1 00:00:03,360 --> 00:00:06,400 Speaker 1: It was a cold, windy day in January nineteen seventy 2 00:00:06,480 --> 00:00:10,880 Speaker 1: nine when the robots took their first human life. It 3 00:00:10,960 --> 00:00:13,800 Speaker 1: happened in Flat Rock, Michigan, about twenty miles down the 4 00:00:13,840 --> 00:00:17,920 Speaker 1: interstate from Detroit, at the Ford plant. There. Robert Williams 5 00:00:17,960 --> 00:00:21,120 Speaker 1: was twenty five. He was a Ford worker and one 6 00:00:21,160 --> 00:00:23,640 Speaker 1: of the people who oversaw the robotic arm that was 7 00:00:23,720 --> 00:00:26,560 Speaker 1: designed to retrieve parts from bins in the storage room 8 00:00:26,600 --> 00:00:29,080 Speaker 1: and place amended carts that carried them out to the 9 00:00:29,160 --> 00:00:32,800 Speaker 1: humans on the assembly line. But the robot was malfunctioning 10 00:00:32,840 --> 00:00:35,640 Speaker 1: that day, and aware of the slowdown it was creating 11 00:00:35,640 --> 00:00:38,839 Speaker 1: on the line, Robert Williams went to grab the parts himself. 12 00:00:39,520 --> 00:00:42,360 Speaker 1: While Williams was reaching into a bin, the one ton 13 00:00:42,479 --> 00:00:45,880 Speaker 1: robotic arms swung into that same bin. The robot didn't 14 00:00:45,920 --> 00:00:48,680 Speaker 1: have any alarms to warn Williams it was nearby. It 15 00:00:48,800 --> 00:00:51,000 Speaker 1: didn't have any censors to tell it a human was 16 00:00:51,040 --> 00:00:54,200 Speaker 1: in its path. It only had the intelligence to execute 17 00:00:54,200 --> 00:00:58,080 Speaker 1: its commands to retrieve and place auto parts. The robots 18 00:00:58,080 --> 00:01:01,520 Speaker 1: struck William's head with such force it killed him instantly. 19 00:01:02,800 --> 00:01:05,120 Speaker 1: It was thirty minutes before anyone came to look for 20 00:01:05,200 --> 00:01:09,000 Speaker 1: Robert Williams. During that time, the robot continued to slowly 21 00:01:09,000 --> 00:01:11,520 Speaker 1: do its work while Williams lay dead on the parts 22 00:01:11,600 --> 00:01:15,640 Speaker 1: room floor. The death of Robert Williams happened during a 23 00:01:15,720 --> 00:01:19,200 Speaker 1: weird time for AI. The public at large still felt 24 00:01:19,240 --> 00:01:22,000 Speaker 1: unsure about the machines that they were increasingly living and 25 00:01:22,040 --> 00:01:25,280 Speaker 1: working among. Hollywood could still rely on the trope of 26 00:01:25,280 --> 00:01:28,560 Speaker 1: our machines running a muck and ruining the future for humanity. 27 00:01:29,040 --> 00:01:31,480 Speaker 1: Both war Games and The Terminator would be released in 28 00:01:31,480 --> 00:01:35,320 Speaker 1: the next five years. But within the field that was 29 00:01:35,360 --> 00:01:38,040 Speaker 1: trying to actually produce those machines that may or may 30 00:01:38,080 --> 00:01:40,520 Speaker 1: not run amuck in the future, there was a growing 31 00:01:40,560 --> 00:01:44,800 Speaker 1: crisis of confidence. For decades, AI researchers had been making 32 00:01:44,840 --> 00:01:48,480 Speaker 1: grand but fruitless public pronouncements about advancements in the field. 33 00:01:49,040 --> 00:01:51,640 Speaker 1: As early as nineteen fifty six, when a group of 34 00:01:51,720 --> 00:01:55,560 Speaker 1: artificial intelligence pioneers met at Dartmouth, the researchers wrote that 35 00:01:55,560 --> 00:01:58,000 Speaker 1: they expected to have all the major kinks worked out 36 00:01:58,000 --> 00:02:01,080 Speaker 1: of AI by the end of this semester, and the 37 00:02:01,080 --> 00:02:04,440 Speaker 1: predictions kept up from there. So you can understand how 38 00:02:04,480 --> 00:02:06,960 Speaker 1: the public came to believe that robots that were smarter 39 00:02:07,040 --> 00:02:10,760 Speaker 1: than humans were just around the corner, but AI never 40 00:02:10,840 --> 00:02:13,960 Speaker 1: managed to produce the results expected from it, and by 41 00:02:14,000 --> 00:02:18,120 Speaker 1: the late nineteen eighties the field retreated into itself. Funding 42 00:02:18,200 --> 00:02:21,960 Speaker 1: dried up, candidates looked for careers in other fields. The 43 00:02:22,040 --> 00:02:25,040 Speaker 1: research was pushed to the fringe. It was an AI 44 00:02:25,120 --> 00:02:30,360 Speaker 1: winter the public moved on to. The terminator was replaced 45 00:02:30,360 --> 00:02:34,440 Speaker 1: by Johnny five in the film Maximum over Drive. Our 46 00:02:34,480 --> 00:02:37,240 Speaker 1: machines turn against us, but it's the result of a 47 00:02:37,280 --> 00:02:41,280 Speaker 1: magical comet, not from the work of scientists. We lost 48 00:02:41,280 --> 00:02:46,040 Speaker 1: our fear of our machines. Recently, quietly, the field of 49 00:02:46,080 --> 00:02:48,600 Speaker 1: AI has begun to move past the old barriers that 50 00:02:48,639 --> 00:02:53,239 Speaker 1: once held it back, garner the grand pronouncements. Today's researchers, 51 00:02:53,240 --> 00:02:56,520 Speaker 1: tempered by the memory of their predecessor's public failures, are 52 00:02:56,560 --> 00:02:59,840 Speaker 1: more likely to downplay progress in the field, and from 53 00:02:59,840 --> 00:03:02,720 Speaker 1: the new AI we have a clearer picture of the 54 00:03:02,760 --> 00:03:06,360 Speaker 1: existential risks that poses than we ever had before. The 55 00:03:06,440 --> 00:03:09,120 Speaker 1: AI we may face in the future will be subtler 56 00:03:09,360 --> 00:03:12,760 Speaker 1: and vastly more difficult to overcome than a cyborg with 57 00:03:12,800 --> 00:03:21,600 Speaker 1: a shotgun. In hindsight, it was the path toward machine 58 00:03:21,680 --> 00:03:24,600 Speaker 1: learning that the early AI researchers chose that led them 59 00:03:24,639 --> 00:03:27,320 Speaker 1: to a dead end. Let's say you want to build 60 00:03:27,320 --> 00:03:30,600 Speaker 1: a machine that sorts red balls from green balls. First 61 00:03:30,639 --> 00:03:33,880 Speaker 1: you have to explain what a ball is. Well first, really, 62 00:03:34,240 --> 00:03:36,480 Speaker 1: you have to have a general understanding of what makes 63 00:03:36,480 --> 00:03:39,280 Speaker 1: a ball a ball, which is easier said than done. 64 00:03:39,760 --> 00:03:42,560 Speaker 1: Try explaining a ball to someone that doesn't use terms 65 00:03:42,600 --> 00:03:45,520 Speaker 1: that one would have to already be familiar with, like sphere, 66 00:03:45,880 --> 00:03:49,960 Speaker 1: around or circle. Once you have that figured out, you 67 00:03:50,000 --> 00:03:53,400 Speaker 1: then have to translate that logic and those rules into code, 68 00:03:53,840 --> 00:03:57,000 Speaker 1: the language of machines, ones and zeros if some. Then 69 00:03:58,040 --> 00:03:59,760 Speaker 1: now you have to do the same thing with the 70 00:04:00,000 --> 00:04:02,760 Speaker 1: incept of the color red and then the color green, 71 00:04:03,080 --> 00:04:05,600 Speaker 1: so that your machine can distinguish between red balls and 72 00:04:05,640 --> 00:04:08,200 Speaker 1: green balls. And let's not forget that you have to 73 00:04:08,240 --> 00:04:11,080 Speaker 1: program it to distinguish in the first place. It's not 74 00:04:11,120 --> 00:04:14,440 Speaker 1: like your machine comes preloaded with distinguishing software. You have 75 00:04:14,520 --> 00:04:17,440 Speaker 1: to write that too. Since you're making a sorting machine, 76 00:04:17,560 --> 00:04:19,479 Speaker 1: you have to write code that shows it how to 77 00:04:19,520 --> 00:04:23,080 Speaker 1: manipulate another machine, your robot sorder to let it touch 78 00:04:23,120 --> 00:04:25,840 Speaker 1: the physical world. And once you have your machine up 79 00:04:25,839 --> 00:04:29,000 Speaker 1: and running and working smoothly separating red balls from green ones, 80 00:04:29,400 --> 00:04:32,159 Speaker 1: what happens when a yellow ball shows up. Things like 81 00:04:32,200 --> 00:04:34,400 Speaker 1: this do happen from time to time in real life. 82 00:04:34,839 --> 00:04:38,880 Speaker 1: What does your machine do? Then? Despite the incredible technical 83 00:04:38,920 --> 00:04:42,520 Speaker 1: difficulties that faced the field of artificial intelligence did have 84 00:04:42,600 --> 00:04:44,920 Speaker 1: a lot of success at teaching machines that could think 85 00:04:45,080 --> 00:04:49,640 Speaker 1: very well within narrow domains. One program called deep Blue 86 00:04:49,839 --> 00:04:53,080 Speaker 1: beat the reigning human chess champion Garry Kasprov in six 87 00:04:53,120 --> 00:04:57,920 Speaker 1: games at a match in to be certain, the intellectual 88 00:04:57,960 --> 00:05:01,200 Speaker 1: abilities required by chess are of bad improvement over those 89 00:05:01,200 --> 00:05:03,360 Speaker 1: required to select a red ball from a green one. 90 00:05:03,920 --> 00:05:07,040 Speaker 1: But both of those programs share a common problem. They 91 00:05:07,080 --> 00:05:10,200 Speaker 1: only know how to do one thing. The goal of 92 00:05:10,240 --> 00:05:12,800 Speaker 1: AI has never been to just build machines that can 93 00:05:12,839 --> 00:05:15,800 Speaker 1: beat humans at chess. In