1 00:00:11,400 --> 00:00:14,400 FAINT GARBLED VOICES 2 00:00:14,400 --> 00:00:17,520 Intelligence is the ability to understand. 3 00:00:17,520 --> 00:00:19,760 We passed on what we know to machines. 4 00:00:19,760 --> 00:00:22,160 The rise of artificial intelligence is happening fast, 5 00:00:22,160 --> 00:00:23,880 but some fear the new technology 6 00:00:23,880 --> 00:00:26,280 might have more problems than anticipated. 7 00:00:26,280 --> 00:00:27,800 We will not control it. 8 00:01:46,640 --> 00:01:52,000 Artificial intelligence is simply non-biological intelligence, 9 00:01:52,000 --> 00:01:55,760 and intelligence itself is simply the ability to accomplish goals. 10 00:01:58,040 --> 00:02:00,400 I'm convinced that AI will ultimately be 11 00:02:00,400 --> 00:02:02,800 either the best thing ever to happen to humanity, 12 00:02:02,800 --> 00:02:04,480 or the worst thing ever to happen. 13 00:02:05,960 --> 00:02:07,680 We can use it to solve 14 00:02:07,680 --> 00:02:11,360 all of today's and tomorrow's greatest problems - 15 00:02:11,360 --> 00:02:15,440 cure diseases, deal with climate change, 16 00:02:15,440 --> 00:02:17,360 lift everybody out of poverty. 17 00:02:19,040 --> 00:02:22,200 But we could use exactly the same technology 18 00:02:22,200 --> 00:02:25,120 to create a brutal global dictatorship, 19 00:02:25,120 --> 00:02:29,400 with unprecedented surveillance and inequality and suffering. 20 00:02:31,000 --> 00:02:34,120 That's why this is the most important conversation of our time. 21 00:02:39,200 --> 00:02:43,000 Artificial intelligence is everywhere 22 00:02:43,000 --> 00:02:45,920 because we now have thinking machines. 23 00:02:47,040 --> 00:02:50,360 If you go on social media or online, 24 00:02:50,360 --> 00:02:52,440 there's an artificial intelligence engine 25 00:02:52,440 --> 00:02:54,920 that decides what to recommend. 26 00:02:55,960 --> 00:02:57,240 If you go on Facebook 27 00:02:57,240 --> 00:02:59,640 and you're just scrolling through your friends' posts, 28 00:02:59,640 --> 00:03:01,760 there's an artificial intelligence engine 29 00:03:01,760 --> 00:03:05,440 that's picking which one to show you first and which one to bury. 30 00:03:05,440 --> 00:03:07,360 If you try to get insurance, 31 00:03:07,360 --> 00:03:11,800 there's an AI engine trying to figure out how risky you are. 32 00:03:11,800 --> 00:03:14,560 And if you apply for a job, 33 00:03:14,560 --> 00:03:18,600 it's quite possible that an AI engine looks at the resume. 34 00:03:25,960 --> 00:03:27,560 We are made of data. 35 00:03:28,880 --> 00:03:31,200 Every one of us is made of data. 36 00:03:32,720 --> 00:03:36,440 In terms of how we behave, how we talk, how we love, 37 00:03:36,440 --> 00:03:37,680 what we do every day. 38 00:03:40,120 --> 00:03:45,440 So computer scientists are developing deep learning algorithms 39 00:03:45,440 --> 00:03:51,200 that can learn to identify, classify and predict patterns 40 00:03:51,200 --> 00:03:53,440 within massive amounts of data. 41 00:04:05,960 --> 00:04:09,960 We are facing a form of precision surveillance. 42 00:04:09,960 --> 00:04:13,320 You could call it algorithmic surveillance, 43 00:04:13,320 --> 00:04:16,400 and it means that you cannot go unrecognised. 44 00:04:17,760 --> 00:04:21,040 You are always under the watch of algorithms. 45 00:04:25,720 --> 00:04:29,000 Almost all the AI development on the planet today 46 00:04:29,000 --> 00:04:31,720 is done by a handful of big technology companies 47 00:04:31,720 --> 00:04:33,440 or by a few large governments. 48 00:04:36,600 --> 00:04:40,720 If we look at what AI is mostly being developed for, 49 00:04:40,720 --> 00:04:43,240 I would say it's, er... 50 00:04:43,240 --> 00:04:46,200 ..killing, spying and brainwashing. 51 00:04:46,200 --> 00:04:48,320 So, I mean, we have military AI, 52 00:04:48,320 --> 00:04:50,920 we have a whole surveillance apparatus 53 00:04:50,920 --> 00:04:53,720 being built using AI by major governments. 54 00:04:53,720 --> 00:04:55,840 And we have an advertising industry 55 00:04:55,840 --> 00:05:00,520 which is oriented toward recognising what ads to try to sell to someone. 56 00:05:03,800 --> 00:05:06,760 We humans have come to a fork in the road now. 57 00:05:08,160 --> 00:05:11,240 The AI we have today is very narrow. 58 00:05:12,480 --> 00:05:15,080 The Holy Grail of AI research ever since the beginning 59 00:05:15,080 --> 00:05:19,040 is to make AI that can do everything better than us. 60 00:05:19,040 --> 00:05:20,880 We basically built a god. 61 00:05:22,280 --> 00:05:24,680 It's going to revolutionise life as we know it. 62 00:05:27,680 --> 00:05:29,440 It's incredibly important 63 00:05:29,440 --> 00:05:32,280 to take a step back and think carefully about this. 64 00:05:33,960 --> 00:05:36,040 What sort of society do we want? 65 00:05:39,400 --> 00:05:42,040 So we're in this historic transformation. 66 00:05:43,440 --> 00:05:45,880 Like, we're raising this new creature. 67 00:05:45,880 --> 00:05:48,720 You have a new offspring of sorts. 68 00:05:50,720 --> 00:05:54,320 But, just like actual offspring, 69 00:05:54,320 --> 00:05:57,880 you don't get to control everything it's going to do. 70 00:06:31,320 --> 00:06:34,680 We are living at this privileged moment 71 00:06:34,680 --> 00:06:38,320 where, for the first time, we will see, probably, 72 00:06:38,320 --> 00:06:42,160 that AI is really going to outcompete humans 73 00:06:42,160 --> 00:06:45,440 in many, many, if not all important fields. 74 00:06:48,760 --> 00:06:51,080 Everything is going to change. 75 00:06:51,080 --> 00:06:54,520 A new form of life is emerging. 76 00:07:19,720 --> 00:07:25,760 When I was a boy, I thought, "How can I maximise my impact?" 77 00:07:26,880 --> 00:07:31,040 And then it was clear that I have to build something 78 00:07:31,040 --> 00:07:34,760 that learns to become smarter than myself, 79 00:07:34,760 --> 00:07:36,360 such that I can retire 80 00:07:36,360 --> 00:07:39,200 and the smarter thing can further self-improve 81 00:07:39,200 --> 00:07:41,720 and solve all the problems that I cannot solve. 82 00:07:46,160 --> 00:07:50,040 Multiplying that tiny little bit of creativity that I have 83 00:07:50,040 --> 00:07:51,280 into infinity. 84 00:07:52,800 --> 00:07:55,840 And that's what has been driving me since then. 85 00:08:13,880 --> 00:08:15,600 How am I trying to build 86 00:08:15,600 --> 00:08:18,160 a general-purpose artificial intelligence? 87 00:08:20,200 --> 00:08:22,240 If you want to be intelligent, 88 00:08:22,240 --> 00:08:27,680 you have to recognise speech and video and handwriting and faces 89 00:08:27,680 --> 00:08:29,320 and all kinds of things. 90 00:08:29,320 --> 00:08:32,160 And there we have made a lot of progress. 91 00:08:34,040 --> 00:08:36,680 The LSTM neural networks, 92 00:08:36,680 --> 00:08:41,240 which we developed in our labs in Munich and in Switzerland, 93 00:08:41,240 --> 00:08:43,920 that's now used for speech recognition 94 00:08:43,920 --> 00:08:47,400 and translation and video recognition. 95 00:08:47,400 --> 00:08:50,360 They are now in everybody's smartphone. 96 00:08:50,360 --> 00:08:52,320 Almost 1 billion iPhones 97 00:08:52,320 --> 00:08:55,320 and in over 2 billion Android phones. 98 00:08:56,720 --> 00:09:02,080 So we are generating all kinds of useful by-products 99 00:09:02,080 --> 00:09:04,000 on the way to the general goal. 100 00:09:14,600 --> 00:09:15,960 The main goal... 101 00:09:17,280 --> 00:09:21,840 ..is an artificial general intelligence, an AGI, 102 00:09:21,840 --> 00:09:25,800 that can learn to improve the learning algorithm itself. 103 00:09:27,640 --> 00:09:31,600 So it basically can learn to improve the way it learns 104 00:09:31,600 --> 00:09:36,880 and it can also recursively improve the way it learns the way it learns 105 00:09:36,880 --> 00:09:38,560 without any limitations, 106 00:09:38,560 --> 00:09:42,480 except for the basic fundamental limitations of computability. 107 00:09:49,320 --> 00:09:52,480 One of my favourite robots is this one here. 108 00:09:52,480 --> 00:09:57,080 We use this robot for our studies of artificial curiosity, 109 00:09:57,080 --> 00:10:02,480 where we are trying to teach this robot to teach itself. 110 00:10:07,520 --> 00:10:08,760 What is a baby doing? 111 00:10:08,760 --> 00:10:12,400 A baby is curiously exploring its world. 112 00:10:14,720 --> 00:10:17,200 That's how it learns how gravity works 113 00:10:17,200 --> 00:10:19,920 and how certain things topple and so on. 114 00:10:21,360 --> 00:10:24,920 And as it learns to ask questions about the world, 115 00:10:24,920 --> 00:10:27,160 and as it learns to answer these questions, 116 00:10:27,160 --> 00:10:29,680 it becomes a more and more general problem-solver. 117 00:10:31,120 --> 00:10:32,720 And so our artificial systems 118 00:10:32,720 --> 00:10:35,720 are also learning to ask all kinds of questions, 119 00:10:35,720 --> 00:10:39,240 not just slavishly try to answer 120 00:10:39,240 --> 00:10:41,560 the questions given to them by humans. 121 00:10:45,560 --> 00:10:49,840 You have to give AI the freedom to invent its own tasks. 122 00:10:52,400 --> 00:10:55,160 If you don't do that, it's not going to become very smart. 123 00:10:56,760 --> 00:10:57,960 On the other hand, 124 00:10:57,960 --> 00:11:00,680 it's really hard to predict what they are going to do. 125 00:11:24,360 --> 00:11:27,400 I feel that technology is a force of nature. 126 00:11:30,240 --> 00:11:32,000 I feel like there is a lot of similarity 127 00:11:32,000 --> 00:11:34,320 between technology and biological evolution. 128 00:11:43,440 --> 00:11:44,760 Playing God. 129 00:11:48,200 --> 00:11:50,960 Scientists have been accused of playing God for a while. 130 00:11:52,400 --> 00:11:55,360 But there is a real sense 131 00:11:55,360 --> 00:11:58,760 in which we are creating something...very different 132 00:11:58,760 --> 00:12:00,840 from anything we've created so far. 133 00:12:19,280 --> 00:12:20,920 SCANNER BEEPS 134 00:12:24,400 --> 00:12:27,600 I was interested in the concept of AI from a relatively early age. 135 00:12:29,720 --> 00:12:32,920 At some point, I got especially interested in machine learning. 136 00:12:36,280 --> 00:12:38,080 What is experience? 137 00:12:38,080 --> 00:12:41,440 What is learning? What is thinking? 138 00:12:41,440 --> 00:12:42,840 How does the brain work? 139 00:12:44,640 --> 00:12:46,560 These questions are philosophical, 140 00:12:46,560 --> 00:12:49,680 but it looks like we can come up with algorithms 141 00:12:49,680 --> 00:12:53,280 that both do useful things and help us answer these questions. 142 00:12:54,640 --> 00:12:57,600 Like, it's almost like applied philosophy. 143 00:13:22,840 --> 00:13:26,080 Artificial general intelligence, AGI. 144 00:13:27,520 --> 00:13:31,840 A computer system that can do any job or any task 145 00:13:31,840 --> 00:13:34,920 that a human does, but only better. 146 00:13:47,920 --> 00:13:50,720 Yeah, I mean, we definitely will be able to create... 147 00:13:52,120 --> 00:13:54,800 ..completely autonomous beings with their own goals. 148 00:13:58,560 --> 00:14:00,280 And it will be very important... 