1 00:00:25,020 --> 00:00:28,000 - Intelligence is the ability to understand. 2 00:00:28,000 --> 00:00:30,220 We passed on what we know to machines. 3 00:00:30,220 --> 00:00:31,760 - The rise of artificial intelligence 4 00:00:31,760 --> 00:00:34,210 is happening fast, but some fear the new technology 5 00:00:34,210 --> 00:00:36,960 might have more problems than anticipated. 6 00:00:36,960 --> 00:00:38,510 - We will not control it. 7 00:01:37,720 --> 00:01:40,290 - Artificially intelligent algorithms are here, 8 00:01:40,290 --> 00:01:41,930 but this is only the beginning. 9 00:01:51,380 --> 00:01:54,850 - In the age of AI, data is the new oil. 10 00:01:57,180 --> 00:01:59,230 - Today, Amazon, Google and Facebook 11 00:01:59,230 --> 00:02:01,440 are richer and more powerful than any companies 12 00:02:01,440 --> 00:02:03,840 that have ever existed throughout human history. 13 00:02:05,000 --> 00:02:08,690 - A handful of people working at a handful of technology companies 14 00:02:08,690 --> 00:02:11,360 steer what a billion people are thinking today. 15 00:02:12,710 --> 00:02:16,840 - This technology is changing: What does it mean to be human? 16 00:03:11,870 --> 00:03:16,340 - Artificial intelligence is simply non-biological intelligence. 17 00:03:17,310 --> 00:03:20,910 And intelligence itself is simply the ability to accomplish goals. 18 00:03:23,440 --> 00:03:26,030 I'm convinced that AI will ultimately be either 19 00:03:26,030 --> 00:03:29,780 the best thing ever to happen to humanity, or the worst thing ever to happen. 20 00:03:31,240 --> 00:03:34,930 We can use it to solve all of today's and tomorrow's 21 00:03:34,930 --> 00:03:40,620 greatest problems; cure diseases, deal with climate change, 22 00:03:40,780 --> 00:03:42,840 lift everybody out of poverty. 23 00:03:44,380 --> 00:03:47,460 But, we could use exactly the same technology 24 00:03:47,460 --> 00:03:51,780 to create a brutal global dictatorship with unprecedented 25 00:03:51,780 --> 00:03:54,600 surveillance and inequality and suffering. 26 00:03:56,230 --> 00:03:59,260 That's why this is the most important conversation of our time. 27 00:04:04,380 --> 00:04:07,590 - Artificial intelligence is everywhere 28 00:04:08,450 --> 00:04:11,500 because we now have thinking machines. 29 00:04:12,420 --> 00:04:15,840 If you go on social media or online, 30 00:04:15,840 --> 00:04:20,290 there's an artificial intelligence engine that decides what to recommend. 31 00:04:21,290 --> 00:04:25,400 If you go on Facebook and you're just scrolling through your friends' posts,- 32 00:04:25,400 --> 00:04:29,100 there's an AI engine that's picking which one to show you first 33 00:04:29,100 --> 00:04:30,840 - and which one to bury. 34 00:04:30,840 --> 00:04:34,190 If you try to get insurance, there is an AI engine 35 00:04:34,190 --> 00:04:36,270 trying to figure out how risky you are. 36 00:04:37,170 --> 00:04:40,710 And if you apply for a job, it's quite possible 37 00:04:40,710 --> 00:04:43,430 that an AI engine looks at the resume. 38 00:04:51,350 --> 00:04:53,500 - We are made of data. 39 00:04:54,240 --> 00:04:59,520 Every one of us is made of data -in terms of how we behave, 40 00:04:59,770 --> 00:05:03,000 how we talk, how we love, what we do every day. 41 00:05:05,410 --> 00:05:09,120 So, computer scientists are developing deep learning 42 00:05:09,120 --> 00:05:14,120 algorithms that can learn to identify, classify, 43 00:05:14,570 --> 00:05:18,680 and predict patterns within massive amounts of data. 44 00:05:31,390 --> 00:05:35,190 We are facing a form of precision surveillance, 45 00:05:35,190 --> 00:05:38,750 you could call it algorithmic surveillance, 46 00:05:38,750 --> 00:05:41,810 and it means that you cannot go unrecognized. 47 00:05:43,120 --> 00:05:46,110 You are always under the watch of algorithms. 48 00:05:51,070 --> 00:05:54,380 - Almost all the AI development on the planet today 49 00:05:54,380 --> 00:05:57,050 is done by a handful of big technology companies 50 00:05:57,050 --> 00:05:58,770 or by a few large governments. 51 00:06:01,930 --> 00:06:06,170 If we look at what AI is mostly being developed for, 52 00:06:06,170 --> 00:06:11,170 I would say it's killing, spying, and brainwashing. 53 00:06:11,570 --> 00:06:13,730 So, I mean, we have military AI, 54 00:06:13,730 --> 00:06:16,850 we have a whole surveillance apparatus being built 55 00:06:16,850 --> 00:06:19,060 using AI by major governments, 56 00:06:19,060 --> 00:06:21,730 and we have an advertising industry which is oriented 57 00:06:21,730 --> 00:06:25,870 toward recognizing what ads to try to sell to someone. 58 00:06:29,080 --> 00:06:32,090 - We humans have come to a fork in the road now. 59 00:06:33,530 --> 00:06:36,670 The AI we have today is very narrow. 60 00:06:37,820 --> 00:06:40,550 The holy grail of AI research ever since the beginning 61 00:06:40,550 --> 00:06:43,300 is to make AI that can do everything better than us, 62 00:06:44,660 --> 00:06:46,280 and we've basically built a God. 63 00:06:47,640 --> 00:06:49,930 It's going to revolutionize life as we know it. 64 00:06:53,040 --> 00:06:55,760 It's incredibly important to take a step back 65 00:06:55,760 --> 00:06:57,660 and think carefully about this. 66 00:06:59,380 --> 00:07:01,510 What sort of society do we want? 67 00:07:04,760 --> 00:07:07,430 - So, we're in this historic transformation. 68 00:07:08,890 --> 00:07:11,260 Like we're raising this new creature. 69 00:07:11,260 --> 00:07:14,230 We have a new offspring of sorts. 70 00:07:16,120 --> 00:07:19,630 But just like actual offspring, 71 00:07:19,630 --> 00:07:22,990 you don't get to control everything it's going to do. 72 00:07:57,000 --> 00:08:00,300 We are living at this privileged moment where, 73 00:08:00,300 --> 00:08:05,270 for the first time, we will see probably that AI 74 00:08:05,270 --> 00:08:08,860 is really going to outcompete humans in many, many, 75 00:08:08,860 --> 00:08:10,460 if not all, important fields. 76 00:08:14,300 --> 00:08:16,450 - Everything is going to change. 77 00:08:16,450 --> 00:08:19,520 A new form of life is emerging. 78 00:08:45,000 --> 00:08:50,900 When I was a boy, I thought, how can I maximize my impact? 79 00:08:52,050 --> 00:08:56,270 And then it was clear that I have to build something 80 00:08:56,270 --> 00:09:00,050 that learns to become smarter than myself, 81 00:09:00,050 --> 00:09:01,680 such that I can retire, 82 00:09:01,680 --> 00:09:04,480 and the smarter thing can further self-improve 83 00:09:04,480 --> 00:09:06,950 and solve all the problems that I cannot solve. 84 00:09:11,180 --> 00:09:14,520 Multiplying that tiny little bit of creativity 85 00:09:14,520 --> 00:09:16,370 that I have into infinity, 86 00:09:18,270 --> 00:09:20,780 and that's what has been driving me since then. 87 00:09:39,220 --> 00:09:40,680 How am I trying to build 88 00:09:40,680 --> 00:09:43,320 a general purpose artificial intelligence? 89 00:09:45,540 --> 00:09:49,690 If you want to be intelligent, you have to recognize speech, 90 00:09:49,690 --> 00:09:54,550 video and handwriting, and faces, and all kinds of things, 91 00:09:54,550 --> 00:09:57,550 and there we have made a lot of progress. 92 00:09:59,290 --> 00:10:02,010 See, LSTM, neural networks, 93 00:10:02,010 --> 00:10:05,890 which we developed in our labs in Munich and in Switzerland, 94 00:10:05,890 --> 00:10:10,460 and it's now used for speech recognition and translation, 95 00:10:10,460 --> 00:10:12,640 and video recognition. 96 00:10:12,640 --> 00:10:16,660 They are now in everybody's smartphone, almost one billion 97 00:10:16,660 --> 00:10:20,660 iPhones and in over two billion Android phones. 98 00:10:21,990 --> 00:10:26,990 So, we are generating all kinds of useful by-products 99 00:10:27,490 --> 00:10:29,040 on the way to the general goal. 100 00:10:40,000 --> 00:10:44,990 The main goal, some Artificial General Intelligence, 101 00:10:44,990 --> 00:10:51,400 an AGI that can learn to improve the learning algorithm itself. 102 00:10:52,960 --> 00:10:56,940 So, it basically can learn to improve the way it learns 103 00:10:56,940 --> 00:11:00,940 and it can also recursively improve the way it learns, 104 00:11:00,940 --> 00:11:04,460 the way it learns without any limitations except for 105 00:11:04,460 --> 00:11:07,970 the basic fundamental limitations of computability. 106 00:11:14,590 --> 00:11:17,880 One of my favorite robots is this one here. 107 00:11:17,880 --> 00:11:21,380 We use this robot for our studies of artificial curiosity. 108 00:11:22,640 --> 00:11:27,620 Where we are trying to teach this robot to teach itself. 109 00:11:32,870 --> 00:11:33,960 What is a baby doing? 110 00:11:33,960 --> 00:11:37,820 A baby is curiously exploring its world. 111 00:11:39,970 --> 00:11:42,210 That's how he learns how gravity works 112 00:11:42,210 --> 00:11:45,130 and how certain things topple, and so on. 113 00:11:46,700 --> 00:11:49,980 And as it learns to ask questions about the world, 114 00:11:49,980 --> 00:11:52,640 and as it learns to answer these questions, 115 00:11:52,640 --> 00:11:55,500 it becomes a more and more general problem solver. 