WEBVTT Kind: captions Language: en 00:00:00.967 --> 00:00:03.336 A very unusual time period. 00:00:03.336 --> 00:00:07.107 The celebrated science fiction author and chemistry professor Isaac Asimov, 00:00:07.307 --> 00:00:10.577 once cataloged a history of inventions and scientific discoveries 00:00:10.577 --> 00:00:12.212 throughout all of human history. 00:00:12.212 --> 00:00:15.315 While incomplete, his efforts still reveal something intriguing 00:00:15.315 --> 00:00:16.583 about our current era. 00:00:16.583 --> 00:00:19.786 Out of these 694 pages in Asimov's book, 00:00:19.786 --> 00:00:24.591 553 pages documented inventions and discoveries since 1500. 00:00:24.724 --> 00:00:27.427 Even though his book starts in 4 million BCE. 00:00:27.727 --> 00:00:31.464 In other words, throughout human history, most scientific innovation has come 00:00:31.464 --> 00:00:34.601 relatively recently, within only the last few hundred years. 00:00:34.968 --> 00:00:37.170 Other historical trends paint a similar picture. 00:00:37.504 --> 00:00:41.207 For example, here's a chart of world populations since 10,000 B.C. 00:00:42.275 --> 00:00:43.977 for nearly all of human history. 00:00:43.977 --> 00:00:47.247 Up until quite recently, there weren't very many people on Earth. 00:00:47.580 --> 00:00:51.317 It took until about 1800 for the population to reach 1 billion people. 00:00:51.451 --> 00:00:53.219 And just 200 years later, 00:00:53.219 --> 00:00:56.089 a blink of an eye compared to how long our species has been around. 00:00:56.222 --> 00:00:58.491 Earth reached 6 billion people. 00:00:58.491 --> 00:01:01.928 Economic historian Bradford DeLong attempted to piece together 00:01:01.928 --> 00:01:05.534 the total world economic production over the last million years. 00:01:05.534 --> 00:01:09.369 By its nature, his reconstruction of the historical data is speculative, 00:01:09.602 --> 00:01:12.705 but the rough story it tells is consistent with the aforementioned 00:01:12.705 --> 00:01:15.041 historical trends in population and technology 00:01:15.208 --> 00:01:17.343 in the millennia preceding the current era. 00:01:17.343 --> 00:01:21.414 Economic growth by which we mean growth in how much valuable stuff humanity 00:01:21.414 --> 00:01:22.782 as a whole can produce. 00:01:22.782 --> 00:01:24.150 Was extremely slow. 00:01:24.484 --> 00:01:26.519 Now growth is much faster. 00:01:26.519 --> 00:01:29.923 Bradford DeLong's data provides historians a quantitative account 00:01:29.923 --> 00:01:32.258 of what they already know from reading narratives written 00:01:32.258 --> 00:01:35.161 in the distant past. For nearly all of human history, 00:01:35.428 --> 00:01:38.164 people lived similarly to the way their grandparents lived, 00:01:38.431 --> 00:01:40.266 unlike what we expect today. 00:01:40.266 --> 00:01:43.603 Most people did not see major changes in living standards, 00:01:43.603 --> 00:01:46.573 technology and economic production over their lifetimes. 00:01:46.906 --> 00:01:49.976 To be sure, people were aware that empires rose and fell. 00:01:50.110 --> 00:01:53.379 Infectious disease ravaged communities and wars were fought. 00:01:53.613 --> 00:01:56.783 Individual humans saw profound change in their own lives 00:01:56.983 --> 00:01:59.152 through the births and deaths of those they loved. 00:01:59.152 --> 00:02:01.187 Cultural change and migration. 00:02:01.187 --> 00:02:04.057 But the idea of a qualitatively different mode of life 00:02:04.290 --> 00:02:08.094 with electricity, computers, and the prospect of thermonuclear war 00:02:08.394 --> 00:02:11.798 that's all come extremely recently on historical timescales 00:02:12.165 --> 00:02:13.900 as new technologies were developed. 00:02:13.900 --> 00:02:17.370 Quality of life shot up in various ways for ten year olds. 00:02:17.637 --> 00:02:20.840 Life expectancy was once under 60 all over the world. 00:02:21.141 --> 00:02:24.577 Now, in many nations, a ten year old can expect to live to the age of 80. 00:02:24.944 --> 00:02:27.639 With progress in automating food production, 00:02:27.639 --> 00:02:30.025 fewer people now are required to grow food. 00:02:30.025 --> 00:02:33.458 As a result, our time has been freed to pursue different activities. 00:02:33.458 --> 00:02:34.621 For example, going to school. 00:02:35.021 --> 00:02:37.056 And it's not just technology that changed. 00:02:37.056 --> 00:02:40.927 In the past, people took for granted some social institutions that had existed 00:02:40.927 --> 00:02:42.228 for thousands of years, 00:02:42.228 --> 00:02:46.466 such as the monarchy and chattel slavery. In the midst of the Industrial Revolution. 00:02:46.666 --> 00:02:48.935 These institutions began to vanish. 00:02:48.935 --> 00:02:53.673 Writing in 1763, the eminent British economist Adam Smith wrote 00:02:53.673 --> 00:02:56.843 that slavery, quote: "Takes place in all societies 00:02:56.843 --> 00:02:59.946 at the beginning and proceeds from that tyrannical disposition, 00:02:59.946 --> 00:03:02.682 which may almost be said to be natural to mankind." 00:03:03.183 --> 00:03:06.152 While Adam Smith personally found the practice repugnant, 00:03:06.152 --> 00:03:09.088 he nonetheless was pessimistic about the future of slavery. 00:03:09.389 --> 00:03:12.725 Predicting that, quote: "It is indeed almost impossible 00:03:12.725 --> 00:03:15.295 that it should ever be totally or generally abolished." 00:03:15.562 --> 00:03:18.498 And yet, mere decades after Adam Smith wrote those lines, 00:03:18.698 --> 00:03:21.301 Britain outlawed slavery and launched a campaign 00:03:21.301 --> 00:03:23.536 to end the practice in its colonies around the world. 00:03:23.703 --> 00:03:25.438 By the end of the 20th century, 00:03:25.438 --> 00:03:28.208 every nation in the world had formally abolished slavery. 00:03:28.608 --> 00:03:32.312 Another scholar from his era, Thomas Malthus, made a similar blunder 00:03:32.812 --> 00:03:35.415 at the end of the 18th century, Malthus was concerned 00:03:35.415 --> 00:03:37.417 by the population growth he saw in his time. 00:03:37.750 --> 00:03:41.821 He reasoned that historically, excess population growth had always outstripped 00:03:41.821 --> 00:03:44.490 the food supply, leading to famine and mass death. 00:03:44.891 --> 00:03:47.994 As a consequence, Malthus predicted that recent high population 00:03:47.994 --> 00:03:50.797 growth in England would inevitably result in a catastrophe. 00:03:51.164 --> 00:03:53.800 But what might have been true about all the centuries before 00:03:53.800 --> 00:03:56.095 the 18th century evidently came to an end 00:03:56.095 --> 00:03:58.728 shortly after Malthus' his pessimistic prediction. 00:03:58.977 --> 00:04:02.970 Historians now recognize that rather than famine becoming more frequent, 00:04:02.970 --> 00:04:05.044 food in Britain became more widely available 00:04:05.044 --> 00:04:08.848 in the 19th and 20th centuries, despite unprecedented population growth. 