149 00:14:00,280 --> 00:14:05,200 Especially as these beings become much smarter than humans, 150 00:14:05,200 --> 00:14:09,200 it's going to be important to...to have these beings... 151 00:14:10,720 --> 00:14:13,760 That the goals of these beings be aligned with our goals. 152 00:14:17,240 --> 00:14:19,960 That's what we're trying to do at OpenAI - 153 00:14:19,960 --> 00:14:22,240 be at the forefront of research 154 00:14:22,240 --> 00:14:26,600 and steer the research, steer the initial conditions, 155 00:14:26,600 --> 00:14:30,640 so to maximise the chance that the future will be good for humans. 156 00:14:46,240 --> 00:14:47,680 Now, AI is a great thing, 157 00:14:47,680 --> 00:14:51,680 because AI will solve all the problems that we have today. 158 00:14:53,640 --> 00:14:55,320 It will solve employment. 159 00:14:56,440 --> 00:14:58,040 It will solve disease. 160 00:14:59,280 --> 00:15:01,320 It will solve poverty. 161 00:15:03,680 --> 00:15:06,240 But it will also create new problems. 162 00:15:09,240 --> 00:15:10,960 I think that... 163 00:15:12,480 --> 00:15:13,680 Hm. 164 00:15:15,280 --> 00:15:19,200 ..the problem of fake news is going to be 1,000 million times worse. 165 00:15:21,200 --> 00:15:23,520 Cyber attacks will become much more extreme. 166 00:15:25,360 --> 00:15:28,240 You will have totally automated AI weapons. 167 00:15:30,920 --> 00:15:32,760 I think AI has the potential to create 168 00:15:32,760 --> 00:15:34,560 infinitely stable dictatorships. 169 00:15:40,560 --> 00:15:43,480 You're going to see dramatically more intelligent systems 170 00:15:43,480 --> 00:15:46,240 in 10 or 15 years from now. 171 00:15:46,240 --> 00:15:48,400 And I think it's highly likely that those systems 172 00:15:48,400 --> 00:15:51,920 will have completely astronomical impact on society. 173 00:15:54,000 --> 00:15:55,880 Will humans actually benefit? 174 00:15:57,280 --> 00:15:59,120 And who will benefit, who will not? 175 00:16:19,160 --> 00:16:24,120 In 2012, IBM estimated that an average person 176 00:16:24,120 --> 00:16:29,640 is generating 500 megabytes of digital footprints every single day. 177 00:16:29,640 --> 00:16:32,560 Imagine that you wanted to back up one day worth of data 178 00:16:32,560 --> 00:16:35,880 that humanity is leaving behind on paper. 179 00:16:35,880 --> 00:16:38,440 How tall would be the stack of paper 180 00:16:38,440 --> 00:16:42,640 that contains just one day worth of data that humanity's producing? 181 00:16:44,360 --> 00:16:47,280 It's like from the Earth to the sun four times over. 182 00:16:49,280 --> 00:16:55,200 In 2025, we'll be generating 62 gigabytes of data 183 00:16:55,200 --> 00:16:57,360 per person per day. 184 00:17:11,680 --> 00:17:15,040 We are leaving a ton of digital footprints 185 00:17:15,040 --> 00:17:17,080 while going through our lives. 186 00:17:19,640 --> 00:17:22,480 They provide computer algorithms with a fairly good idea 187 00:17:22,480 --> 00:17:26,920 about who we are, what we want, what we're doing. 188 00:17:30,560 --> 00:17:34,040 In my work, I looked at different types of digital footprints. 189 00:17:34,040 --> 00:17:38,440 I looked at Facebook likes, I looked at language, credit card records, 190 00:17:38,440 --> 00:17:41,280 web-browsing histories, search records. 191 00:17:42,480 --> 00:17:46,600 Then each time, I found that if you get enough of this data, 192 00:17:46,600 --> 00:17:49,320 you can accurately predict future behaviour 193 00:17:49,320 --> 00:17:52,280 and reveal important intimate traits. 194 00:17:53,680 --> 00:17:56,000 This can be used in great ways, 195 00:17:56,000 --> 00:17:59,360 but it can also be used to manipulate people. 196 00:18:04,680 --> 00:18:08,440 Facebook is delivering daily information 197 00:18:08,440 --> 00:18:11,000 to 2 billion people or more. 198 00:18:13,400 --> 00:18:17,400 If you slightly change the functioning of Facebook engine, 199 00:18:17,400 --> 00:18:24,240 you can move the opinions, and hence the votes, of millions of people. 200 00:18:24,240 --> 00:18:26,240 What do we want? Brexit! 201 00:18:26,240 --> 00:18:28,520 When do we want it? Now! 202 00:18:30,480 --> 00:18:33,200 A politician wouldn't be able to figure out 203 00:18:33,200 --> 00:18:37,440 which message each one of his or her voters would like, 204 00:18:37,440 --> 00:18:40,720 but a computer can see what political message 205 00:18:40,720 --> 00:18:43,640 would be particularly convincing for you. 206 00:18:46,040 --> 00:18:49,840 Ladies and gentlemen, it's my privilege to speak to you today 207 00:18:49,840 --> 00:18:52,920 about the power of big data and psychographics 208 00:18:52,920 --> 00:18:54,800 in the electoral process. 209 00:18:54,800 --> 00:18:57,880 Data firm Cambridge Analytica secretly harvested 210 00:18:57,880 --> 00:18:59,680 the personal information 211 00:18:59,680 --> 00:19:02,800 of 50 million unsuspecting Facebook users. 212 00:19:02,800 --> 00:19:06,400 CHANTING: USA! USA! USA! 213 00:19:06,400 --> 00:19:07,480 The data firm, 214 00:19:07,480 --> 00:19:10,320 hired by Donald Trump's presidential election campaign, 215 00:19:10,320 --> 00:19:12,760 used secretly-obtained information 216 00:19:12,760 --> 00:19:15,520 to directly target potential American voters. 217 00:19:15,520 --> 00:19:18,760 With that, they say they can predict the personality 218 00:19:18,760 --> 00:19:22,080 of every single adult in the United States. 219 00:19:22,080 --> 00:19:23,640 Tonight, we're hearing from 220 00:19:23,640 --> 00:19:26,440 Cambridge Analytica whistle-blower Christopher Wylie. 221 00:19:26,440 --> 00:19:29,680 What we worked on was data-harvesting programmes 222 00:19:29,680 --> 00:19:31,440 where we would pull data 223 00:19:31,440 --> 00:19:33,120 and run that data through algorithms 224 00:19:33,120 --> 00:19:35,000 that could profile their personality traits 225 00:19:35,000 --> 00:19:36,560 and other psychological attributes 226 00:19:36,560 --> 00:19:39,160 to exploit mental vulnerabilities 227 00:19:39,160 --> 00:19:41,120 that our algorithms show that they had. 228 00:19:50,040 --> 00:19:52,440 Cambridge Analytica mentioned once, 229 00:19:52,440 --> 00:19:55,160 or said that their models were based on my work. 230 00:19:57,120 --> 00:20:00,840 But Cambridge Analytica is just one of the hundreds of companies 231 00:20:00,840 --> 00:20:04,920 that are using such methods to target voters. 232 00:20:06,720 --> 00:20:10,680 You know, I would be asked questions by journalists such as, you know, 233 00:20:10,680 --> 00:20:16,160 "So how do you feel about electing Trump and supporting Brexit?" 234 00:20:16,160 --> 00:20:18,440 And, you know, how do you answer to such a question? 235 00:20:20,040 --> 00:20:26,200 But I guess that I have to deal with being blamed for all of it. 236 00:20:42,400 --> 00:20:47,080 How tech started was as a democratising force. 237 00:20:47,080 --> 00:20:48,320 As a force for good, 238 00:20:48,320 --> 00:20:51,720 as an ability for humans to interact with each other without gatekeepers. 239 00:20:54,920 --> 00:20:57,120 There's never been a bigger experiment 240 00:20:57,120 --> 00:20:59,640 in communications for the human race. 241 00:21:00,920 --> 00:21:04,320 What happens when everybody gets to have their say? 242 00:21:04,320 --> 00:21:06,680 You would assume that it would be for the better, 243 00:21:06,680 --> 00:21:09,440 that there'd be more democracy. there'd be more discussion, 244 00:21:09,440 --> 00:21:11,160 there'd be more tolerance. 245 00:21:11,160 --> 00:21:14,280 But what's happened is that these systems have been hijacked. 246 00:21:15,920 --> 00:21:19,440 We stand for connecting every person. 247 00:21:19,440 --> 00:21:21,440 For a global community. 248 00:21:22,480 --> 00:21:26,200 One company, Facebook, is responsible for the communications 249 00:21:26,200 --> 00:21:28,120 of a lot of the human race. 250 00:21:30,120 --> 00:21:31,800 Same thing with Google. 251 00:21:31,800 --> 00:21:34,800 Everything you want to know about the world comes from them. 252 00:21:36,240 --> 00:21:39,840 This is a global information economy 253 00:21:39,840 --> 00:21:42,200 that is controlled by a small group of people. 254 00:21:42,200 --> 00:21:44,240 BELL RINGS 255 00:21:44,240 --> 00:21:46,400 CHEERING 256 00:21:49,000 --> 00:21:53,440 The world's richest companies are all technology companies. 257 00:21:53,440 --> 00:21:59,160 Google, Apple, Microsoft, Amazon, Facebook. 258 00:22:00,520 --> 00:22:05,680 It's staggering how, in probably just ten years, 259 00:22:05,680 --> 00:22:09,560 that the entire corporate power structure 260 00:22:09,560 --> 00:22:14,080 are basically in the business of trading electrons. 261 00:22:15,920 --> 00:22:21,880 These little bits and bytes are really the new currency. 262 00:22:28,040 --> 00:22:31,720 The way that data is monetised is happening all around us, 263 00:22:31,720 --> 00:22:33,360 even if it's invisible to us. 264 00:22:35,640 --> 00:22:38,840 Google has every amount of information available. 265 00:22:38,840 --> 00:22:40,320 They track people 266 00:22:40,320 --> 00:22:41,760 by their GPS location. 267 00:22:41,760 --> 00:22:44,880 They know exactly what your search history has been. 268 00:22:44,880 --> 00:22:47,720 They know your political preferences. 269 00:22:47,720 --> 00:22:51,080 Your search history alone can tell you everything about an individual, 270 00:22:51,080 --> 00:22:54,040 from their health problems to their sexual preferences. 271 00:22:54,040 --> 00:22:56,560 So Google's reach is unlimited. 272 00:23:03,360 --> 00:23:06,320 So we've seen Google and Facebook 273 00:23:06,320 --> 00:23:09,680 rise into these large surveillance machines, 274 00:23:09,680 --> 00:23:12,880 and they're both actually ad brokers. 275 00:23:12,880 --> 00:23:16,600 It sounds really mundane, but they're high-tech ad brokers. 276 00:23:17,560 --> 00:23:19,240 And the reason they're so profitable 277 00:23:19,240 --> 00:23:22,640 is that they're using artificial intelligence 278 00:23:22,640 --> 00:23:24,440 to process all this data about you... 279 00:23:26,280 --> 00:23:29,320 ..and then to match you with the advertiser 280 00:23:29,320 --> 00:23:34,360 that wants to reach people like you, for whatever message. 281 00:23:37,560 --> 00:23:39,440 One of the problems with technology 282 00:23:39,440 --> 00:23:42,560 is that it's been developed to be addictive. 283 00:23:42,560 --> 00:23:44,600 The way these companies design these things 284 00:23:44,600 --> 00:23:47,840 is in order to pull you in and engage you. 285 00:23:47,840 --> 00:23:51,320 They want to become essentially a slot machine of attention. 286 00:23:53,200 --> 00:23:54,800 So you're always paying attention. 287 00:23:54,800 --> 00:23:56,480 You're always jacked into The Matrix. 288 00:23:56,480 --> 00:23:58,080 You're always checking. 289 00:24:01,280 --> 00:24:03,440 When somebody controls what you read, 290 00:24:03,440 --> 00:24:05,360 they also control what you think. 291 00:24:06,920 --> 00:24:09,360 You get more of what you've seen before and liked before, 292 00:24:09,360 --> 00:24:12,400 because this gives more traffic and that gives more ads. 293 00:24:14,200 --> 00:24:17,840 But it also locks you into your echo chamber. 