116 00:11:56,540 --> 00:11:59,160 And so, our artificial systems are also learning 117 00:11:59,160 --> 00:12:03,320 to ask all kinds of questions, not just slavishly 118 00:12:03,320 --> 00:12:07,010 try to answer the questions given to them by humans. 119 00:12:10,880 --> 00:12:14,920 You have to give AI the freedom to invent its own tasks. 120 00:12:17,830 --> 00:12:20,610 If you don't do that, it's not going to become very smart. 121 00:12:22,160 --> 00:12:26,780 On the other hand, it's really hard to predict what they are going to do. 122 00:12:49,560 --> 00:12:52,910 - I feel that technology is a force of nature. 123 00:12:55,420 --> 00:13:00,020 I feel like there is a lot of similarity between technology and biological evolution. 124 00:13:08,850 --> 00:13:10,140 Playing God. 125 00:13:13,470 --> 00:13:16,500 Scientists have been accused of playing God for a while, 126 00:13:17,960 --> 00:13:22,410 - but there is a real sense in which we are creating something 127 00:13:23,380 --> 00:13:26,250 very different from anything we've created so far. 128 00:13:49,790 --> 00:13:53,250 I was interested in the concept of AI from a relatively early age. 129 00:13:55,080 --> 00:13:58,420 At some point, I got especially interested in machine learning. 130 00:14:01,690 --> 00:14:03,520 What is experience? 131 00:14:03,520 --> 00:14:04,710 What is learning? 132 00:14:04,710 --> 00:14:05,720 What is thinking? 133 00:14:06,890 --> 00:14:08,150 How does the brain work? 134 00:14:10,140 --> 00:14:11,940 These questions are philosophical, 135 00:14:11,940 --> 00:14:15,190 but it looks like we can come up with algorithms that 136 00:14:15,190 --> 00:14:18,530 both do useful things and help us answer these questions. 137 00:14:20,000 --> 00:14:22,250 Like it's almost like applied philosophy. 138 00:14:48,060 --> 00:14:51,410 Artificial General Intelligence, AGI. 139 00:14:52,610 --> 00:14:57,210 A computer system that can do any job or any task 140 00:14:57,210 --> 00:15:00,380 that a human does, but only better. 141 00:15:13,180 --> 00:15:15,810 Yeah, I mean, we definitely will be able to create 142 00:15:17,340 --> 00:15:20,150 completely autonomous beings with their own goals. 143 00:15:24,060 --> 00:15:25,630 And it will be very important, 144 00:15:25,630 --> 00:15:30,440 especially as these beings become much smarter than humans, 145 00:15:30,440 --> 00:15:34,720 it's going to be important to have these beings, 146 00:15:36,220 --> 00:15:39,330 that the goals of these beings be aligned with our goals. 147 00:15:42,600 --> 00:15:45,340 That's what we're trying to do at OpenAI. 148 00:15:45,340 --> 00:15:49,700 Be at the forefront of research and steer the research, 149 00:15:49,700 --> 00:15:53,940 steer their initial conditions so to maximize the chance 150 00:15:53,940 --> 00:15:56,100 that the future will be good for humans. 151 00:16:11,660 --> 00:16:14,370 Now, AI is a great thing because AI will solve 152 00:16:14,370 --> 00:16:16,440 all the problems that we have today. 153 00:16:19,080 --> 00:16:22,580 It will solve employment, it will solve disease, 154 00:16:24,670 --> 00:16:26,560 it will solve poverty, 155 00:16:29,000 --> 00:16:31,360 but it will also create new problems. 156 00:16:34,580 --> 00:16:36,320 I think that... 157 00:16:40,490 --> 00:16:43,120 The problem of fake news is going to be a thousand, 158 00:16:43,120 --> 00:16:44,360 a million times worse. 159 00:16:46,440 --> 00:16:48,820 Cyberattacks will become much more extreme. 160 00:16:50,640 --> 00:16:53,470 You will have totally automated AI weapons. 161 00:16:56,100 --> 00:16:59,980 I think AI has the potential to create infinitely stable dictatorships. 162 00:17:05,820 --> 00:17:08,850 You're gonna see dramatically more intelligent systems 163 00:17:08,850 --> 00:17:12,620 in 10 or 15 years from now, and I think it's highly likely 164 00:17:12,620 --> 00:17:17,120 that those systems will have completely astronomical impact on society. 165 00:17:19,480 --> 00:17:21,160 Will humans actually benefit? 166 00:17:22,860 --> 00:17:24,860 And who will benefit, who will not? 167 00:17:44,560 --> 00:17:47,660 - In 2012, IBM estimated that 168 00:17:47,660 --> 00:17:52,090 an average person is generating 500 megabytes 169 00:17:52,090 --> 00:17:55,110 of digital footprints every single day. 170 00:17:55,110 --> 00:17:57,390 Imagine that you wanted to back-up one day worth 171 00:17:57,390 --> 00:18:01,260 of data that humanity is leaving behind, on paper. 172 00:18:01,260 --> 00:18:04,370 How tall will be the stack of paper that contains 173 00:18:04,370 --> 00:18:07,820 just one day worth of data that humanity is producing? 174 00:18:09,610 --> 00:18:12,670 It's like from the earth to the sun, four times over. 175 00:18:14,550 --> 00:18:20,340 In 2025, we'll be generating 62 gigabytes of data 176 00:18:20,640 --> 00:18:22,380 per person, per day. 177 00:18:36,970 --> 00:18:42,380 We're leaving a ton of digital footprints while going through our lives. 178 00:18:44,790 --> 00:18:49,380 They provide computer algorithms with a fairly good idea about who we are, 179 00:18:49,380 --> 00:18:52,280 what we want, what we are doing. 180 00:18:56,020 --> 00:18:59,320 In my work, I looked at different types of digital footprints. 181 00:18:59,320 --> 00:19:02,650 I looked at Facebook likes, I looked at language, 182 00:19:02,650 --> 00:19:06,890 credit card records, web browsing histories, search records. 183 00:19:07,740 --> 00:19:12,010 and each time I found that if you get enough of this data, 184 00:19:12,010 --> 00:19:14,370 you can accurately predict future behavior 185 00:19:14,370 --> 00:19:17,560 and reveal important intimate traits. 186 00:19:19,140 --> 00:19:21,270 This can be used in great ways, 187 00:19:21,270 --> 00:19:24,670 but it can also be used to manipulate people. 188 00:19:29,910 --> 00:19:33,880 Facebook is delivering daily information 189 00:19:33,880 --> 00:19:36,460 to two billion people or more. 190 00:19:38,750 --> 00:19:42,840 If you slightly change the functioning of the Facebook engine, 191 00:19:42,840 --> 00:19:46,500 you can move the opinions and hence, 192 00:19:46,500 --> 00:19:49,870 the votes of millions of people. 193 00:19:50,720 --> 00:19:52,680 - Brexit! - When do we want it? 194 00:19:52,720 --> 00:19:54,140 - Now! 195 00:19:56,070 --> 00:19:58,690 - A politician wouldn't be able to figure out 196 00:19:58,690 --> 00:20:02,830 which message each one of his or her voters would like, 197 00:20:02,830 --> 00:20:06,010 but a computer can see what political message 198 00:20:06,010 --> 00:20:08,920 would be particularly convincing for you. 199 00:20:11,320 --> 00:20:12,640 - Ladies and gentlemen, 200 00:20:12,640 --> 00:20:16,170 it's my privilege to speak to you today about the power 201 00:20:16,170 --> 00:20:20,430 of big data and psychographics in the electoral process. 202 00:20:20,430 --> 00:20:22,360 - Data from Cambridge Analytica 203 00:20:22,360 --> 00:20:25,070 secretly harvested the personal information 204 00:20:25,070 --> 00:20:28,320 of 50 million unsuspecting Facebook users. 205 00:20:28,320 --> 00:20:31,690 USA! 206 00:20:31,690 --> 00:20:35,500 - The data firm hired by Donald Trump's presidential election campaign 207 00:20:35,870 --> 00:20:41,080 used secretly obtained information to directly target potential American voters. 208 00:20:41,080 --> 00:20:42,940 - With that, they say they can predict 209 00:20:42,940 --> 00:20:46,760 the personality of every single adult in the United States. 210 00:20:47,470 --> 00:20:49,460 - Tonight we're hearing from Cambridge Analytica 211 00:20:49,460 --> 00:20:51,650 whistleblower, Christopher Wiley. 212 00:20:51,650 --> 00:20:54,990 - What we worked on was data harvesting programs 213 00:20:54,990 --> 00:20:59,200 where we would pull data and run that data through algorithms that could profile 214 00:20:59,200 --> 00:21:01,930 their personality traits and other psychological attributes 215 00:21:01,930 --> 00:21:06,580 to exploit mental vulnerabilities that our algorithms showed that they had. 216 00:21:15,500 --> 00:21:17,570 - Cambridge Analytica mentioned once 217 00:21:17,570 --> 00:21:20,460 or said that their models were based on my work, 218 00:21:22,510 --> 00:21:25,140 but Cambridge Analytica is just one of the hundreds 219 00:21:25,140 --> 00:21:30,040 of companies that are using such methods to target voters. 220 00:21:32,130 --> 00:21:35,760 You know, I would be asked questions by journalists such as, 221 00:21:35,760 --> 00:21:37,220 "So how do you feel about 222 00:21:38,810 --> 00:21:41,990 "electing Trump and supporting Brexit?" 223 00:21:41,990 --> 00:21:43,940 How do you answer to such question? 