MIT AGI: Future of Intelligence

In this AI video ...

Welcome to MIT Course 6S099 Artificial General Intelligence. Today we have Ray Kurzweil. He is one of the world’s leading inventors, thinkers, and futurists, with a 30-year track record of accurate predictions, called the Restless Genius by the Wall Street Journal and the ultimate thinking machine by Ford’s magazine. He was selected as one of the top entrepreneurs by Inc. Magazine, which described him as the rightful heir to Thomas Edison. PBS selected him as one of the 16 revolutionaries who made America. Ray was the principal investigator of the first CCD flatbed scanner, the first omnifot optical character recognition, the first point of speech reading machines for the blind, the first text-to-speech synthesizer, the first music synthesizer capable of recreating the grand piano and other orchestral instruments, and the first commercially marketed large vocabulary speech recognition. Among where he’s many honors, he received a Grammy Award for outstanding achievements in music technology. He’s the recipient of the National Medal of Technology, was inducted into the National Mentors Hall of Fame, holds 21 honorary doctorates and honors from three U.S. presidents. Ray has written five national bestselling books, including The New York Times Best Seller’s The Singularity is Near, from 2005, and Had to Create a Mind from 2012. He’s co-founder and chancellor of Singularity University, and a director of engineering at Google, heading up a team developing machine intelligence and natural language understanding. Please give Ray a warm welcome. It’s good to be back. I’ve been in this lecture home many times, and walked the infinite corridor. I came here as an undergraduate in 1965. Within a year of my being here, they started a new major called Computer Science. It did not get its own course number. It’s 6-1, even biotechnology recently got its own course number. How many of you are CS majors? How many of you do work in deep learning? How many of you have heard of deep learning? I came here first in 1962 when I was 14. I became excited about artificial intelligence. It had only gotten its name six years earlier, in 1956 Dartmouth Conference by Marvin Minsky and John McCarthy. I wrote Minsky a letter. There was no email back then, and he invited me up. He spent all day with me. He had nothing else to do. He was a consummate educator. I then… The AI field had already bifurcated into two-waring camps. The symbolic school, which Minsky was associated with, and the connection to school was not widely known. In fact, I think it’s still not widely known that Minsky actually invented the neural net in 1953, but he had become negative about it. Larger because there was a lot of hype that these giant brains could solve any problem. The first popular neural net, the perceptron, was being promulgated by Frank Rosenblatt at Cornell. Minsky said, where are you going now? I said to see Rosenblatt at Cornell, so I don’t bother doing that. I went there and Rosenblatt was touting the perceptron that it ultimately would be able to solve any problem. I brought some printed letters that had the camera, and it did a perfect job of recognizing them as long as they were a carrier. Ten different types of style didn’t work at all. He said, but don’t worry. We can take the output of the perceptron. A feeder is the input to another perceptron, and take the output of that and feed it to a third layer. If we add more layers, it’ll get smarter and smarter and generalize. That’s interesting. I think if you tried that, well, no, but it’s high on our research agenda. Things did not move quite as quickly back then as they do now. He died nine years later. Never having tried that idea turns out to be remarkably prescient. I mean, he never tried multi-layer neural nets. And all the excitement we see now about deep learning comes from a combination of two things. Many layer neural nets and the law of accelerating returns, which I’ll get to a little bit later, which is basically the exponential growth of computing, so that we can run these massive nets and handle massive amounts of data. It would be decades before that idea was tried. Several decades later, three-level neural nets were tried. They were a little bit better. They could deal with multiple type styles. Still weren’t very flexible. It’s not hard to add other layers. It’s a very straightforward concept. There was a math problem that disappearing gradient or the exploding gradient, which I’m sure many of you are familiar with. Basically, you need to take maximum advantage of the range of values in the gradients, not let them explode or disappear and lose the resolution. That’s a fairly straightforward mathematical transformation. With that insight, we could now go to 100 layer neural nets. And that’s behind sort of all the fantastic gains that we’ve seen recently. Alpha Go trained on every online game and then became a fair go player. It then trained itself by playing itself and sort past the best human. Alpha Go zero started with no human input at all, within hours of iteration, sort past, Alpha Go. Also, sort past the best chess programs. They had another innovation. Basically, you need to