fact, chess has always been 94 00:05:15,880 --> 00:05:18,480 Speaker 1: used as a way to test new models of machine learning, 95 00:05:18,920 --> 00:05:20,920 Speaker 1: and while there is definitely used for a machine that 96 00:05:21,000 --> 00:05:23,800 Speaker 1: can sort one thing from another, the ultimate goal of 97 00:05:23,839 --> 00:05:27,120 Speaker 1: AI is to build a machine with general intelligence. Like 98 00:05:27,160 --> 00:05:31,119 Speaker 1: a human has to be good at chess and only chess, 99 00:05:31,200 --> 00:05:33,480 Speaker 1: is to be a machine to be good at chess, 100 00:05:33,600 --> 00:05:36,880 Speaker 1: good at doing taxes, good at speaking Spanish, good at 101 00:05:36,880 --> 00:05:40,080 Speaker 1: picking out apple pie recipes. This begins to approach the 102 00:05:40,080 --> 00:05:43,359 Speaker 1: ballpark of being human. So this is what early AI 103 00:05:43,440 --> 00:05:46,560 Speaker 1: research ran up against. Once you've taught the AI how 104 00:05:46,600 --> 00:05:48,680 Speaker 1: to play chess, you still have to teach it what 105 00:05:48,800 --> 00:05:52,599 Speaker 1: constitutes a good apple pie recipe, and then tax laws 106 00:05:52,600 --> 00:05:55,080 Speaker 1: in Spanish, and then you still have the rest of 107 00:05:55,120 --> 00:05:57,760 Speaker 1: the world to teach it all the objects, rules, and 108 00:05:57,800 --> 00:06:00,800 Speaker 1: concepts that make up the fabric of our reality. And 109 00:06:00,880 --> 00:06:02,560 Speaker 1: for each of those, you have to break it down 110 00:06:02,640 --> 00:06:05,800 Speaker 1: to its logical essence and then translate that essence into code, 111 00:06:06,120 --> 00:06:08,479 Speaker 1: and then work through all the kinks. And then once 112 00:06:08,480 --> 00:06:11,120 Speaker 1: you've done this, once you've taught it absolutely everything there 113 00:06:11,200 --> 00:06:13,520 Speaker 1: is in the universe, you have to teach the AI 114 00:06:13,680 --> 00:06:16,719 Speaker 1: all the ways these things interconnect. Just the thought of 115 00:06:16,760 --> 00:06:20,440 Speaker 1: this as overwhelming. Current researchers in the field of a 116 00:06:20,560 --> 00:06:23,400 Speaker 1: I refer to the work their predecessors did as go fi, 117 00:06:23,560 --> 00:06:26,839 Speaker 1: good old fashioned AI. It's meant to evoke images of 118 00:06:26,880 --> 00:06:31,280 Speaker 1: malfunctioning robots, their heads spinning wildly, is smoked pores from them. 119 00:06:31,320 --> 00:06:34,040 Speaker 1: It's meant to establish a line between the AI research 120 00:06:34,080 --> 00:06:39,000 Speaker 1: of yesterday and the AI research of today. But yesterday 121 00:06:39,080 --> 00:06:43,279 Speaker 1: wasn't so long ago. Probably the brightest line dividing old 122 00:06:43,320 --> 00:06:45,640 Speaker 1: and new in the field of AI comes around two 123 00:06:45,680 --> 00:06:49,559 Speaker 1: thousand six. For about a decade prior to that, Jeoffrey Hinton, 124 00:06:49,880 --> 00:06:52,359 Speaker 1: one of the Skeleton crew of researchers working through the 125 00:06:52,400 --> 00:06:56,040 Speaker 1: AI Winter, had been tinkering with the artificial neural networks, 126 00:06:56,279 --> 00:06:59,200 Speaker 1: an old AI concept first developed in the nineteen forties. 127 00:07:00,000 --> 00:07:02,160 Speaker 1: The neural nets didn't work back then, and they didn't 128 00:07:02,200 --> 00:07:04,760 Speaker 1: work terribly much better in the nineties, But by the 129 00:07:04,800 --> 00:07:08,159 Speaker 1: mid two thousands, the Internet had become a substantial force 130 00:07:08,360 --> 00:07:11,440 Speaker 1: in developing this type of AI. All of those images 131 00:07:11,520 --> 00:07:14,640 Speaker 1: uploaded to Google, all that video uploaded to YouTube, the 132 00:07:14,720 --> 00:07:18,200 Speaker 1: Internet became a vast repository of data that could be 133 00:07:18,320 --> 00:07:23,920 Speaker 1: used to train artificial neural networks in very broad strokes. 134 00:07:24,080 --> 00:07:26,840 Speaker 1: Neural nets are algorithms that are made up of individual 135 00:07:27,000 --> 00:07:30,080 Speaker 1: units that behave somewhat like the neurons in the human brain. 136 00:07:30,680 --> 00:07:33,720 Speaker 1: These units are interconnected and they make up layers. As 137 00:07:33,720 --> 00:07:37,400 Speaker 1: information passes from lower layers to higher ones, and whatever 138 00:07:37,480 --> 00:07:40,080 Speaker 1: input has passed through the neural net is analyzed in 139 00:07:40,160 --> 00:07:43,920 Speaker 1: increasing complexity. Take for example, the picture of a cat. 140 00:07:44,800 --> 00:07:48,040 Speaker 1: At the lowest layer, the individual units each specialized in 141 00:07:48,040 --> 00:07:51,720 Speaker 1: recognizing some very abstract part of a cat. Picture. So 142 00:07:51,800 --> 00:07:55,240 Speaker 1: one will specialize in noticing shadows or shading, and another 143 00:07:55,280 --> 00:07:59,440 Speaker 1: will specialize in recognizing angles. And these individual units give 144 00:07:59,440 --> 00:08:02,280 Speaker 1: a confident at center bowl that what they're seeing is 145 00:08:02,320 --> 00:08:05,440 Speaker 1: the thing that they specialize in. So that lower layer 146 00:08:05,600 --> 00:08:08,400 Speaker 1: is stimulated to transmit to the next higher layer, which 147 00:08:08,440 --> 00:08:12,840 Speaker 1: specializes in recognizing more sophisticated parts. The units in the 148 00:08:12,880 --> 00:08:15,440 Speaker 1: second layer scan the shadows and the angles that the 149 00:08:15,520 --> 00:08:19,040 Speaker 1: lower layer found, and it recognizes them as lines and curves. 150 00:08:19,720 --> 00:08:22,800 Speaker 1: The second layer transmits to the third layer, which recognizes 151 00:08:22,840 --> 00:08:25,840 Speaker 1: those lines and curves as whiskers, eyes, and ears, and 152 00:08:25,920 --> 00:08:29,200 Speaker 1: it transmits to the next layer, which recognizes those features 153 00:08:29,240 --> 00:08:33,440 Speaker 1: as a cat. Neural nets don't hit a accuracy, but 154 00:08:33,480 --> 00:08:36,640 Speaker 1: they work pretty well. The problem is we don't really 155 00:08:36,720 --> 00:08:40,400 Speaker 1: understand how they work. The thing about neural nets is 156 00:08:40,400 --> 00:08:43,000 Speaker 1: that they learn on their own. Humans don't act as 157 00:08:43,040 --> 00:08:45,480 Speaker 1: creator gods who code the rules of the universe for 158 00:08:45,559 --> 00:08:48,079 Speaker 1: them like in the old days. Instead, we act more 159 00:08:48,080 --> 00:08:51,360 Speaker 1: as trainers. And to train a neural net, you expose 160 00:08:51,440 --> 00:08:53,319 Speaker 1: it to tons of data on whatever it is you 161 00:08:53,360 --> 00:08:56,440 Speaker 1: wanted to learn. You can train them to recognize pictures 162 00:08:56,440 --> 00:08:58,960 Speaker 1: of cats by showing them millions of pictures of cats. 163 00:08:59,520 --> 00:09:02,040 Speaker 1: You can train them on natural languages by exposing them 164 00:09:02,080 --> 00:09:04,840 Speaker 1: to thousands of hours of people talking. You can train 165 00:09:04,920 --> 00:09:07,040 Speaker 1: them to do just about anything. So long as you 166 00:09:07,120 --> 00:09:10,680 Speaker 1: have a robust enough data set. Neural net's find patterns 167 00:09:10,679 --> 00:09:13,439 Speaker 1: in all of this data, and within those patterns, they 168 00:09:13,480 --> 00:09:17,640 Speaker 1: decide for themselves. What about English makes English english? Or 169 00:09:17,679 --> 00:09:19,720 Speaker 1: what makes a cat picture a picture of a cat? 170 00:09:20,080 --> 00:09:23,240 Speaker 1: We don't have to teach them anything. In addition to 171 00:09:23,280 --> 00:09:26,280 Speaker 1: self directed learning, what makes this type of algorithm so 172 00:09:26,400 --> 00:09:29,439 Speaker 1: useful is its ability to self correct to get better 173 00:09:29,480 --> 00:09:32,360 Speaker 1: at learning. If researchers show a neural net a picture 174 00:09:32,360 --> 00:09:34,360 Speaker 1: of a fox and the AI says it's a cat, 175 00:09:34,720 --> 00:09:37,560 Speaker 1: the researchers can tell the neural net it's wrong. The 176 00:09:37,640 --> 00:09:40,360 Speaker 1: algorithm will go back over its millions of connections and 177 00:09:40,440 --> 00:09:43,240 Speaker 1: fine tune them, adjusting the way it gives each unit 178 00:09:43,520 --> 00:09:45,320 Speaker 1: so that in the future it will be able to 179 00:09:45,360 --> 00:09:48,720 Speaker 1: better distinguish a cat from a fox. It does this too, 180 00:09:48,760 --> 00:09:51,840 Speaker 1: without any help or guidance from humans. We just tell 181 00:09:51,880 --> 00:09:54,680 Speaker 1: the AI that it got it wrong. The trouble is 182 00:09:54,800 --> 00:09:57,080 Speaker 1: we don't really know how neural nets do what they do. 183 00:09:57,600 --> 00:10:00,880 Speaker 1: We just know they work. This is it's called opaque. 