224 00:21:45,400 --> 00:21:51,740 I guess that I have to deal with being blamed for all of it. 225 00:22:07,700 --> 00:22:12,480 - How tech started was as a democratizing force, 226 00:22:12,480 --> 00:22:15,110 as a force for good, as an ability for humans 227 00:22:15,110 --> 00:22:17,900 to interact with each other without gatekeepers. 228 00:22:20,210 --> 00:22:25,110 There's never been a bigger experiment in communications for the human race. 229 00:22:26,330 --> 00:22:29,060 What happens when everybody gets to have their say? 230 00:22:29,780 --> 00:22:31,850 You would assume that it would be for the better, 231 00:22:31,850 --> 00:22:34,540 that there would be more democracy, there would be more discussion, 232 00:22:34,540 --> 00:22:37,660 there would be more tolerance, but what's happened is that 233 00:22:37,660 --> 00:22:39,720 these systems have been hijacked. 234 00:22:41,300 --> 00:22:43,860 - We stand for connecting every person. 235 00:22:44,960 --> 00:22:46,620 For a global community. 236 00:22:47,780 --> 00:22:50,500 - One company, Facebook, is responsible 237 00:22:50,500 --> 00:22:53,540 for the communications of a lot of the human race. 238 00:22:55,650 --> 00:22:57,360 Same thing with Google. 239 00:22:57,360 --> 00:23:00,540 Everything you want know about the world comes from them. 240 00:23:01,540 --> 00:23:05,120 This is global information economy 241 00:23:05,120 --> 00:23:07,830 that is controlled by a small group of people. 242 00:23:14,400 --> 00:23:17,920 - The world's richest companies are all technology companies. 243 00:23:18,770 --> 00:23:24,500 Google, Apple, Microsoft, Amazon, Facebook. 244 00:23:25,880 --> 00:23:29,040 It's staggering how, 245 00:23:29,040 --> 00:23:31,770 in probably just 10 years, 246 00:23:31,770 --> 00:23:34,810 that the entire corporate power structure 247 00:23:34,810 --> 00:23:39,460 are basically in the business of trading electrons. 248 00:23:41,240 --> 00:23:47,060 These little bits and bytes are really the new currency. 249 00:23:53,560 --> 00:23:55,800 - The way that data is monetized 250 00:23:55,800 --> 00:23:58,820 is happening all around us, even if it's invisible to us. 251 00:24:00,960 --> 00:24:04,280 Google has every amount of information available. 252 00:24:04,280 --> 00:24:07,110 They track people by their GPS location. 253 00:24:07,110 --> 00:24:10,080 They know exactly what your search history has been. 254 00:24:10,080 --> 00:24:12,850 They know your political preferences. 255 00:24:12,850 --> 00:24:14,840 Your search history alone can tell you 256 00:24:14,840 --> 00:24:17,570 everything about an individual from their health problems 257 00:24:17,570 --> 00:24:19,480 to their sexual preferences. 258 00:24:19,480 --> 00:24:21,880 So, Google's reach is unlimited. 259 00:24:28,750 --> 00:24:31,620 - So we've seen Google and Facebook 260 00:24:31,620 --> 00:24:34,150 rise into these large surveillance machines 261 00:24:35,040 --> 00:24:38,200 and they're both actually ad brokers. 262 00:24:38,200 --> 00:24:42,040 It sounds really mundane, but they're high tech ad brokers. 263 00:24:42,910 --> 00:24:46,080 And the reason they're so profitable is that they're using 264 00:24:46,080 --> 00:24:49,750 artificial intelligence to process all this data about you, 265 00:24:51,560 --> 00:24:54,660 and then to match you with the advertiser 266 00:24:54,660 --> 00:24:59,660 that wants to reach people like you, - for whatever message. 267 00:25:03,000 --> 00:25:07,880 - One of the problems with technology is that it's been developed to be addictive. 268 00:25:07,880 --> 00:25:09,910 The way these companies design these things 269 00:25:09,910 --> 00:25:12,010 is in order to pull you in and engage you. 270 00:25:13,180 --> 00:25:16,820 They want to become essentially a slot machine of attention. 271 00:25:18,600 --> 00:25:21,770 So you're always paying attention, you're always jacked into the matrix, 272 00:25:21,770 --> 00:25:23,520 you're always checking. 273 00:25:26,580 --> 00:25:30,450 - When somebody controls what you read, they also control what you think. 274 00:25:32,180 --> 00:25:34,610 You get more of what you've seen before and liked before, 275 00:25:34,610 --> 00:25:37,890 because this gives more traffic and that gives more ads, 276 00:25:39,410 --> 00:25:43,160 but it also locks you into your echo chamber. 277 00:25:43,160 --> 00:25:46,350 And this is what leads to this polarization that we see today. 278 00:25:46,460 --> 00:25:49,300 Jair Bolsonaro! 279 00:25:49,970 --> 00:25:52,590 - Jair Bolsonaro, Brazil's right-wing 280 00:25:52,590 --> 00:25:56,000 populist candidate sometimes likened to Donald Trump, 281 00:25:56,000 --> 00:25:58,640 winning the presidency Sunday night in that country's 282 00:25:58,640 --> 00:26:01,840 most polarizing election in decades. 283 00:26:01,840 --> 00:26:02,840 - Bolsonaro! 284 00:26:04,160 --> 00:26:09,550 - What we are seeing around the world is upheaval and polarization and conflict 285 00:26:10,770 --> 00:26:15,000 that is partially pushed by algorithms 286 00:26:15,000 --> 00:26:19,070 that's figured out that political extremes, 287 00:26:19,070 --> 00:26:22,180 tribalism, and sort of shouting for your team, 288 00:26:22,180 --> 00:26:25,290 and feeling good about it, is engaging. 289 00:26:30,420 --> 00:26:33,150 - Social media may be adding to the attention 290 00:26:33,150 --> 00:26:35,490 to hate crimes around the globe. 291 00:26:35,490 --> 00:26:38,120 - It's about how people can become radicalized 292 00:26:38,120 --> 00:26:41,740 by living in the fever swamps of the Internet. 293 00:26:41,740 --> 00:26:45,960 - So is this a key moment for the tech giants? Are they now prepared to take responsibility 294 00:26:45,960 --> 00:26:48,840 as publishers for what they share with the world? 295 00:26:49,790 --> 00:26:52,750 - If you deploy a powerful potent technology 296 00:26:52,750 --> 00:26:56,670 at scale, and if you're talking about Google and Facebook, 297 00:26:56,670 --> 00:26:59,530 you're deploying things at scale of billions. 298 00:26:59,530 --> 00:27:02,770 If your artificial intelligence is pushing polarization, 299 00:27:02,770 --> 00:27:05,090 you have global upheaval potentially. 300 00:27:05,720 --> 00:27:09,240 White lives matter! 301 00:27:09,320 --> 00:27:13,620 Black lives matter! 302 00:28:00,620 --> 00:28:04,040 - Artificial General Intelligence, AGI. 303 00:28:06,350 --> 00:28:08,320 Imagine your smartest friend, 304 00:28:09,600 --> 00:28:11,950 with 1,000 friends, just as smart, 305 00:28:14,680 --> 00:28:17,580 and then run them at a 1,000 times faster than real time. 306 00:28:17,580 --> 00:28:19,940 So it means that in every day of our time, 307 00:28:19,940 --> 00:28:22,510 they will do three years of thinking. 308 00:28:22,510 --> 00:28:25,520 Can you imagine how much you could do 309 00:28:26,580 --> 00:28:31,480 if, for every day, you could do three years' worth of work? 310 00:28:52,440 --> 00:28:55,720 It wouldn't be an unfair comparison to say 311 00:28:55,720 --> 00:28:59,780 that what we have right now is even more exciting than 312 00:28:59,780 --> 00:29:02,510 the quantum physicists of the early 20th century. 313 00:29:02,510 --> 00:29:04,240 They discovered nuclear power. 314 00:29:05,740 --> 00:29:08,380 I feel extremely lucky to be taking part in this. 315 00:29:15,310 --> 00:29:18,820 Many machine learning experts, who are very knowledgeable and experienced, 316 00:29:18,820 --> 00:29:20,820 have a lot of skepticism about AGI. 317 00:29:22,350 --> 00:29:25,900 About when it would happen, and about whether it could happen at all. 318 00:29:31,460 --> 00:29:35,740 But right now, this is something that just not that many people have realized yet. 319 00:29:36,500 --> 00:29:41,240 That the speed of computers, for neural networks, for AI, 320 00:29:41,240 --> 00:29:45,240 are going to become maybe 100,000 times faster 321 00:29:45,240 --> 00:29:46,980 in a small number of years. 322 00:29:49,330 --> 00:29:52,100 The entire hardware industry for a long time 323 00:29:52,100 --> 00:29:54,660 didn't really know what to do next, 324 00:29:55,660 --> 00:30:00,900 but with artificial neural networks, now that they actually work, 325 00:30:00,900 --> 00:30:03,420 you have a reason to build huge computers. 326 00:30:04,530 --> 00:30:06,930 You can build a brain in silicon, it's possible. 327 00:30:14,450 --> 00:30:18,710 The very first AGIs will be basically very, 328 00:30:18,710 --> 00:30:22,850 very large data centers packed with specialized 329 00:30:22,850 --> 00:30:25,660 neural network processors working in parallel. 330 00:30:28,070 --> 00:30:30,920 Compact, hot, power hungry package, 331 00:30:32,140 --> 00:30:35,540 consuming like 10 million homes' worth of energy. 332 00:30:54,140 --> 00:30:55,640 A roast beef sandwich. 333 00:30:55,640 --> 00:30:58,290 Yeah, something slightly different. 334 00:30:58,290 --> 00:30:59,200 Just this once. 335 00:31:04,060 --> 00:31:06,260 Even the very first AGIs 336 00:31:06,260 --> 00:31:09,060 will be dramatically more capable than humans. 