evaluate the quality of the board at each point. They used another 100 layer neural nets to do that evaluation. So there’s still a problem in the field, which is there’s a motto that life begins at a billion examples. One of the reasons I met Google is we have a billion examples. For example, there’s pictures of dogs and cats that are labeled. So you’ve got a picture of a cat and a cat, and then you can learn from it, and you need a lot of them. Alpha Go trained on a million online moves. That’s how many we had of Master Games. And that only created a sort of fair go player. A good amateur could defeat it. So they worked around that in the case of Go by basically generating an infinite amount of data by having the system play itself. I had a chat with Davis House of Us. What kind of situations can you do that with? You have to have some way of simulating the world. So Go or chess, or even though Go is considered a difficult game, it’s the definition of it exists on one page. So you can simulate it. That applies to math. I mean, math axioms are going to be contained on a page or two. It’s not very complicated. It gets more difficult when you have real life situations, like biology. So we have biological simulators, but the simulators are not perfect. So learning from the simulators will only be as good as the simulators. That’s actually the key to being able to do deep learning on biology. Autonomous vehicles, you need real life data. So the Waymo systems have gone 3.5 million miles. That’s enough data to then create a very good simulator. So the simulators are really quite realistic because they had a lot of real world experience. And they’ve gone a billion miles in the simulator. But we don’t always have that opportunity to either create the data or have the data around. Humans can learn from a small number of examples. Your significant other, your professor, your boss, your investor, can tell you something once or twice. And you might actually learn from that. Some humans have been reported to do that. And that’s kind of the remaining advantage of humans. Now, there’s actually no back propagation in the human brain. It doesn’t use deep learning. It uses a different architecture. That same year in 1962 I wrote a paper how I thought the human brain worked. There was actually very little neuroscience to go on. There was one neuroscience test, Vernon Mountcastle, that had something relevant to say, which is he did. I mean, there was the common wisdom at the time. And there’s still a lot of neurosciences to say this. So we have all these different regions of the brain. They do different things. They must be different. There’s V1 in the back of the head where the optic nerve spills into that can tell that that’s a curved line. That’s a straight line. There’s these simple feature extractions on visual images. It’s actually a large part of the neocortex. There’s a fused-shaped gyrus up here, which can recognize faces. We know that because if it gets knocked out through injury or stroke, people can’t recognize faces. They will learn it again with a different region of the neocortex. It’s a famous frontal cortex, which does language and poetry and music. So these must work on different principles. He did autopsies on the neocortex in all these different regions and found they all looked the same. They had the same repeating pattern, same interconnections. He said neocortex is neocortex. I had that hint. Otherwise, I could actually observe human brains in action, which I did from time to time. There’s a lot of hints that you can get that way. For example, if I ask you to recite the alphabet, you actually don’t do it from A to Z. You do it as a sequence of sequences, ABCD, EFG, H-I-JK. We learn things that forward sequences of sequences. Forward, because if I ask you to recite the alphabet backwards, you can’t do it unless you learn that as a new sequence. So these are all interesting hints. I wrote a paper that the neocortex is organized as a hierarchy of modules in each module can learn a simple pattern. And that’s how I got to meet President Johnson and that initiated a half century of thinking about this issue. I came to MIT to study with Marvin Inzky. Actually, I came for two reasons. One, that Inzky became my mentor, which was a mentorship that lasted for over 50 years. The fact that MIT was so advanced that I actually had a computer, which the other colleges I considered didn’t have, it was an IBM 1794 or 32K of 36-bit words. So it’s 150K of core storage, two microseconds cycle time, two cycles for instructions, so a quarter of a mip. And thousands of students and professors share that one machine. In 2012, I wrote a book about this thesis. It’s now actually an explosion of neuroscience evidence to support it. The European Brain Reverse Engineering Project has identified a repeating module of about 100 neurons. It’s repeated 300 million times. It’s about 30 billion neurons in the neocortex. The neocortex is the outer layer of the brain, that’s part where we do our thinking. And they can see in each module axons coming in from another module. And then the output ax, the single output axon of that module goes as the input to another module. So we can see it organized as a