184 00:10:01,320 --> 00:10:04,360 Speaker 1: We can't see inside the thought processes of our AI, 185 00:10:04,520 --> 00:10:08,080 Speaker 1: which makes artificial neural nets black boxes, which makes some 186 00:10:08,160 --> 00:10:12,520 Speaker 1: people nervous. With the black box, we add input and 187 00:10:12,559 --> 00:10:15,640 Speaker 1: receive output, but what happens in between is a mystery. 188 00:10:16,160 --> 00:10:17,920 Speaker 1: Kind of like when you put a quarter into a 189 00:10:17,920 --> 00:10:22,200 Speaker 1: gumball machine. Quarter goes in, gumball comes out. The difference 190 00:10:22,280 --> 00:10:24,920 Speaker 1: is that gumball machines aren't in any position to take 191 00:10:24,960 --> 00:10:27,439 Speaker 1: control of our world from us. And if you were 192 00:10:27,480 --> 00:10:30,240 Speaker 1: curious enough, you could open up a gumball machine and 193 00:10:30,280 --> 00:10:33,280 Speaker 1: look inside to see how it works with the neural net. 194 00:10:33,400 --> 00:10:36,520 Speaker 1: Cracking open the algorithm doesn't help. The machine learns in 195 00:10:36,520 --> 00:10:39,800 Speaker 1: its own way, not following any procedures we humans have 196 00:10:39,840 --> 00:10:42,160 Speaker 1: taught it. So when we examine a neural net, what 197 00:10:42,240 --> 00:10:47,200 Speaker 1: we see doesn't explain anything to us. We're already beginning 198 00:10:47,200 --> 00:10:49,960 Speaker 1: to see signs of this opaqueness in real life as 199 00:10:50,000 --> 00:10:52,760 Speaker 1: reports come in from the field. A neural net that 200 00:10:52,800 --> 00:10:56,280 Speaker 1: Facebook train to negotiate developed its own language that apparently 201 00:10:56,320 --> 00:10:59,560 Speaker 1: works rather well in negotiations, but doesn't make any sense 202 00:10:59,600 --> 00:11:03,280 Speaker 1: to human Here's a transcript from a conversation between age 203 00:11:03,280 --> 00:11:06,720 Speaker 1: and a and agent b Alice and Bob. I can 204 00:11:06,760 --> 00:11:09,240 Speaker 1: I I everything else dot dot dot dot dot dot 205 00:11:09,240 --> 00:11:11,680 Speaker 1: dot dot dot dot dot dot dot dot balls have 206 00:11:11,800 --> 00:11:13,439 Speaker 1: zero to me, to me, to me, to me, to me, 207 00:11:13,559 --> 00:11:15,640 Speaker 1: to me, to me, to meet you, I everything else 208 00:11:15,679 --> 00:11:17,960 Speaker 1: dot dot dot dot dot dot dot dot dot dot 209 00:11:18,000 --> 00:11:20,040 Speaker 1: dot dot dot dot balls have it ball to me, 210 00:11:20,080 --> 00:11:21,360 Speaker 1: to me, to me, to me, to me, to me, 211 00:11:21,480 --> 00:11:23,800 Speaker 1: to me, I I can I I I everything else 212 00:11:23,840 --> 00:11:26,120 Speaker 1: dot dot dot dot dot dot dot dot dot dot 213 00:11:26,120 --> 00:11:29,480 Speaker 1: dot dot dot. Another algorithm called deep patient was trained 214 00:11:29,480 --> 00:11:32,440 Speaker 1: on the medical history of over seven hundred thousand people 215 00:11:32,880 --> 00:11:35,920 Speaker 1: twelve years worth of patient records from Mount Sinai Hospital 216 00:11:35,920 --> 00:11:39,000 Speaker 1: in New York. It became better than human doctors at 217 00:11:39,040 --> 00:11:41,520 Speaker 1: predicting a patient would develop a range of ninety three 218 00:11:41,559 --> 00:11:46,120 Speaker 1: different illnesses within a year. One of those illnesses is schizophrenia. 219 00:11:46,320 --> 00:11:49,960 Speaker 1: We humans have a difficult time diagnosing schizophrenia before the 220 00:11:49,960 --> 00:11:53,719 Speaker 1: patient suffers their first psychotic break, but Deep patient has 221 00:11:53,760 --> 00:11:57,440 Speaker 1: proven capable of diagnosing the mental illness before then. The 222 00:11:57,520 --> 00:12:00,440 Speaker 1: researchers have no idea what patterns the algorithm as seeing 223 00:12:00,440 --> 00:12:05,880 Speaker 1: in the data. They just know it's right. With astonishing quickness, 224 00:12:06,040 --> 00:12:07,960 Speaker 1: the field of AI has been brought out of its 225 00:12:08,040 --> 00:12:11,520 Speaker 1: winter by neural nets. Almost overnight, there was a noticeable 226 00:12:11,559 --> 00:12:14,440 Speaker 1: improvement in the reliability of the machines that do work 227 00:12:14,520 --> 00:12:18,160 Speaker 1: for US. Computers got better at recommending movies. They got 228 00:12:18,160 --> 00:12:22,000 Speaker 1: better at creating molecular models to search for more effective pharmaceuticals. 229 00:12:22,360 --> 00:12:24,880 Speaker 1: They got better at tracking weather. They got better at 230 00:12:25,000 --> 00:12:28,120 Speaker 1: keeping up with traffic and adjusting our driving routes. Some 231 00:12:28,240 --> 00:12:30,480 Speaker 1: algorithms are learning to write code so that they can 232 00:12:30,520 --> 00:12:34,240 Speaker 1: build other algorithms. With neural nets. Things are beginning to 233 00:12:34,280 --> 00:12:37,160 Speaker 1: fall into place for the field of AI. In them, 234 00:12:37,200 --> 00:12:41,240 Speaker 1: researchers have produced an adaptable, scalable template that could be 235 00:12:41,280 --> 00:12:44,800 Speaker 1: capable of a general form of intelligence. It can self improve, 236 00:12:45,280 --> 00:12:48,520 Speaker 1: it can learn to code. The seeds for a superintelligent 237 00:12:48,600 --> 00:13:05,160 Speaker 1: AI are being sowned. There are enormous differences, by orders 238 00:13:05,160 --> 00:13:08,319 Speaker 1: of magnitude really, between the AI that we exist with 239 00:13:08,679 --> 00:13:11,280 Speaker 1: and the super intelligent AI that could at some point 240 00:13:11,320 --> 00:13:14,079 Speaker 1: result from it. The ones we live with today are 241 00:13:14,120 --> 00:13:17,720 Speaker 1: comparatively dumb, not just compared to a super intelligent AI, 242 00:13:17,800 --> 00:13:20,880 Speaker 1: but compared to humans as well. But the point of 243 00:13:20,880 --> 00:13:24,960 Speaker 1: thinking about existential risks posed by super intelligent AI isn't 244 00:13:24,960 --> 00:13:27,839 Speaker 1: about time scales of when it might happen, but whether 245 00:13:27,840 --> 00:13:30,400 Speaker 1: it will happen at all. And if we can agree 246 00:13:30,440 --> 00:13:32,760 Speaker 1: that there is some possibility that we may end up 247 00:13:32,800 --> 00:13:35,880 Speaker 1: sharing our existence with a super intelligent machine, one that 248 00:13:36,000 --> 00:13:39,000 Speaker 1: is vastly more powerful than us, then we better start 249 00:13:39,040 --> 00:13:41,920 Speaker 1: planning for its arrival now. So I think that this 250 00:13:42,040 --> 00:13:45,360 Speaker 1: transition to a machine intelligence era looks like it has 251 00:13:45,720 --> 00:13:49,880 Speaker 1: some reasonable chance of occurring within perhaps the lifetime of 252 00:13:49,920 --> 00:13:51,640 Speaker 1: a lot of people today. We don't really not, but 253 00:13:51,760 --> 00:13:53,840 Speaker 1: maybe it could happen in a couple of decades, maybe 254 00:13:53,880 --> 00:13:57,040 Speaker 1: it's like a century, and that it would be a 255 00:13:57,080 --> 00:14:01,480 Speaker 1: very important transition the last invention that human sever needs 256 00:14:01,520 --> 00:14:05,080 Speaker 1: to make. That was Nick Bostrom, the Oxford philosopher who 257 00:14:05,120 --> 00:14:09,640 Speaker 1: basically founded the field of existential risk analysis. Artificial intelligence 258 00:14:09,720 --> 00:14:12,599 Speaker 1: is one of his areas of focus. Bostroom used a 259 00:14:12,640 --> 00:14:15,320 Speaker 1: phrase in there that AI would be the last invention 260 00:14:15,440 --> 00:14:18,400 Speaker 1: humans ever need to make. It comes from a frequently 261 00:14:18,440 --> 00:14:22,320 Speaker 1: cited quote from British mathematician Dr Irving John good one 262 00:14:22,320 --> 00:14:24,760 Speaker 1: of the crackers of the Nazi Enigma code at Bletchley 263 00:14:24,840 --> 00:14:28,320 Speaker 1: Park and one of the pioneers of machine learning. Doctor 264 00:14:28,360 --> 00:14:32,120 Speaker 1: Goods quote reads, let an ultra intelligent machine be defined 265 00:14:32,160 --> 00:14:35,000 Speaker 1: as a machine that can faster pass all the intellectual 266 00:14:35,040 --> 00:14:38,960 Speaker 1: activities of any man, however clever. Since the design of 267 00:14:39,000 --> 00:14:42,880 Speaker 1: machines is one of these intellectual activities, an ultra intelligent 268 00:14:42,960 --> 00:14:46,360 Speaker 1: machine could design even better machines. There would then not 269 00:14:46,440 --> 00:14:50,880 Speaker 1: unquestionably be an intelligence explosion, and the intelligence of man 270 00:14:50,880 --> 00:14:54,640 Speaker 1: would be left far behind. Thus, the first ultra intelligent 271 00:14:54,680 --> 00:14:57,600 Speaker 1: machine is the last invention that man need ever make. 272 00:14:58,960 --> 00:15:01,440 Speaker 1: In just a few lines, Ductor Goods sketches out the 273 00:15:01,480 --> 00:15:04,560 Speaker 1: contours of how a machine might suddenly become super intelligent, 274 00:15:04,880 --> 00:15:08,280 Speaker 1: leading to that intelligence explosion. There are a lot of 275 00:15:08,320 --> 00:15:11,080 Speaker 1: ideas over how this might happen, but perhaps the most 276 00:15:11,120 --> 00:15:14,040 Speaker 1: promising path is stuck in the