294 00:24:17,840 --> 00:24:20,960 And this is what leads to this polarisation that we see today. 295 00:24:20,960 --> 00:24:23,520 CHANTING 296 00:24:24,680 --> 00:24:28,360 Jair Bolsonaro, Brazil's right-wing populist candidate, 297 00:24:28,360 --> 00:24:32,680 sometimes likened to Donald Trump, winning the presidency Sunday night, 298 00:24:32,680 --> 00:24:36,200 in that country's most polarising election in decades. 299 00:24:36,200 --> 00:24:37,640 Bolsonaro! 300 00:24:38,800 --> 00:24:41,160 What we're seeing around the world 301 00:24:41,160 --> 00:24:44,160 is upheaval and polarisation and conflict... 302 00:24:45,560 --> 00:24:49,760 ..that is partially pushed by algorithms 303 00:24:49,760 --> 00:24:53,800 that's figured out that political extremes, 304 00:24:53,800 --> 00:24:56,840 tribalism and sort of shouting for your team 305 00:24:56,840 --> 00:24:59,440 and feeling good about it, is engaging. 306 00:24:59,440 --> 00:25:02,000 SINGING 307 00:25:05,000 --> 00:25:07,680 Social media may be adding to the attention 308 00:25:07,680 --> 00:25:10,320 to hate crimes around the globe. 309 00:25:10,320 --> 00:25:12,760 It's about how people can become radicalised 310 00:25:12,760 --> 00:25:16,080 by living in the fever swamps of the Internet. 311 00:25:16,080 --> 00:25:18,800 So is this a key moment for the tech giants? 312 00:25:18,800 --> 00:25:21,720 Are they now prepared to take responsibility as publishers 313 00:25:21,720 --> 00:25:24,400 for what they share with the world? 314 00:25:24,400 --> 00:25:28,800 If you deploy a powerful, potent technology at scale - 315 00:25:28,800 --> 00:25:31,440 and if you're talking about Google and Facebook, 316 00:25:31,440 --> 00:25:34,200 you're deploying things at a scale of billions - 317 00:25:34,200 --> 00:25:37,400 if your artificial intelligence is pushing polarisation, 318 00:25:37,400 --> 00:25:40,080 you have global upheaval, potentially. 319 00:25:40,080 --> 00:25:44,200 CHANTING: White lives matter! White lives matter! White lives matter! 320 00:25:44,200 --> 00:25:48,440 CHANTING: Black lives matter! Black lives matter! Black lives matter! 321 00:25:55,000 --> 00:25:58,360 SCREAMING 322 00:26:23,280 --> 00:26:29,600 QUIET CLASSICAL PIANO 323 00:26:35,200 --> 00:26:38,600 Artificial general intelligence, AGI. 324 00:26:40,920 --> 00:26:42,840 Imagine your smartest friend. 325 00:26:44,160 --> 00:26:46,800 With 1,000 friends, just as smart. 326 00:26:49,000 --> 00:26:52,240 And then run them at 1,000-times faster than real-time. 327 00:26:52,240 --> 00:26:54,560 So it means that, in every day of our time, 328 00:26:54,560 --> 00:26:56,880 they will do three years of thinking. 329 00:26:56,880 --> 00:26:59,440 Can you imagine how much... 330 00:26:59,440 --> 00:27:01,400 ..you could do... 331 00:27:01,400 --> 00:27:02,920 ..if for every day... 332 00:27:03,960 --> 00:27:06,040 ..you could do three years' worth of work? 333 00:27:27,120 --> 00:27:32,240 It wouldn't be an unfair comparison to say that what we have right now 334 00:27:32,240 --> 00:27:35,600 is even more exciting than, you know, the quantum physicists 335 00:27:35,600 --> 00:27:38,640 of the early 20th century when they discovered nuclear power. 336 00:27:40,320 --> 00:27:42,800 I feel extremely lucky to be taking part in. 337 00:27:49,880 --> 00:27:52,760 Many machine learning experts, people who are very knowledgeable, 338 00:27:52,760 --> 00:27:55,400 very experienced, have a lot of scepticism about AGI. 339 00:27:57,120 --> 00:28:00,480 About when it could happen and about whether it could happen at all. 340 00:28:05,880 --> 00:28:08,000 But right now, this is something 341 00:28:08,000 --> 00:28:10,840 that just not that many people have realised yet, 342 00:28:10,840 --> 00:28:15,800 that the speed of computers for neural networks, for AI, 343 00:28:15,800 --> 00:28:19,880 are going to become maybe 20,000 times faster 344 00:28:19,880 --> 00:28:21,600 in a small number of years. 345 00:28:23,920 --> 00:28:25,960 The entire hardware industry, 346 00:28:25,960 --> 00:28:30,320 for a long time, didn't really know what to do next. 347 00:28:30,320 --> 00:28:35,640 But with artificial neural networks, now that they actually work, 348 00:28:35,640 --> 00:28:38,960 you have a reason to build huge computers. 349 00:28:38,960 --> 00:28:41,480 You can build a brain in silicon. It's possible. 350 00:28:49,160 --> 00:28:52,360 But the very first AGIs will be 351 00:28:52,360 --> 00:28:55,440 basically very, very large data centres, 352 00:28:55,440 --> 00:28:59,200 packed with specialised neural network processors, 353 00:28:59,200 --> 00:29:00,520 working in parallel. 354 00:29:02,040 --> 00:29:05,520 A compact, hot, power-hungry package. 355 00:29:06,760 --> 00:29:10,000 Consuming, like, 10 million homes' worth of energy. 356 00:29:27,200 --> 00:29:29,760 Do you have a... a roast beef sandwich? 357 00:29:29,760 --> 00:29:32,560 Roast beef sandwich. Yeah. Something slightly different. 358 00:29:32,560 --> 00:29:33,920 Yes. Just this once. 359 00:29:38,720 --> 00:29:41,120 But even the very first AGIs 360 00:29:41,120 --> 00:29:43,840 will be dramatically more capable than humans. 361 00:29:45,720 --> 00:29:47,960 Humans will no longer be economically useful 362 00:29:47,960 --> 00:29:49,120 for nearly any task. 363 00:29:50,880 --> 00:29:53,880 Why would you want to hire a human if you could just get a computer 364 00:29:53,880 --> 00:29:56,640 that's going to do it much better and much more cheaply? 365 00:30:03,520 --> 00:30:06,680 AGI is going to be, like, without question, 366 00:30:06,680 --> 00:30:10,120 the most important technology in the history of the planet 367 00:30:10,120 --> 00:30:11,520 by a huge margin. 368 00:30:14,120 --> 00:30:16,240 It's going to be bigger than electricity, 369 00:30:16,240 --> 00:30:18,720 nuclear and Internet combined. 370 00:30:20,440 --> 00:30:23,240 In fact, you could say that the whole purpose of all human science, 371 00:30:23,240 --> 00:30:25,120 the purpose of computer science, the endgame, 372 00:30:25,120 --> 00:30:28,880 this is the endgame, to build this. And it's going to be built. 373 00:30:28,880 --> 00:30:31,640 It's going to be a new life form. It's going to be... 374 00:30:33,640 --> 00:30:35,280 It's going to make us obsolete. 375 00:30:40,280 --> 00:30:43,880 TELEGRAPH BEEPS IN MORSE CODE 376 00:30:43,880 --> 00:30:45,800 RADIO: We got programmes to go down here, 377 00:30:45,800 --> 00:30:47,840 where we got that AK-47 fire, over. 378 00:30:47,840 --> 00:30:50,960 INDISTINCT RADIO CHATTER 379 00:30:56,720 --> 00:31:00,320 European manufacturers know the Americans have invested heavily 380 00:31:00,320 --> 00:31:02,000 in the necessary hardware... 381 00:31:02,000 --> 00:31:04,440 ..to step into a brave new world of power, 382 00:31:04,440 --> 00:31:06,680 performance and productivity. 383 00:31:06,680 --> 00:31:10,160 All of the images you are about to see on the large screen 384 00:31:10,160 --> 00:31:14,120 will be generated by what's in that Macintosh. 385 00:31:14,120 --> 00:31:16,920 It's my honour and privilege to introduce to you 386 00:31:16,920 --> 00:31:18,800 the Windows 95 development team. 387 00:31:20,400 --> 00:31:22,880 Human physical labour has been mostly obsolete 388 00:31:22,880 --> 00:31:25,160 for getting on for a century. 389 00:31:26,200 --> 00:31:30,040 Routine human mental labour is rapidly becoming obsolete. 390 00:31:30,040 --> 00:31:34,000 And that's why we're seeing a lot of middle class jobs disappearing. 391 00:31:35,480 --> 00:31:37,240 STEVE JOBS: Every once in a while, 392 00:31:37,240 --> 00:31:40,920 a revolutionary product comes along that changes everything. 393 00:31:40,920 --> 00:31:43,720 Today, Apple is reinventing the phone. 394 00:31:43,720 --> 00:31:46,360 EXCITED SCREAMS 395 00:31:55,520 --> 00:31:58,960 Machine intelligence is already all around us. 396 00:31:58,960 --> 00:32:01,880 The list of things that we humans can do better than machines 397 00:32:01,880 --> 00:32:04,360 is actually shrinking pretty fast. 398 00:32:10,480 --> 00:32:14,720 Driverless cars are great, they probably will reduce accidents, 399 00:32:14,720 --> 00:32:18,480 except alongside with that in the United States, 400 00:32:18,480 --> 00:32:20,920 you're going to lose 10 million jobs. 401 00:32:20,920 --> 00:32:24,560 What are you going to do with 10 million unemployed people? 402 00:32:29,520 --> 00:32:33,200 The risk for social conflict and tensions 403 00:32:33,200 --> 00:32:36,920 if you exacerbate inequalities, is very, very high. 404 00:32:46,280 --> 00:32:51,320 AGI can, by definition, do all jobs better than we can do. 405 00:32:51,320 --> 00:32:52,520 People who are saying, 406 00:32:52,520 --> 00:32:55,480 "Ah, there'll always be jobs that humans can do better than machines," 407 00:32:55,480 --> 00:32:58,880 are simply betting against science in saying there'll never be AGI. 408 00:33:05,480 --> 00:33:06,960 What we're seeing now 409 00:33:06,960 --> 00:33:11,760 is like a train hurtling down a dark tunnel at breakneck speed. 410 00:33:11,760 --> 00:33:14,040 And it looks like we're sleeping at the wheel. 411 00:33:53,800 --> 00:33:56,200 BUTTONS BEEP 412 00:34:03,760 --> 00:34:07,520 A large fraction of the digital footprints we are leaving behind 413 00:34:07,520 --> 00:34:09,240 are digital images. 414 00:34:10,400 --> 00:34:13,760 And specifically what's really interesting to me as a psychologist 415 00:34:13,760 --> 00:34:16,000 are digital images of our faces. 416 00:34:19,240 --> 00:34:21,920 Here you can see the difference in the facial outline 417 00:34:21,920 --> 00:34:24,360 of an average gay and average straight face. 418 00:34:24,360 --> 00:34:30,200 And you can see that straight men have slightly broader jaws. 419 00:34:30,200 --> 00:34:33,320 Gay women have slightly larger jaws, 420 00:34:33,320 --> 00:34:34,920 compared with straight women. 421 00:34:37,120 --> 00:34:39,200 Computer algorithms can reveal 422 00:34:39,200 --> 00:34:41,560 our political views or sexual orientation 423 00:34:41,560 --> 00:34:45,600 or intelligence, just based on the picture of our faces. 424 00:34:46,800 --> 00:34:50,120 Even a human brain can distinguish between gay and straight men 425 00:34:50,120 --> 00:34:51,520 with some accuracy. 426 00:34:51,520 --> 00:34:52,800 Again, now, it turns out 427 00:34:52,800 --> 00:34:56,280 that the computer can do it with much higher accuracy. 428 00:34:56,280 --> 00:34:59,440 What you're seeing here is an accuracy 429 00:34:59,440 --> 00:35:03,040 of off-the-shelf facial recognition software. 430 00:35:04,120 --> 00:35:09,120 This is terrible news for gay men and women all around the world. 431 00:35:09,120 --> 00:35:10,520 And not only gay men and women, 432 00:35:10,520 --> 00:35:12,560 because the same algorithms can be used to detect 433 00:35:12,560 --> 00:35:14,240 other intimate traits. 434 00:35:14,240 --> 00:35:17,120 Think being a member of the opposition, 435 00:35:17,120 --> 00:35:19,880 or being a liberal, or being an atheist. 