hierarchy. It’s not a physical hierarchy. The hierarchy comes from these connections. The neocortex is a very thin structure. It’s actually one module thick. There’s six layers of neurons, but it constitutes one module. And we can see that it learns a simple pattern. And various reasons I cite in the book, the pattern recognition model that’s using is basically a hidden Markov model. How many of you have worked with Markov models? Okay. That’s usually no hands go up when I ask that question. But Markov model is not, it is learned, but it’s not back propagation. It can learn local features. So it’s very good for speech recognition. And the speech recognition I work I did in the 80s used these Markov models. I became the standard approach because it can deal with local variations. And the fact that a vowel is stretched, you can learn that in a Markov model. It doesn’t learn long distance relationships. That’s handled by the hierarchy. And something we don’t fully understand yet is exactly how the neocortex creates that hierarchy. But we have figured out how it can connect this module to this module. Does it then grow? I mean, there’s no virtual communication or wireless communication. It’s actually a connection. So does it grow and act on from one place to another, which could be inches apart? Actually, all these connections are there from birth, like the streets and avenues of Manhattan. There’s vertical and horizontal connections. So if it decides, and how it makes that decision, it’s still not fully understood, that it wants to connect this module to this module. There’s already a vertical, horizontal, and a vertical connection that just activates them. We can actually see that now, and it can see that happening in real time on non-invasive brain scans. So there’s a tremendous amount of evidence that, in fact, the neocortex is a hierarchy of modules that each module learns a simple sequential pattern. And even though the patterns we perceive don’t seem like sequences, they may seem three-dimensional or even more complicated, they are, in fact, represented as sequences. But the complexity comes in with the hierarchy. So the neocortex emerged 200 million years ago with mammals. All mammals have a neocortex. It’s one of the distinguishing features of mammals. These first mammals were small. They were rodents, but they were capable of a new type of thinking. Other non-mallion animals had fixed behaviors, but those fixed behaviors were very well adapted for their ecological niche. But these new mammals could invent a new behavior. So creativity and innovation was one feature of the neocortex. So a mouse is escaping a predator. Its usual escape path is blocked. It will invent a new behavior to deal with it. Probably wouldn’t work, but if it did work, it would remember it, and it would have a new behavior. And that behavior could spread virally through the community. Another mouse watching this would say to itself, hmm, that was really clever going around that rock. I wouldn’t remember to do that. And it would have a new behavior. Didn’t help these early mammals that much, because, as I say, the non-mallion animals were very well adapted to their niches, and nothing much happened for 135 million years. But then 65 million years ago, something did happen. There was a sudden violent change to the environment. We now call it the Cretaceous Extinction Event. There’s been debate as to whether it was a meteor or an asteroid. I mean, a meteor or a volcanic eruption, the asteroid or meteor hypothesis is in the ascendancy. But if you dig down to an area of rock reflecting 65 million years ago, the geologists will explain that it shows a very violent sudden change to the environment. And we see it all around the globe. So it was a worldwide phenomenon. The reason we call it an extinction event is that’s when the dinosaurs went extinct. That’s when 75% of all the animal and plant species went extinct. And that’s when mammals overtook their ecological niche. So to anthropomorphize biological evolution, said to itself, hmm, the sneacortex is pretty good stuff and it began to grow it. So now mammals got bigger. Their brains got bigger. They had an even faster pasting a biological fraction of their body. The neocortex got bigger even faster than that, and developed these curvatures that are distinctive of a primate brain, basically to increase its surface area. But if you stretched it out, the human neocortex is still a flat structure. It’s about the size of a table napkin, just as thin. And it basically created primates, which became dominance in their ecological niche. Then something else happened two million years ago. Biological evolution decided to increase the neocortex further and increase the size of the enclosure and basically filled up the frontal cortex with our big skulls with more neocortex. And up until recently it was felt that, as I said, that this was, the frontal cortex was different because it does these qualitatively different things. But we now realize that it’s really just additional neocortex. So remember what we did with it. We were already doing a very good job of being primates. So we put it at the top of the neocortical hierarchy. And we increased the size of