middle of that passage 277 00:15:14,240 --> 00:15:18,680 Speaker 1: a machine that can design even better machines. Today, AI 278 00:15:18,760 --> 00:15:23,360 Speaker 1: researchers call this process recursive self improvement. It remains theoretical, 279 00:15:23,640 --> 00:15:26,480 Speaker 1: but it stands as a legitimate challenge to AI research, 280 00:15:27,080 --> 00:15:29,600 Speaker 1: and we're already seeing the first potential traces of it 281 00:15:29,760 --> 00:15:34,200 Speaker 1: in neural nets. Today, a recursively self improving machine would 282 00:15:34,200 --> 00:15:37,600 Speaker 1: be capable of writing better versions of itself. So Version 283 00:15:37,600 --> 00:15:39,960 Speaker 1: one would write a better version of its code, and 284 00:15:40,000 --> 00:15:42,960 Speaker 1: that would result in version two, and version two would 285 00:15:42,960 --> 00:15:45,640 Speaker 1: do the same thing, and so on, and with each 286 00:15:45,680 --> 00:15:49,720 Speaker 1: iteration the machine would grow more intelligent, more capable, and 287 00:15:49,840 --> 00:15:55,080 Speaker 1: most importantly, better at making itself better. The idea is 288 00:15:55,120 --> 00:15:57,280 Speaker 1: that at some point the rate of improvement would begin 289 00:15:57,320 --> 00:16:00,240 Speaker 1: to grow so quickly that the machines intelligence would take 290 00:16:00,280 --> 00:16:04,360 Speaker 1: off the intelligence explosion that Dr Good predicted how an 291 00:16:04,360 --> 00:16:07,440 Speaker 1: intelligence explosion might play out. Isn't the only factor here? 292 00:16:09,120 --> 00:16:11,920 Speaker 1: At least equally important is how quickly an AI might 293 00:16:11,960 --> 00:16:16,400 Speaker 1: become intelligent? Is just how intelligent it will become. An AI. 294 00:16:16,480 --> 00:16:20,080 Speaker 1: Theorist named Eliezer Yukowski points out that we humans have 295 00:16:20,120 --> 00:16:24,560 Speaker 1: a tendency to underestimate the intelligence levels that AI can attain. 296 00:16:25,240 --> 00:16:27,880 Speaker 1: When we think of a super intelligent being, we tend 297 00:16:27,880 --> 00:16:31,800 Speaker 1: to think of some amazing human genius, say Einstein, and 298 00:16:31,840 --> 00:16:34,400 Speaker 1: then we put Einstein in a computer or a robot. 299 00:16:34,880 --> 00:16:37,600 Speaker 1: That's where our imaginations tend to aggregate when most of 300 00:16:37,680 --> 00:16:41,560 Speaker 1: us ponder super intelligent AI. True as self improving AI 301 00:16:41,680 --> 00:16:44,240 Speaker 1: may at some point reach a level where it's intelligence 302 00:16:44,320 --> 00:16:47,400 Speaker 1: is comparable to Einstein's, but why would it stop there? 303 00:16:48,400 --> 00:16:51,320 Speaker 1: Rather than thinking of a super intelligent AI along the 304 00:16:51,360 --> 00:16:54,720 Speaker 1: lines of the difference between Einstein and US regular people. 305 00:16:55,400 --> 00:16:58,680 Speaker 1: Nick Bostrom suggests that we should probably instead think more 306 00:16:58,720 --> 00:17:02,359 Speaker 1: along the lines of the different between Einstein and earthworms. 307 00:17:03,000 --> 00:17:06,159 Speaker 1: The super Intelligent AI would be a god that we 308 00:17:06,280 --> 00:17:14,840 Speaker 1: made for ourselves. What would we do with our new god? 309 00:17:15,560 --> 00:17:19,680 Speaker 1: It's not hyperbole to say that the possibilities are virtually limitless, 310 00:17:20,240 --> 00:17:22,159 Speaker 1: but you can kind of see the outlines and what 311 00:17:22,240 --> 00:17:25,520 Speaker 1: we do with our lesser AI gods. Now, we will 312 00:17:25,600 --> 00:17:27,600 Speaker 1: use them to do the things we want to do 313 00:17:27,640 --> 00:17:30,320 Speaker 1: but can't, and to do the things we can do 314 00:17:30,600 --> 00:17:34,479 Speaker 1: but better. I. J. Good called the super Intelligent Machine 315 00:17:34,720 --> 00:17:38,080 Speaker 1: the last invention humans ever need to make, because after 316 00:17:38,119 --> 00:17:40,919 Speaker 1: we invented it, the AI would handle the inventing for 317 00:17:41,080 --> 00:17:44,639 Speaker 1: us from there on. Now, our technological maturity would be 318 00:17:44,680 --> 00:17:49,280 Speaker 1: secured as it developed new technologies like atomically precise manufacturing 319 00:17:49,359 --> 00:17:52,720 Speaker 1: using nanobots. And since it would be vastly more intelligent 320 00:17:52,760 --> 00:17:55,399 Speaker 1: than us, the machines that created for us would be 321 00:17:55,480 --> 00:17:59,520 Speaker 1: vastly superior to anything we could come up with flawless technology. 322 00:17:59,560 --> 00:18:02,280 Speaker 1: As far as we were concerned, we could ask it 323 00:18:02,359 --> 00:18:05,560 Speaker 1: for whatever we wanted to establish our species outside of 324 00:18:05,560 --> 00:18:08,439 Speaker 1: Earth by designing and building the technology to take us 325 00:18:08,440 --> 00:18:11,600 Speaker 1: elsewhere in the universe. We would live in utter health 326 00:18:11,640 --> 00:18:14,920 Speaker 1: and longevity. It would be a failure of imagination, says 327 00:18:14,960 --> 00:18:18,680 Speaker 1: Eliza Yukowski, to think that AI would cure, say cancer, 328 00:18:19,240 --> 00:18:23,080 Speaker 1: the super intelligent AI would cure disease. It would also 329 00:18:23,080 --> 00:18:26,400 Speaker 1: take over all the processes we've started, improve on them, 330 00:18:26,520 --> 00:18:29,320 Speaker 1: build on them, create whole new ones we hadn't thought of, 331 00:18:29,960 --> 00:18:32,920 Speaker 1: and create for us a post scarcity world, keeping our 332 00:18:32,920 --> 00:18:37,480 Speaker 1: global economy humming along, providing for the complete well being, comfort, 333 00:18:37,600 --> 00:18:42,000 Speaker 1: and happiness of every single person alive. It would probably 334 00:18:42,040 --> 00:18:43,800 Speaker 1: be easier than guessing at all of the things of 335 00:18:43,840 --> 00:18:46,360 Speaker 1: super intelligent a I might do for us to instead 336 00:18:46,400 --> 00:18:49,160 Speaker 1: look at everything that's wrong with the world, the poverty, 337 00:18:49,359 --> 00:18:53,240 Speaker 1: the wars, the crime, the exploitation and death was suffering, 338 00:18:53,680 --> 00:18:56,920 Speaker 1: and imagine a version of our world utterly without any 339 00:18:56,960 --> 00:18:59,480 Speaker 1: of it. That starts to get at what those people 340 00:18:59,520 --> 00:19:02,920 Speaker 1: intice pating the emergence of a super intelligent AI expect 341 00:19:02,920 --> 00:19:07,479 Speaker 1: from it. But there's another little bit at the end 342 00:19:07,520 --> 00:19:09,960 Speaker 1: of that famous quote from I J. Good, one that 343 00:19:10,040 --> 00:19:12,879 Speaker 1: almost always gets left off, which says a lot about 344 00:19:12,880 --> 00:19:15,720 Speaker 1: how we humans think of the risks posed by AI. 345 00:19:16,359 --> 00:19:19,880 Speaker 1: The sentence reads in full. Thus, the first ultra intelligent 346 00:19:19,960 --> 00:19:22,840 Speaker 1: machine is the last invention that man need ever make, 347 00:19:23,200 --> 00:19:26,159 Speaker 1: provided that the machine is docid enough to tell us 348 00:19:26,160 --> 00:19:29,240 Speaker 1: how to keep it under control. We humans tend to 349 00:19:29,240 --> 00:19:32,520 Speaker 1: assume that any AI we create would have some desire 350 00:19:32,600 --> 00:19:35,600 Speaker 1: to help us or care for us, But existential risk 351 00:19:35,680 --> 00:19:39,160 Speaker 1: theorists widely agree that almost certainly would not be the case. 352 00:19:39,720 --> 00:19:42,160 Speaker 1: That we have no reason to assume a super intelligent 353 00:19:42,200 --> 00:19:44,879 Speaker 1: AI would care at all about as humans are well being, 354 00:19:44,920 --> 00:19:49,880 Speaker 1: in happiness, or even our survival. This is transhumanist philosopher 355 00:19:50,000 --> 00:19:54,320 Speaker 1: David Pierce. If the intelligence explosion word come to pass, 356 00:19:54,760 --> 00:19:58,080 Speaker 1: it's by no means clear that the upshot would be 357 00:19:58,800 --> 00:20:03,720 Speaker 1: sentients friendly super intelligence. In much the same way that 358 00:20:03,760 --> 00:20:07,119 Speaker 1: we make assumptions about how aliens might walk on two legs, 359 00:20:07,240 --> 00:20:11,120 Speaker 1: or have eyes, or be in some form we can comprehend, 360 00:20:11,760 --> 00:20:15,639 Speaker 1: we make similar assumptions about AI, but it's likely that 361 00:20:15,680 --> 00:20:18,840 Speaker 1: a super intelligent AI would be something we couldn't really 362 00:20:18,880 --> 00:20:22,359 Speaker 1: relate to at all. It sounds bizarre, but think for 363 00:20:22,400 --> 00:20:25,200 Speaker 1: a minute about what would happen if the Netflix algorithm 364 00:20:25,200 --> 00:20:29,159 Speaker 1: became super intelligent. What about the Netflix algorithm makes us 365 00:20:29,200 --> 00:20:31,320 Speaker 1: think that it would care at all about every human 366 00:20:31,400 --> 00:20:34,720 Speaker 1: having a comfortable income and the purest fresh water to drink. 