00:04:09.249 --> 00:04:13.119 What Adam Smith and Thomas Malthus failed to see was that they were living 00:04:13.119 --> 00:04:17.442 in the very beginning of a historically atypical time of rapid change. 00:04:17.442 --> 00:04:20.479 A period we now refer to as the Industrial Revolution. 00:04:20.693 --> 00:04:23.997 The effects of the Industrial Revolution have been dramatic, reshaping 00:04:23.997 --> 00:04:28.201 not only how we live, but also our ideas about what to expect in the future. 00:04:28.935 --> 00:04:31.738 The 21st century could be much weirder than we imagine. 00:04:32.505 --> 00:04:36.042 It's easy to fault historical figures at the time of the Industrial Revolution 00:04:36.042 --> 00:04:38.678 for failing to see what was to come in the next few centuries. 00:04:39.012 --> 00:04:40.480 But their method of reasoning 00:04:40.480 --> 00:04:43.650 of looking at the past and extrapolating past trends outwards 00:04:43.850 --> 00:04:47.720 is something we still commonly do today to gauge our expectations of the future. 00:04:48.054 --> 00:04:51.658 In the last several decades, society has become accustomed to the global 00:04:51.658 --> 00:04:55.261 economy growing at a steady rate of about 2% to 4% per year. 00:04:55.561 --> 00:04:57.130 When people imagine the future, 00:04:57.130 --> 00:05:00.800 they often implicitly extrapolate this rate of change continuing indefinitely. 00:05:01.200 --> 00:05:03.703 We can call this perspective: "Business as usual." 00:05:03.936 --> 00:05:07.740 The idea that change in the coming decades will look more or less like change 00:05:07.740 --> 00:05:09.208 in the last few decades. 00:05:09.208 --> 00:05:11.911 Under business as usual, the world in 50 years 00:05:11.911 --> 00:05:14.614 looks a lot like our current world, but with some modifications. 00:05:14.881 --> 00:05:18.318 The world in 2072 would look just about as strange 00:05:18.318 --> 00:05:21.287 as someone from 1972 looking at our current world. 00:05:21.521 --> 00:05:24.791 Which is to say there will be more technology, different ways of 00:05:24.791 --> 00:05:28.594 communicating and socializing with others, distinct popular social movements 00:05:28.828 --> 00:05:32.598 and novel economic circumstances, but nothing too out of the ordinary 00:05:32.732 --> 00:05:36.703 like the prospect of mind uploading or building Dyson spheres around the sun. 00:05:37.203 --> 00:05:40.473 Parents sometimes take this perspective when imagining what life will one day 00:05:40.473 --> 00:05:42.008 be like for their children. 00:05:42.008 --> 00:05:44.977 Policymakers often take this perspective when crafting policy 00:05:45.178 --> 00:05:48.648 so that their proposed rules will be robust to new developments in the future. 00:05:49.040 --> 00:05:51.217 And workers who save for their retirement often 00:05:51.217 --> 00:05:54.053 take this perspective when they prepare for the challenges of growing old. 00:05:54.387 --> 00:05:57.090 Contrast business as usual with another perspective, 00:05:57.290 --> 00:05:59.625 which we can call: "The radical change thesis." 00:05:59.892 --> 00:06:02.829 Perhaps like Adam Smith and Thomas Malthus in their time, 00:06:02.829 --> 00:06:05.631 we are failing to see something really big on the horizon. 00:06:05.832 --> 00:06:07.600 Under the radical change thesis, 00:06:07.600 --> 00:06:10.670 the world by the end of the 21st century will look so different 00:06:10.670 --> 00:06:13.573 as to make it almost unrecognizable to people living today. 00:06:13.806 --> 00:06:18.211 Technologies that might seem unimaginable to us now, like advanced nanotechnology, 00:06:18.378 --> 00:06:22.849 cures for aging, interstellar spaceflight, and fully realistic virtual reality. 00:06:22.949 --> 00:06:24.815 Could only be a few decades away. 00:06:24.815 --> 00:06:27.420 As opposed to many centuries or thousands of years away. 00:06:27.954 --> 00:06:30.957 It's sensible to be skeptical of the radical change thesis. 00:06:31.257 --> 00:06:34.394 Scientific research happens slowly, and there's even some evidence 00:06:34.394 --> 00:06:37.663 that the rate of technological progress has slowed down in recent decades. 00:06:37.964 --> 00:06:41.401 At the same time, the radical change thesis makes intuitive sense 00:06:41.401 --> 00:06:43.302 from a long view perspective. 00:06:43.302 --> 00:06:45.405 Recall the trend in economic growth. 00:06:45.405 --> 00:06:48.341 Here, we've zoomed in on the last hundred years of economic growth. 00:06:48.741 --> 00:06:51.444 This view justifies the business as usual perspective. 00:06:51.744 --> 00:06:55.314 In the last 100 years, economic production and technology 00:06:55.314 --> 00:06:58.618 underlying economic production has grown at a fairly constant rate. 00:06:58.951 --> 00:07:01.521 Given the regularity of economic growth in recent decades. 00:07:01.721 --> 00:07:04.791 It makes a lot of sense to expect that affairs will continue to change 00:07:04.791 --> 00:07:07.126 at the current rate for the foreseeable future. 00:07:07.126 --> 00:07:08.928 But now zoom out. 00:07:08.928 --> 00:07:12.064 What we see is not the regular and predictable trend we saw before, 00:07:12.064 --> 00:07:14.600 but something far more dynamic and uncertain. 00:07:14.600 --> 00:07:17.270 And rather than looking like change is about to slow down, 00:07:17.270 --> 00:07:20.072 it looks more likely that change will continue to speed up. 00:07:20.072 --> 00:07:23.576 In a classic 1993 paper, economist Michael Kremer 00:07:23.576 --> 00:07:26.612 proposed that the best fit to this long run economic data 00:07:26.612 --> 00:07:30.149 is not the familiar slow rate of exponential change that we're used to, 00:07:30.149 --> 00:07:32.945 but rather what's called hyperbolic growth. 00:07:32.945 --> 00:07:34.620 Over a long enough time horizon, 00:07:34.620 --> 00:07:37.857 hyperbolic growth is much more dramatic than exponential growth, 00:07:37.857 --> 00:07:40.326 making it consistent with the radical change thesis. 00:07:40.326 --> 00:07:43.463 But it can also be a bit counterintuitive for people to imagine. 00:07:43.463 --> 00:07:45.198 So here's an analogy. 00:07:45.198 --> 00:07:49.335 Imagine a pile of gold that spontaneously grows over the course of a day. 00:07:49.335 --> 00:07:51.818 This pile of gold represents the total size of the 00:07:51.818 --> 00:07:53.473 world economy, over time. 00:07:53.473 --> 00:07:55.765 Which we assume grows hyperbolically. 00:07:55.765 --> 00:07:57.944 At 12 a.m. the beginning of the day. 00:07:57.944 --> 00:08:00.046 There's only one piece of gold in the pile. 00:08:00.213 --> 00:08:05.384 After 12 hours, the gold pile doubles in size so that there are now two pieces. 00:08:05.384 --> 00:08:08.888 Then at 6 p.m., there are four gold pieces. 