436 00:35:21,400 --> 00:35:24,680 Being an atheist is also punishable by death 437 00:35:24,680 --> 00:35:27,080 in Saudi Arabia, for instance. 438 00:35:34,480 --> 00:35:38,000 My mission as an academic is to warn people about the dangers 439 00:35:38,000 --> 00:35:43,040 of algorithms being able to reveal our intimate traits. 440 00:35:44,520 --> 00:35:48,800 The problem is that when people receive bad news, 441 00:35:48,800 --> 00:35:51,200 they very often choose to dismiss them. 442 00:35:52,600 --> 00:35:55,160 Well, it's a bit scary when you start receiving 443 00:35:55,160 --> 00:35:57,040 death threats from one day to another. 444 00:35:57,040 --> 00:35:59,360 And I received quite a few death threats. 445 00:36:00,480 --> 00:36:02,240 But as a scientist, 446 00:36:02,240 --> 00:36:05,480 I have to basically show what is possible. 447 00:36:08,280 --> 00:36:11,760 So what I'm really interested in now is to try to see 448 00:36:11,760 --> 00:36:15,720 whether we can predict other traits from people's faces. 449 00:36:21,000 --> 00:36:26,360 Now, if you can detect depression from face or suicidal thoughts, 450 00:36:26,360 --> 00:36:31,840 maybe a CCTV system on the train station can save some lives. 451 00:36:33,800 --> 00:36:37,040 What if we could predict that someone is more prone 452 00:36:37,040 --> 00:36:38,320 to commit a crime? 453 00:36:39,680 --> 00:36:42,640 You probably had a school counsellor, a psychologist 454 00:36:42,640 --> 00:36:46,160 hired there to identify children that potentially 455 00:36:46,160 --> 00:36:49,680 may have some behavioural problems. 456 00:36:52,080 --> 00:36:55,280 So now imagine if you could predict with high accuracy that someone 457 00:36:55,280 --> 00:36:58,360 is likely to commit a crime in the future from the language used, 458 00:36:58,360 --> 00:37:00,720 from their face, from their facial expressions, 459 00:37:00,720 --> 00:37:02,360 from the likes on Facebook. 460 00:37:06,960 --> 00:37:08,720 I'm not developing new methods, 461 00:37:08,720 --> 00:37:11,760 I'm just describing something or testing something 462 00:37:11,760 --> 00:37:13,640 in an academic environment. 463 00:37:15,200 --> 00:37:20,480 But there obviously is a chance that while warning people against 464 00:37:20,480 --> 00:37:22,680 risks of new technologies, 465 00:37:22,680 --> 00:37:25,040 I may also give some people new ideas. 466 00:37:45,800 --> 00:37:49,000 We haven't yet seen the future in terms of the ways 467 00:37:49,000 --> 00:37:56,080 in which the new data-driven society is going to really evolve. 468 00:37:58,240 --> 00:38:02,480 The tech companies want to get every possible bit of information 469 00:38:02,480 --> 00:38:06,480 that they can collect on everyone to facilitate business. 470 00:38:08,200 --> 00:38:11,680 The police and the military want to do the same thing 471 00:38:11,680 --> 00:38:13,560 to facilitate security. 472 00:38:16,560 --> 00:38:21,120 The interests that the two have in common are immense, 473 00:38:21,120 --> 00:38:26,200 and so the extent of collaboration between what you might call 474 00:38:26,200 --> 00:38:31,240 a military tech complex is growing dramatically. 475 00:38:35,200 --> 00:38:39,600 The CIA, for a very long time, has maintained a close connection 476 00:38:39,600 --> 00:38:41,840 with Silicon Valley. 477 00:38:41,840 --> 00:38:44,960 Their venture capital firm, known as In-Q-Tel, 478 00:38:44,960 --> 00:38:47,680 makes seed investments to start-up companies 479 00:38:47,680 --> 00:38:52,120 developing breakthrough technology that the CIA hopes to deploy. 480 00:38:53,440 --> 00:38:56,760 Palantir, the big data analytics firm, 481 00:38:56,760 --> 00:38:59,640 one of their first seed investments was from In-Q-Tel. 482 00:39:03,480 --> 00:39:06,600 In-Q-Tel struck gold in Palantir 483 00:39:06,600 --> 00:39:10,760 in helping to create a private vendor 484 00:39:10,760 --> 00:39:16,480 that has intelligence and artificial intelligence capabilities 485 00:39:16,480 --> 00:39:19,000 that the government can't even compete with. 486 00:39:21,000 --> 00:39:22,320 Good evening. 487 00:39:22,320 --> 00:39:24,440 I'm Peter Thiel. 488 00:39:24,440 --> 00:39:28,760 I'm not a politician, but neither is Donald Trump. 489 00:39:28,760 --> 00:39:30,720 He is a builder. 490 00:39:30,720 --> 00:39:33,680 And it's time to rebuild America. 491 00:39:33,680 --> 00:39:35,960 CHEERING AND APPLAUSE 492 00:39:35,960 --> 00:39:38,520 Peter Thiel, the founder of Palantir, 493 00:39:38,520 --> 00:39:41,120 was a Donald Trump transition adviser 494 00:39:41,120 --> 00:39:43,800 and a close friend and donor. 495 00:39:45,800 --> 00:39:49,440 Trump was elected largely on the promise to deport 496 00:39:49,440 --> 00:39:52,280 millions of immigrants. 497 00:39:52,280 --> 00:39:56,840 The only way you can do that is with a lot of intelligence - 498 00:39:56,840 --> 00:39:59,720 and that's where Palantir comes in. 499 00:40:03,440 --> 00:40:09,280 They ingest huge troves of data which include where you live, 500 00:40:09,280 --> 00:40:13,640 where you work, who you know, who your neighbours are, 501 00:40:13,640 --> 00:40:19,000 who your family is, where you have visited, where you stay, 502 00:40:19,000 --> 00:40:20,920 your social media profile. 503 00:40:23,960 --> 00:40:29,440 Palantir gets all of that and is remarkably good at structuring it 504 00:40:29,440 --> 00:40:35,120 in a way that helps law enforcement, immigration authorities 505 00:40:35,120 --> 00:40:40,800 or intelligence agencies of any kind track you, find you 506 00:40:40,800 --> 00:40:44,120 and learn everything there is to know about you. 507 00:41:23,120 --> 00:41:26,640 We're putting AI in charge now of ever more important decisions 508 00:41:26,640 --> 00:41:29,600 that affect people's lives. 509 00:41:29,600 --> 00:41:32,880 Old-school AI used to have its intelligence programmed in 510 00:41:32,880 --> 00:41:35,200 by humans who understood how it worked. 511 00:41:35,200 --> 00:41:38,840 But today, powerful AI systems have just learned for themselves, 512 00:41:38,840 --> 00:41:42,200 and we have no clue, really, how they work, 513 00:41:42,200 --> 00:41:44,240 which makes it really hard to trust them. 514 00:41:49,080 --> 00:41:51,560 This isn't some futuristic technology. 515 00:41:51,560 --> 00:41:52,600 This is now. 516 00:41:54,280 --> 00:41:59,240 AI might help determine where a fire department is built in a community 517 00:41:59,240 --> 00:42:00,600 or where a school is built. 518 00:42:00,600 --> 00:42:05,440 It might decide whether you get bail or whether you stay in jail. 519 00:42:05,440 --> 00:42:07,760 It might decide where the police are going to be. 520 00:42:07,760 --> 00:42:09,880 It might decide whether you're going to be 521 00:42:09,880 --> 00:42:11,840 under additional police scrutiny. 522 00:42:18,040 --> 00:42:21,640 It's popular now in the US to do predictive policing. 523 00:42:21,640 --> 00:42:24,320 So what they do is they use an algorithm to figure out 524 00:42:24,320 --> 00:42:26,480 where crime will be... 525 00:42:26,480 --> 00:42:28,000 ..and then they use that to tell 526 00:42:28,000 --> 00:42:29,920 where we should send police officers. 527 00:42:31,160 --> 00:42:35,240 So, that's based on a measurement of crime rate, 528 00:42:35,240 --> 00:42:37,000 so we know that there is bias. 529 00:42:37,000 --> 00:42:40,200 Black people and Hispanic people are pulled over or stopped 530 00:42:40,200 --> 00:42:42,800 by the police officers more frequently than white people are, 531 00:42:42,800 --> 00:42:44,480 so we have this biased data going in. 532 00:42:44,480 --> 00:42:46,520 And then what happens is you use that to say, 533 00:42:46,520 --> 00:42:48,240 "Oh, here's where the cops should go." 534 00:42:48,240 --> 00:42:52,360 Well, the cops go to those neighbourhoods - and guess what they do? They arrest people. 535 00:42:52,360 --> 00:42:55,680 And then it feeds back biased data into the system. 536 00:42:55,680 --> 00:42:57,640 And that's called a feedback loop. 537 00:43:10,720 --> 00:43:16,320 Predictive policing leads, at the extremes, 538 00:43:16,320 --> 00:43:21,680 to experts saying, "Show me your baby and I will tell you 539 00:43:21,680 --> 00:43:23,760 "whether she's going to be a criminal." 540 00:43:25,840 --> 00:43:27,440 Now that we can predict it, 541 00:43:27,440 --> 00:43:33,840 we're going to then surveil those kids much more closely 542 00:43:33,840 --> 00:43:38,440 and we're going to jump on them at the first sign of a problem. 543 00:43:38,440 --> 00:43:41,800 And that's going to make for more effective policing. 544 00:43:41,800 --> 00:43:45,880 It does, but it's going to make for a really grim society, 545 00:43:45,880 --> 00:43:50,960 and it's reinforcing - dramatically - existing injustices. 546 00:43:57,480 --> 00:44:02,400 Imagine a world in which networks of CCTV cameras, 547 00:44:02,400 --> 00:44:04,280 drone surveillance cameras, 548 00:44:04,280 --> 00:44:08,280 have sophisticated face recognition technologies 549 00:44:08,280 --> 00:44:11,520 and are connected to other government surveillance databases. 550 00:44:12,720 --> 00:44:17,520 We will have the technology in place to have all of our movements 551 00:44:17,520 --> 00:44:20,440 comprehensively tracked and recorded. 552 00:44:22,400 --> 00:44:25,840 What that also means is that we will have created 553 00:44:25,840 --> 00:44:29,520 a surveillance time machine that will allow governments 554 00:44:29,520 --> 00:44:33,440 and powerful corporations to essentially hit rewind on our lives. 555 00:44:33,440 --> 00:44:36,480 We might not be under any suspicion now. 556 00:44:36,480 --> 00:44:39,760 And five years from now, they might want to know more about us 557 00:44:39,760 --> 00:44:43,520 and can then recreate - granularly - everything we've done, 558 00:44:43,520 --> 00:44:46,080 everyone we've seen, everyone we've been around 559 00:44:46,080 --> 00:44:47,840 over that entire period. 560 00:44:50,520 --> 00:44:53,920 That's an extraordinary amount of power 561 00:44:53,920 --> 00:44:56,200 for us to cede to anyone. 562 00:44:57,600 --> 00:45:00,040 And it's a world that I think has been difficult 563 00:45:00,040 --> 00:45:02,440 for people to imagine. 564 00:45:02,440 --> 00:45:06,680 But we've already built the architecture to enable that. 565 00:45:41,840 --> 00:45:45,720 I'm a political reporter, and I'm very interested in the ways 566 00:45:45,720 --> 00:45:50,120 powerful industries use their political power to influence 567 00:45:50,120 --> 00:45:51,800 the public policy process. 568 00:45:55,720 --> 00:45:59,200 The large tech companies and their lobbyists get together 569 00:45:59,200 --> 00:46:02,120 behind closed doors and are able to craft policies 570 00:46:02,120 --> 00:46:03,760 that we all have to live under. 571 00:46:05,520 --> 00:46:07,800 That's true for surveillance policies, 572 00:46:07,800 --> 00:46:10,320 for policies in terms of data collection, 573 00:46:10,320 --> 00:46:12,920 but also increasingly important when it comes to 574 00:46:12,920 --> 00:46:15,200 military and foreign policy. 575 00:46:18,720 --> 00:46:20,800 Starting in 2016, 576 00:46:20,800 --> 00:46:24,880 the Defense Department formed the Defense Innovation Board. 