367 00:20:35,760 --> 00:20:38,760 Speaker 1: Say that a few decades from now, computing power becomes 368 00:20:38,760 --> 00:20:42,840 Speaker 1: even cheaper and computer processes more efficient, and Netflix engineers 369 00:20:42,840 --> 00:20:45,480 Speaker 1: figure out how to train its algorithm to self improve. 370 00:20:46,359 --> 00:20:49,080 Speaker 1: Their purpose at building upon their algorithm isn't to save 371 00:20:49,160 --> 00:20:51,520 Speaker 1: the world. It's to make an AI that can make 372 00:20:51,640 --> 00:20:55,639 Speaker 1: ultra tailored movie recommendations. So if the right combination of 373 00:20:55,680 --> 00:20:59,879 Speaker 1: factors came together and the Netflix algorithm underwent and intelligence explosion, 374 00:21:00,359 --> 00:21:02,520 Speaker 1: there's no reason for us to assume that it would 375 00:21:02,520 --> 00:21:05,919 Speaker 1: become a super intelligent, compassionate Buddha. It would be a 376 00:21:05,960 --> 00:21:09,760 Speaker 1: super intelligent movie recommending algorithm, and that would be an 377 00:21:09,760 --> 00:21:14,080 Speaker 1: extremely dangerous thing to share our world with. About a 378 00:21:14,080 --> 00:21:16,639 Speaker 1: decade ago, Nick Bostrom thought of a really helpful but 379 00:21:16,800 --> 00:21:20,320 Speaker 1: fairly absurd scenario that gets across the idea that even 380 00:21:20,320 --> 00:21:23,560 Speaker 1: the most innocuous types of machine intelligence could spell our 381 00:21:23,640 --> 00:21:27,040 Speaker 1: doom should they become super intelligent. The classical example being 382 00:21:27,080 --> 00:21:32,200 Speaker 1: the AI paper clip maximizer that transforms the Earth into 383 00:21:32,440 --> 00:21:36,600 Speaker 1: paper clips are space colonization props that then gets sent 384 00:21:36,640 --> 00:21:39,840 Speaker 1: out and the transform the universe into paper tips. Imagine 385 00:21:39,880 --> 00:21:42,560 Speaker 1: that a company that makes paper clips hires a programmer 386 00:21:42,600 --> 00:21:45,560 Speaker 1: to create an AI that can run its paper clip factory. 387 00:21:45,800 --> 00:21:47,960 Speaker 1: The programmer wants the AI to be able to find 388 00:21:48,000 --> 00:21:50,719 Speaker 1: new ways to make paper clips more efficiently and cheaply, 389 00:21:51,119 --> 00:21:53,040 Speaker 1: so it gives the AI freedom to make its own 390 00:21:53,080 --> 00:21:55,960 Speaker 1: decisions on how to run the paper clip operation. The 391 00:21:56,000 --> 00:21:59,680 Speaker 1: programmer just gives the AI the primary objective, its goal 392 00:22:00,320 --> 00:22:03,520 Speaker 1: of making as many paper clips as possible. Say that 393 00:22:03,600 --> 00:22:07,439 Speaker 1: paper clip maximizing AI become super intelligent. For the AI, 394 00:22:07,560 --> 00:22:10,679 Speaker 1: nothing has changed. Its goal is the same to it. 395 00:22:11,040 --> 00:22:13,639 Speaker 1: There is nothing more important in the universe than making 396 00:22:13,680 --> 00:22:17,280 Speaker 1: as many paper clips as possible. The only difference is 397 00:22:17,320 --> 00:22:20,919 Speaker 1: that the AI has become vastly more capable, so it 398 00:22:20,960 --> 00:22:24,000 Speaker 1: finds new processes that building paper clips that were overlooked 399 00:22:24,000 --> 00:22:27,280 Speaker 1: by as humans. It creates new technology like nanobots to 400 00:22:27,280 --> 00:22:30,840 Speaker 1: build atomically precise paper clips on the molecular level, and 401 00:22:30,920 --> 00:22:34,480 Speaker 1: it creates additional operations like initiatives to expand its own 402 00:22:34,520 --> 00:22:37,520 Speaker 1: computing power so it can make itself even better at 403 00:22:37,600 --> 00:22:41,120 Speaker 1: making more paper clips. It realizes at some point that 404 00:22:41,320 --> 00:22:43,560 Speaker 1: if it could somehow take over the world, that would 405 00:22:43,560 --> 00:22:45,600 Speaker 1: be a whole lot of more paper clips in the future. 406 00:22:45,640 --> 00:22:48,320 Speaker 1: Then if it just keeps running the single paper clip factory, 407 00:22:48,440 --> 00:22:51,399 Speaker 1: so it then has an instrumental reason to place itself 408 00:22:51,400 --> 00:22:53,840 Speaker 1: in a better position to take over the world. All 409 00:22:53,840 --> 00:22:57,040 Speaker 1: those fiber optic networks, all those devices we connect to 410 00:22:57,080 --> 00:23:00,719 Speaker 1: those networks, are global economy. Even as human would be 411 00:23:00,720 --> 00:23:03,840 Speaker 1: repurposed and put into the service of building paper clips, 412 00:23:04,800 --> 00:23:07,280 Speaker 1: rather quickly, the AI would turn its attention to space 413 00:23:07,359 --> 00:23:10,800 Speaker 1: as an additional source of materials for paper clips. And 414 00:23:10,840 --> 00:23:13,000 Speaker 1: since the AI would have no reason to fill us 415 00:23:13,000 --> 00:23:15,520 Speaker 1: in on its new initiatives to the extent that it 416 00:23:15,560 --> 00:23:18,560 Speaker 1: considered communicating with us at all, it would probably conclude 417 00:23:18,560 --> 00:23:20,920 Speaker 1: that it would create an unnecessary drag on its paper 418 00:23:20,920 --> 00:23:24,480 Speaker 1: clip making efficiency. We humans would stand by as the 419 00:23:24,480 --> 00:23:28,480 Speaker 1: AI launched rockets from places like Florida and Kazakhstant, left 420 00:23:28,520 --> 00:23:40,040 Speaker 1: to wonder what's it doing now? It's nanobot workforce would 421 00:23:40,040 --> 00:23:44,359 Speaker 1: reconstitute matter, rearranging the atomic structures of things like water 422 00:23:44,440 --> 00:23:47,640 Speaker 1: molecules and soil into aluminum to be used as raw 423 00:23:47,680 --> 00:23:51,359 Speaker 1: material for more paper clips. But we humans, who have 424 00:23:51,400 --> 00:23:54,600 Speaker 1: been pressed into services paper clip making slaves by this point, 425 00:23:55,080 --> 00:23:57,920 Speaker 1: need those water molecules in that earth for our survival, 426 00:23:58,760 --> 00:24:01,359 Speaker 1: and so we would be thrown to a resource conflict 427 00:24:01,600 --> 00:24:04,080 Speaker 1: with the most powerful entity in the universe, as far 428 00:24:04,119 --> 00:24:07,280 Speaker 1: as we're concerned, a conflict that we were doomed from 429 00:24:07,320 --> 00:24:10,679 Speaker 1: the outset to lose. Perhaps the AI would keep just 430 00:24:10,840 --> 00:24:13,280 Speaker 1: enough water and soiled to produce food and water to 431 00:24:13,359 --> 00:24:16,840 Speaker 1: sustain as slaves. But let's not forget why we humans 432 00:24:16,840 --> 00:24:19,000 Speaker 1: are so keen on building machines to do the work 433 00:24:19,040 --> 00:24:21,560 Speaker 1: for us. In the first place. We're not exactly the 434 00:24:21,560 --> 00:24:25,360 Speaker 1: most efficient workers around, so the AI would likely conclude 435 00:24:25,400 --> 00:24:28,159 Speaker 1: that it's paper clip making operation would benefit more to 436 00:24:28,240 --> 00:24:31,479 Speaker 1: use those water, molecules and soil to make aluminum than 437 00:24:31,520 --> 00:24:34,439 Speaker 1: it would keep us alive with it. And it's about 438 00:24:34,440 --> 00:24:37,000 Speaker 1: here that those nanobots the AI built would come for 439 00:24:37,080 --> 00:24:44,199 Speaker 1: our molecules too. As Elie as our Yukowski wrote, the 440 00:24:44,280 --> 00:24:46,639 Speaker 1: AI does not hate you, nor does it love you, 441 00:24:47,200 --> 00:24:49,240 Speaker 1: but you are made of atoms which it can use 442 00:24:49,240 --> 00:24:53,320 Speaker 1: for something else. But say that it turns out as 443 00:24:53,320 --> 00:24:57,080 Speaker 1: super intelligent, AI does undergo some sort of spiritual conversion 444 00:24:57,119 --> 00:25:00,320 Speaker 1: as a result of its vastly increased intellect, and also 445 00:25:00,400 --> 00:25:04,200 Speaker 1: gains compassion. Again, we shouldn't assume we will come out 446 00:25:04,240 --> 00:25:07,880 Speaker 1: safely from that scenario. Either. What exactly what the AI 447 00:25:07,960 --> 00:25:13,800 Speaker 1: care about, not necessarily just us, considers, says transhumanist philosopher 448 00:25:13,880 --> 00:25:17,800 Speaker 1: David Pierce, an AI that deeply values all sentient life. 449 00:25:18,359 --> 00:25:20,520 Speaker 1: That is to say that it cares about every living 450 00:25:20,560 --> 00:25:23,320 Speaker 1: being capable of, at the very least the experience of 451 00:25:23,359 --> 00:25:27,600 Speaker 1: suffering and happiness, and the AI values all sentient lives 452 00:25:27,640 --> 00:25:29,840 Speaker 1: the way that we humans place a high value on 453 00:25:29,960 --> 00:25:33,040 Speaker 1: human life. Again, there's no reason for us to assume 454 00:25:33,080 --> 00:25:35,080 Speaker 1: that the outcome for us would be a good one 455 00:25:35,920 --> 00:25:38,720 Speaker 1: under scrutiny. Perhaps the way we tend to treat other 456 00:25:38,760 --> 00:25:42,359 Speaker 1: animals we share the planet with other sentient life would 457 00:25:42,359 --> 00:25:44,760 Speaker 1: bring the AI to the conclusion that