00:08:08.888 --> 00:08:12.592 3 hours later, at 9 p.m., there are eight gold pieces 00:08:12.592 --> 00:08:16.062 and only one and a half hours after that 10:30 p.m.. 00:08:16.062 --> 00:08:18.164 There are 16 gold pieces. 00:08:18.164 --> 00:08:20.766 From the perspective of someone watching at 10:30 p.m.. 00:08:20.900 --> 00:08:24.203 It would be tempting to think that growth over the next few hours will be similar 00:08:24.203 --> 00:08:25.805 to growth in the last few hours, 00:08:25.805 --> 00:08:28.908 and that by midnight the next day, an hour and a half later, 00:08:28.908 --> 00:08:32.512 the gold pile will double in size yet again to 32 pieces. 00:08:32.512 --> 00:08:34.280 But this intuition would be wrong. 00:08:34.280 --> 00:08:37.725 Notice that the time it takes for the gold pile to double in size 00:08:37.725 --> 00:08:40.629 cut in half each time, the size of the pile doubles 00:08:40.629 --> 00:08:42.121 As midnight draws nearer. 00:08:42.121 --> 00:08:45.128 The pile will continue to double in size again and again, 00:08:45.128 --> 00:08:47.460 more frequently each time than the last. 00:08:47.460 --> 00:08:50.229 In fact, exactly at midnight, the pile will reach 00:08:50.229 --> 00:08:53.432 what's called a singularity and grow to be infinitely large. 00:08:53.933 --> 00:08:57.406 We can perform a similar exercise with the real world economy. 00:08:57.406 --> 00:09:01.007 In 2020, David Roodman found that after fitting a hyperbolic function 00:09:01.007 --> 00:09:02.845 to historical economic data. 00:09:02.845 --> 00:09:06.012 The economy is expected to grow to be infinitely large 00:09:06.012 --> 00:09:08.748 as we approach roughly the year 2047. 00:09:08.915 --> 00:09:11.250 A general result, he says, is fairly robust 00:09:11.250 --> 00:09:13.953 to extrapolating the trend at different periods in history. 00:09:14.453 --> 00:09:16.322 Now it's important not to take this headline 00:09:16.322 --> 00:09:20.026 result too seriously, especially the specific year 2047. 00:09:20.359 --> 00:09:23.029 Physical resource limits prohibit the economy from becoming 00:09:23.029 --> 00:09:27.967 infinitely large, and moreover, long run historical data is notoriously unreliable. 00:09:28.267 --> 00:09:31.971 Using historical statistics alone, especially those from the distant past, 00:09:32.038 --> 00:09:34.774 it's very difficult to predict what the future will really be like. 00:09:35.308 --> 00:09:37.843 The business as usual perspective might still be correct 00:09:37.843 --> 00:09:41.314 for the foreseeable future, or something else entirely might happen. 00:09:41.314 --> 00:09:43.411 Maybe civilization itself will collapse 00:09:43.411 --> 00:09:46.953 as we become consigned to fighting endless wars, famines and pandemics. 00:09:46.953 --> 00:09:48.955 Humanity might even go extinct. 00:09:48.955 --> 00:09:51.657 These possibilities notwithstanding, there do exist 00:09:51.657 --> 00:09:54.794 concrete reasons to think that the world could be fundamentally different 00:09:54.794 --> 00:09:58.864 by the end of the 21st century, in a way consistent with the radical change thesis. 00:09:59.231 --> 00:10:01.200 Infinite growth is out of the question. 00:10:01.200 --> 00:10:04.203 But one technology is clearly visible on the horizon, 00:10:04.203 --> 00:10:07.907 which arguably has the potential to change our civilization dramatically. 00:10:08.641 --> 00:10:10.209 "The Duplicator" 00:10:10.209 --> 00:10:14.013 before we try to predict what technology in the coming decades could precipitate 00:10:14.013 --> 00:10:15.781 explosive economic growth. 00:10:15.781 --> 00:10:18.851 Let's first start by examining a hypothetical example. 00:10:18.851 --> 00:10:21.587 A technology that would be sufficient to have such an effect. 00:10:22.088 --> 00:10:25.858 The machine is The Duplicator from the comic Calvin and Hobbes. 00:10:25.858 --> 00:10:28.361 Here's how it works: When someone steps in 00:10:28.361 --> 00:10:30.634 two exact replicas step out. 00:10:30.634 --> 00:10:34.667 The replicas keep all of the memories, personalities and talents of the original. 00:10:34.667 --> 00:10:38.771 Basically, it splits you into two people, each of whom remember an identical past. 00:10:39.038 --> 00:10:40.973 To constrain our expectations a bit, 00:10:40.973 --> 00:10:43.576 let's imagine that using The Duplicator isn't free. 00:10:43.909 --> 00:10:45.878 We can only use it to replicate people. 00:10:45.878 --> 00:10:49.015 And it costs a lot of money to build and operate a Duplicator. 00:10:49.015 --> 00:10:51.951 Yet even with these constraints, it seems safe to assume 00:10:51.951 --> 00:10:55.021 that The Duplicator would have a very large impact on the world. 00:10:55.021 --> 00:10:58.524 Even without knowing exactly how it will be used, there will presumably 00:10:58.524 --> 00:11:01.927 be people who want to use the duplicated to clone themselves and other people. 00:11:02.328 --> 00:11:05.665 Unlike with ordinary population growth in which it takes 18 years 00:11:05.665 --> 00:11:08.768 and a considerable amount of effort to create one productive worker. 00:11:08.768 --> 00:11:12.371 The Duplicator would allow us to create productive humans almost instantly 00:11:12.371 --> 00:11:14.974 and without having to pay the costs of educating and raising them. 00:11:15.307 --> 00:11:17.410 Consider the case of scientific innovation. 00:11:17.410 --> 00:11:20.079 If we could clone our most productive scientists 00:11:20.079 --> 00:11:22.782 say Newton, Galileo or von Neumann 00:11:22.782 --> 00:11:25.551 Just imagine how much faster our civilization could innovate. 00:11:25.951 --> 00:11:29.121 And it's not just raw scientific innovation that could increase 00:11:29.121 --> 00:11:32.491 With a machine that can duplicate people, all industries could benefit. 00:11:32.491 --> 00:11:35.294 The most talented rocket engineers, for example, could be cloned 00:11:35.294 --> 00:11:38.130 and directly planted into places where they're needed the most 00:11:38.130 --> 00:11:42.034 and we could duplicate the most talented doctors, musicians and architects. 00:11:42.034 --> 00:11:44.170 As a consequence, The Duplicator would left 00:11:44.170 --> 00:11:46.839 a crucial bottleneck to civilization wide output. 00:11:47.273 --> 00:11:51.310 Unless we impose some strict controls on how often people could use the duplicator. 00:11:51.310 --> 00:11:54.213 It appears likely that there would be a population explosion. 00:11:54.213 --> 00:11:58.317 But, more than just a population explosion, it would be a productivity explosion. 00:11:58.317 --> 00:11:59.518 Since the new duplicates 00:11:59.518 --> 00:12:02.521 could be direct copies of the most productive people on earth, 00:12:02.521 --> 00:12:05.491 who in turn, would use their talents to invent even more productive 00:12:05.491 --> 00:12:08.868 technologies and potentially even better duplicators 00:12:08.868 --> 00:12:10.196 Would the Duplicator be enough 00:12:10.196 --> 00:12:13.463 to sustain the hyperbolic growth trend in gross world product, 00:12:13.463 --> 00:12:15.534 and enable us to approach the economic singularity 00:12:15.534 --> 00:12:16.836 we discussed earlier? 