577 00:46:24,880 --> 00:46:28,760 That's a special body created to bring top tech executives 578 00:46:28,760 --> 00:46:30,960 into closer contact with the military. 579 00:46:33,920 --> 00:46:36,520 Eric Schmidt, former chairman of Alphabet, 580 00:46:36,520 --> 00:46:38,040 the parent company of Google, 581 00:46:38,040 --> 00:46:41,880 became the chairman of the Defense Innovation Board. 582 00:46:41,880 --> 00:46:44,720 And one of their first priorities was to say, 583 00:46:44,720 --> 00:46:46,520 "We need more artificial intelligence 584 00:46:46,520 --> 00:46:48,520 "integrated into the military." 585 00:46:50,640 --> 00:46:53,600 I've worked with a group of volunteers 586 00:46:53,600 --> 00:46:56,360 over the last couple of years to take a look at innovation 587 00:46:56,360 --> 00:46:58,360 in the overall military. 588 00:46:58,360 --> 00:47:02,760 And my summary conclusion is that we have fantastic people 589 00:47:02,760 --> 00:47:05,160 who are trapped in a very bad system. 590 00:47:07,960 --> 00:47:10,160 From the Department of Defense's perspective, 591 00:47:10,160 --> 00:47:12,320 where I really started to get interested in it, 592 00:47:12,320 --> 00:47:15,400 when we started to think about unmanned systems 593 00:47:15,400 --> 00:47:20,880 and how robotic and unmanned systems would start to change war. 594 00:47:20,880 --> 00:47:24,920 The smarter you made the unmanned systems and robots, 595 00:47:24,920 --> 00:47:28,320 the more powerful you might be able to make your military. 596 00:47:30,240 --> 00:47:32,880 Under Secretary of Defense Robert Work 597 00:47:32,880 --> 00:47:35,120 put together a major memo known as 598 00:47:35,120 --> 00:47:38,240 the Algorithmic Warfare Cross-Functional Team, 599 00:47:38,240 --> 00:47:40,080 better known as Project Maven. 600 00:47:42,400 --> 00:47:44,280 Eric Schmidt gave a number of speeches 601 00:47:44,280 --> 00:47:48,720 and media appearances where he said this effort was designed to increase 602 00:47:48,720 --> 00:47:52,880 fuel efficiency in the Air Force to help with the logistics. 603 00:47:52,880 --> 00:47:55,920 But behind closed doors, there was another parallel effort. 604 00:48:01,400 --> 00:48:04,400 Late in 2017, as part of Project Maven, 605 00:48:04,400 --> 00:48:06,320 Google, Eric Schmidt's firm, 606 00:48:06,320 --> 00:48:10,720 was tapped to secretly work on another part of Project Maven, 607 00:48:10,720 --> 00:48:15,320 and that was to take the vast volumes of image data 608 00:48:15,320 --> 00:48:19,880 vacuumed up by drones operating in Iraq and Afghanistan 609 00:48:19,880 --> 00:48:24,400 and to teach an AI to quickly identify targets on the battlefield. 610 00:48:27,320 --> 00:48:31,360 We have a sensor and the sensor can do full motion video 611 00:48:31,360 --> 00:48:33,120 of an entire city. 612 00:48:33,120 --> 00:48:37,760 And we would have three seven-person teams working constantly 613 00:48:37,760 --> 00:48:41,000 and they could process 15% of the information. 614 00:48:41,000 --> 00:48:43,680 The other 85% of the information was just left 615 00:48:43,680 --> 00:48:45,800 on the cutting room floor. 616 00:48:45,800 --> 00:48:49,960 So we said, "Hey, AI and machine learning would help us process 617 00:48:49,960 --> 00:48:52,160 "100% of the information." 618 00:49:00,120 --> 00:49:03,480 Google has long had the motto, "Don't be evil." 619 00:49:03,480 --> 00:49:06,920 They have created a public image that they are devoted 620 00:49:06,920 --> 00:49:09,840 to public transparency. 621 00:49:09,840 --> 00:49:13,760 But for Google to slowly transform into a defence contractor, 622 00:49:13,760 --> 00:49:16,520 they maintained the utmost secrecy. 623 00:49:16,520 --> 00:49:19,960 You had Google entering into this contract with most of the employees, 624 00:49:19,960 --> 00:49:22,040 even employees who were working on the programme, 625 00:49:22,040 --> 00:49:23,800 completely left in the dark. 626 00:49:36,480 --> 00:49:38,080 Usually within Google, 627 00:49:38,080 --> 00:49:40,760 anyone in the company is allowed to know 628 00:49:40,760 --> 00:49:43,520 about any other project that's happening in some other part 629 00:49:43,520 --> 00:49:44,600 of the company. 630 00:49:45,840 --> 00:49:49,040 With Project Maven, the fact that it was kept secret 631 00:49:49,040 --> 00:49:51,240 I think was kind of alarming to people, 632 00:49:51,240 --> 00:49:53,240 cos that's not the norm at Google. 633 00:49:55,320 --> 00:49:57,760 When the story was first revealed, 634 00:49:57,760 --> 00:50:00,160 it started off a firestorm within Google. 635 00:50:00,160 --> 00:50:03,160 You had a number of employees quitting in protest, 636 00:50:03,160 --> 00:50:06,600 others signing a petition objecting to this work. 637 00:50:08,640 --> 00:50:10,200 You have to really say, like, 638 00:50:10,200 --> 00:50:12,240 "I don't want to be part of this any more." 639 00:50:13,640 --> 00:50:16,440 There are companies called defence contractors 640 00:50:16,440 --> 00:50:20,600 and Google should just not be one of those companies, 641 00:50:20,600 --> 00:50:24,880 because people need to trust Google for Google to work. 642 00:50:26,760 --> 00:50:28,160 Good morning. 643 00:50:28,160 --> 00:50:29,720 Welcome to Google I/O. 644 00:50:29,720 --> 00:50:31,800 CHEERING 645 00:50:31,800 --> 00:50:34,520 We've seen emails that show that Google 646 00:50:34,520 --> 00:50:37,360 simply continued to mislead their employees, 647 00:50:37,360 --> 00:50:41,080 that the drone targeting programme was only a minor effort 648 00:50:41,080 --> 00:50:44,760 that could at most be worth $9 million to the firm, which is, 649 00:50:44,760 --> 00:50:49,000 you know, drops in the bucket for a gigantic company like Google. 650 00:50:49,000 --> 00:50:52,000 But from internal emails that we obtained, 651 00:50:52,000 --> 00:50:55,920 Google was expecting Project Maven would ramp up to 652 00:50:55,920 --> 00:51:00,600 as much as $250 million and that this entire effort 653 00:51:00,600 --> 00:51:04,160 would provide Google with special Defence Department certification 654 00:51:04,160 --> 00:51:07,320 to make them available for even bigger defence contracts, 655 00:51:07,320 --> 00:51:09,200 some worth as much as $10 billion. 656 00:51:20,440 --> 00:51:25,040 The pressure for Google to compete for military contracts has come 657 00:51:25,040 --> 00:51:28,320 at a time when its competitors are also shifting their culture. 658 00:51:30,600 --> 00:51:34,680 Amazon similarly pitching the military and law enforcement. 659 00:51:34,680 --> 00:51:36,800 IBM and other leading firms, 660 00:51:36,800 --> 00:51:40,600 they're pitching law enforcement and military. 661 00:51:40,600 --> 00:51:44,040 To stay competitive, Google has slowly transformed. 662 00:51:49,880 --> 00:51:55,120 The Defense Science Board said of all of the technological advances 663 00:51:55,120 --> 00:51:56,920 that are happening right now, 664 00:51:56,920 --> 00:52:01,680 the single most important thing was artificial intelligence 665 00:52:01,680 --> 00:52:05,400 and the autonomous operations that it would lead. 666 00:52:05,400 --> 00:52:07,120 Are we investing enough? 667 00:52:12,520 --> 00:52:19,840 Once we develop what are known as autonomous lethal weapons - 668 00:52:19,840 --> 00:52:23,480 in other words, weapons that are not controlled at all, 669 00:52:23,480 --> 00:52:26,280 they are genuinely autonomous - 670 00:52:26,280 --> 00:52:28,680 you've only got to get a president who says, 671 00:52:28,680 --> 00:52:31,240 "The hell with international law. We've got these weapons. 672 00:52:31,240 --> 00:52:33,320 "We're going to do what we want with them." 673 00:52:37,040 --> 00:52:38,440 We're very close. 674 00:52:38,440 --> 00:52:40,920 When you have the hardware already set up 675 00:52:40,920 --> 00:52:43,120 and all you have to do is flip a switch 676 00:52:43,120 --> 00:52:45,040 to make it fully autonomous, 677 00:52:45,040 --> 00:52:47,480 what is it there that's stopping you from doing that? 678 00:52:49,360 --> 00:52:54,600 There's something really to be feared in war machine speed. 679 00:52:54,600 --> 00:52:57,560 What if you're a machine and you've run millions and millions 680 00:52:57,560 --> 00:53:00,800 of different war scenarios and you have a team of drones 681 00:53:00,800 --> 00:53:02,720 and you've delegated control to half of them, 682 00:53:02,720 --> 00:53:05,440 and you're collaborating in real time? 683 00:53:05,440 --> 00:53:08,920 What happens when that swarm of drones is tasked 684 00:53:08,920 --> 00:53:10,440 with engaging a city? 685 00:53:12,160 --> 00:53:14,560 How will they take over that city? 686 00:53:14,560 --> 00:53:17,520 The answer is, we won't know until it happens. 687 00:53:24,640 --> 00:53:30,760 We do not want an AI system to decide what human it would attack, 688 00:53:30,760 --> 00:53:34,240 but we're going up against authoritarian competitors. 689 00:53:34,240 --> 00:53:39,120 So in my view, an authoritarian regime will have less problem 690 00:53:39,120 --> 00:53:44,120 delegating authority to a machine to make lethal decisions. 691 00:53:44,120 --> 00:53:47,480 So how that plays out remains to be seen. 692 00:54:11,960 --> 00:54:16,440 Almost the gift of AI now is that it will force us collectively 693 00:54:16,440 --> 00:54:18,880 to think through, at a very basic level, 694 00:54:18,880 --> 00:54:20,960 what does it mean to be human? 695 00:54:23,160 --> 00:54:25,040 What do I do as a human better 696 00:54:25,040 --> 00:54:27,840 than a certain super smart machine can do? 697 00:54:31,040 --> 00:54:35,680 First, we create our technology, and then it recreates us. 698 00:54:35,680 --> 00:54:40,320 We need to make sure that we don't miss some of the things 699 00:54:40,320 --> 00:54:42,240 that make us so beautifully human. 700 00:54:46,360 --> 00:54:49,000 Once we build intelligent machines, 701 00:54:49,000 --> 00:54:51,440 the philosophical vocabulary we have available 702 00:54:51,440 --> 00:54:55,360 to think about ourselves as human increasingly fails us. 703 00:54:57,920 --> 00:55:01,200 If I ask you to write up a list of all the terms you have available 704 00:55:01,200 --> 00:55:05,680 to describe yourself as human, there are not so many terms - 705 00:55:05,680 --> 00:55:11,400 culture, history, sociality, maybe politics, 706 00:55:11,400 --> 00:55:15,040 civilisation, subjectivity. 707 00:55:17,000 --> 00:55:20,240 All of these terms ground in two positions - 708 00:55:20,240 --> 00:55:23,640 that humans are more than mere animals 709 00:55:23,640 --> 00:55:26,440 and that humans are more than mere machines. 710 00:55:29,560 --> 00:55:32,360 But if machines truly think, 711 00:55:32,360 --> 00:55:36,360 there is a large set of key philosophical questions 712 00:55:36,360 --> 00:55:38,040 in which what is at stake is... 713 00:55:39,480 --> 00:55:41,920 ..who are we? What's our place in the world? 714 00:55:41,920 --> 00:55:44,040 What is the world? How is it structured? 715 00:55:44,040 --> 00:55:46,920 Do the categories that we have relied on, 716 00:55:46,920 --> 00:55:49,240 do they still work? Are they wrong? 