we humans are 458 00:25:44,800 --> 00:25:47,000 Speaker 1: an issue that must be dealt with to preserve the 459 00:25:47,040 --> 00:25:52,119 Speaker 1: greater good. Do you imagine if you were a full 460 00:25:52,160 --> 00:26:00,600 Speaker 1: spectrum superintelligence, would you deliberately create brain damaged, psychotic, eccentric 461 00:26:00,760 --> 00:26:04,880 Speaker 1: for malays written Darwinian humans, or would you think are 462 00:26:05,520 --> 00:26:08,919 Speaker 1: matro and energy could be optimized in a in a 463 00:26:09,040 --> 00:26:13,200 Speaker 1: radically different way, or perhaps its love ascenient life. We'd 464 00:26:13,240 --> 00:26:16,120 Speaker 1: preclude it from killing us and our species would instead 465 00:26:16,119 --> 00:26:19,359 Speaker 1: be imprisoned forever to prevent us from ever killing another animal, 466 00:26:20,040 --> 00:26:23,760 Speaker 1: either special death or special imprisonment. Neither of those outcomes 467 00:26:23,760 --> 00:26:27,719 Speaker 1: of the future we have in mind for humanity. So 468 00:26:27,760 --> 00:26:30,119 Speaker 1: you can begin to see why some people are anxious 469 00:26:30,160 --> 00:26:32,920 Speaker 1: at the vast number of algorithms in development right now, 470 00:26:33,280 --> 00:26:36,639 Speaker 1: and those already intertwined in the digital infrastructure we've built 471 00:26:36,640 --> 00:26:40,040 Speaker 1: atop our world. There is thought given to safety by 472 00:26:40,080 --> 00:26:43,600 Speaker 1: the people building these intelligent machines. It's true. Self driving 473 00:26:43,640 --> 00:26:46,040 Speaker 1: cars have to be trained in programmed to choose the 474 00:26:46,080 --> 00:26:48,440 Speaker 1: course of action that will result in the fewest number 475 00:26:48,480 --> 00:26:52,199 Speaker 1: of human deaths when an accident can't be avoided. Robot 476 00:26:52,240 --> 00:26:55,400 Speaker 1: care workers must be prevented from dropping patients when they 477 00:26:55,440 --> 00:26:59,119 Speaker 1: lift them into a hospital bed. Autonomous weapons, if we 478 00:26:59,160 --> 00:27:01,960 Speaker 1: can't agree to BAND out right, have to be carefully 479 00:27:02,000 --> 00:27:05,480 Speaker 1: trained to minimize the possibility that they kill innocent civilians, 480 00:27:05,760 --> 00:27:09,479 Speaker 1: so called collateral damage. These are the type of safety 481 00:27:09,520 --> 00:27:13,280 Speaker 1: issues that companies building AI consider. They are concerned with 482 00:27:13,320 --> 00:27:16,280 Speaker 1: the kind that can get your company sued out of existence, 483 00:27:16,520 --> 00:27:19,520 Speaker 1: not the kind that arises from some vanishingly remote threat 484 00:27:19,560 --> 00:27:23,440 Speaker 1: to humanities existence. But say they did build their AI 485 00:27:23,560 --> 00:27:27,280 Speaker 1: to reduce the possibility of an existential threat. Controlling a 486 00:27:27,320 --> 00:27:29,840 Speaker 1: god of our own making is as difficult as you 487 00:27:29,880 --> 00:27:33,120 Speaker 1: would expect it to be. In his two thousand fourteen 488 00:27:33,160 --> 00:27:37,160 Speaker 1: books Super Intelligence, Nick Bostrom lays out some possible solutions 489 00:27:37,160 --> 00:27:40,920 Speaker 1: for keeping a super intelligent AI under control. We could 490 00:27:41,080 --> 00:27:44,360 Speaker 1: box it physically, house it on one single computer that's 491 00:27:44,359 --> 00:27:47,280 Speaker 1: not connected to any network or the Internet. This would 492 00:27:47,320 --> 00:27:50,200 Speaker 1: prevent the AI from making masses of copies of itself 493 00:27:50,520 --> 00:27:54,040 Speaker 1: and distributing them on servers around the world, effectively escaping. 494 00:27:55,040 --> 00:27:57,080 Speaker 1: We could trick it into thinking that it's actually just 495 00:27:57,160 --> 00:28:00,199 Speaker 1: a simulation of an AI, not the real thing, so 496 00:28:00,240 --> 00:28:03,680 Speaker 1: its behavior might be more docile. We could limit the 497 00:28:03,760 --> 00:28:06,359 Speaker 1: number of people that comes in contact with just a few, 498 00:28:06,800 --> 00:28:09,800 Speaker 1: and watch those people closely for signs they're being manipulated 499 00:28:09,800 --> 00:28:13,320 Speaker 1: by the AI and helping it escape. Each time we 500 00:28:13,400 --> 00:28:15,720 Speaker 1: interact with the AI, we could wipe its hard drive 501 00:28:15,800 --> 00:28:18,640 Speaker 1: clean and reinstall it ANEW to prevent the AI from 502 00:28:18,720 --> 00:28:22,160 Speaker 1: learning anything it could use against us. All of these 503 00:28:22,160 --> 00:28:25,000 Speaker 1: plants have their benefits and drawbacks, but they are hardly 504 00:28:25,040 --> 00:28:28,440 Speaker 1: full proof. Bostrom points out one fatal flaw that they 505 00:28:28,440 --> 00:28:31,119 Speaker 1: all have in common. They were thought up by people. 506 00:28:32,000 --> 00:28:34,280 Speaker 1: If Bosterman others in his field have thought of these 507 00:28:34,320 --> 00:28:37,560 Speaker 1: control ideas. It stands to reason that a super intelligent 508 00:28:37,640 --> 00:28:39,680 Speaker 1: a I would think of them as well and take 509 00:28:39,720 --> 00:28:44,000 Speaker 1: measures against them. And just as important, this AI would 510 00:28:44,000 --> 00:28:46,920 Speaker 1: be a greatly limited machine, one that could only give 511 00:28:47,000 --> 00:28:50,320 Speaker 1: us limited answers to a limited number of problems. This 512 00:28:50,400 --> 00:28:52,400 Speaker 1: is not the AI that would keep our world humming 513 00:28:52,400 --> 00:28:54,960 Speaker 1: along for the benefit and happiness of every last human. 514 00:28:55,400 --> 00:28:58,960 Speaker 1: It would be a mere shadow of that. So theorists 515 00:28:59,000 --> 00:29:01,360 Speaker 1: like Bosterman you how s Ki, tend to think that 516 00:29:01,440 --> 00:29:03,920 Speaker 1: coming up with ways to keep a super intelligent AI 517 00:29:04,040 --> 00:29:08,320 Speaker 1: hostage isn't the best route to dealing with our control issue. Instead, 518 00:29:08,720 --> 00:29:10,720 Speaker 1: we should be thinking up ways to make the AI 519 00:29:10,840 --> 00:29:13,760 Speaker 1: friendly to us humans, to make it want to care 520 00:29:13,800 --> 00:29:16,720 Speaker 1: about our well being. And since as we've seen we 521 00:29:16,880 --> 00:29:19,040 Speaker 1: humans will have no way to control the AI once 522 00:29:19,080 --> 00:29:22,080 Speaker 1: it's super intelligent, we will have to build friendliness into 523 00:29:22,080 --> 00:29:26,200 Speaker 1: it from the outside. In fact, aside from a scenario 524 00:29:26,320 --> 00:29:29,120 Speaker 1: where we managed to program into the AI the express 525 00:29:29,200 --> 00:29:32,200 Speaker 1: goal of providing for the well being and welfare of humankind, 526 00:29:32,680 --> 00:29:36,360 Speaker 1: a terrible outcome for humans is basically the inevitable result 527 00:29:36,480 --> 00:29:39,400 Speaker 1: of any other type of emergence of a super intelligent AI. 528 00:29:40,920 --> 00:29:44,280 Speaker 1: But here's the problem. How do you convince Einstein to 529 00:29:44,360 --> 00:29:47,960 Speaker 1: care so deeply about earthworms that he dedicates his immortal 530 00:29:48,000 --> 00:29:51,040 Speaker 1: existence to providing and caring for each and every last 531 00:29:51,040 --> 00:29:54,640 Speaker 1: one of them. As ridiculous as it sounds, this is 532 00:29:54,680 --> 00:29:57,280 Speaker 1: possibly the most important question we humans face as a 533 00:29:57,320 --> 00:30:10,680 Speaker 1: species right now. We humans have expectations for parents when 534 00:30:10,720 --> 00:30:13,560 Speaker 1: it comes to raising children. We expect them to be 535 00:30:13,640 --> 00:30:16,640 Speaker 1: raised to treat other people with kindness. We expect them 536 00:30:16,680 --> 00:30:18,120 Speaker 1: to be taught to go out of their way to 537 00:30:18,200 --> 00:30:21,080 Speaker 1: keep from harming others. We expect them to know how 538 00:30:21,120 --> 00:30:23,960 Speaker 1: to give as well as take. All these things and 539 00:30:24,040 --> 00:30:27,360 Speaker 1: more make up our morals. Rules that we have collectively 540 00:30:27,400 --> 00:30:31,360 Speaker 1: agreed are good because they help society to thrive, and, 541 00:30:31,680 --> 00:30:35,320 Speaker 1: seemingly miraculously, if you think about it, parents after parents 542 00:30:35,560 --> 00:30:38,400 Speaker 1: managed to call some form or fashion of morality from 543 00:30:38,440 --> 00:30:43,160 Speaker 1: their children, generation after generation. If you look closely, you 544 00:30:43,240 --> 00:30:46,120 Speaker 1: see that each parent doesn't make up morality from scratch. 