00:12:16.836 --> 00:12:17.652 Maybe. 00:12:17.652 --> 00:12:21.107 In academic models of economic growth, there are roughly three inputs 00:12:21.107 --> 00:12:24.730 to economic production that determine the overall size of the economy. 00:12:24.730 --> 00:12:29.582 These inputs are: population, capital and technology. 00:12:29.582 --> 00:12:31.851 Population is the most intuitive input. 00:12:31.851 --> 00:12:34.954 It just means how many workers there are in a society. 00:12:34.954 --> 00:12:38.157 Capital is another term for equipment and supplies 00:12:38.157 --> 00:12:42.328 like machines, roads and buildings that allow workers to produce stuff. 00:12:42.328 --> 00:12:45.297 Technology is what joins these two inputs together. 00:12:45.297 --> 00:12:48.567 It refers to the inherent efficiency of labor and includes 00:12:48.567 --> 00:12:52.338 the quality of tools and the knowledge of how to create stuff in the first place. 00:12:52.338 --> 00:12:54.740 A simple model of economic growth is the following: 00:12:54.740 --> 00:12:58.077 Over time, people have children and the population gets larger. 00:12:58.077 --> 00:13:00.579 With more people, the economy also grows. 00:13:00.579 --> 00:13:04.150 But the economy also grows faster than the population because people work 00:13:04.150 --> 00:13:07.219 to accumulate new capital and invent new technology, 00:13:07.219 --> 00:13:09.255 at the same time the population is growing. 00:13:09.255 --> 00:13:12.291 Most importantly, people come up with new ideas. 00:13:12.291 --> 00:13:14.193 These new ideas are shared with others 00:13:14.193 --> 00:13:16.629 and can be used by everyone to increase efficiency. 00:13:17.029 --> 00:13:19.999 The process of economic growth looks a lot like a feedback loop. 00:13:20.332 --> 00:13:21.801 Start with a set of people. 00:13:21.801 --> 00:13:25.237 These people innovate and produce capital, which makes them more productive, 00:13:25.237 --> 00:13:28.808 and they also produce resources which enables them to produce more people. 00:13:28.808 --> 00:13:30.843 The next generation can rinse and repeat 00:13:30.843 --> 00:13:33.312 with each generation more productive than the last. 00:13:33.312 --> 00:13:37.324 Not only producing more people, but also more stuff per capita. 00:13:37.324 --> 00:13:39.585 In the absence of increased economic efficiency. 00:13:39.585 --> 00:13:41.720 The population will grow exponentially, 00:13:41.720 --> 00:13:45.024 but when the total production per person also increases 00:13:45.024 --> 00:13:48.460 This process implies super exponential economic growth. 00:13:48.460 --> 00:13:52.097 Recall the analogy of the gold pile that grows in size from earlier. 00:13:52.097 --> 00:13:56.068 The underlying dynamic is that the gold pile doubles in size every interval, 00:13:56.068 --> 00:13:59.071 but the interval in which the pile doubles is not fixed. 00:13:59.071 --> 00:14:01.482 Metaphorically, the gold pile becomes more efficient, 00:14:01.482 --> 00:14:03.742 at doubling in size during each interval. 00:14:03.742 --> 00:14:07.046 Giving rise to a process much faster than exponential growth. 00:14:07.046 --> 00:14:08.973 Many standard models of economic growth 00:14:08.973 --> 00:14:11.016 carry the same implication about the world economy. 00:14:11.650 --> 00:14:14.882 However, we know our current economy isn't exploding in size 00:14:14.882 --> 00:14:16.856 in the way predicted by this simple model. 00:14:16.856 --> 00:14:18.190 So, why not? 00:14:18.190 --> 00:14:21.660 One reason could be that our population growth is slowing down. 00:14:21.660 --> 00:14:22.995 Since the 1960s, 00:14:22.995 --> 00:14:26.932 population growth worldwide actually peaked before entering a decline. 00:14:26.932 --> 00:14:30.936 Not coincidentally, world economic growth has fallen since that decade. 00:14:30.936 --> 00:14:34.773 The underlying reason behind the fall in population growth and economic growth, 00:14:34.773 --> 00:14:38.653 also called the "demographic transition," is still a matter of debate. 00:14:38.653 --> 00:14:41.480 But regardless of its causes, the effect of the demographic 00:14:41.480 --> 00:14:45.551 transition has been that most mainstream economic and population forecasters 00:14:45.551 --> 00:14:48.754 do not anticipate explosive growth in the 21st century. 00:14:48.754 --> 00:14:52.124 But hypothetically, if something like the Duplicator were invented, 00:14:52.124 --> 00:14:54.493 then the potential for explosive growth could return. 00:14:55.294 --> 00:14:57.363 The most important invention ever? 00:14:57.363 --> 00:15:00.532 It's unlikely that humans will soon build the Duplicator. 00:15:00.532 --> 00:15:02.635 What's more likely however, is the invention 00:15:02.635 --> 00:15:05.337 of artificial intelligence that can automate labor. 00:15:05.337 --> 00:15:09.141 As with the Duplicator, A.I. could be copied and used as a worker. 00:15:09.141 --> 00:15:11.877 Vastly increasing total economic productivity. 00:15:11.877 --> 00:15:13.479 In fact, the potential for A.I. 00:15:13.479 --> 00:15:15.648 is even more profound than the duplicator. 00:15:15.648 --> 00:15:18.717 That's because A.I. have a number of advantages over humans. 00:15:18.717 --> 00:15:21.487 That could enable them to be far more productive in principle. 00:15:21.487 --> 00:15:24.490 These advantages include being able to think faster, 00:15:24.490 --> 00:15:26.272 save their current memory state, 00:15:26.272 --> 00:15:28.627 copy and transfer themselves across the internet, 00:15:28.627 --> 00:15:31.297 make improvements to their own software, and much more. 00:15:31.697 --> 00:15:33.666 Under the assumption that the development of A.I. 00:15:33.666 --> 00:15:35.998 will have a similar effect on the long term future 00:15:35.998 --> 00:15:37.770 as the Duplicator hypothetically would. 00:15:38.103 --> 00:15:41.940 The question of whether the 21st century will have explosive growth 00:15:41.941 --> 00:15:45.077 becomes a question of predicting when A.I. will arrive. 00:15:45.077 --> 00:15:48.105 If advanced A.I. is indeed invented later this century, 00:15:48.105 --> 00:15:52.551 it might mean the 21st century is the most important century in history. 00:15:52.551 --> 00:15:54.753 When will advanced AI arrive? 00:15:54.753 --> 00:15:57.856 If you've paid any attention to the field of AI in recent years, 00:15:57.856 --> 00:16:00.859 you've probably noticed that we've made some startling developments. 