337 00:31:10,970 --> 00:31:14,730 Humans will no longer be economically useful for nearly any task. 338 00:31:16,200 --> 00:31:17,880 Why would you want to hire a human, 339 00:31:17,880 --> 00:31:21,980 if you could just get a computer that's going to do it much better and much more cheaply? 340 00:31:28,900 --> 00:31:31,020 AGI is going to be like, without question, 341 00:31:31,880 --> 00:31:34,660 the most important technology in the history 342 00:31:34,660 --> 00:31:36,600 of the planet by a huge margin. 343 00:31:39,410 --> 00:31:42,550 It's going to be bigger than electricity, nuclear, 344 00:31:42,550 --> 00:31:44,140 and the Internet combined. 345 00:31:45,740 --> 00:31:47,620 In fact, you could say that the whole purpose 346 00:31:47,620 --> 00:31:49,660 of all human science, the purpose of computer science, 347 00:31:49,660 --> 00:31:52,550 the End Game, this is the End Game, to build this. 348 00:31:52,550 --> 00:31:54,130 And it's going to be built. 349 00:31:54,130 --> 00:31:56,000 It's going to be a new life form. 350 00:31:56,000 --> 00:31:57,090 It's going to be... 351 00:31:59,190 --> 00:32:00,740 It's going to make us obsolete. 352 00:32:22,070 --> 00:32:24,520 - European manufacturers know the Americans 353 00:32:24,520 --> 00:32:27,460 have invested heavily in the necessary hardware. 354 00:32:27,460 --> 00:32:29,780 - Step into a brave new world of power, 355 00:32:29,780 --> 00:32:31,630 performance and productivity. 356 00:32:32,360 --> 00:32:35,480 - All of the images you are about to see on the large screen 357 00:32:35,480 --> 00:32:39,030 will be generated by what's in that Macintosh. 358 00:32:39,760 --> 00:32:42,210 - It's my honor and privilege to introduce to you 359 00:32:42,210 --> 00:32:44,700 the Windows 95 Development Team. 360 00:32:45,640 --> 00:32:49,040 - Human physical labor has been mostly obsolete for 361 00:32:49,040 --> 00:32:50,620 getting on for a century. 362 00:32:51,390 --> 00:32:55,300 Routine human mental labor is rapidly becoming obsolete 363 00:32:55,300 --> 00:32:58,980 and that's why we're seeing a lot of the middle class jobs disappearing. 364 00:33:01,010 --> 00:33:02,340 - Every once in a while, 365 00:33:02,340 --> 00:33:06,370 a revolutionary product comes along that changes everything. 366 00:33:06,370 --> 00:33:09,190 Today, Apple is reinventing the phone. 367 00:33:20,730 --> 00:33:24,270 - Machine intelligence is already all around us. 368 00:33:24,270 --> 00:33:27,620 The list of things that we humans can do better than machines 369 00:33:27,620 --> 00:33:29,600 is actually shrinking pretty fast. 370 00:33:35,730 --> 00:33:37,400 - Driverless cars are great. 371 00:33:37,400 --> 00:33:40,010 They probably will reduce accidents. 372 00:33:40,010 --> 00:33:43,670 Except, alongside with that, in the United States, 373 00:33:43,670 --> 00:33:46,410 you're going to lose 10 million jobs. 374 00:33:46,480 --> 00:33:49,980 What are you going to do with 10 million unemployed people? 375 00:33:54,900 --> 00:33:58,660 - The risk for social conflict and tensions, 376 00:33:58,660 --> 00:34:02,080 if you exacerbate inequalities, is very, very high. 377 00:34:11,560 --> 00:34:13,870 - AGI can, by definition, 378 00:34:13,870 --> 00:34:16,700 do all jobs better than we can do. 379 00:34:16,700 --> 00:34:18,610 People who are saying, "Oh, there will always be jobs 380 00:34:18,610 --> 00:34:21,120 "that humans can do better than machines," are simply 381 00:34:21,120 --> 00:34:24,180 betting against science and saying there will never be AGI. 382 00:34:30,860 --> 00:34:33,630 - What we are seeing now is like a train hurtling 383 00:34:33,630 --> 00:34:37,880 down a dark tunnel at breakneck speed and it looks like 384 00:34:37,880 --> 00:34:39,660 we're sleeping at the wheel. 385 00:35:29,170 --> 00:35:34,710 - A large fraction of the digital footprints we're leaving behind are digital images. 386 00:35:35,690 --> 00:35:39,340 And specifically, what's really interesting to me as a psychologist 387 00:35:39,340 --> 00:35:41,470 are digital images of our faces. 388 00:35:44,500 --> 00:35:47,260 Here you can see the difference in the facial outline 389 00:35:47,260 --> 00:35:50,130 of an average gay and an average straight face. 390 00:35:50,130 --> 00:35:52,660 And you can see that straight men 391 00:35:52,660 --> 00:35:55,380 have slightly broader jaws. 392 00:35:55,680 --> 00:36:00,580 Gay women have slightly larger jaws, compared with straight women. 393 00:36:02,510 --> 00:36:05,670 Computer algorithms can reveal our political views 394 00:36:05,670 --> 00:36:08,280 or sexual orientation, or intelligence, 395 00:36:08,280 --> 00:36:11,120 just based on the picture of our faces. 396 00:36:12,070 --> 00:36:16,760 Even a human brain can distinguish between gay and straight men with some accuracy. 397 00:36:16,920 --> 00:36:21,700 Now it turns out that the computer can do it with much higher accuracy. 398 00:36:21,700 --> 00:36:25,200 What you're seeing here is an accuracy of 399 00:36:25,200 --> 00:36:29,410 off-the-shelf facial recognition software. 400 00:36:29,410 --> 00:36:31,640 This is terrible news 401 00:36:31,640 --> 00:36:34,320 for gay men and women all around the world. 402 00:36:34,320 --> 00:36:35,680 And not only gay men and women, 403 00:36:35,680 --> 00:36:38,300 because the same algorithms can be used to detect other 404 00:36:38,300 --> 00:36:42,350 intimate traits, think being a member of the opposition, 405 00:36:42,350 --> 00:36:45,090 or being a liberal, or being an atheist. 406 00:36:46,850 --> 00:36:50,070 Being an atheist is also punishable by death 407 00:36:50,070 --> 00:36:52,610 in Saudi Arabia, for instance. 408 00:36:59,780 --> 00:37:04,440 My mission as an academic is to warn people about the dangers of algorithms 409 00:37:04,440 --> 00:37:08,290 being able to reveal our intimate traits. 410 00:37:09,780 --> 00:37:14,060 The problem is that when people receive bad news, 411 00:37:14,060 --> 00:37:16,260 they very often choose to dismiss them. 412 00:37:17,810 --> 00:37:22,250 Well, it's a bit scary when you start receiving death threats from one day to another, 413 00:37:22,250 --> 00:37:24,830 and I've received quite a few death threats, - 414 00:37:25,910 --> 00:37:30,870 -but as a scientist, I have to basically show what is possible. 415 00:37:33,590 --> 00:37:36,900 So what I'm really interested in now is to try to see 416 00:37:36,900 --> 00:37:41,080 whether we can predict other traits from people's faces. 417 00:37:46,270 --> 00:37:48,580 Now, if you can detect depression from a face, 418 00:37:48,580 --> 00:37:53,580 or suicidal thoughts, maybe a CCTV system 419 00:37:53,690 --> 00:37:56,970 on the train station can save some lives. 420 00:37:59,130 --> 00:38:03,830 What if we could predict that someone is more prone to commit a crime? 421 00:38:04,960 --> 00:38:07,150 You probably had a school counselor, 422 00:38:07,150 --> 00:38:10,380 a psychologist hired there to identify children 423 00:38:10,380 --> 00:38:14,830 that potentially may have some behavioral problems. 424 00:38:17,330 --> 00:38:20,080 So now imagine if you could predict with high accuracy 425 00:38:20,080 --> 00:38:22,590 that someone is likely to commit a crime in the future 426 00:38:22,590 --> 00:38:24,780 from the language use, from the face, 427 00:38:24,780 --> 00:38:27,580 from the facial expressions, from the likes on Facebook. 428 00:38:32,310 --> 00:38:35,210 I'm not developing new methods, I'm just describing 429 00:38:35,210 --> 00:38:38,890 something or testing something in an academic environment. 430 00:38:40,590 --> 00:38:42,790 But there obviously is a chance that, 431 00:38:42,790 --> 00:38:47,790 while warning people against risks of new technologies, 432 00:38:47,960 --> 00:38:50,560 I may also give some people new ideas. 433 00:39:11,140 --> 00:39:13,830 - We haven't yet seen the future in terms of 434 00:39:13,830 --> 00:39:18,780 the ways in which the new data-driven society 435 00:39:18,780 --> 00:39:21,380 is going to really evolve. 436 00:39:23,540 --> 00:39:26,740 The tech companies want to get every possible bit 437 00:39:26,740 --> 00:39:30,170 of information that they can collect on everyone 438 00:39:30,170 --> 00:39:31,690 to facilitate business. 439 00:39:33,470 --> 00:39:37,090 The police and the military want to do the same thing 440 00:39:37,090 --> 00:39:38,830 to facilitate security. 441 00:39:41,960 --> 00:39:46,440 The interests that the two have in common are immense, 442 00:39:46,440 --> 00:39:51,220 and so the extent of collaboration between what you might 443 00:39:51,220 --> 00:39:56,820 call the Military-Tech Complex is growing dramatically. 444 00:40:00,400 --> 00:40:03,010 - The CIA, for a very long time, 445 00:40:03,010 --> 00:40:06,120 has maintained a close connection with Silicon Valley. 446 00:40:07,110 --> 00:40:10,270 Their venture capital firm known as In-Q-Tel, 447 00:40:10,270 --> 00:40:13,870 makes seed investments to start-up companies developing 448 00:40:13,870 --> 00:40:17,490 breakthrough technology that the CIA hopes to deploy. 