the hierarchy. It was maybe 20% more neocortex, but it doubled and tripled the number of levels. Because as you go up the hierarchy, it’s kind of like a pyramid, there’s fewer modules. And that was the enabling factor for us to invent language and art, music. Every human culture we’ve ever discovered has music. No primate culture really has music. It’s debate about that, but it’s really true. Invention, technology, technology required another evolutionary adaptation, which is this humble appendage here. No other animal has that. If you look at the chimpanzee, it looks like they have a similar hand, but the thumb is actually down here. It doesn’t work very well if you watch them trying to grab a stick. So we could imagine creative solutions. Yeah, I could take that branch and strip off the leaves and put a point on it, and we could actually carry out these ideas and create tools, and then use tools to create new tools, and start a whole nother evolutionary process of tool making. And that all came with the neocortex. So Larry Page read my book in 2012, and liked it, so I met with him in S&P for an investment. In a company, I started actually a couple of weeks earlier to develop those ideas commercially, because that’s how I went about things as a serial entrepreneur. And he said, well, we’ll invest, but let me give you a better idea. Why don’t you do it here at Google? We have a billion pictures of dogs and cats, and we got a lot of other data and lots of computers and lots of talent, all of which is true. And he said, well, I don’t know. I just started this company to develop this. It’s a developed, this is by your company. And how you going to value a company that hasn’t done anything, and just started a couple of weeks ago, and he said, we can value anything. So I took my first job five years ago, and I’ve been basically applying this model, this hierarchical model, to understanding language, which I think really is the holy grail of AI. I think touring was correct in designating basically text communication as what we now call a touring complete problem that requires. There’s no simple NLP tricks that you can apply to pass a valid touring test with emphasis on the word valid, Mitch K. Per and I had a six month debate on what the rules should be, because if you read Turing’s 1950 paper, he describes this in a few paragraphs, and doesn’t really describe how to go about it. But if it’s a valid touring test, meaning it’s really convincing you through interrogation and dialogue that it’s a human, that requires a full range of human intelligence. And I think that test has to test of time. We’re making very good progress on that. I mean, just last week you may have read that two systems passed a paragraph comprehension test. It’s really very impressive, when I came to Google, we were trying to pass these paragraph comprehension tests. We aced the first grade test. Second grade test, we kind of got average performance. And the third grade test had too much inference. Already you had to know some common sense knowledge, as it’s called, and make implications of things that were in different parts of the paragraph. And there’s too much inference, and it really didn’t, didn’t work. So this is now adult level. It just slightly surpassed average human performance. But we’ve seen that once something, an AI does something at average human levels, it doesn’t take long for it to surpassed average human levels. I think it’ll take longer in language than it did in simple games like Go, but it’s actually very impressive that it surpasses now average human performance. It used an LSTM, long short temporal memory. But if you look at the adult test, in order to answer these questions, it has to put together inferences and implications of several different things in the paragraph. With some common sense knowledge, it’s not explicitly stated. So that’s, I think, a pretty impressive milestone. So I’ve been developing, I’ve got a team of about 45 people, and we’ve been developing this hierarchical model. We don’t use Markov models, because we can use deep learning for each module, and so we create an embedding for each word, and we create an embedding for each sense. This, we have a, I can talk about it, because we have a published paper on it. It can take into consideration context. If you use Smart Reply on, if you use Gmail on your phone, and you’ll see it gives you three suggestions for responses. That’s called Smart Reply. They’re simple suggestions, but it has to actually understand, perhaps, a complicated email. And the quality of the suggestions is really quite good, quite on point. That’s for my team using this kind of hierarchical model. So instead of Markov models, it uses embeddings, because we can use back propagation, we might as well use it. But I think what’s missing from deep learning is this hierarchical aspect of understanding, because the world is hierarchical. That’s why evolution developed a hierarchical brain structure to understand the natural hierarchy in the world. And there’s several problems with big deep neural nets. One is the fact that you really do need a billion examples, and we don’t, sometimes we can generate them as in the case of go, or if we have a