00:16:00.859 --> 00:16:02.027 In the last ten years, 00:16:02.027 --> 00:16:04.196 we've seen the rise of artificial neural networks 00:16:04.196 --> 00:16:06.919 that can match human performance in image classification, 00:16:06.919 --> 00:16:08.968 generate photorealistic images, 00:16:08.968 --> 00:16:10.956 beat the top Go players in the world, 00:16:10.956 --> 00:16:12.671 drive cars autonomously, 00:16:12.671 --> 00:16:16.408 reach grand master level at the real time strategy game Starcraft 2, 00:16:16.408 --> 00:16:20.212 learn how to play Atari games from scratch using only the raw pixel data, 00:16:20.212 --> 00:16:22.948 and write rudimentary poetry and fiction stories. 00:16:22.948 --> 00:16:25.354 If the next decade is anything like the last, 00:16:25.354 --> 00:16:28.988 we're sure to see even more impressive developments in the near-term future. 00:16:28.988 --> 00:16:29.989 Looking further, 00:16:29.989 --> 00:16:32.889 the holy grail of the field of artificial intelligence 00:16:32.889 --> 00:16:36.996 is so-called artificial general intelligence or AGI. 00:16:36.996 --> 00:16:38.564 A system capable of performing 00:16:38.564 --> 00:16:42.301 not only narrow tasks like game playing and image classification, 00:16:42.301 --> 00:16:45.838 but the entire breadth of tasks that humans are capable of performing. 00:16:45.838 --> 00:16:47.072 This would include the ability 00:16:47.072 --> 00:16:50.109 to reason abstractly and prove new mathematical theorems, 00:16:50.109 --> 00:16:55.280 perform scientific research, write books, and invent new goods and services. 00:16:55.280 --> 00:16:57.351 In short, an AGI, if created, 00:16:57.351 --> 00:16:59.818 would be at least as useful as a human worker 00:16:59.818 --> 00:17:04.156 and would in theory have the ability to automate almost any type of human labor. 00:17:04.156 --> 00:17:07.526 We know that AGI is possible in principle because human brains 00:17:07.526 --> 00:17:10.729 are already a type of general intelligence created by evolution. 00:17:10.729 --> 00:17:14.333 Many cognitive scientists and neuroscientists regularly analogized 00:17:14.333 --> 00:17:15.501 the brain to a computer 00:17:15.501 --> 00:17:18.871 and believe that everything we do is a consequence of algorithms implicit 00:17:18.871 --> 00:17:22.341 in the structure of our biological neural networks within our brains. 00:17:22.708 --> 00:17:25.490 Unless the human brain performs literal magic, 00:17:25.490 --> 00:17:27.769 then we have good reason to believe that one day 00:17:27.769 --> 00:17:30.349 we should be able to replicate its abilities in a computer. 00:17:30.783 --> 00:17:33.752 However, knowing that AGI is possible is one thing. 00:17:34.086 --> 00:17:36.789 Knowing when it will be developed and deployed in the real world 00:17:37.022 --> 00:17:38.690 is another thing entirely. 00:17:38.690 --> 00:17:40.626 Since the beginning of their discipline, A.I. 00:17:40.626 --> 00:17:43.162 researchers have experienced notorious difficulty 00:17:43.162 --> 00:17:44.930 predicting the future of the field. 00:17:44.930 --> 00:17:49.168 In the early days of the 1960s, the field was plagued with over-optimism, 00:17:49.168 --> 00:17:52.171 with some prominent researchers declaring that human like A.I. 00:17:52.171 --> 00:17:53.972 was due within 20 years. 00:17:53.972 --> 00:17:56.275 After these predictions failed to come to fruition. 00:17:56.275 --> 00:17:58.285 Most expectations became more modest. 00:17:58.650 --> 00:18:02.548 in the 1980s, the field entered what is now known as an A.I. winter 00:18:02.548 --> 00:18:06.452 and many researchers became more focused on short term commercial applications 00:18:06.452 --> 00:18:10.556 rather than the long term goal of creating something that rivals the human brain. 00:18:10.556 --> 00:18:12.758 These days, there's a wide variety of opinions 00:18:12.758 --> 00:18:15.761 among researchers about when AGI will be developed. 00:18:15.761 --> 00:18:20.341 The most comprehensive survey to date was by Grace, et al from 2016. 00:18:20.341 --> 00:18:23.388 A.I. experts were asked when they expected, quote: 00:18:23.388 --> 00:18:25.504 "for any occupation, machines could be built 00:18:25.504 --> 00:18:28.907 to carry out the task better and more cheaply than human workers." 00:18:28.907 --> 00:18:31.610 The median response was the year 2061, 00:18:31.610 --> 00:18:34.513 with considerable variation in individual responses. 00:18:34.813 --> 00:18:37.716 These results show that most researchers take the possibility 00:18:37.716 --> 00:18:40.886 of AGI being developed by the end of the century quite seriously. 00:18:41.120 --> 00:18:44.456 However, it is worth noting that responses were very sensitive 00:18:44.456 --> 00:18:46.758 to the exact phrasing of the question being asked. 00:18:47.059 --> 00:18:50.629 A subset of the researchers were asked when they expect AI to automate 00:18:50.629 --> 00:18:56.168 all human labor, and the median guess for that question was the year 2136. 00:18:56.168 --> 00:18:57.436 It's clear that a significant 00:18:57.436 --> 00:19:01.273 minority of researchers are unconvinced that AGI will arrive any time soon. 00:19:02.007 --> 00:19:04.629 To see more about this survey of A.I. researchers, 00:19:04.629 --> 00:19:06.945 check out Rob Miles' video about it on his channel. 00:19:07.379 --> 00:19:09.630 ...or my channel ...really, I mean 00:19:09.630 --> 00:19:11.881 I'm Rob Miles, I do the voice for this channel, 00:19:11.881 --> 00:19:14.019 but I also have my own channel. Rob Miles AI 00:19:14.019 --> 00:19:15.154 ....anyway.... 00:19:15.154 --> 00:19:19.091 In general, forecasting when A.I. will be developed is extremely difficult. 00:19:19.091 --> 00:19:21.560 Ideally, we could create a measure that represents 00:19:21.560 --> 00:19:24.863 how much progress there has been in the field of artificial intelligence. 00:19:24.863 --> 00:19:28.500 Then we could plot that measure on a graph and extrapolate outwards 00:19:28.500 --> 00:19:31.097 until we're expected to reach a critical threshold 00:19:31.097 --> 00:19:33.305 identified with the development of AGI. 00:19:33.305 --> 00:19:34.907 The problem with this approach is that it's 00:19:34.907 --> 00:19:37.743 very hard to find a robust measure of progress in AI. 00:19:38.110 --> 00:19:41.180 Luckily, a recent report from researcher Ajeya Cotra 00:19:41.180 --> 00:19:42.863 takes us part of the way there 00:19:42.863 --> 00:19:45.651 instead of trying to extrapolate progress in AI directly, 00:19:45.918 --> 00:19:49.354 her report tries to estimate when something like AGI might be developed 00:19:49.354 --> 00:19:52.858 based on trends in how affordable it is to build increasingly large 00:19:52.858 --> 00:19:55.794 AI models trained using increasingly difficult tasks. 