717 00:55:54,240 --> 00:55:56,920 Many people think of intelligence as something mysterious 718 00:55:56,920 --> 00:56:00,800 that can only exist inside of biological organisms like us. 719 00:56:00,800 --> 00:56:05,040 But intelligence is all about information processing. 720 00:56:05,040 --> 00:56:07,120 It doesn't matter whether the intelligence is processed 721 00:56:07,120 --> 00:56:10,560 by carbon atoms inside of cells in brains in people 722 00:56:10,560 --> 00:56:12,920 or by silicon atoms in computers. 723 00:56:15,440 --> 00:56:18,840 Part of the success of AI recently has come from stealing 724 00:56:18,840 --> 00:56:22,240 great ideas from evolution. 725 00:56:22,240 --> 00:56:24,160 We noticed that the brain, for example, 726 00:56:24,160 --> 00:56:27,720 has all these neurons inside connected in complicated ways. 727 00:56:27,720 --> 00:56:30,360 So we stole that idea and abstracted it 728 00:56:30,360 --> 00:56:34,080 into artificial neural networks in computers. 729 00:56:34,080 --> 00:56:37,520 And that's what's revolutionised machine intelligence. 730 00:56:42,440 --> 00:56:44,840 If we one day get artificial general intelligence, 731 00:56:44,840 --> 00:56:47,760 then by definition, AI can also do better 732 00:56:47,760 --> 00:56:49,880 the job of AI programming. 733 00:56:51,760 --> 00:56:54,880 And that means that further progress in making AI will be dominated 734 00:56:54,880 --> 00:56:57,600 not by human programmers, but by AI. 735 00:57:00,240 --> 00:57:05,320 Recursively self-improving AI could leave human intelligence far behind, 736 00:57:05,320 --> 00:57:07,480 creating superintelligence. 737 00:57:09,680 --> 00:57:12,720 It's going to be the last invention we ever need to make, 738 00:57:12,720 --> 00:57:15,280 because it can then invent everything else 739 00:57:15,280 --> 00:57:17,640 much faster than we could. 740 00:57:17,640 --> 00:57:19,680 MACHINE MURMURS AND STUTTERS 741 00:57:32,160 --> 00:57:34,160 MACHINE SPEAKS INDISTINCTLY 742 00:57:59,960 --> 00:58:01,960 MACHINE SPEEDS UP 743 00:58:29,160 --> 00:58:33,880 There is a future that we all need to talk about. 744 00:58:33,880 --> 00:58:36,960 Some of the fundamental questions about the future 745 00:58:36,960 --> 00:58:41,080 of artificial intelligence, not just where it's going, 746 00:58:41,080 --> 00:58:44,320 but what it means for society to go there. 747 00:58:45,440 --> 00:58:49,360 It is not what computers can do, 748 00:58:49,360 --> 00:58:52,200 but what computers should do. 749 00:58:52,200 --> 00:58:56,840 As the generation of people that is bringing AI to the future, 750 00:58:56,840 --> 00:59:00,840 we are the generation that will answer this question 751 00:59:00,840 --> 00:59:02,760 first and foremost. 752 00:59:09,280 --> 00:59:11,960 We haven't created the human-level thinking machine yet, 753 00:59:11,960 --> 00:59:14,040 but we get closer and closer. 754 00:59:15,880 --> 00:59:19,400 Maybe we'll get to human-level AI in five years from now, 755 00:59:19,400 --> 00:59:22,280 or maybe it'll take 50 or 100 years from now. 756 00:59:22,280 --> 00:59:23,600 It almost doesn't matter. 757 00:59:23,600 --> 00:59:26,120 Like, these are all really, really soon 758 00:59:26,120 --> 00:59:30,480 in terms of the overall history of humanity. 759 00:59:30,480 --> 00:59:32,560 AI SINGS OPERATICALLY 760 00:59:35,800 --> 00:59:36,840 Very nice. 761 00:59:50,480 --> 00:59:54,040 So, the AI field is extremely international. 762 00:59:54,040 --> 00:59:56,120 China is up-and-coming 763 00:59:56,120 --> 01:00:00,040 and it's starting to rival the US, Europe and Japan 764 01:00:00,040 --> 01:00:04,360 in terms of putting a lot of processing power behind AI 765 01:00:04,360 --> 01:00:07,760 and gathering a lot of data to help AI learn. 766 01:00:10,720 --> 01:00:15,080 We have a young generation of Chinese researchers now. 767 01:00:15,080 --> 01:00:18,320 Nobody knows where the next revolution is going to come from. 768 01:00:24,880 --> 01:00:28,520 China always wants to become a superpower in the world. 769 01:00:30,680 --> 01:00:33,160 The Chinese government thinks AI gave them the chance 770 01:00:33,160 --> 01:00:38,840 to become one of the most advanced technology-wise, business-wise. 771 01:00:38,840 --> 01:00:42,160 So the Chinese government look at this as a huge opportunity. 772 01:00:43,880 --> 01:00:48,480 Like, they raise a flag and said, "That's good field. 773 01:00:48,480 --> 01:00:50,920 "The companies should jump into it." 774 01:00:50,920 --> 01:00:52,840 Then China's commercial world, as a company, 775 01:00:52,840 --> 01:00:55,400 just, "OK, come and raise the flag. That's good. 776 01:00:55,400 --> 01:00:57,200 "Let's put the money into it." 777 01:00:58,640 --> 01:01:02,800 Chinese tech giants like Baidu, like Tencent, like Alibaba, 778 01:01:02,800 --> 01:01:06,360 they put a lot of their investment into the AI field. 779 01:01:08,080 --> 01:01:11,200 So we see China's AI development is booming. 780 01:01:18,240 --> 01:01:21,960 In China, everybody has Alipay and WeChat Pay, 781 01:01:21,960 --> 01:01:25,520 so the mobile payment is everywhere. 782 01:01:25,520 --> 01:01:29,120 And with that, they can do, like, a lot of AI analysis 783 01:01:29,120 --> 01:01:31,760 to know, like, your spending habits, 784 01:01:31,760 --> 01:01:33,840 like, your credit rating. 785 01:01:35,880 --> 01:01:40,800 Face recognition technology is widely adopted in China, 786 01:01:40,800 --> 01:01:43,720 in the airport, in the train station. 787 01:01:43,720 --> 01:01:46,200 So in the future, maybe in just a few months, 788 01:01:46,200 --> 01:01:49,760 you don't need a paper ticket to board a train. 789 01:01:49,760 --> 01:01:50,920 Only your face. 790 01:01:59,440 --> 01:02:04,400 We generate the world's biggest platform of facial recognition. 791 01:02:05,880 --> 01:02:12,880 We have 300,000 developers using our platform. 792 01:02:12,880 --> 01:02:16,200 A lot of it is the selfies camera apps. 793 01:02:16,200 --> 01:02:18,800 It makes you look more beautiful. 794 01:02:21,280 --> 01:02:25,560 There is millions and millions of camera in the world. 795 01:02:25,560 --> 01:02:29,600 Each camera, from my point, is a data generator. 796 01:02:33,560 --> 01:02:37,480 In the machine's eye, your face would change into the features 797 01:02:37,480 --> 01:02:42,240 and it would turn your face into a paragraph of the code... 798 01:02:42,240 --> 01:02:43,800 ..so we can detect, 799 01:02:43,800 --> 01:02:48,160 "How old are you?", "Are you male or female?", and your emotions. 800 01:02:51,600 --> 01:02:54,880 This whole thing is about what kind of thing you are looking at. 801 01:02:54,880 --> 01:02:57,560 So then we can track your eyeballs, 802 01:02:57,560 --> 01:03:00,080 so if you are focusing on some product, 803 01:03:00,080 --> 01:03:04,440 we can track that so that we can know which kind of people 804 01:03:04,440 --> 01:03:06,320 like which kind of product. 805 01:04:20,480 --> 01:04:22,680 The Chinese government is using multiple 806 01:04:22,680 --> 01:04:24,400 different kinds of technologies, 807 01:04:24,400 --> 01:04:27,440 whether it's AI, whether it's big data platforms, 808 01:04:27,440 --> 01:04:29,560 facial recognition, voice recognition, 809 01:04:29,560 --> 01:04:33,560 essentially to monitor what the population is doing. 810 01:04:36,440 --> 01:04:40,400 I think the Chinese government has made very clear its intent 811 01:04:40,400 --> 01:04:46,000 to gather massive amounts of data about people to socially engineer 812 01:04:46,000 --> 01:04:47,920 a dissent-free society. 813 01:04:50,400 --> 01:04:55,000 The logic behind the Chinese government's social credit system, 814 01:04:55,000 --> 01:05:00,600 it's to take the idea that whether you are credit worthy 815 01:05:00,600 --> 01:05:06,080 for a financial loan and adding to it a very political dimension 816 01:05:06,080 --> 01:05:08,800 to say, are you a trustworthy human being? 817 01:05:10,760 --> 01:05:12,560 What you've said online, 818 01:05:12,560 --> 01:05:14,520 have you ever been critical of the authorities? 819 01:05:14,520 --> 01:05:16,280 Do you have a criminal record? 820 01:05:17,800 --> 01:05:21,200 And all of that information is packaged up together 821 01:05:21,200 --> 01:05:26,640 to rate you in ways that if you have performed well, in their view, 822 01:05:26,640 --> 01:05:31,960 you'll have easier access to certain kinds of state services or benefits, 823 01:05:31,960 --> 01:05:35,000 but that if you haven't done very well, you're going 824 01:05:35,000 --> 01:05:36,640 to be penalised or restricted. 825 01:05:40,560 --> 01:05:43,800 There's no way for people to challenge those designations, 826 01:05:43,800 --> 01:05:46,920 or in some cases even know that they've been put in that category. 827 01:05:46,920 --> 01:05:51,360 And it's not until they try to access some kind of state service 828 01:05:51,360 --> 01:05:54,440 or buy a plane ticket or get a passport or enrol their kids 829 01:05:54,440 --> 01:05:57,520 in school that they come to learn that they've been labelled 830 01:05:57,520 --> 01:06:01,320 in this way and that there are negative consequences for them 831 01:06:01,320 --> 01:06:03,120 as a result. 832 01:06:19,520 --> 01:06:23,400 We've spent the better part of the last one or two years 833 01:06:23,400 --> 01:06:27,760 looking at abuses of surveillance technology across China. 834 01:06:27,760 --> 01:06:30,800 And a lot of that work has taken us to Xinjiang... 835 01:06:32,160 --> 01:06:36,680 ..the northwestern region of China that has a more than half population 836 01:06:36,680 --> 01:06:41,120 of Turkic Muslims, Uyghurs, Kazakhs, Hui. 837 01:06:42,840 --> 01:06:45,800 This is a region and a population the Chinese government has 838 01:06:45,800 --> 01:06:49,120 long considered to be politically suspect or disloyal. 839 01:06:52,240 --> 01:06:54,920 We came to find information about what's called 840 01:06:54,920 --> 01:06:57,480 the Integrated Joint Operations Platform, 841 01:06:57,480 --> 01:06:59,880 which is a predictive policing programme. 842 01:06:59,880 --> 01:07:04,720 And that's one of the programmes that has been spitting out lists 843 01:07:04,720 --> 01:07:07,800 of names of people to be subjected to political re-education. 844 01:07:13,600 --> 01:07:17,320 A number of our interviewees for the report we just released 845 01:07:17,320 --> 01:07:20,360 about the political education camps in Xinjiang 846 01:07:20,360 --> 01:07:24,840 just painted an extraordinary portrait of a surveillance state. 847 01:07:28,080 --> 01:07:30,240 A region awash in surveillance cameras 848 01:07:30,240 --> 01:07:33,160 for facial recognition purposes, 849 01:07:33,160 --> 01:07:38,600 checkpoints, body scanners, QR codes outside people's homes. 850 01:07:40,840 --> 01:07:45,240 It really is the stuff of dystopian movies that we've all gone to 851 01:07:45,240 --> 01:07:47,760 and thought, "Wow, that would be a creepy world to live in." 852 01:07:47,760 --> 01:07:51,680 Yeah, well, 13 million Turkic Muslims in China are living 853 01:07:51,680 --> 01:07:53,960 in that reality right now. 854 01:08:05,480 --> 01:08:08,280 The Intercept reports that Google is planning to launch 855 01:08:08,280 --> 01:08:11,080 a censored version of its search engine in China. 