really good simulator, as in the case of autonomous vehicles, we might the case yet in biology very often, you don’t have a billion examples. We shouldn’t have billions of examples of language, but they’re not annotated. And how would you annotate it anyway with more language that we can’t understand in the first place? So there’s kind of a chicken and an egg problem. So I believe this hierarchical structure is needed. Another criticism of deep neural nets, they don’t explain themselves very well. It’s a big black box that gives you pretty remarkable answers, I mean, in the case of these games, Demis described it’s playing in both Go and Chess is almost an alien intelligence, because we do things that we’re shocking to human experts, like sacrificing a queen and a bishop at the same time, or in close succession, which shocked everybody, but then went on to win, or early in a Go game, putting a piece at the corner of the board, which is kind of crazy to most experts, because you really want to start controlling territory, and yet on reflection, that was the brilliant move that enabled it to win that game. But it doesn’t really explain how it does these things. So if you have a hierarchy, it’s much better to explaining it, because you could look at the content of the modules in the hierarchy, and they’ll explain what they’re doing. And just to end on the first application, the first application of applying this to health and medicine, this will get into high gear, and we’re going to really see us break out of the linear extension to longevity that we’ve experienced. I believe we’re only about a decade away from longevity escape velocity, we’re adding more time than is going by, not just the infant life expectancy, but to your remaining life expectancy. I think if someone is diligent, they can be there already. I think anybody else could be there, not just the one-time life at longevity escape velocity. Now, a word on what life expectancy means, it used to be assumed that not much would happen. So whatever your life expectancy is, with or without scientific progress, it really didn’t matter. Now it matters a lot. So life expectancy really means, how long would you live, likelihood if they were not continued scientific progress. But that’s a very inaccurate assumption. The scientific progress is extremely rapid. I mean, just as an AI in biotech, there are advances now every week. It’s quite stunning. Now you could have a computed life expectancy, let’s say, 30 years, 50 years, 70 years. From now, you could still be hit by the proverbial bus tomorrow. We’re working on that with self-driving vehicles. But we’ll get to a point. I think if you’re diligent, you could be there now in terms of basically advancing your own statistical life expectancy, at least to keep pace with the passage of time. I think it would be there for most of the population, at least if they’re diligent within about a decade. So if we can hang in there, we may get to see the remarkable century ahead. Thank you very much. For a question, please raise your hand, we’ll get your mic. Hi. So you mentioned both neural network models and symbolic models. And I was wondering how far have you been thinking about combining these two approaches, creating a symbiosis between neural models and symbolic ones? I don’t think we want to use symbolic models as they’ve been used. How many are familiar with the psych project? That was a very diligent effort in Texas to define all of common sense reasoning. And it kind of collapsed on itself and became impossible to debug. Because you fix one thing and it breaks three other things. That complexity ceiling has become typical of trying to define things through logical rules. Now it does seem that humans can understand logical rules. We have logical rules written down for things like law and game playing and so on. But you can actually define a connection system to have such a high reliability on a certain type of action that it looks like it’s a symbolic rule, even though it’s represented in a connectionist way. And connection systems can both capture the soft edges, because many things in life are not sharply defined. They can also generate exceptions. So you don’t want to sacrifice your queen in chess. Except certain situations that might be a good idea. So you can capture that kind of complexity. So we do want to be able to learn from accumulated human wisdom that looks like it’s symbolic, but I think we’ll do it with a connection system. But again, I think the connection systems should develop a sense of hierarchy and not just be one big massive neural net. So I understand how we want to use the neocortex to extract useful stuff and commercialize that. But I’m wondering how our middle brain and the organs that are below the neocortex will be useful for turning that into what you want to do. The cerebellum is an interesting case in point. It actually has more neurons than the neocortex. And it’s used to govern most of our behavior. Some things, if you write a signature that’s actually controlled by the cerebellum, so a simple sequence is stored in the cerebellum. But there’s not any reasoning to it. It’s basically a script. And most of our movement now has actually been