00:19:56.128 --> 00:19:57.696 We can ground estimates of the needed 00:19:57.696 --> 00:20:02.367 size of the models in what she calls biological anchors, estimates from biology 00:20:02.367 --> 00:20:05.737 that inform how much computation and effort more generally 00:20:05.737 --> 00:20:09.141 will be required to develop software that can do what human brains do. 00:20:09.575 --> 00:20:12.511 The result very roughly coincides with the median responses 00:20:12.511 --> 00:20:15.847 from the expert survey predicting that humans will likely develop 00:20:15.847 --> 00:20:20.152 AI that can cheaply automate nearly all human labor by the end of the century. 00:20:20.152 --> 00:20:22.588 In fact, more likely than not by 2060. 00:20:22.921 --> 00:20:25.591 To understand how her model comes to this conclusion, 00:20:25.591 --> 00:20:28.360 we need to first get a sense of how current progress in A.I. 00:20:28.360 --> 00:20:29.394 is made. 00:20:29.628 --> 00:20:32.350 Ajeya Cotra's report on AI timelines 00:20:32.731 --> 00:20:33.932 Since 2012, 00:20:33.932 --> 00:20:37.569 the field of AI has been revolutionized by developments in deep learning. 00:20:37.569 --> 00:20:40.739 Deep learning involves training large artificial neural networks 00:20:40.739 --> 00:20:43.475 to learn tasks using an enormous amount of data 00:20:43.475 --> 00:20:46.078 without the aid of hand-crafted rules and heuristics. 00:20:46.278 --> 00:20:49.548 And it's been successful at cracking a set of traditionally hard problems 00:20:49.548 --> 00:20:52.551 in the field, such as those involving vision and natural language. 00:20:52.884 --> 00:20:56.221 One way to visualize deep learning is to imagine a digital brain 00:20:56.521 --> 00:20:58.490 called an artificial neural network 00:20:58.490 --> 00:21:01.827 that tries a massive amount of trial and error at a given task. 00:21:01.827 --> 00:21:05.142 For example, predicting the next character in a sequence of text. 00:21:05.142 --> 00:21:10.302 The neural network starts out randomly initialized and so will generate gibberish at first. 00:21:10.302 --> 00:21:13.805 During training, however, the neural network will be given a set of problems 00:21:14.006 --> 00:21:17.075 and will be asked to provide a solution to each of those problems. 00:21:17.075 --> 00:21:19.211 If the neural network solutions are incorrect, 00:21:19.211 --> 00:21:21.380 its inner workings are slightly rewired 00:21:21.546 --> 00:21:24.149 with the intention of it providing a better answer next time. 00:21:24.483 --> 00:21:28.053 Over time, the neural network should become better at the assigned task. 00:21:28.353 --> 00:21:30.622 If it does, then we say that it's learning. 00:21:30.622 --> 00:21:33.692 In her report, Ajeya Cotra discusses a deep connection 00:21:33.692 --> 00:21:36.461 between progress in hardware and progress in AI. 00:21:36.862 --> 00:21:40.205 The idea is simple with greater access to computation, 00:21:40.205 --> 00:21:43.902 A.I. researchers can train larger and more general deep learning models. 00:21:43.902 --> 00:21:46.272 In fact, the rise of deep learning in the last decade 00:21:46.272 --> 00:21:49.041 has largely been attributed to the falling cost of computation, 00:21:49.274 --> 00:21:53.045 especially given progress in graphics processing units, which are used heavily 00:21:53.045 --> 00:21:54.313 in A.I. research. 00:21:54.313 --> 00:21:57.749 In a nutshell, Cotra tries to predict when it will become affordable 00:21:57.749 --> 00:22:01.653 for companies or governments to train extremely large, deep learning models, 00:22:01.653 --> 00:22:03.355 with the effect of automating labor 00:22:03.355 --> 00:22:06.458 across the wide variety of tasks that humans are capable of learning. 00:22:06.825 --> 00:22:09.394 In building this forecast, she makes some assumptions 00:22:09.394 --> 00:22:11.596 about our ability to scale deep learning models 00:22:11.596 --> 00:22:13.198 to reach human level performance 00:22:13.198 --> 00:22:16.702 and the continued growth in computing hardware over the coming decades. 00:22:16.702 --> 00:22:17.602 Over time, 00:22:17.602 --> 00:22:20.605 researchers have found increasingly effective, deep learning models 00:22:20.806 --> 00:22:21.907 that are capable of learning 00:22:21.907 --> 00:22:26.044 to perform a wider variety of tasks and more efficiently than before. 00:22:26.044 --> 00:22:29.581 It seems plausible that this process will continue until deep learning models 00:22:29.581 --> 00:22:32.284 are capable of automating any aspect of human cognition. 00:22:32.617 --> 00:22:35.253 Even if future AI is not created via deep learning. 00:22:35.420 --> 00:22:38.423 Cotra's model may still be useful because of how it puts 00:22:38.423 --> 00:22:41.193 a soft upper bound on when humans will create AGI. 00:22:41.526 --> 00:22:43.929 She readily admits that if some more clever 00:22:43.929 --> 00:22:47.699 and more efficient paradigm for designing AI than deep learning is discovered, 00:22:47.899 --> 00:22:51.269 then the dates predicted by this model may end up being too conservative. 00:22:51.703 --> 00:22:52.537 Cotra's model 00:22:52.537 --> 00:22:56.074 extrapolates the rate at which progress in computing hardware will continue 00:22:56.074 --> 00:22:59.911 and the economy will continue to steadily get larger, making it more feasible 00:22:59.911 --> 00:23:03.849 for people to train extremely large, deep learning models on very large datasets. 00:23:04.282 --> 00:23:07.619 By their nature, deep learning algorithms are extremely data 00:23:07.619 --> 00:23:09.154 and computation hungry. 00:23:09.154 --> 00:23:12.424 It can often take a lot of money and many gigabytes of training data 00:23:12.424 --> 00:23:15.260 for deep learning algorithms to learn tasks that most humans 00:23:15.260 --> 00:23:18.497 find relatively easy, like manipulating a Rubik's cube 00:23:18.497 --> 00:23:21.166 or spotting simple, logical errors in natural language. 00:23:21.500 --> 00:23:24.302 This is partly why it has only recently become practically 00:23:24.302 --> 00:23:27.472 possible to train neural networks to perform these tasks. 00:23:27.472 --> 00:23:30.976 Historically, progress in computing power was extraordinarily rapid, 00:23:31.343 --> 00:23:33.445 according to data from William Nordhaus. 00:23:33.445 --> 00:23:38.717 Between 1950 and 2010, the price of computation dropped by over 100 billion. 00:23:39.151 --> 00:23:43.188 Put another way, that means that if you needed to perform a calculation in 1950 00:23:43.188 --> 00:23:44.489 that would cost the same amount 00:23:44.489 --> 00:23:47.192 as the Apollo program, that would later send humans to the moon, 00:23:47.492 --> 00:23:50.996 then the equivalent calculation would cost less than $10 00:23:50.996 --> 00:23:55.100 using computers in 2010, when adjusted for inflation over the same time period. 00:23:55.534 --> 00:23:56.701 Since 2010, 00:23:56.701 --> 00:23:59.504 the rate of progress in computing hardware has slowed down 00:23:59.504 --> 00:24:03.575 from this spectacular pace, but still continues to improve relatively steadily. 