856 01:08:11,080 --> 01:08:13,720 Google's search for new markets leads it to China 857 01:08:13,720 --> 01:08:16,440 despite Beijing's rules on censorship. 858 01:08:16,440 --> 01:08:20,640 Tell us more about why you felt it was your ethical responsibility 859 01:08:20,640 --> 01:08:22,120 to resign, because you talk about 860 01:08:22,120 --> 01:08:25,800 being complicit in censorship and oppression and surveillance. 861 01:08:25,800 --> 01:08:29,280 There is a Chinese venture company that has to be set up 862 01:08:29,280 --> 01:08:31,120 for Google to operate in China. 863 01:08:31,120 --> 01:08:33,720 And the question is, to what degree did they get to control 864 01:08:33,720 --> 01:08:37,160 the blacklist and to what degree would they have just unfettered 865 01:08:37,160 --> 01:08:39,920 access to surveilling Chinese citizens? 866 01:08:39,920 --> 01:08:41,880 And the fact that Google refuses to respond 867 01:08:41,880 --> 01:08:43,720 to human rights organisations on this, 868 01:08:43,720 --> 01:08:46,160 I think, should be extremely disturbing to everyone. 869 01:08:51,120 --> 01:08:53,800 Due to my conviction that dissent is fundamental 870 01:08:53,800 --> 01:08:55,520 to functioning democracies, 871 01:08:55,520 --> 01:08:58,560 I am forced to resign in order to avoid contributing to 872 01:08:58,560 --> 01:09:02,960 or profiting from the erosion of protections for dissidents. 873 01:09:02,960 --> 01:09:05,360 The United Nations is currently reporting 874 01:09:05,360 --> 01:09:07,880 that between 200,000 and one million Uyghurs 875 01:09:07,880 --> 01:09:10,680 have been disappeared into re-education camps. 876 01:09:10,680 --> 01:09:13,840 And there is a serious argument that Google would be complicit, 877 01:09:13,840 --> 01:09:16,600 should it want to surveil the version of Search in China. 878 01:09:19,880 --> 01:09:25,680 Dragonfly is a project meant to launch Search in China 879 01:09:25,680 --> 01:09:27,880 under Chinese government regulations, 880 01:09:27,880 --> 01:09:31,680 which include censoring sensitive content, 881 01:09:31,680 --> 01:09:34,160 basic queries on human rights. 882 01:09:34,160 --> 01:09:37,600 Information about political representatives is blocked. 883 01:09:37,600 --> 01:09:41,480 Information about student protests is blocked, 884 01:09:41,480 --> 01:09:43,880 and that's one small part of it. 885 01:09:43,880 --> 01:09:46,400 Perhaps a deeper concern is the surveillance side of this. 886 01:09:50,720 --> 01:09:53,880 When I raised the issue with my managers, with my colleagues, 887 01:09:53,880 --> 01:09:55,680 there was a lot of concern. 888 01:09:55,680 --> 01:09:57,960 But everyone just said, "I don't know anything." 889 01:10:02,480 --> 01:10:06,520 And then when there was a meeting, finally, there was essentially 890 01:10:06,520 --> 01:10:09,600 no addressing the serious concerns associated with it. 891 01:10:11,640 --> 01:10:15,560 So then I filed my formal resignation, not just to my manager, 892 01:10:15,560 --> 01:10:17,600 but I actually distributed it company-wide. 893 01:10:17,600 --> 01:10:20,240 And that's the letter that I was reading from. 894 01:10:24,880 --> 01:10:27,440 Personally, I haven't slept well. 895 01:10:27,440 --> 01:10:30,000 I've had pretty horrific headaches. 896 01:10:30,000 --> 01:10:33,200 Wake up in the middle of the night just sweating. 897 01:10:34,560 --> 01:10:37,680 With that said, what I've found since speaking out 898 01:10:37,680 --> 01:10:42,400 is just how positive the global response to this has been. 899 01:10:45,080 --> 01:10:47,440 Engineers should demand to know 900 01:10:47,440 --> 01:10:50,280 what the uses of their technical contributions are 901 01:10:50,280 --> 01:10:53,320 and to have a seat at the table in those ethical decisions. 902 01:11:01,400 --> 01:11:04,080 Most citizens don't really understand what it means 903 01:11:04,080 --> 01:11:07,240 to be in a very large scale prescriptive technology, 904 01:11:07,240 --> 01:11:10,480 where someone has already pre-divided the work 905 01:11:10,480 --> 01:11:12,960 and all you know about is your little piece, 906 01:11:12,960 --> 01:11:16,000 and almost certainly you don't understand how it fits in. 907 01:11:19,000 --> 01:11:23,800 So I think it's worth drawing the analogy to physicists' work 908 01:11:23,800 --> 01:11:25,640 on the atomic bomb. 909 01:11:27,880 --> 01:11:32,160 In fact, that's actually the community I came out of. 910 01:11:34,000 --> 01:11:35,960 I wasn't a nuclear scientist by any means, 911 01:11:35,960 --> 01:11:38,400 but I was an applied mathematician. 912 01:11:38,400 --> 01:11:42,240 And my PhD programme was largely funded to train people 913 01:11:42,240 --> 01:11:44,520 to work in weapons labs. 914 01:11:47,040 --> 01:11:52,280 One could certainly argue that there is an existential threat. 915 01:11:52,280 --> 01:11:56,520 And whoever is leading in AI will lead militarily. 916 01:12:06,120 --> 01:12:09,040 China fully expects to pass the United States 917 01:12:09,040 --> 01:12:11,200 as the number one economy in the world 918 01:12:11,200 --> 01:12:15,360 and it believes that AI will make that jump more quickly 919 01:12:15,360 --> 01:12:17,280 and more dramatically. 920 01:12:18,680 --> 01:12:22,360 And they also see it as being able to leapfrog the United States 921 01:12:22,360 --> 01:12:24,320 in terms of military power. 922 01:12:24,320 --> 01:12:27,600 CHANTING 923 01:12:32,200 --> 01:12:34,280 Their plan is very simple. 924 01:12:34,280 --> 01:12:38,040 We want to catch the United States in these technologies by 2020. 925 01:12:38,040 --> 01:12:42,520 We want to surpass the United States in these technologies by 2025. 926 01:12:42,520 --> 01:12:44,840 And we want to be the world leader in AI 927 01:12:44,840 --> 01:12:47,440 and autonomous technologies by 2030. 928 01:12:49,760 --> 01:12:51,560 It is a national plan. 929 01:12:51,560 --> 01:12:56,480 It is backed up by at least $150 billion in investments. 930 01:12:56,480 --> 01:12:59,600 So this is definitely a race. 931 01:13:24,800 --> 01:13:26,960 AI is a little bit like fire. 932 01:13:28,000 --> 01:13:31,040 Fire was invented 700,000 years ago. 933 01:13:32,120 --> 01:13:34,280 And it has its pros and cons. 934 01:13:36,680 --> 01:13:41,440 People realised you can use fire to keep warm at night and to cook. 935 01:13:43,120 --> 01:13:45,600 But they also realise that you can... 936 01:13:47,160 --> 01:13:48,960 ..kill other people with it. 937 01:13:54,160 --> 01:13:59,920 Fire also has this AI-like quality of growing in a wildfire 938 01:13:59,920 --> 01:14:02,320 without further human ado. 939 01:14:04,720 --> 01:14:10,160 But the advantages outweigh the disadvantages by so much 940 01:14:10,160 --> 01:14:13,200 that we are not going to stop its development. 941 01:14:23,720 --> 01:14:25,640 Europe is waking up. 942 01:14:27,000 --> 01:14:31,840 Lots of companies in Europe are realising that the next wave 943 01:14:31,840 --> 01:14:36,520 of AI will be much bigger than the current wave. 944 01:14:38,000 --> 01:14:42,880 The next wave of AI will be about robots. 945 01:14:44,560 --> 01:14:50,040 All these machines that make things, that produce stuff, 946 01:14:50,040 --> 01:14:53,880 that build other machines, they are going to become smart. 947 01:15:01,320 --> 01:15:03,560 In the not so distant future, 948 01:15:03,560 --> 01:15:07,600 we will have robots that we can teach like we teach kids. 949 01:15:09,960 --> 01:15:14,840 For example, I will talk to a little robot and I will say, 950 01:15:14,840 --> 01:15:17,280 "Look here, robot, look here. 951 01:15:17,280 --> 01:15:19,400 "Let's assemble a smartphone. 952 01:15:19,400 --> 01:15:22,040 "We take a slab of plastic like that and we take 953 01:15:22,040 --> 01:15:23,440 "the screwdriver like that. 954 01:15:23,440 --> 01:15:27,120 "And now we screw in everything like this. 955 01:15:27,120 --> 01:15:29,880 "No, no, not like this. Like this. 956 01:15:29,880 --> 01:15:32,320 "Look, robot, look. Like this." 957 01:15:33,520 --> 01:15:36,600 And he will fail a couple of times, but rather quickly, 958 01:15:36,600 --> 01:15:41,680 he will learn to do the same thing much better than I could do it. 959 01:15:41,680 --> 01:15:43,480 And then we stop the learning 960 01:15:43,480 --> 01:15:45,760 and we make a million copies and sell it. 961 01:16:07,160 --> 01:16:10,360 Regulation of AI sounds like an attractive idea, 962 01:16:10,360 --> 01:16:12,640 but I don't think it's possible. 963 01:16:14,960 --> 01:16:17,440 One of the reasons why it won't work 964 01:16:17,440 --> 01:16:21,600 is the sheer curiosity of scientists. 965 01:16:21,600 --> 01:16:24,360 They don't give a damn for regulation. 966 01:16:27,160 --> 01:16:31,120 Military powers won't give a damn for regulations either. 967 01:16:31,120 --> 01:16:33,960 They will say, "If we, the Americans, don't do it, 968 01:16:33,960 --> 01:16:36,760 "the Chinese will do it," and the Chinese will say, 969 01:16:36,760 --> 01:16:39,680 "Oh, if we don't do it, then the Russians will do it." 970 01:16:42,440 --> 01:16:45,600 No matter what kind of political regulation is out there, 971 01:16:45,600 --> 01:16:49,560 all these military industrial complexes, 972 01:16:49,560 --> 01:16:53,480 they will almost by definition have to ignore that, 973 01:16:53,480 --> 01:16:56,680 because they want to avoid falling behind. 974 01:17:07,840 --> 01:17:10,560 A programme developed by the company Open AI 975 01:17:10,560 --> 01:17:14,040 can write coherent and credible stories just like human beings. 976 01:17:14,040 --> 01:17:18,880 It's one small step for machine, one giant leap for machinekind. 977 01:17:18,880 --> 01:17:22,240 IBM's newest artificial intelligence system took on experienced 978 01:17:22,240 --> 01:17:26,320 human debaters and won a live debate. 979 01:17:26,320 --> 01:17:29,520 Computer generated videos known as deepfakes are being used 980 01:17:29,520 --> 01:17:32,520 to put women's faces on pornographic videos. 981 01:17:36,920 --> 01:17:41,120 Artificial intelligence evolves at a very crazy pace. 982 01:17:42,400 --> 01:17:44,840 You know, it's, like, progressing so fast. 983 01:17:44,840 --> 01:17:48,840 In some ways, we're only at the beginning right now. 984 01:17:48,840 --> 01:17:50,880 You have so many potential applications. 985 01:17:50,880 --> 01:17:52,120 It's a gold mine. 986 01:17:54,400 --> 01:17:58,160 Well, since 2012, when deep learning became, like, a big game changer 987 01:17:58,160 --> 01:17:59,680 in the computer vision community, 988 01:17:59,680 --> 01:18:03,400 we were one of the first to actually adopt deep learning 989 01:18:03,400 --> 01:18:06,400 and apply it in the field of computer graphics. 990 01:18:08,960 --> 01:18:12,240 A lot of our research is funded by government, 991 01:18:12,240 --> 01:18:14,840 military, intelligence agencies. 