migrated from the cerebellum to the neocortex. Cerebellum is still there. Some people’s entire cerebellum is destroyed through disease. They still functioned fairly normally. Their movement might be a little erratic as our movement. It’s largely controlled by the neocortex, but some of the subtlety is a kind of pre-programmed script, and so they’ll look a little clumsy, but they’re actually function-okay. A lot of other areas of the brain control autonomic functions like breathing. But I’m thinking really is controlled by the neocortex. The terms of mastering intelligence, I think, of the neocortex is the brain region we want to study. I’m curious what you think might happen after the singularity is reached in terms of this exponential growth of information. Do you think it will continue or will there be a whole paradigm shift? What do you predict? Well, in the singularity’s near, talk about the atomic limits based on molecular computing as we understand it. It can actually go well past 2045 and actually go to trillions of trillions of times greater computational capacity than we have today. So I don’t see that stopping anytime soon and we’ll go way beyond what we can imagine. It becomes an interesting discussion what the impact on human civilization will be. So to take it maybe slightly more mundane issue that comes up is, oh, it’s going to eliminate most jobs or all jobs. A point I make is it’s not the first time in human history you’ve done that. How many jobs are can 1900 exist today? That was the feeling of the lawdites which was an actual society that formed in 1800 after the automation of the textile industry in England. They looked at all these jobs going away and felt, oh, employment is going to be just limited to an elite. Indeed, those jobs did go away but new jobs were created. So if I were a pressure futures in 1900, I would say, well, 38% of you work on farms and 25% work on factories. That’s two thirds of the working force. When I predict by 2015, 115 years from now it’s going to be 2% on farms and 9% factories and everybody would go, oh my god, we’re going to be out of work. And I say, well, don’t worry, for all these jobs we eliminate through automation, we’re going to invent new jobs and people say, oh, really what new jobs? And I’d say, well, I don’t know, we haven’t invented them yet. That’s the political problem. We can see jobs very clearly going away fairly soon, like driving a car or truck. And the new jobs haven’t been invented. I mean, just look at the last five or six years. Many, a lot of the increase in employment has been through mobile app-related types of ways of making money that just weren’t contemplated even six years ago. If I really pression, I would say, well, you’re going to get jobs creating mobile apps and websites and doing data analytics and self-driving cars. Cars, what’s a car? And nobody would have any idea what I’m talking about. Now the new job, some people say, yeah, we created new jobs, but it’s not as many. Actually, we’ve gone from 24 million jobs in 1900 to 142 million jobs today, from 30% of the population to 45% of the population. The new jobs fail 11 times as much in constant dollars. And they’re more interesting. I mean, as I talk to people starting out their career now, they really want a career that gives them some life definition and purpose and gratification. We’re moving up Maslow’s hierarchy. 100 years ago, you were happy if you had a backbreak in job that put food on your family’s table. We couldn’t do these new jobs without enhancing our intelligence. So we’ve been doing that well for most of the last 100 years through education. We’ve expanded K through 12 and constant dollars to 10 fold. We’ve gone from 38,000 college students in 1870 to 15 million today. More recently, we have brain extenders and not yet connected directly in our brain, but they’re very close at hand. When I was here at MIT, I had to take my bicycle across campus to get to the computer and show an idea to get in the building. Now we carry them in our pockets and on our belts. They’re going to go inside our bodies and brains. I think that’s a really important distinction. So we’re basically going to be continued to enhance our capability through merging with AI. And that’s the ultimate answer to the kind of dystopian view we see in futures movies, where it’s the AI versus a brave band of humans for control of humanity. We don’t have one or two AI’s in the world today. We have several billion, three billion smartphones at last count. It’ll be six billion in just a couple of years, according to the projections. So we’re already deeply integrated with this. And I think that’s going to continue. It’s going to continue to do things which you can’t even imagine today. Just as we are doing today, things we couldn’t imagine even 20 years ago. You showed many graphs that goes through exponential growth, but I haven’t seen one that isn’t. So I would be very interested in hearing things. You haven’t seen one that what? That is not exponential. So tell me about regions that you’ve investigated that have not seen exponential growth and why do you think that’s the case? Well, price performance and capacity of information technology invariably