00:24:03.875 --> 00:24:07.946 By projecting outwards current trends, we can forecast when various computing 00:24:07.946 --> 00:24:09.414 milestones will be reached. 00:24:09.414 --> 00:24:12.117 For example, we can predict when it will become affordable to train 00:24:12.117 --> 00:24:15.954 a deep learning model with 1 billion petaflops or 1 trillion petaflops. 00:24:16.388 --> 00:24:20.321 Of course, simply knowing how much computation will be available in the future, 00:24:20.321 --> 00:24:23.228 can't tell us how powerful deep learning models will be. 00:24:23.228 --> 00:24:25.330 For that, we need to understand the relationship 00:24:25.330 --> 00:24:28.433 between the amount of computation used to train a deep learning model 00:24:28.633 --> 00:24:30.101 and its performance. 00:24:30.101 --> 00:24:32.904 Here's where Ajeya Cotra's model becomes a little tricky. 00:24:32.904 --> 00:24:36.408 Roughly speaking, there are two factors that determine how much computation 00:24:36.408 --> 00:24:39.160 it takes to train a deep learning model on some task 00:24:39.160 --> 00:24:40.846 holding other factors fixed. 00:24:40.846 --> 00:24:42.481 The size of the deep learning model 00:24:42.481 --> 00:24:45.951 and how much trial and error the model needs to learn the task. 00:24:45.951 --> 00:24:49.788 Let's consider the first factor the size of the deep learning model. 00:24:49.788 --> 00:24:52.257 Larger models are sort of like larger brains. 00:24:52.257 --> 00:24:55.994 They're able to learn more stuff, execute more complex instructions, 00:24:55.994 --> 00:24:58.530 and pick up on more nuanced patterns in the world. 00:24:58.864 --> 00:25:01.624 In the animal kingdom, we often, but not always, 00:25:01.624 --> 00:25:05.470 consider animals with larger brains, such as dogs, to be more generally 00:25:05.470 --> 00:25:08.744 intelligent than animals with smaller brains, such as fish. 00:25:08.744 --> 00:25:11.176 For many complex tasks, like writing code, 00:25:11.176 --> 00:25:13.632 you need a fairly large minimum model size 00:25:13.632 --> 00:25:15.780 Perhaps almost the size of a mouse's brain 00:25:15.780 --> 00:25:18.383 to learn the task to any reasonable degree of performance. 00:25:18.583 --> 00:25:21.786 And it will still be significantly worse than human programmers. 00:25:21.786 --> 00:25:25.557 The task we're most interested in is the task of automating human labor, 00:25:25.690 --> 00:25:29.427 at least the key components of human labor most important for advancing science 00:25:29.427 --> 00:25:30.395 and technology. 00:25:30.395 --> 00:25:33.298 It seems plausible that we'd need to use substantially larger models 00:25:33.298 --> 00:25:36.868 than any we've trained so far, if we want to train on this task 00:25:36.868 --> 00:25:38.203 Cotra roughly anchors 00:25:38.203 --> 00:25:42.340 "the size of model that would be capable of learning the task of automating R&D" 00:25:42.340 --> 00:25:44.109 with the size of the human brain. 00:25:44.109 --> 00:25:46.077 Give or take a few orders of magnitude. 00:25:46.077 --> 00:25:49.381 In fact, this approach of using biological anchors to guide our A.I. 00:25:49.381 --> 00:25:50.982 forecasts has precedent. 00:25:50.982 --> 00:25:54.619 In 1997, the computer scientist Hans Moravec, 00:25:54.619 --> 00:25:56.477 tried to predict when computer hardware 00:25:56.477 --> 00:25:58.723 would become available that would rival the human brain. 00:25:59.024 --> 00:26:02.360 Later work by Ray Kurzweil mirrored Moravec's approach. 00:26:02.360 --> 00:26:04.362 After looking into these older estimates, 00:26:04.362 --> 00:26:07.120 Cotra felt that the loose brain anchor made sense 00:26:07.120 --> 00:26:10.201 and seemed broadly consistent with machine learning performance so far. 00:26:10.502 --> 00:26:14.606 She ultimately made the guess that current algorithms are about 1/10 as efficient 00:26:14.606 --> 00:26:17.876 as the human brain, with very wide uncertainty around this estimate. 00:26:18.343 --> 00:26:20.111 Now we come to the second factor. 00:26:20.111 --> 00:26:23.748 How much trial and error would a brain sized model need to learn 00:26:23.748 --> 00:26:26.651 tasks required to automate science and technology R&D? 00:26:27.085 --> 00:26:30.922 This is the hardest and most uncertain open question in this entire analysis. 00:26:31.323 --> 00:26:34.426 How much trial and error is needed to train this model is a question 00:26:34.426 --> 00:26:38.263 of how efficient deep learning models will be at learning tasks of this form. 00:26:38.530 --> 00:26:41.866 And we only have limited and very indirect evidence about that question. 00:26:42.334 --> 00:26:44.469 The possibilities span a wide range. 00:26:44.469 --> 00:26:48.306 For example, perhaps deep learning models will be as efficient as human children, 00:26:48.306 --> 00:26:51.142 taking only the equivalent of 20 years of experience, 00:26:51.409 --> 00:26:55.413 sped up greatly within a computer, to learn how to be a scientist or engineer. 00:26:55.413 --> 00:26:57.816 Cotra thinks this is unlikely, but not impossible. 00:26:58.183 --> 00:27:00.452 On the other extreme, perhaps deep learning models 00:27:00.452 --> 00:27:03.888 will be as inefficient as evolution, taking the equivalent of hundreds 00:27:03.888 --> 00:27:08.259 of trillions of lifetimes of experience to evolve into a scientist or engineer. 00:27:08.493 --> 00:27:11.096 Again, Cotra thinks this is unlikely but possible. 00:27:11.463 --> 00:27:14.132 Cotra thinks the answer is likely to lie somewhere in between. 00:27:14.466 --> 00:27:17.669 That training deep learning models to automate science and engineering 00:27:17.669 --> 00:27:21.272 will require much more computation than raising a human child to be a scientist, 00:27:21.439 --> 00:27:24.509 but much less computation than simulating natural selection 00:27:24.509 --> 00:27:26.411 until it involves scientists. 00:27:26.411 --> 00:27:27.379 In her report, 00:27:27.379 --> 00:27:31.082 Cotra considers several anchors that lie in between these two extremes 00:27:31.349 --> 00:27:34.919 and proceeds by placing a subjective probability distribution over all of them. 00:27:35.387 --> 00:27:39.257 She arrived at the prediction that there's a roughly 50% chance that it will become 00:27:39.257 --> 00:27:42.327 economically feasible to train the relevant type of advanced A.I. 00:27:42.327 --> 00:27:43.962 by 2052. 00:27:43.962 --> 00:27:46.665 The uncertainty in her calculation was quite large, however, 00:27:46.898 --> 00:27:49.467 reflecting the uncertainty in her assumptions. 00:27:49.467 --> 00:27:52.804 Here's a graph showing how the resulting probability distribution changes 00:27:52.804 --> 00:27:54.179 under different assumptions. 00:27:54.539 --> 00:27:56.641 What does this all ultimately mean? 00:27:56.641 --> 00:27:59.711 Now you might be thinking, let's say these predictions are right. 