449 00:40:18,660 --> 00:40:22,100 Palantir, the big data analytics firm, 450 00:40:22,100 --> 00:40:25,100 one of their first seed investments was from In-Q-Tel. 451 00:40:28,950 --> 00:40:31,780 - In-Q-Tel has struck gold in Palantir 452 00:40:31,780 --> 00:40:35,990 in helping to create a private vendor 453 00:40:35,990 --> 00:40:40,850 that has intelligence and artificial intelligence 454 00:40:40,850 --> 00:40:44,460 capabilities that the government can't even compete with. 455 00:40:46,370 --> 00:40:48,720 - Good evening, I'm Peter Thiel. 456 00:40:49,740 --> 00:40:54,210 I'm not a politician, but neither is Donald Trump. 457 00:40:54,210 --> 00:40:58,620 He is a builder and it's time to rebuild America. 458 00:41:01,280 --> 00:41:03,940 - Peter Thiel, the founder of Palantir, 459 00:41:03,940 --> 00:41:06,680 was a Donald Trump transition advisor 460 00:41:06,680 --> 00:41:09,180 and a close friend and donor. 461 00:41:11,080 --> 00:41:13,700 Trump was elected largely on the promise 462 00:41:13,700 --> 00:41:17,560 to deport millions of immigrants. 463 00:41:17,560 --> 00:41:22,240 The only way you can do that is with a lot of intelligence 464 00:41:22,240 --> 00:41:25,010 and that's where Palantir comes in. 465 00:41:28,680 --> 00:41:33,340 They ingest huge troves of data, which include, 466 00:41:33,340 --> 00:41:37,370 where you live, where you work, who you know, 467 00:41:37,370 --> 00:41:40,930 who your neighbors are, who your family is, 468 00:41:40,930 --> 00:41:44,480 where you have visited, where you stay, 469 00:41:44,480 --> 00:41:46,320 your social media profile. 470 00:41:49,250 --> 00:41:53,240 Palantir gets all of that and is remarkably good 471 00:41:53,240 --> 00:41:58,240 at structuring it in a way that helps law enforcement, 472 00:41:58,460 --> 00:42:02,450 immigration authorities or intelligence agencies 473 00:42:02,450 --> 00:42:06,210 of any kind, track you, find you, 474 00:42:06,210 --> 00:42:09,550 and learn everything there is to know about you. 475 00:42:48,430 --> 00:42:51,040 - We're putting AI in charge now of evermore 476 00:42:51,040 --> 00:42:53,860 important decisions that affect people's lives. 477 00:42:54,810 --> 00:42:58,050 Old-school AI used to have its intelligence programmed in 478 00:42:58,050 --> 00:43:01,420 by humans who understood how it worked, but today, 479 00:43:01,420 --> 00:43:04,080 powerful AI systems have just learned for themselves, 480 00:43:04,080 --> 00:43:07,480 and we have no clue really how they work, 481 00:43:07,480 --> 00:43:09,680 which makes it really hard to trust them. 482 00:43:14,270 --> 00:43:17,820 - This isn't some futuristic technology, this is now. 483 00:43:19,470 --> 00:43:23,220 AI might help determine where a fire department 484 00:43:23,220 --> 00:43:25,900 is built in a community or where a school is built. 485 00:43:25,900 --> 00:43:28,380 It might decide whether you get bail, 486 00:43:28,380 --> 00:43:30,680 or whether you stay in jail. 487 00:43:30,680 --> 00:43:32,930 It might decide where the police are going to be. 488 00:43:32,930 --> 00:43:37,140 It might decide whether you're going to be under additional police scrutiny. 489 00:43:43,340 --> 00:43:46,140 - It's popular now in the US to do predictive policing. 490 00:43:46,980 --> 00:43:50,900 So what they do is they use an algorithm to figure out where crime will be, 491 00:43:51,760 --> 00:43:55,240 - and they use that to tell where we should send police officers. 492 00:43:56,430 --> 00:43:59,420 So that's based on a measurement of crime rate. 493 00:44:00,640 --> 00:44:02,330 So we know that there is bias. 494 00:44:02,330 --> 00:44:05,090 Black people and Hispanic people are pulled over, 495 00:44:05,090 --> 00:44:07,180 and stopped by the police officers more frequently 496 00:44:07,180 --> 00:44:09,820 than white people are, so we have this biased data going in, 497 00:44:09,820 --> 00:44:11,820 and then what happens is you use that to say, 498 00:44:11,820 --> 00:44:13,640 "Oh, here's where the cops should go." 499 00:44:13,640 --> 00:44:17,320 Well, the cops go to those neighborhoods, and guess what they do, they arrest people. 500 00:44:17,680 --> 00:44:21,160 And then it feeds back biased data into the system, 501 00:44:21,160 --> 00:44:23,060 and that's called a feedback loop. 502 00:44:35,990 --> 00:44:40,650 - Predictive policing leads at the extremes 503 00:44:41,590 --> 00:44:45,910 to experts saying, "Show me your baby, 504 00:44:45,910 --> 00:44:48,860 "and I will tell you whether she's going to be a criminal." 505 00:44:51,300 --> 00:44:55,780 Now that we can predict it, we're going to then surveil 506 00:44:55,780 --> 00:45:01,100 those kids much more closely and we're going to jump on them 507 00:45:01,100 --> 00:45:03,700 at the first sign of a problem. 508 00:45:03,700 --> 00:45:06,700 And that's going to make for more effective policing. 509 00:45:07,000 --> 00:45:11,250 It does, but it's going to make for a really grim society 510 00:45:11,250 --> 00:45:16,250 and it's reinforcing dramatically existing injustices. 511 00:45:22,960 --> 00:45:27,810 - Imagine a world in which networks of CCTV cameras, 512 00:45:27,810 --> 00:45:30,930 drone surveillance cameras, have sophisticated 513 00:45:30,930 --> 00:45:34,440 face recognition technologies and are connected 514 00:45:34,440 --> 00:45:37,020 to other government surveillance databases. 515 00:45:38,080 --> 00:45:41,190 We will have the technology in place to have 516 00:45:41,190 --> 00:45:45,690 all of our movements comprehensively tracked and recorded. 517 00:45:47,680 --> 00:45:50,330 What that also means is that we will have 518 00:45:50,330 --> 00:45:52,940 created a surveillance time machine 519 00:45:52,940 --> 00:45:56,620 that will allow governments and powerful corporations 520 00:45:56,620 --> 00:45:58,710 to essentially hit rewind on our lives. 521 00:45:58,710 --> 00:46:01,780 We might not be under any suspicion now 522 00:46:01,780 --> 00:46:03,860 and five years from now, they might want to know 523 00:46:03,860 --> 00:46:08,100 more about us, and can then recreate granularly 524 00:46:08,100 --> 00:46:10,260 everything we've done, everyone we've seen, 525 00:46:10,260 --> 00:46:13,050 everyone we've been around over that entire period. 526 00:46:15,820 --> 00:46:19,320 That's an extraordinary amount of power 527 00:46:19,320 --> 00:46:21,660 for us to seed to anyone. 528 00:46:22,910 --> 00:46:25,370 And it's a world that I think has been difficult 529 00:46:25,370 --> 00:46:29,180 for people to imagine, but we've already built 530 00:46:29,180 --> 00:46:31,660 the architecture to enable that. 531 00:47:07,140 --> 00:47:10,290 - I'm a political reporter and I'm very interested 532 00:47:10,290 --> 00:47:14,670 in the ways powerful industries use their political power 533 00:47:14,670 --> 00:47:17,180 to influence the public policy process. 534 00:47:21,000 --> 00:47:24,420 The large tech companies and their lobbyists get together 535 00:47:24,420 --> 00:47:27,530 behind closed doors and are able to craft policies 536 00:47:27,530 --> 00:47:29,340 that we all have to live under. 537 00:47:30,790 --> 00:47:35,780 That's true for surveillance policies, for policies in terms of data collection, 538 00:47:35,780 --> 00:47:40,540 but also increasingly important when it comes to military and foreign policy. 539 00:47:43,930 --> 00:47:50,220 Starting in 2016, the Defense Department formed the Defense Innovation Board. 540 00:47:50,220 --> 00:47:54,010 That's a special body created to bring top tech executives 541 00:47:54,010 --> 00:47:56,420 into closer contact with the military. 542 00:47:59,140 --> 00:48:01,750 Eric Schmidt, former chairman of Alphabet, 543 00:48:01,750 --> 00:48:03,380 the parent company of Google, 544 00:48:03,380 --> 00:48:07,240 became the chairman of the Defense Innovation Board, 545 00:48:07,240 --> 00:48:09,930 and one of their first priorities was to say, 546 00:48:09,930 --> 00:48:13,850 "We need more artificial intelligence integrated into the military." 547 00:48:15,930 --> 00:48:19,190 - I've worked with a group of volunteers over the 548 00:48:19,190 --> 00:48:22,150 last couple of years to take a look at innovation in the 549 00:48:22,150 --> 00:48:26,780 overall military, and my summary conclusion is that we have 550 00:48:26,780 --> 00:48:30,460 fantastic people who are trapped in a very bad system. 551 00:48:33,290 --> 00:48:35,700 - From the Department of Defense's perspective, 552 00:48:35,700 --> 00:48:37,700 where I really started to get interested in it 553 00:48:37,700 --> 00:48:40,980 was when we started thinking about Unmanned Systems and 554 00:48:40,980 --> 00:48:45,980 how robotic and unmanned systems would start to change war. 555 00:48:46,160 --> 00:48:50,320 The smarter you made the Unmanned Systems and robots, 556 00:48:50,320 --> 00:48:53,570 the more powerful you might be able to make your military. 