follows a exponential. When it impacts human society, it can be linear. So for example, the growth of democracy has been linear, but still pretty steady. You can count the number of democracies on the fingers of one hand a century ago. Two centuries ago, you can count the number of democracies in the world on the fingers of one finger. Now there are dozens of them that this and it’s become a kind of a consensus that that’s how we should be governed. So the, and I attribute all this to the growth and information technology communication in particular for progression of social and cultural institutions. But information technology because it ultimately depends on a vanishingly small energy and material requirement grows exponentially and will for a long time. There’s recently a criticism that well, chess scores have, this actually are mockably straight linear progression. So humans, it’s like 2800 and it just sort past that in 1997 with the blue and it’s kept going and are mockably straight and saying, well, this is linear, not exponential, but the chess score is a logarithmic measurement. So it really is exponential progression. So philosophers like to think a lot about the meaning of things, especially in the 20th century, so for instance Martin Haidega gave a couple of speeches and lectures on the relationship of human society to technology. And he particularly distinguished between the mode of thinking which is calculating thinking and a mode of thinking which is reflective thinking or meditative thinking. And he posed this question, what is the meaning and purpose of technological development? And he couldn’t find an answer, he recommended to remain open to what he called this an openness to the mystery. I wonder whether you have any thoughts on this, is there a meaning of purpose to technological development and is there a way for humans to access that meaning? We started using technology to shore up weaknesses in our own capabilities. So physically, I mean, who here could build this building? So we’ve leveraged the power of our muscles with machines. And we’re in fact very bad at doing things that the simplest computers can do, like factor numbers or even just multiply to eight digit numbers. Computers can do that. Trivially, we can’t do it. So we originally started using computers to make up for that weakness. I think the essence of what I’ve been writing about is to master the unique strengths of humanity, creating, loving expressions and poetry and music and the kinds of things we associate with the better qualities of humanity with machines. That’s the two promise of AI. They were not there yet, but we’re making pretty stunning progress. Just in the last year, there’s so many milestones that are really significant, including in language. But I think of technology as an expression of humanity. It’s part of who we are, and the human species is already a biological, technological civilization. And it’s part of who we are. And AI is part of humans. So AI is human, and it’s part of the technological expression of humanity. And we use technology to extend our reach. I couldn’t reach that fruit at that higher branch a thousand years ago. So we invented a tool to extend our physical reach. We now extend our mental reach. We can access all of human knowledge with a few key strokes. And we’re going to make ourselves literally smarter by merging with AI. Hi. First of all, honor to hear you speak here. So I first read the singularity as near nine years ago or so. And it changed the way I thought entirely. But something I think it caused me to oversteeply discount was tail risk in geopolitics, in systems that span the entire globe. And my concern is that there are, there is obviously the possibility of tail risk, existential level events swamping all of these trends that are otherwise war proof, climate proof, you name it. So my question for you is what steps do you think we can take in designing engineered systems in designing social and economic institutions to kind of minimize our exposure to these tail risks and survive to make it to a beautiful, mind-filled future? Yeah. Well, the world was first introduced to a human-made existential risk. When I was in elementary school, we would have these civil defense drills get under our desk and put our hands behind our head to protect this from a thermonuclear war. And it worked. We made it through. But that was really the first introduction to an existential risk. And those weapons are still there, by the way, and they’re still on a hair trigger and they don’t get that much attention. There’s been a lot of discussion, much of which I’ve been in the forefront of initiating the existential risks of what sometimes referred to as GNR, G for genetics, which is biotechnology and phenanotechnology and Grego robotics, which is AI. And I’ve been accused of being an optimist. And I think you have to be an optimist to be an entrepreneur. If you knew all the problems, you were going to encounter, you’d never start any project. But I’ve written a lot about the downsides. I remain optimistic. There are specific paradigms and not foolproof that we can follow to keep these technologies safe. So for example, over 40 years ago, some visionaries