00:27:59.711 --> 00:28:02.814 AGI arrives later this century, and as a result, 00:28:02.814 --> 00:28:06.017 there's a productivity explosion changing human life as we know it. 00:28:06.284 --> 00:28:08.353 What does it ultimately mean for us? 00:28:08.353 --> 00:28:09.821 Should we be excited? 00:28:09.821 --> 00:28:11.222 Should we be scared? 00:28:11.222 --> 00:28:13.925 At the very least, our expectations should be calibrated 00:28:13.925 --> 00:28:15.660 to the magnitude of the event. 00:28:15.660 --> 00:28:19.230 If your main concern from AGI is that you might lose your job to a robot, 00:28:19.531 --> 00:28:22.834 then your vision of what the future will be like might be too parochial. 00:28:23.234 --> 00:28:27.272 If AGI arrives later this century, it could mark a fundamental transformation 00:28:27.272 --> 00:28:28.373 in our species, 00:28:28.373 --> 00:28:32.911 not merely a societal shift, a new fad, or the invention of a few new gadgets. 00:28:33.244 --> 00:28:37.382 As science fiction author Vernor Vinge put it in a famous 1993 essay, 00:28:37.716 --> 00:28:40.952 this change will be a throwing away of all the human rules. 00:28:41.152 --> 00:28:42.721 Perhaps in the blink of an eye, 00:28:42.721 --> 00:28:45.686 an exponential runaway beyond any hope of control. 00:28:45.959 --> 00:28:47.192 Developments that were thought 00:28:47.192 --> 00:28:50.196 might only happen in a million years, if ever, 00:28:50.196 --> 00:28:52.363 will likely happen in the next century. 00:28:52.363 --> 00:28:54.132 One way of thinking about this event 00:28:54.132 --> 00:28:56.915 is to imagine that humanity is on a cliff's edge... 00:28:56.915 --> 00:28:58.636 Ready to slide off into the abyss. 00:28:59.070 --> 00:28:59.738 In the abyss. 00:28:59.738 --> 00:29:02.340 We have only the vaguest sense as to what lies beyond. 00:29:02.640 --> 00:29:04.190 But we can give our best guesses. 00:29:04.476 --> 00:29:06.845 A good guess is that humans, or our descendants, 00:29:07.112 --> 00:29:09.414 will colonize the galaxy and beyond. 00:29:09.414 --> 00:29:12.884 That means, roughly speaking, the history of life looks a lot like this 00:29:13.184 --> 00:29:14.352 with our present time 00:29:14.352 --> 00:29:17.789 tightly packed between the time humans first came into existence 00:29:17.956 --> 00:29:20.759 and the time of the first galaxy wide civilization. 00:29:21.326 --> 00:29:25.730 Our future could be extraordinarily vast, filled with technological wonders. 00:29:25.997 --> 00:29:28.907 A giant number of people and strange beings 00:29:28.907 --> 00:29:32.437 and forms of social and political organization we can scarcely imagine. 00:29:32.804 --> 00:29:35.173 The future could be extremely bright... 00:29:35.173 --> 00:29:36.674 or it could be filled with horrors. 00:29:37.108 --> 00:29:38.910 One harrowing possibility 00:29:38.910 --> 00:29:41.780 is that the development of AGI will be much like the development 00:29:41.780 --> 00:29:43.194 of human life on earth, 00:29:43.194 --> 00:29:46.050 which wasn't necessarily so great for the other animals. 00:29:46.584 --> 00:29:48.887 The worry here is one of value misalignment. 00:29:49.120 --> 00:29:53.291 It is not guaranteed that even though humans will be the ones to develop A.I. 00:29:53.625 --> 00:29:55.860 That A.I. will be aligned with human values. 00:29:56.361 --> 00:29:59.130 In fact, my channel is about how advanced A.I. 00:29:59.130 --> 00:30:01.332 could turn out to be misaligned with human values, 00:30:01.633 --> 00:30:03.968 which could surely lead to bad outcomes. 00:30:03.968 --> 00:30:07.038 This prospect raises a variety of technical challenges, 00:30:07.038 --> 00:30:10.441 prompting the need for more research on how to make sure future A.I. 00:30:10.441 --> 00:30:12.477 is compatible with human values. 00:30:12.477 --> 00:30:14.612 In addition to understanding what's to come, 00:30:14.612 --> 00:30:17.949 we also have a responsibility to make sure things go right. 00:30:17.949 --> 00:30:21.419 When approaching what may be the most important event in human history. 00:30:21.419 --> 00:30:24.322 Perhaps the only appropriate emotion is vigilance. 00:30:24.722 --> 00:30:28.293 If we are on the edge of a radical transformation of our civilization. 00:30:28.293 --> 00:30:32.686 The actions we take today could have profound effects on the long-term future, 00:30:32.686 --> 00:30:35.700 reverberating through billions of years of future history. 00:30:36.301 --> 00:30:39.771 Even if we're not literally among the most important people who ever live, 00:30:39.771 --> 00:30:43.441 our actions may still reach far farther than we might have otherwise expected. 00:30:43.775 --> 00:30:47.863 The astronomer, Carl Sagan, once wrote that Earth is a pale blue dot 00:30:47.863 --> 00:30:49.914 suspended in the void of space 00:30:49.914 --> 00:30:52.483 whose inhabitants exaggerate their importance. 00:30:52.984 --> 00:30:54.118 He asked us to: 00:30:54.118 --> 00:30:58.024 "think of the rivers of blood spilled by all those generals and emperors, 00:30:58.024 --> 00:31:00.091 so that, in glory and triumph, 00:31:00.091 --> 00:31:04.135 they could become the momentary masters of a fraction of a dot." 00:31:04.351 --> 00:31:06.698 Carl Sagan's intention was to get us to understand 00:31:06.698 --> 00:31:10.468 the gravity of our decisions by virtue of our cosmic insignificance. 00:31:10.802 --> 00:31:13.771 However, in the light of the fact that the 21st century could be 00:31:13.771 --> 00:31:16.007 the most important century in human history, 00:31:16.007 --> 00:31:18.270 we can turn this sentiment on its head. 00:31:18.270 --> 00:31:22.247 If indeed humans will, sometime in he foreseeable future, reach the stars. 00:31:22.614 --> 00:31:25.358 Our choices now are cosmically significant, 00:31:25.358 --> 00:31:29.220 despite currently only being tiny inhabitants of a pale blue dot. 00:31:29.988 --> 00:31:33.854 Right now our civilization is tiny, but we could be the seed 00:31:33.854 --> 00:31:37.244 to something that will soon become almost incomprehensibly big 00:31:37.244 --> 00:31:39.130 and measured on astronomical scales. 00:31:39.597 --> 00:31:42.400 The vast majority of planets and stars could be dead, 00:31:42.734 --> 00:31:44.836 but we might fill them with value. 00:31:44.836 --> 00:31:48.231 Much of what we might have previously thought impossible this century 00:31:48.231 --> 00:31:49.474 may soon become possible. 00:31:49.874 --> 00:31:53.327 Truly, mind-bending, technological and economic change 00:31:53.327 --> 00:31:55.313 may all play a role in our lifetimes. 00:31:55.747 --> 00:31:58.082 Rather than pondering our unimportance. 00:31:58.416 --> 00:32:02.153 Perhaps it's wiser to try to understand our true place in history. 00:32:02.153 --> 00:32:03.051 And quickly. 00:32:03.354 --> 00:32:05.323 There might be only so much time left 00:32:05.523 --> 00:32:08.964 before we take the deep plunge into the abyss.