557 00:48:55,540 --> 00:48:57,240 - Under Secretary of Defense, 558 00:48:57,240 --> 00:49:00,420 Robert Work put together a major memo known as 559 00:49:00,420 --> 00:49:03,680 the Algorithmic Warfare Cross-Functional Team, 560 00:49:03,680 --> 00:49:05,540 better known as Project Maven. 561 00:49:07,830 --> 00:49:10,760 Eric Schmidt gave a number of speeches and media appearances 562 00:49:10,760 --> 00:49:14,310 where he said this effort was designed to increase fuel 563 00:49:14,310 --> 00:49:18,240 efficiency in the Air Force, to help with the logistics, 564 00:49:18,240 --> 00:49:21,300 but behind closed doors there was another parallel effort. 565 00:49:26,710 --> 00:49:29,760 Late in 2017 as part of Project Maven, 566 00:49:29,760 --> 00:49:33,710 Google, Eric Schmidt's firm, was tasked to secretly work 567 00:49:33,710 --> 00:49:36,100 on another part of Project Maven, 568 00:49:36,100 --> 00:49:41,480 and that was to take the vast volumes of image data vacuumed up 569 00:49:41,480 --> 00:49:47,060 by drones operating in Iraq and Afghanistan and to teach an AI 570 00:49:47,060 --> 00:49:49,690 to quickly identify targets on the battlefield. 571 00:49:52,600 --> 00:49:58,470 - We have a sensor and the sensor can do full motion video of an entire city. 572 00:49:58,470 --> 00:50:01,690 And we would have three seven-person teams working 573 00:50:01,690 --> 00:50:06,360 constantly and they could process 15% of the information. 574 00:50:06,360 --> 00:50:08,950 The other 85% of the information was just left 575 00:50:08,950 --> 00:50:14,040 on the cutting room floor, so we said, "Hey, AI and machine learning 576 00:50:14,040 --> 00:50:17,810 "would help us process 100% of the information." 577 00:50:25,380 --> 00:50:28,760 - Google has long had the motto, "Don't be evil." 578 00:50:28,760 --> 00:50:30,630 They have created a public image 579 00:50:30,630 --> 00:50:35,080 that they are devoted to public transparency. 580 00:50:35,080 --> 00:50:39,200 But for Google to slowly transform into a defense contractor, 581 00:50:39,200 --> 00:50:41,720 they maintained the utmost secrecy. 582 00:50:41,720 --> 00:50:45,280 And you had Google entering into this contract with most of the employees, 583 00:50:45,280 --> 00:50:49,020 even employees who were working on the program completely left in the dark. 584 00:51:01,840 --> 00:51:05,050 - Usually within Google, anyone in the company 585 00:51:05,050 --> 00:51:08,100 is allowed to know about any other project that's happening 586 00:51:08,100 --> 00:51:09,800 in some other part of the company. 587 00:51:11,180 --> 00:51:14,300 With Project Maven, the fact that it was kept secret, 588 00:51:14,300 --> 00:51:18,740 I think was alarming to people because that's not the norm at Google. 589 00:51:20,620 --> 00:51:22,940 - When this story was first revealed, 590 00:51:22,940 --> 00:51:25,540 it set off a firestorm within Google. 591 00:51:25,540 --> 00:51:28,420 You had a number of employees quitting in protests, 592 00:51:28,420 --> 00:51:31,900 others signing a petition objecting to this work. 593 00:51:34,130 --> 00:51:37,680 - You have to really say, "I don't want to be part of this anymore." 594 00:51:38,820 --> 00:51:41,750 There are companies called defense contractors 595 00:51:41,750 --> 00:51:45,830 and Google should just not be one of those companies 596 00:51:45,830 --> 00:51:50,040 because people need to trust Google for Google to work. 597 00:51:52,070 --> 00:51:55,710 - Good morning and welcome to Google I/O. 598 00:51:57,310 --> 00:52:00,160 - We've seen emails that show that Google simply 599 00:52:00,160 --> 00:52:03,350 continued to mislead their employees that the drone 600 00:52:03,350 --> 00:52:07,390 targeting program was only a minor effort that could at most 601 00:52:07,390 --> 00:52:11,170 be worth $9 million to the firm, which is drops in the bucket 602 00:52:11,170 --> 00:52:13,540 for a gigantic company like Google. 603 00:52:14,400 --> 00:52:17,300 But from internal emails that we obtained, 604 00:52:17,300 --> 00:52:21,530 Google was expecting Project Maven would ramp up to as much 605 00:52:21,530 --> 00:52:26,530 as $250 million, and that this entire effort would provide 606 00:52:26,630 --> 00:52:29,900 Google with Special Defense Department certification to make 607 00:52:29,900 --> 00:52:32,660 them available for even bigger defense contracts, 608 00:52:32,660 --> 00:52:34,980 some worth as much as $10 billion. 609 00:52:45,710 --> 00:52:49,750 The pressure for Google to compete for military contracts 610 00:52:49,750 --> 00:52:53,780 has come at a time when its competitors are also shifting their culture. 611 00:52:55,850 --> 00:53:00,070 Amazon, similarly pitching the military and law enforcement. 612 00:53:00,070 --> 00:53:02,280 IBM and other leading firms, 613 00:53:02,280 --> 00:53:04,890 they're pitching law enforcement and military. 614 00:53:05,840 --> 00:53:09,420 To stay competitive, Google has slowly transformed. 615 00:53:15,150 --> 00:53:18,910 - The Defense Science Board said of all of the 616 00:53:18,910 --> 00:53:22,170 technological advances that are happening right now, 617 00:53:22,170 --> 00:53:26,780 the single most important thing was artificial intelligence 618 00:53:26,780 --> 00:53:30,780 and the autonomous operations that it would lead. 619 00:53:30,780 --> 00:53:32,330 Are we investing enough? 620 00:53:37,810 --> 00:53:41,220 - Once we develop what are known as 621 00:53:41,220 --> 00:53:45,820 autonomous lethal weapons, in other words, 622 00:53:45,820 --> 00:53:51,510 weapons that are not controlled at all, they are genuinely autonomous, 623 00:53:51,510 --> 00:53:54,000 you've only got to get a president who says, 624 00:53:54,000 --> 00:53:56,410 "The hell with international law, we've got these weapons. 625 00:53:56,410 --> 00:53:58,560 "We're going to do what we want with them." 626 00:54:02,540 --> 00:54:03,960 - We're very close. 627 00:54:03,960 --> 00:54:06,340 When you have the hardware already set up 628 00:54:06,340 --> 00:54:10,360 and all you have to do is flip a switch to make it fully autonomous, 629 00:54:10,360 --> 00:54:13,240 what is it there that's stopping you from doing that? 630 00:54:14,670 --> 00:54:19,080 There's something really to be feared in war at machine speed. 631 00:54:19,930 --> 00:54:24,370 What if you're a machine and you've run millions and millions of different war scenarios 632 00:54:24,370 --> 00:54:28,060 and you have a team of drones and you've delegated control to half of them, 633 00:54:28,060 --> 00:54:30,790 and you're collaborating in real time? 634 00:54:30,790 --> 00:54:35,600 What happens when that swarm of drones is tasked with engaging a city? 635 00:54:37,490 --> 00:54:39,870 How will they take over that city? 636 00:54:39,870 --> 00:54:42,760 The answer is we won't know until it happens. 637 00:54:50,020 --> 00:54:53,910 - We do not want an AI system to decide 638 00:54:53,910 --> 00:54:56,180 what human it would attack, 639 00:54:56,180 --> 00:54:59,500 but we're going up against authoritarian competitors. 640 00:54:59,500 --> 00:55:02,450 So in my view, an authoritarian regime 641 00:55:02,450 --> 00:55:05,710 will have less problem delegating authority 642 00:55:05,710 --> 00:55:08,460 to a machine to make lethal decisions. 643 00:55:09,410 --> 00:55:12,660 So how that plays out remains to be seen. 644 00:55:37,270 --> 00:55:41,020 - Almost the gift of AI now is that it will force us 645 00:55:41,020 --> 00:55:44,410 collectively to think through at a very basic level, 646 00:55:44,410 --> 00:55:46,310 what does it mean to be human? 647 00:55:48,620 --> 00:55:50,510 What do I do as a human better 648 00:55:50,510 --> 00:55:52,970 than a certain super smart machine can do? 649 00:55:56,300 --> 00:56:01,000 First, we create our technology and then it recreates us. 650 00:56:01,000 --> 00:56:05,630 We need to make sure that we don't miss some of the things 651 00:56:05,630 --> 00:56:07,520 that make us so beautiful human. 652 00:56:11,840 --> 00:56:14,380 - Once we build intelligent machines, 653 00:56:14,380 --> 00:56:16,920 the philosophical vocabulary we have available to think 654 00:56:16,920 --> 00:56:20,680 about ourselves as human increasingly fails us. 655 00:56:23,160 --> 00:56:25,950 If I ask you to write up a list of all the terms you have 656 00:56:25,950 --> 00:56:28,960 available to describe yourself as human, 657 00:56:28,960 --> 00:56:31,100 there are not so many terms. 658 00:56:31,100 --> 00:56:38,140 Culture, history, sociality, maybe politics, civilization, 659 00:56:38,820 --> 00:56:45,090 subjectivity, all of these terms ground in two positions 660 00:56:45,670 --> 00:56:48,020 that humans are more than mere animals 661 00:56:49,000 --> 00:56:52,200 and that humans are more than mere machines. 662 00:56:55,140 --> 00:56:59,290 But if machines truly think there is a large set 663 00:56:59,290 --> 00:57:03,520 of key philosophical questions in which what is at stake is: 664 00:57:04,880 --> 00:57:09,080 Who are we? What is our place in the world? What is the world? How is it structured? 