992 01:18:18,520 --> 01:18:23,480 The way we create these photo-real mappings, usually the way it works 993 01:18:23,480 --> 01:18:26,520 is that we need two subjects, a source and a target, 994 01:18:26,520 --> 01:18:28,480 and I can do a face replacement. 995 01:18:33,360 --> 01:18:37,320 One of the applications is, for example, I want to manipulate 996 01:18:37,320 --> 01:18:40,680 someone's face saying things that he did not. 997 01:18:43,760 --> 01:18:47,800 It can be used for creative things, for funny content, but obviously, 998 01:18:47,800 --> 01:18:51,800 it can also be used for just simply manipulate videos 999 01:18:51,800 --> 01:18:53,160 and generate fake news. 1000 01:18:55,800 --> 01:18:58,920 This can be very dangerous. 1001 01:18:58,920 --> 01:19:01,440 If it gets into the wrong hands, 1002 01:19:01,440 --> 01:19:04,120 it can get out of control very quickly. 1003 01:19:08,080 --> 01:19:11,240 We're entering an era in which our enemies can make it look like 1004 01:19:11,240 --> 01:19:13,840 anyone is saying anything at any point in time. 1005 01:19:13,840 --> 01:19:16,200 Even if they would never say those things. 1006 01:19:16,200 --> 01:19:19,120 Moving forward, we need to be more vigilant 1007 01:19:19,120 --> 01:19:21,760 with what we trust from the internet. 1008 01:19:21,760 --> 01:19:25,720 It may sound basic, but how we move forward 1009 01:19:25,720 --> 01:19:29,640 in the age of information is going to be the difference 1010 01:19:29,640 --> 01:19:33,080 between whether we survive or whether we become some kind 1011 01:19:33,080 --> 01:19:34,880 of fucked-up dystopia. 1012 01:19:36,760 --> 01:19:41,360 OVERLAPPING VOICES BABBLE 1013 01:21:05,480 --> 01:21:09,880 One criticism that is frequently raised against my work 1014 01:21:09,880 --> 01:21:15,600 is saying that, hey, you know, there were stupid ideas in the past, 1015 01:21:15,600 --> 01:21:17,920 like phrenology or physiognomy. 1016 01:21:19,520 --> 01:21:24,040 There were people claiming that you can read a character 1017 01:21:24,040 --> 01:21:26,680 of a person just based on their face. 1018 01:21:28,160 --> 01:21:30,760 People would say, this is rubbish - 1019 01:21:30,760 --> 01:21:35,880 we know it was just thinly veiled racism and superstition. 1020 01:21:39,560 --> 01:21:43,960 But the fact that someone made a claim in the past 1021 01:21:43,960 --> 01:21:49,360 and tried to support this claim with invalid reasoning 1022 01:21:49,360 --> 01:21:53,640 doesn't automatically invalidate the claim. 1023 01:21:58,520 --> 01:22:01,760 Of course, people should have rights to their privacy when it comes 1024 01:22:01,760 --> 01:22:05,400 to sexual orientation or political views. 1025 01:22:06,640 --> 01:22:09,040 But I'm also afraid that in the current 1026 01:22:09,040 --> 01:22:12,520 technological environment, this is essentially impossible. 1027 01:22:17,280 --> 01:22:19,720 People should realise there's no going back. 1028 01:22:19,720 --> 01:22:22,840 There's no running away from the algorithms. 1029 01:22:25,880 --> 01:22:31,080 The sooner we accept the inevitable and inconvenient truth 1030 01:22:31,080 --> 01:22:34,120 that privacy is gone... 1031 01:22:36,280 --> 01:22:40,920 ..the sooner we can actually start thinking about how to make sure 1032 01:22:40,920 --> 01:22:46,240 that our societies are ready for the post-privacy age. 1033 01:23:09,120 --> 01:23:11,840 While speaking about facial recognition, 1034 01:23:11,840 --> 01:23:14,880 in my deep thoughts, I sometimes get 1035 01:23:14,880 --> 01:23:18,160 to the very dark era of our history... 1036 01:23:20,320 --> 01:23:23,840 ..when the people had to live in the system 1037 01:23:23,840 --> 01:23:28,160 where some part of the society was accepted 1038 01:23:28,160 --> 01:23:31,480 and some part of the society was abused to death. 1039 01:23:36,360 --> 01:23:40,840 What would Mengele do to have such an instrument in his hands? 1040 01:23:45,320 --> 01:23:49,280 It would be very quick and efficient for selection. 1041 01:23:52,880 --> 01:23:57,200 And this is the apocalyptic vision. 1042 01:24:41,440 --> 01:24:45,120 TRAIN WHOOSHES 1043 01:24:50,680 --> 01:24:55,440 So in the near future, the entire story of you will exist 1044 01:24:55,440 --> 01:24:58,400 in a vast array of connected databases 1045 01:24:58,400 --> 01:25:03,080 of faces, genomes, behaviours and emotion. 1046 01:25:05,720 --> 01:25:11,080 So you will have a digital avatar of yourself online, which records 1047 01:25:11,080 --> 01:25:13,640 how well you are doing as a citizen, 1048 01:25:13,640 --> 01:25:16,640 what kind of relationship do you have, 1049 01:25:16,640 --> 01:25:20,800 what kind of political orientation and sexual orientation. 1050 01:25:24,480 --> 01:25:28,440 Based on all of those data, those algorithms will be able 1051 01:25:28,440 --> 01:25:32,960 to manipulate your behaviour with an extreme precision. 1052 01:25:32,960 --> 01:25:36,240 Changing how we think, 1053 01:25:36,240 --> 01:25:39,560 and probably in the future, how we feel. 1054 01:26:00,400 --> 01:26:05,360 The beliefs and desires of the first AGIs will be extremely important. 1055 01:26:06,720 --> 01:26:09,720 So it's important to programme them correctly. 1056 01:26:10,920 --> 01:26:16,480 I think that if this is not done, then the nature of evolution, 1057 01:26:16,480 --> 01:26:20,120 of natural selection, would favour those systems, 1058 01:26:20,120 --> 01:26:22,560 prioritise their own survival above all else. 1059 01:26:26,240 --> 01:26:31,040 It's not that it's going to actively hate humans and want to harm them. 1060 01:26:32,520 --> 01:26:36,160 But it is going to be too powerful. 1061 01:26:36,160 --> 01:26:38,640 And I think a good analogy would be the way 1062 01:26:38,640 --> 01:26:40,120 humans treat animals. 1063 01:26:41,760 --> 01:26:44,400 It's not that we hate animals. I think humans love animals 1064 01:26:44,400 --> 01:26:46,120 and have a lot of affection for them. 1065 01:26:46,120 --> 01:26:52,400 But when the time comes to build a highway between two cities, 1066 01:26:52,400 --> 01:26:54,840 we're not asking the animals for permission. 1067 01:26:54,840 --> 01:26:57,280 We just do it because it's important for us. 1068 01:26:58,960 --> 01:27:01,720 And I think by default, that's the kind of relationship 1069 01:27:01,720 --> 01:27:07,600 that's going to be between us and AGIs which are truly autonomous 1070 01:27:07,600 --> 01:27:10,000 and operating on their own behalf. 1071 01:27:21,640 --> 01:27:24,200 If you have an arms race dynamics 1072 01:27:24,200 --> 01:27:28,280 between multiple teams trying to do the AGI first, 1073 01:27:28,280 --> 01:27:31,440 they will have less time to make sure 1074 01:27:31,440 --> 01:27:35,120 that the AGI they will build do care deeply for humans. 1075 01:27:38,080 --> 01:27:41,120 Because the way I imagine it is that there is an avalanche. 1076 01:27:41,120 --> 01:27:43,920 There is an avalanche of AGI development. 1077 01:27:43,920 --> 01:27:46,920 Imagine you have this huge unstoppable force. 1078 01:27:49,880 --> 01:27:53,680 And I think it's pretty likely that the surface of the Earth 1079 01:27:53,680 --> 01:27:56,720 will be covered with solar panels and data centres. 1080 01:28:00,720 --> 01:28:03,040 Given these kinds of concerns, 1081 01:28:03,040 --> 01:28:07,800 it will be important that AGI is somehow built 1082 01:28:07,800 --> 01:28:10,760 as a cooperation between multiple countries. 1083 01:28:12,280 --> 01:28:15,440 The future is going to be good for the AIs regardless. 1084 01:28:15,440 --> 01:28:18,640 It would be nice if it were good for humans as well. 1085 01:28:41,080 --> 01:28:44,720 Is there a lot of responsibility weighing on my shoulders? 1086 01:28:44,720 --> 01:28:46,280 Not really. 1087 01:28:47,360 --> 01:28:51,360 Was there a lot of responsibility on the shoulders 1088 01:28:51,360 --> 01:28:53,960 of the parents of Einstein? 1089 01:28:53,960 --> 01:28:58,280 The parents somehow made him, but they had no way of predicting 1090 01:28:58,280 --> 01:29:02,360 what he would do and how he would change the world. 1091 01:29:02,360 --> 01:29:07,520 And so you can't really hold them responsible for that. 1092 01:29:32,400 --> 01:29:35,200 So I'm not a very human-centric person. 1093 01:29:36,560 --> 01:29:41,080 I think I'm a little stepping stone in the evolution of the universe 1094 01:29:41,080 --> 01:29:42,840 towards higher complexity. 1095 01:29:45,560 --> 01:29:49,840 But it's also clear to me that I'm not the crown of creation 1096 01:29:49,840 --> 01:29:53,720 and that humankind as a whole is not the crown of creation. 1097 01:29:55,920 --> 01:29:59,760 But we are setting the stage for something that is bigger than us, 1098 01:29:59,760 --> 01:30:01,160 that transcends us. 1099 01:30:03,760 --> 01:30:07,640 And that will go out there in a way where humans cannot follow 1100 01:30:07,640 --> 01:30:12,000 and transform the entire universe, or at least the reachable universe. 1101 01:30:16,200 --> 01:30:19,640 So I find beauty and awe 1102 01:30:19,640 --> 01:30:23,880 and see myself as part of this much grander theme. 1103 01:30:48,800 --> 01:30:50,480 AI is inevitable. 1104 01:30:52,280 --> 01:30:58,000 We need to make sure we have the necessary human regulation 1105 01:30:58,000 --> 01:31:02,800 to prevent the weaponization of artificial intelligence. 1106 01:31:02,800 --> 01:31:08,600 We don't need any more weaponization of such a powerful tool. 1107 01:31:11,760 --> 01:31:14,720 One of the most critical things, I think, is the need 1108 01:31:14,720 --> 01:31:16,840 for international governance. 1109 01:31:18,640 --> 01:31:20,760 We have an imbalance of power here. 1110 01:31:20,760 --> 01:31:23,880 So now we have corporations with more power, might and ability 1111 01:31:23,880 --> 01:31:25,120 than entire countries. 1112 01:31:25,120 --> 01:31:28,600 How do we make sure that people's voices are getting heard? 1113 01:31:32,560 --> 01:31:36,400 It can't be a law-free zone. It can't be a rights-free zone. 1114 01:31:36,400 --> 01:31:40,320 We can't embrace all of these wonderful new technologies 1115 01:31:40,320 --> 01:31:42,240 for the 21st century 1116 01:31:42,240 --> 01:31:47,000 without trying to bring with us the package 1117 01:31:47,000 --> 01:31:51,360 of human rights that we fought so hard to achieve, 1118 01:31:51,360 --> 01:31:53,520 and that remains so fragile. 1119 01:32:03,080 --> 01:32:06,880 AI isn't good and it isn't evil either. 1120 01:32:06,880 --> 01:32:09,720 It's just going to amplify the desires and goals 1121 01:32:09,720 --> 01:32:11,280 of whoever controls it. 1122 01:32:11,280 --> 01:32:13,280 And the AI today is under the control 1123 01:32:13,280 --> 01:32:15,840 of a very, very small group of people. 1124 01:32:18,880 --> 01:32:22,560 The most important question that we humans have to ask ourselves 1125 01:32:22,560 --> 01:32:26,160 at this point in history requires no technical knowledge. 1126 01:32:26,160 --> 01:32:27,720 It's the question of, 1127 01:32:27,720 --> 01:32:31,600 what sort of future society do we want to create 1128 01:32:31,600 --> 01:32:34,280 with all this technology we're making? 1129 01:32:36,120 --> 01:32:39,400 What do we want the role of humans to be in this world?