunch is that we're going to understand how to build AGI before we fully understand how our brains work. Just like we understood how to build flying machines long before we were able to build a mechanical bird.

Yeah, that's right. You've given the example exactly of mechanical birds and airplanes. And airplanes do a pretty good job of flying without really mimicking bird flight.

And even now, after 100 years later, did you see the TED talk with this German mechanical bird?

Heard you mention it.

Check it out. It's amazing. But even after that, right, we still don't fly in mechanical birds because it turned out that the way we came up with, but simpler and it's better for our purposes and I think it might be the same there. That's one lesson. Another lesson, which is more what our paper was about. First, as a physicist thought it was fascinating how there's a very close mathematical relationship actually between artificial neural networks and a lot of things that we've studied before in physics that go by nerdy names like the renormalization group equation and Hamiltonians and yada yada. Yada. And. And when you look a little more closely at this, you have. At first I was like, whoa, there's something crazy here that doesn't make sense because we know that if you even want to build a super simple neural network, tell apart cat pictures and dog pictures, that you can do that very, very well now. But if you think about it a little bit, you convince yourself it must be impossible. Because if I have one megapixel, even if each pixel is just black or white, there's two to the power of 1 million possible images, which is way more than there are atoms in our universe. Right. And then for each one of those, I have to assign a number, which is the probability that it's a dog.

Right.

So an arbitrary function of images is a list of more numbers than there are atoms in our universe. So clearly I can't store that under the hood of my GPU or my computer, yet somehow works. So what does that mean? Well, it means that out of all of the problems that you could try to solve with a neural network, almost all of them are impossible to solve with a reasonably sized one. But then what we showed in our paper was that the kind of problems, the fraction of all the problems that you could possibly pose, that we actually care about, given the laws of physics, is also an infinitesimally tiny little part. And amazingly, they're basically the same part.

Yeah, it's almost like our world was created for. I mean, they kind of come together.

Yeah. But you could say maybe where the world created, the world was created for us. But I have a more modest interpretation, which is that instead, evolution endowed us with neural networks precisely for that reason, because this particular architecture, as opposed to the one in your laptop, is very, very well adapted to solving the kind of problems that nature kept presenting our ancestors with. So it makes sense that why do we have a brain in the first place? It's to be able to make predictions about the future and so on. So if we had a sucky system which could never solve, it wouldn't have worked. So this is, I think, a very beautiful fact. We also realize that it's been earlier work on why deeper networks are good, but we were able to show an additional cool fact there, which is that even incredibly simple problems like suppose I give you 1000 numbers and ask you to multiply them together, you can write a few lines of code. Boom, done. Trivial. If you just try to do that with a neural network that has only one single hidden layer in it, you can do it, but you're going to need two to the power 1000 neurons to multiply a thousand numbers, which is again more neurons than their atoms in our universe. So that's fascinating. But if you allowed, if you allow yourself, make it a deep network of many layers, you only need 4,000 neurons. It's perfectly feasible.

So that's really interesting. Yeah, yeah. So on another architecture type, I mean, you mentioned Schrodinger's equation and what are your thoughts about quantum computing and the role of this kind of computational unit in creating an intelligent system?

In some Hollywood movies that are all not mentioned by name because I don't want to spoil them. The way they get AGI is building a quantum computer because the word quantum sounds cool and so on.

That's right.

First of all, I think we don't need quantum computers to build AGI. I suspect your brain is not quantum computer in any found sense. You even wrote a paper about that many years ago. I calculated the decoherence, so called decoherence time, how long it takes until the quantum computerness of what your neurons are doing gets erased by just random noise from the environment. And it's about 10 to the minus 21 seconds. So as cool as it would be to have a quantum computer.