665 00:57:09,080 --> 00:57:12,190 Do the categories that we have relied on 666 00:57:12,190 --> 00:57:14,510 - Do they still work? Were they wrong? 667 00:57:19,500 --> 00:57:22,220 - Many people think of intelligence as something mysterious 668 00:57:22,220 --> 00:57:26,190 that can only exist inside of biological organisms, like us, 669 00:57:26,190 --> 00:57:29,260 but intelligence is all about information processing. 670 00:57:30,280 --> 00:57:32,230 It doesn't matter whether the intelligence is processed 671 00:57:32,230 --> 00:57:35,940 by carbon atoms inside of cells and brains, and people, 672 00:57:35,940 --> 00:57:38,020 or by silicon atoms in computers. 673 00:57:40,700 --> 00:57:43,320 Part of the success of AI recently has come 674 00:57:43,320 --> 00:57:47,540 from stealing great ideas from evolution. 675 00:57:47,540 --> 00:57:49,320 We noticed that the brain, for example, 676 00:57:49,320 --> 00:57:52,920 has all these neurons inside connected in complicated ways. 677 00:57:52,920 --> 00:57:55,660 So we stole that idea and abstracted it 678 00:57:55,660 --> 00:57:58,350 into artificial neural networks in computers, 679 00:57:59,330 --> 00:58:02,800 and that's what has revolutionized machine intelligence. 680 00:58:07,690 --> 00:58:10,300 If we one day get Artificial General Intelligence, 681 00:58:10,300 --> 00:58:14,330 then by definition, AI can also do better the job of AI 682 00:58:14,330 --> 00:58:18,930 programming and that means that further progress in making 683 00:58:18,930 --> 00:58:22,950 AI will be dominated not by human programmers, but by AI. 684 00:58:25,450 --> 00:58:29,090 Recursively self-improving AI could leave human intelligence 685 00:58:29,090 --> 00:58:32,990 far behind, creating super intelligence. 686 00:58:34,970 --> 00:58:37,970 It's gonna be the last invention we ever need to make, 687 00:58:37,970 --> 00:58:40,520 because it can then invent everything else 688 00:58:40,520 --> 00:58:42,010 much faster than we could. 689 00:59:54,040 --> 00:59:59,040 - There is a future that we all need to talk about. 690 00:59:59,110 --> 01:00:02,430 Some of the fundamental questions about the future 691 01:00:02,430 --> 01:00:06,280 of artificial intelligence, not just where it's going, 692 01:00:06,280 --> 01:00:09,630 but what it means for society to go there. 693 01:00:10,930 --> 01:00:14,620 It is not what computers can do, 694 01:00:14,620 --> 01:00:17,590 but what computers should do. 695 01:00:17,590 --> 01:00:22,210 As the generation of people that is bringing AI to the future, 696 01:00:22,210 --> 01:00:27,690 we are the generation that will answer this question first and foremost. 697 01:00:34,570 --> 01:00:37,380 - We haven't created the human-level thinking machine yet, 698 01:00:37,380 --> 01:00:39,410 but we get closer and closer. 699 01:00:41,100 --> 01:00:44,780 Maybe we'll get to human-level AI in five years from now 700 01:00:44,780 --> 01:00:47,410 or maybe it'll take 50 or 100 years from now. 701 01:00:47,660 --> 01:00:51,540 It almost doesn't matter. Like these are all really, really soon, 702 01:00:51,540 --> 01:00:56,060 in terms of the overall history of humanity. 703 01:01:01,200 --> 01:01:02,180 Very nice. 704 01:01:15,750 --> 01:01:19,350 So, the AI field is extremely international. 705 01:01:19,350 --> 01:01:23,790 China is up and coming and it's starting to rival the US, 706 01:01:23,790 --> 01:01:26,940 Europe and Japan in terms of putting a lot 707 01:01:26,940 --> 01:01:29,770 of processing power behind AI 708 01:01:29,770 --> 01:01:33,120 and gathering a lot of data to help AI learn. 709 01:01:35,960 --> 01:01:40,540 We have a young generation of Chinese researchers now. 710 01:01:40,540 --> 01:01:43,910 Nobody knows where the next revolution is going to come from. 711 01:01:50,120 --> 01:01:54,290 - China always wanted to become the superpower in the world. 712 01:01:56,120 --> 01:01:59,410 The Chinese government thinks AI gave them the chance to become 713 01:01:59,410 --> 01:02:03,940 one of the most advanced technology wise, business wise. 714 01:02:04,190 --> 01:02:07,780 So the Chinese government look at this as a huge opportunity. 715 01:02:09,140 --> 01:02:13,670 Like they've raised a flag and said, "That's a good field. 716 01:02:13,670 --> 01:02:16,300 "The companies should jump into it." 717 01:02:16,300 --> 01:02:18,300 Then China's commercial world and companies say, 718 01:02:18,300 --> 01:02:20,800 "Okay, the government raised a flag, that's good. 719 01:02:20,800 --> 01:02:22,500 "Let's put the money into it." 720 01:02:23,900 --> 01:02:28,090 Chinese tech giants, like Baidu, like Tencent and like AliBaba, 721 01:02:28,090 --> 01:02:31,800 they put a lot of the investment into the AI field. 722 01:02:33,410 --> 01:02:36,740 So we see that China's AI development is booming. 723 01:02:43,430 --> 01:02:47,370 - In China, everybody has Alipay and WeChat pay, 724 01:02:47,370 --> 01:02:49,500 so mobile payment is everywhere. 725 01:02:50,850 --> 01:02:54,850 And with that, they can do a lot of AI analysis 726 01:02:54,850 --> 01:02:59,340 to know like your spending habits, your credit rating. 727 01:03:01,150 --> 01:03:06,100 Face recognition technology is widely adopted in China, 728 01:03:06,100 --> 01:03:08,200 in airports and train stations. 729 01:03:09,040 --> 01:03:11,570 So, in the future, maybe in just a few months, 730 01:03:11,570 --> 01:03:15,140 you don't need a paper ticket to board a train. 731 01:03:15,140 --> 01:03:15,970 Only your face. 732 01:03:24,620 --> 01:03:29,690 - We generate the world's biggest platform of facial recognition. 733 01:03:31,260 --> 01:03:37,300 We have 300,000 developers using our platform. 734 01:03:38,330 --> 01:03:41,550 A lot of it is selfie camera apps. 735 01:03:41,550 --> 01:03:43,930 It makes you look more beautiful. 736 01:03:46,550 --> 01:03:49,780 There are millions and millions of cameras in the world, 737 01:03:50,760 --> 01:03:54,880 each camera from my point is a data generator. 738 01:03:58,820 --> 01:04:02,810 In a machine's eye, your face will change into the features 739 01:04:02,810 --> 01:04:06,660 and it will turn your face into a paragraph of code. 740 01:04:07,720 --> 01:04:11,980 So we can detect how old you are, if you're male or female, 741 01:04:11,980 --> 01:04:13,440 and your emotions. 742 01:04:16,640 --> 01:04:20,110 Shopping is about what kind of thing you are looking at. 743 01:04:20,110 --> 01:04:25,650 We can track your eyeballs, so if you are focusing on some product, 744 01:04:25,650 --> 01:04:28,690 we can track that so that we can know 745 01:04:28,690 --> 01:04:32,120 which kind of people like which kind of product. 746 01:04:42,450 --> 01:04:45,820 Our mission is to create a platform 747 01:04:45,950 --> 01:04:50,300 that will enable millions of AI developers in China. 748 01:04:51,280 --> 01:04:58,060 We study all the data we can get. 749 01:05:00,440 --> 01:05:04,350 Not just user profiles, 750 01:05:04,350 --> 01:05:08,380 but what you are doing at the moment, 751 01:05:09,460 --> 01:05:12,060 your geographical location. 752 01:05:15,380 --> 01:05:23,580 This platform will be so valuable that we don't even worry about profit now, 753 01:05:23,580 --> 01:05:27,660 because it is definitely there. 754 01:05:29,660 --> 01:05:34,420 China's social credit system is just one of the applications. 755 01:05:45,990 --> 01:05:48,040 - The Chinese government is using multiple 756 01:05:48,040 --> 01:05:50,810 different kinds of technologies, whether it's AI, 757 01:05:50,810 --> 01:05:53,650 whether it's big data platforms, facial recognition, 758 01:05:53,650 --> 01:05:57,030 voice recognition, essentially to monitor 759 01:05:57,030 --> 01:05:58,800 what the population is doing. 760 01:06:01,860 --> 01:06:04,240 I think the Chinese government has made very clear 761 01:06:04,240 --> 01:06:09,240 its intent to gather massive amounts of data about people 762 01:06:09,600 --> 01:06:13,300 to socially engineer a dissent-free society. 763 01:06:16,040 --> 01:06:19,060 The logic behind the Chinese government's 764 01:06:19,060 --> 01:06:23,180 social credit system, it's to take the idea that 765 01:06:23,180 --> 01:06:27,930 whether you are credit worthy for a financial loan 766 01:06:27,930 --> 01:06:31,860 and adding to it a very political dimension to say, 767 01:06:31,860 --> 01:06:34,270 "Are you a trustworthy human being? 768 01:06:36,530 --> 01:06:40,270 "What you've said online, have you ever been critical of the authorities? 769 01:06:40,270 --> 01:06:42,190 "Do you have a criminal record?" 770 01:06:43,720 --> 01:06:47,340 And all that information is packaged up together to rate 771 01:06:47,340 --> 01:06:51,820 you in ways that if you have performed well in their view, 772 01:06:51,820 --> 01:06:56,490 you'll have easier access to certain kinds of state services or benefits. 