recognized the revolutionary potential, both for promise and peril, of biotechnology. Neither of the promise and peril was feasible 40 years ago. But they had a conference at the Asillemar Conference Center in California to develop both professional ethics and strategies to keep biotechnology safe. And they’ve been known as the Asillemar guidelines. They’ve been refined through successive Asillemar conferences, much of that’s baked into law. And in my opinion, it’s worked quite well. For now, as I mentioned, getting profound benefit, it’s a trickle today. It’ll be a flood over the next decade. And the number of people who have been harmed, either through intentional or accidental abuse of biotechnology so far as zero, actually, take that back. There was one boy who died in gene therapy trials, about 12 years ago, and there was congressional hearings in the canceled all research for gene therapy for a number of years. You could do an interesting message thesis and demonstrate that 300,000 people died as a result of that delay, but you can’t name them. They can’t go on CNN, so we don’t know who they are, but it has to do with the balancing of risk. But in large measure, virtually no one has been hurt by biotechnology. That doesn’t mean you can cross an arpharalist. Okay, we took care of that one because the technology keeps getting more sophisticated. CRISPR’s great opportunity, there’s hundreds of trials of CRISPR technologies, to overcome disease, but it could be abused. You can easily describe scenarios, so we have to keep reinventing it. January, we had our first, a silomar conference on AI ethics. So I think this is a good paradigm. It’s not foolproof. I think the best way we can assure a democratic future that includes our ideas of liberty is to practice that in the world today because the future world of the singularity, which is a merger of biological and non-biological intelligence, is not going to come from Mars. I mean, it’s going to emerge from our society today. So if we practice these ideals today, it’s going to have a higher chance of us practicing them as we get more enhanced with technology. If that doesn’t sound like a foolproof solution, it isn’t, but I think that’s the best approach. Turns out, technological solutions, I mean, AI is the most daunting. You can imagine there are technical solutions to biotechnology and nanotechnology. This really knows subroutine you can put in your AI software that will assure that it remains safe, intelligence is inherently not controllable. There’s some AI that’s much smarter than you. That’s out for your destruction. The best way to deal with that is not to get in that situation in the first place. If you are in that situation, find some AI that will be on your side. But basically, it’s going to, I believe we have been headed through technology to a better reality. I think around the world, people really think things are getting worse. And I think that’s because our information about what’s wrong with the world is getting exponentially better. I say, oh, this is the most peaceful time in you and history. And people say, what are you crazy? Didn’t you hear about the event yesterday and last week? And, well, 100 years ago, there could be a battle that wiped out the next village and you wouldn’t even hear about it for months. We have all these graphs on education and literacy has gone from like 10% to 90% over a century in health, wealth, poverty has declined 95% in Asia over the last 12-5 years. It’s talking about the World Bank. All these trends are very smoothly getting better and everybody thinks things are getting worse. But you’re right, like on violence, that curve could be quite disrupted. There’s an existential event. As I say, I’m optimistic, but I think that is something we need to deal with. And a lot of it is not technological. It’s dealing with our social, cultural institutions. So you mentioned also exponential growth of software and ideas, I guess, related to software. So one of the reasons for which you said that all that information technology costs this exponential is because of fundamental properties of matter and energy. But in the case of ideas, why would it have to be exponential? Well, a lot of ideas produce exponential gains. They don’t increase performance linearly. They would actually study during the Obama administration by a scientific advisory board on assessing this question. How much gains on 23 classical engineering problems were gained through hardware improvements over the last decade in software improvements? And there was about a thousand to one improvement. It’s about doubling every year from hardware. There was an average of something like 26,000 to one through software improvements, algorithmic improvements. So we do see both. And apparently, if you come up with an advance, it doubles the performance. So it multiplies by 10. We see basically exponential growth from each innovation. So and we certainly see that in deep learning. The architectures are getting better while we also have more data and more computation and more memory to throw at these algorithms. Thank you very much. Let’s give a very big hand. Thank you for being here.

AI video(s) you might be interested in …