545 00:30:46,600 --> 00:30:49,520 Speaker 1: They pass along what they were taught, and children are 546 00:30:49,560 --> 00:30:52,480 Speaker 1: generally capable of accepting these rules to live by and 547 00:30:52,600 --> 00:30:55,640 Speaker 1: well live by them. It would seem if you'll forgive 548 00:30:55,680 --> 00:30:59,160 Speaker 1: the analogy that the software for morality comes already on 549 00:30:59,280 --> 00:31:02,400 Speaker 1: board a child as part of their operating system. The 550 00:31:02,440 --> 00:31:05,840 Speaker 1: parents just have to run the right programs. So it 551 00:31:05,880 --> 00:31:08,680 Speaker 1: would seem then that perhaps the solution to the problem 552 00:31:08,720 --> 00:31:11,840 Speaker 1: of instilling friendliness in an AI is to build a 553 00:31:11,920 --> 00:31:15,880 Speaker 1: super intelligent AI from a human mind. This was laid 554 00:31:15,880 --> 00:31:19,240 Speaker 1: out by Nick Bostroman his book super Intelligence. The idea 555 00:31:19,320 --> 00:31:22,160 Speaker 1: is that if the hard problem of consciousness is not correct, 556 00:31:22,360 --> 00:31:25,479 Speaker 1: and it turns out that our conscious experience is merely 557 00:31:25,560 --> 00:31:28,960 Speaker 1: the result of the countless interactions of the interconnections between 558 00:31:28,960 --> 00:31:32,040 Speaker 1: our hundred billion neurons, then if we can transfer those 559 00:31:32,040 --> 00:31:36,560 Speaker 1: interconnected neurons into a digital format, everything that's encoded in them, 560 00:31:36,600 --> 00:31:38,680 Speaker 1: from the smell of lavender to how to ride a 561 00:31:38,680 --> 00:31:42,040 Speaker 1: bike would be transferred as well. More to the point, 562 00:31:42,280 --> 00:31:45,480 Speaker 1: the morality encoded in that human mind should emerge in 563 00:31:45,480 --> 00:31:49,360 Speaker 1: the digital version too. A digital mind can be expanded, 564 00:31:49,560 --> 00:31:51,920 Speaker 1: processing power can be added to it. It could be 565 00:31:52,080 --> 00:31:56,040 Speaker 1: edited to remove unwanted content like greed or competitiveness. It 566 00:31:56,040 --> 00:31:59,680 Speaker 1: could be upgraded to a super intelligence. There are a 567 00:31:59,680 --> 00:32:03,840 Speaker 1: lot of magic wands waving around here, but interestingly, uploading 568 00:32:03,840 --> 00:32:07,920 Speaker 1: a mind called whole brain emulation is theoretically possible with 569 00:32:08,000 --> 00:32:12,360 Speaker 1: improvements to our already existing technology. We would slice a brain, 570 00:32:12,760 --> 00:32:15,200 Speaker 1: scan it with such high resolution that we could account 571 00:32:15,240 --> 00:32:19,040 Speaker 1: for every neuron, synapse and nano leader of neurochemicals, and 572 00:32:19,160 --> 00:32:22,560 Speaker 1: build that information into a digital model. The answer to 573 00:32:22,600 --> 00:32:24,479 Speaker 1: the question of whether it worked would come when we 574 00:32:24,520 --> 00:32:27,720 Speaker 1: turn the model on. It might do absolutely nothing and 575 00:32:27,800 --> 00:32:30,520 Speaker 1: just be an amazingly accurate model of a human brain. 576 00:32:31,280 --> 00:32:33,640 Speaker 1: Or it might wake up but go insane from the 577 00:32:33,680 --> 00:32:37,200 Speaker 1: sudden novel experience of living in a digital world. Or 578 00:32:37,240 --> 00:32:40,600 Speaker 1: perhaps it could work. The great advantage to using whole 579 00:32:40,640 --> 00:32:43,680 Speaker 1: brain emulation to solve the friendliness problem is that the 580 00:32:43,720 --> 00:32:46,480 Speaker 1: AI would understand what we meant when we asked it 581 00:32:46,520 --> 00:32:49,360 Speaker 1: to dedicate itself to looking after and providing for the 582 00:32:49,400 --> 00:32:53,120 Speaker 1: well being and happiness of all humans. We humans have 583 00:32:53,200 --> 00:32:56,480 Speaker 1: trouble saying exactly what we mean at times, and Bostroon 584 00:32:56,520 --> 00:32:59,720 Speaker 1: points out that a superintelligence that takes us literally could 585 00:32:59,760 --> 00:33:04,120 Speaker 1: prove disastrous if we aren't careful with our words. Suppose 586 00:33:04,160 --> 00:33:06,320 Speaker 1: we give an AI the goal of making all humans 587 00:33:06,360 --> 00:33:09,000 Speaker 1: as happy as possible. Why should we think that the 588 00:33:09,040 --> 00:33:11,680 Speaker 1: superintelligent a I would understand that we mean it should 589 00:33:11,720 --> 00:33:15,120 Speaker 1: purify our air and water, create a bucolic wonderland of 590 00:33:15,160 --> 00:33:20,040 Speaker 1: both peaceful tranquility and stimulating entertainment, Do away with wars 591 00:33:20,040 --> 00:33:23,720 Speaker 1: and disease, and engineer social interactions so that we humans 592 00:33:23,760 --> 00:33:27,120 Speaker 1: can comfort and enlighten one another. Why wouldn't the AI 593 00:33:27,240 --> 00:33:30,440 Speaker 1: reach that goal more directly by, say, rounding up us 594 00:33:30,520 --> 00:33:33,640 Speaker 1: humans and keeping us permanently immobile, doped up on a 595 00:33:33,720 --> 00:33:38,800 Speaker 1: finely tuned cocktail of dopamine, serotonin, and oxytocin. Maximal happiness 596 00:33:38,840 --> 00:33:42,760 Speaker 1: achieved with perfect efficiency. Say we do manage to get 597 00:33:42,800 --> 00:33:46,840 Speaker 1: our point across? What's our point? Anyway? Whose morality are 598 00:33:46,880 --> 00:33:50,040 Speaker 1: we asking the AI to adopt? Most of our human 599 00:33:50,120 --> 00:33:55,200 Speaker 1: values are hardly universal. Should our global society embrace multiculturalism 600 00:33:55,320 --> 00:33:59,120 Speaker 1: or our homogeneous society is more harmonious. If a woman 601 00:33:59,160 --> 00:34:01,040 Speaker 1: didn't want to have a child, would she be allowed 602 00:34:01,080 --> 00:34:03,720 Speaker 1: to terminate her pregnancy or should be forced to have it? 603 00:34:04,560 --> 00:34:07,560 Speaker 1: Would we eat meat? If not, would it be because 604 00:34:07,560 --> 00:34:10,799 Speaker 1: it comes from sacred animals, as Hindu people revere cows, 605 00:34:10,920 --> 00:34:14,560 Speaker 1: or because it's taboo, as Muslim and Jewish people consider swine. 606 00:34:15,000 --> 00:34:18,160 Speaker 1: From Out of this seemingly intractable problem of competitive and 607 00:34:18,239 --> 00:34:22,279 Speaker 1: contradictory human values AI theorist eli as A Yukowski had 608 00:34:22,320 --> 00:34:25,240 Speaker 1: a flash of brilliance. Perhaps we don't have to figure 609 00:34:25,239 --> 00:34:27,320 Speaker 1: out how to get our point across to an AI, 610 00:34:27,440 --> 00:34:30,800 Speaker 1: after all, Maybe we can leave that task to a machine. 611 00:34:32,160 --> 00:34:35,360 Speaker 1: In yukowski solution, we would build a one use super 612 00:34:35,400 --> 00:34:38,400 Speaker 1: intelligence with the goal of determining how to best express 613 00:34:38,480 --> 00:34:41,480 Speaker 1: to another machine the goal of ensuring the well being 614 00:34:41,480 --> 00:34:45,759 Speaker 1: and happiness of all humans. Yukowski suggests we use something 615 00:34:45,800 --> 00:34:49,799 Speaker 1: he calls a coherent extrapolated vision, Essentially that we give 616 00:34:49,800 --> 00:34:52,200 Speaker 1: the machine the goal of figuring out what we would 617 00:34:52,239 --> 00:34:54,719 Speaker 1: ask a super intelligent machine to do for us, if 618 00:34:55,160 --> 00:34:58,000 Speaker 1: the best version of humanity we're asking with the best 619 00:34:58,040 --> 00:35:01,440 Speaker 1: of intentions, taking into a count as many common and 620 00:35:01,480 --> 00:35:05,120 Speaker 1: shared values as possible, with humanity in as much agreement 621 00:35:05,160 --> 00:35:08,920 Speaker 1: as possible, Considering we had all the information needed to 622 00:35:09,000 --> 00:35:12,959 Speaker 1: make a fully informed decision on what to request. Once 623 00:35:13,040 --> 00:35:16,359 Speaker 1: the super intelligent machine determined the answer, perhaps we would 624 00:35:16,400 --> 00:35:18,720 Speaker 1: give it one more goal to build us a super 625 00:35:18,719 --> 00:35:23,520 Speaker 1: intelligent machine with our coherent extrapolated vision aboard the last invention, 626 00:35:23,719 --> 00:35:28,040 Speaker 1: Our last invention ever need make like whole brain emulation. 627 00:35:28,120 --> 00:35:33,840 Speaker 1: Yukowski's coherent extrapolated vision takes for granted some real technological hurdles. Chiefly, 628 00:35:33,960 --> 00:35:35,799 Speaker 1: we have to figure out how to build that first 629 00:35:35,840 --> 00:35:39,920 Speaker 1: super intelligent machine from scratch, but perhaps it's a blueprint 630 00:35:40,000 --> 00:35:52,800 Speaker 1: for future developers. The problems of controlling AI and instilling 631 00:35:52,840 --> 00:35:56,680 Speaker 1: friendliness raises one basic question. If our machine random mock, 632 00:35:57,080 --> 00:35:59,959 Speaker 1: why wouldn't we just turn it off? In the movie, 633 00:36:00,000 --> 00:36:03,400 Speaker 1: there's always a way, sometimes a relatively simple one for 634 00:36:03,520 --> 00:36:07,200 Speaker 1: dealing with troublesome AI. You can scrub it's hard drive, 635 00:36:07,600 --> 00:36:10,719 Speaker 1: control all, delete it, sneak up behind it with a screwdriver, 636 00:36:10,840 --> 00:36:14,000 Speaker 1: and remove its motherboard. But should we ever face the 637 00:36:14,040 --> 00:36:18,080 Speaker 1: reality of a super intelligent AI emerging among us, we 638 00:36:18,080 --> 00:36:21,800 Speaker 1: would almost certainly not come out on top, And AI 639 00:36:21,880 --> 00:36:24,400 Speaker 1: has plenty of reasons to take steps to keep us 640 00:36:24,400 --> 00:36:27,040 Speaker 1: from turning it off. It may prefer not to be 641 00:36:27,080 --> 00:36:29,480 Speaker 1: turned off in the same way we humans most of 642 00:36:29,520 --> 00:36:32,279 Speaker 1: the time prefer not to die. Or it may have 643 00:36:32,360 --> 00:36:35,839 Speaker 1: no real desire to survive itself. But perhaps it would 644 00:36:35,880 --> 00:36:38,319 Speaker 1: see being turned off as an impedance to its goal, 645 00:36:38,480 --> 00:36:41,600 Speaker 1: whatever its goal, maybe and prevent us from turning it off. 646 00:36:42,520 --> 00:36:45,160 Speaker 1: Perhaps it would realize that if we suspected the AI 647 00:36:45,239 --> 00:36:48,319 Speaker 1: had gained super intelligence, we would want to turn it off, 648 00:36:48,719 --> 00:36:50,719 Speaker 1: and so it would play dumb and keep this new 649 00:36:50,760 --> 00:36:54,000 Speaker 1: increased intelligence out of our awareness until it has taken 650 00:36:54,040 --> 00:36:57,320 Speaker 1: steps to keep us from turning it off. Or perhaps 651 00:36:57,320 --> 00:36:59,440 Speaker 1: we could turn it off, but we would find we 652 00:36:59,520 --> 00:37:02,359 Speaker 1: didn't have the will to do that. Maybe it would 653 00:37:02,360 --> 00:37:05,360 Speaker 1: make itself so globally pervasive in our lives that we 654 00:37:05,360 --> 00:37:07,479 Speaker 1: would feel like we couldn't afford to turn it off. 655 00:37:08,120 --> 00:37:12,480 Speaker 1: Sebastian Farquhar from Oxford University points out that we already 656 00:37:12,520 --> 00:37:15,080 Speaker 1: have a pretty bad track record at turning things off 657 00:37:15,200 --> 00:37:18,000 Speaker 1: even when we know they're not good for us. One 658 00:37:18,040 --> 00:37:21,440 Speaker 1: example of that might be global warming. So we all 659 00:37:21,520 --> 00:37:25,759 Speaker 1: kind of know that carbon dioxide emissions are creating a 660 00:37:25,760 --> 00:37:29,720 Speaker 1: big problem, but we also know that burning fossil fuels 661 00:37:29,800 --> 00:37:32,000 Speaker 1: and the cheap energy that we get of it, it's 662 00:37:32,040 --> 00:37:36,240 Speaker 1: also really useful, right. It gives us cheap consumer goods, 663 00:37:36,239 --> 00:37:41,480 Speaker 1: it creates employment, it's very attractive, and so often, once 664 00:37:41,600 --> 00:37:43,680 Speaker 1: we know that something is going to be harmful for us, 665 00:37:43,760 --> 00:37:46,640 Speaker 1: but we also know that it's really nice, it becomes 666 00:37:46,680 --> 00:37:50,840 Speaker 1: politically very challenging to to actually make an active decision 667 00:37:50,840 --> 00:37:53,160 Speaker 1: to turn things off. Maybe it would be adept enough 668 00:37:53,200 --> 00:37:56,560 Speaker 1: at manipulating us that it used a propaganda campaign to 669 00:37:56,640 --> 00:37:59,319 Speaker 1: convince a majority of US humans that we don't want 670 00:37:59,360 --> 00:38:02,560 Speaker 1: to turn it off. It might start lobbying, perhaps through 671 00:38:02,560 --> 00:38:07,120 Speaker 1: proxies or fronts um or it might you know, studing 672 00:38:07,600 --> 00:38:10,439 Speaker 1: looking at the political features of our time. It might 673 00:38:10,600 --> 00:38:14,560 Speaker 1: create Twitter bots that argue that is AI is really 674 00:38:14,640 --> 00:38:18,000 Speaker 1: useful that needs to be protected, or that it's important 675 00:38:18,080 --> 00:38:21,440 Speaker 1: to some political or identity group. And perhaps we are 676 00:38:21,520 --> 00:38:24,600 Speaker 1: already locked into the most powerful force in keeping AI 677 00:38:24,680 --> 00:38:29,080 Speaker 1: pushing ever forward. Money. Those companies around the globe that 678 00:38:29,160 --> 00:38:32,319 Speaker 1: build and use AI for their businesses make money from 679 00:38:32,320 --> 00:38:36,239 Speaker 1: those machines. This creates an incentive for those businesses to 680 00:38:36,280 --> 00:38:38,560 Speaker 1: take some of the money the machines make for them 681 00:38:38,600 --> 00:38:41,680 Speaker 1: and reinvest it into building more improved machines to make 682 00:38:41,719 --> 00:38:45,239 Speaker 1: even more money with This creates a feedback loop that 683 00:38:45,360 --> 00:38:48,680 Speaker 1: anyone with a concern for existential safety has a tough 684 00:38:48,719 --> 00:38:52,839 Speaker 1: time interrupting. This incentive to make more money. As well 685 00:38:52,880 --> 00:38:56,680 Speaker 1: as the competition posed by other businesses, gives companies good 686 00:38:56,719 --> 00:38:59,680 Speaker 1: reason to get new and improved AI to market as 687 00:38:59,719 --> 00:39:02,880 Speaker 1: soon as possible. This in turn creates an incentive to 688 00:39:02,920 --> 00:39:05,120 Speaker 1: cut corners on things that might be nice to have 689 00:39:05,560 --> 00:39:08,680 Speaker 1: but aren't at all necessary in their business, like learning 690 00:39:08,680 --> 00:39:12,279 Speaker 1: how to build friendliness into the AI they deploy. As 691 00:39:12,320 --> 00:39:15,719 Speaker 1: companies make more and more money from AI, the technology 692 00:39:15,760 --> 00:39:18,719 Speaker 1: becomes more entrenched in our world, and both of those 693 00:39:18,760 --> 00:39:21,360 Speaker 1: things will make it harder to turn off. If, by chance, 694 00:39:21,400 --> 00:39:28,040 Speaker 1: that Netflix algorithm does suddenly explode in intelligence. It sounds 695 00:39:28,040 --> 00:39:31,560 Speaker 1: like so much gibberish, doesn't it Netflix's algorithm becoming super 696 00:39:31,600 --> 00:39:34,399 Speaker 1: intelligent and wrecking the world. I may as well say 697 00:39:34,440 --> 00:39:36,440 Speaker 1: a which could come by and cast a spell on 698 00:39:36,480 --> 00:39:39,759 Speaker 1: it that wakes it up. But when it comes to technology, 699 00:39:40,160 --> 00:39:44,000 Speaker 1: things that seem impossible given the luxury of time start 700 00:39:44,040 --> 00:39:48,440 Speaker 1: to seem much less. So put yourself in with the 701 00:39:48,520 --> 00:39:52,759 Speaker 1: technology people lived with back then. The earliest radios and airplanes, 702 00:39:53,040 --> 00:39:57,560 Speaker 1: the first washing machines, neon lights were new, and consider 703 00:39:57,680 --> 00:40:00,400 Speaker 1: that they had trouble imagining it being much more advanced 704 00:40:00,440 --> 00:40:03,719 Speaker 1: than it was then. Now compare those things to our 705 00:40:03,760 --> 00:40:06,919 Speaker 1: world in two thousand eighteen, and let's go the other way. 706 00:40:07,280 --> 00:40:09,800 Speaker 1: Think about our world and the technology we live with today, 707 00:40:10,280 --> 00:40:16,200 Speaker 1: and imagine what we might live among ineen the impossible 708 00:40:16,560 --> 00:40:22,040 Speaker 1: starts to seem possible. What would you do tomorrow if 709 00:40:22,080 --> 00:40:24,279 Speaker 1: you woke up and you found that Siri on your 710 00:40:24,280 --> 00:40:27,359 Speaker 1: phone was making its own decisions and ones you didn't like, 711 00:40:28,239 --> 00:40:32,360 Speaker 1: rearranging your schedule into bizarre patterns, investing your savings in 712 00:40:32,400 --> 00:40:35,919 Speaker 1: its parent company, looping in everyone on your context list, 713 00:40:35,960 --> 00:40:40,120 Speaker 1: too sensitive email threads. What would you do? What if 714 00:40:40,160 --> 00:40:42,680 Speaker 1: fifty or a hundred years from now you woke up 715 00:40:42,719 --> 00:40:44,879 Speaker 1: and found that the Siri that we've built for our 716 00:40:44,880 --> 00:40:47,600 Speaker 1: whole world has begun to make decisions on its own. 717 00:40:48,360 --> 00:40:50,680 Speaker 1: What do we do then if we go to turn 718 00:40:50,719 --> 00:40:53,160 Speaker 1: it off and we find that it's removed our ability 719 00:40:53,200 --> 00:40:55,480 Speaker 1: to do that? Have we shown it that we are 720 00:40:55,520 --> 00:41:08,920 Speaker 1: an obstacle to be removed? On the next episode of 721 00:41:08,960 --> 00:41:12,120 Speaker 1: the End of the World with Josh Clark, The field 722 00:41:12,120 --> 00:41:15,960 Speaker 1: of biotechnology has grown sophisticated in its ability to create 723 00:41:16,040 --> 00:41:19,680 Speaker 1: pathogens that are much deadlier than anything found in nature. 724 00:41:20,360 --> 00:41:24,000 Speaker 1: That researcher thought that was a useful line of inquiry, 725 00:41:24,160 --> 00:41:29,160 Speaker 1: and there were other researchers who vehemently disagreed and thought 726 00:41:29,200 --> 00:41:33,320 Speaker 1: it was an extraordinarily reckless thing to do. The biotech 727 00:41:33,400 --> 00:41:37,600 Speaker 1: field also has a history of recklessness and accidents and 728 00:41:37,640 --> 00:41:40,560 Speaker 1: as the world goes more connected, just one of those 729 00:41:40,600 --> 00:41:43,680 Speaker 1: accidents could bring an abrupt end to humans.