773 01:06:57,790 --> 01:06:59,710 But if you haven't done very well, 774 01:06:59,710 --> 01:07:02,130 you are going to be penalized or restricted. 775 01:07:05,940 --> 01:07:09,020 There's no way for people to challenge those designations 776 01:07:09,020 --> 01:07:12,100 or, in some cases, even know that they've been put in that category, 777 01:07:12,100 --> 01:07:17,990 and it's not until they try to access some kind of state service or buy a plane ticket, 778 01:07:17,990 --> 01:07:20,460 or get a passport, or enroll their kid in school, 779 01:07:20,460 --> 01:07:23,590 that they come to learn that they've been labeled in this way, 780 01:07:23,590 --> 01:07:27,550 and that there are negative consequences for them as a result. 781 01:07:44,970 --> 01:07:48,660 We've spent the better part of the last one or two years 782 01:07:48,660 --> 01:07:53,180 looking at abuses of surveillance technology across China, 783 01:07:53,180 --> 01:07:56,030 and a lot of that work has taken us to Xinjiang, 784 01:07:57,580 --> 01:08:01,220 the Northwestern region of China that has a more than half 785 01:08:01,220 --> 01:08:05,870 population of Turkic Muslims, Uyghurs, Kazakhs and Hui. 786 01:08:08,430 --> 01:08:10,980 This is a region and a population the Chinese government 787 01:08:10,980 --> 01:08:14,900 has long considered to be politically suspect or disloyal. 788 01:08:17,580 --> 01:08:20,160 We came to find information about what's called 789 01:08:20,160 --> 01:08:23,130 the Integrated Joint Operations Platform, 790 01:08:23,130 --> 01:08:27,190 which is a predictive policing program and that's one 791 01:08:27,190 --> 01:08:30,650 of the programs that has been spitting out lists of names 792 01:08:30,650 --> 01:08:33,560 of people to be subjected to political re-education. 793 01:08:39,090 --> 01:08:42,770 A number of our interviewees for the report we just released 794 01:08:42,770 --> 01:08:46,110 about the political education camps in Xinjiang just 795 01:08:46,110 --> 01:08:50,380 painted an extraordinary portrait of a surveillance state. 796 01:08:53,460 --> 01:08:56,380 A region awash in surveillance cameras 797 01:08:56,380 --> 01:09:00,600 for facial recognition purposes, checkpoints, body scanners, 798 01:09:00,600 --> 01:09:04,030 QR codes outside people's homes. 799 01:09:05,660 --> 01:09:10,690 Yeah, it really is the stuff of dystopian movies that we've all gone to and thought, 800 01:09:10,690 --> 01:09:13,160 "Wow, that would be a creepy world to live in." 801 01:09:13,160 --> 01:09:16,510 Yeah, well, 13 million Turkic Muslims in China 802 01:09:16,510 --> 01:09:18,530 are living in that reality right now. 803 01:09:30,840 --> 01:09:33,060 - The Intercept reports that Google is planning to 804 01:09:33,060 --> 01:09:36,490 launch a censored version of its search engine in China. 805 01:09:36,490 --> 01:09:38,730 - Google's search for new markets leads it 806 01:09:38,730 --> 01:09:42,170 to China, despite Beijing's rules on censorship. 807 01:09:42,170 --> 01:09:44,080 - Tell us more about why you felt it was 808 01:09:44,080 --> 01:09:46,650 your ethical responsibility to resign, 809 01:09:46,650 --> 01:09:51,060 because you talk about being complicit in censorship and oppression, and surveillance. 810 01:09:51,060 --> 01:09:54,220 - There is a Chinese venture company that has to be 811 01:09:54,220 --> 01:09:56,230 set up for Google to operate in China. 812 01:09:56,230 --> 01:09:58,920 And the question is, to what degree did they get to control 813 01:09:58,920 --> 01:10:01,970 the blacklist and to what degree would they have just 814 01:10:01,970 --> 01:10:05,220 unfettered access to surveilling Chinese citizens? 815 01:10:05,220 --> 01:10:07,470 And the fact that Google refuses to respond 816 01:10:07,470 --> 01:10:09,250 to human rights organizations on this, 817 01:10:09,250 --> 01:10:12,060 I think should be extremely disturbing to everyone. 818 01:10:16,640 --> 01:10:21,120 Due to my conviction that dissent is fundamental to functioning democracies 819 01:10:21,120 --> 01:10:25,560 and forced to resign in order to avoid contributing to or profiting from the erosion 820 01:10:25,560 --> 01:10:27,510 of protections for dissidents. 821 01:10:28,560 --> 01:10:30,920 The UN is currently reporting that between 822 01:10:30,920 --> 01:10:34,190 200,000 and one million Uyghurs have been disappeared 823 01:10:34,190 --> 01:10:36,520 into re-education camps. 824 01:10:36,520 --> 01:10:39,290 And there is a serious argument that Google would be complicit 825 01:10:39,290 --> 01:10:42,460 should it launch a surveilled version of search in China. 826 01:10:45,180 --> 01:10:51,380 Dragonfly is a project meant to launch search in China under 827 01:10:51,390 --> 01:10:55,810 Chinese government regulations, which include censoring 828 01:10:55,810 --> 01:10:59,510 sensitive content, basic queries on human rights, 829 01:10:59,510 --> 01:11:03,210 information about political representatives is blocked, 830 01:11:03,210 --> 01:11:07,080 information about student protests is blocked. 831 01:11:07,080 --> 01:11:09,180 And that's one small part of it. 832 01:11:09,180 --> 01:11:12,250 Perhaps a deeper concern is the surveillance side of this. 833 01:11:16,070 --> 01:11:19,580 When I raised the issue with my managers, with my colleagues, 834 01:11:19,580 --> 01:11:23,370 there was a lot of concern, but everyone just said, "I don't know anything." 835 01:11:27,760 --> 01:11:30,200 And then when there was a meeting finally, 836 01:11:30,200 --> 01:11:35,080 there was essentially no addressing the serious concerns associated with it. 837 01:11:37,010 --> 01:11:39,790 So then I filed my formal resignation, 838 01:11:39,790 --> 01:11:42,950 not just to my manager, but I actually distributed it company-wide. 839 01:11:42,950 --> 01:11:45,440 And that's the letter that I was reading from. 840 01:11:50,360 --> 01:11:55,960 Personally, I haven't slept well. I've had pretty horrific headaches, 841 01:11:55,960 --> 01:11:58,520 wake up in the middle of the night just sweating. 842 01:12:00,640 --> 01:12:03,340 With that said, what I found since speaking out 843 01:12:03,340 --> 01:12:07,930 is just how positive the global response to this has been. 844 01:12:10,350 --> 01:12:13,910 Engineers should demand to know what the uses 845 01:12:13,910 --> 01:12:16,280 of their technical contributions are 846 01:12:16,280 --> 01:12:19,340 and to have a seat at the table in those ethical decisions. 847 01:12:27,020 --> 01:12:29,590 Most citizens don't really understand what it means to be 848 01:12:29,590 --> 01:12:32,250 in a very large scale prescriptive technology. 849 01:12:33,720 --> 01:12:36,260 Where someone has already pre-divided the work 850 01:12:36,260 --> 01:12:38,300 and all you know about is your little piece, 851 01:12:38,300 --> 01:12:41,150 and almost certainly you don't understand how it fits in. 852 01:12:44,630 --> 01:12:51,040 So, I think it's worth drawing the analogy to physicists' work on the atomic bomb. 853 01:12:53,840 --> 01:12:56,610 In fact, that's actually the community I came out of. 854 01:12:59,220 --> 01:13:03,460 I wasn't a nuclear scientist by any means, but I was an applied mathematician 855 01:13:04,400 --> 01:13:09,400 and my PhD program was actually funded to train people to work in weapons labs. 856 01:13:12,980 --> 01:13:17,100 One could certainly argue that there is an existential threat 857 01:13:17,820 --> 01:13:22,060 and whoever is leading in AI will lead militarily. 858 01:13:31,420 --> 01:13:34,250 - China fully expects to pass the United States 859 01:13:34,250 --> 01:13:37,480 as the number one economy in the world and it believes that 860 01:13:37,480 --> 01:13:42,450 AI will make that jump more quickly and more dramatically. 861 01:13:43,900 --> 01:13:46,480 And they also see it as being able to leapfrog 862 01:13:46,480 --> 01:13:49,820 the United States in terms of military power. 863 01:13:57,580 --> 01:13:59,600 Their plan is very simple. 864 01:13:59,600 --> 01:14:03,620 We want to catch the United States and these technologies by 2020, 865 01:14:03,620 --> 01:14:07,940 we want to surpass the United States in these technologies by 2025, 866 01:14:07,940 --> 01:14:12,600 and we want to be the world leader in AI and autonomous technologies by 2030. 867 01:14:15,110 --> 01:14:16,850 It is a national plan. 868 01:14:16,850 --> 01:14:21,790 It is backed up by at least $150 billion in investments. 869 01:14:21,790 --> 01:14:25,060 So, this is definitely a race. 870 01:14:50,070 --> 01:14:52,070 - AI is a little bit like fire. 871 01:14:53,360 --> 01:14:56,270 Fire was invented 700,000 years ago, 872 01:14:57,480 --> 01:14:59,820 and it has its pros and cons. 873 01:15:01,960 --> 01:15:06,810 People realized you can use fire to keep warm at night and to cook, 874 01:15:08,520 --> 01:15:14,420 but they also realized that you can kill other people with that. 875 01:15:19,860 --> 01:15:24,300 Fire also has this AI-like quality of growing 876 01:15:24,300 --> 01:15:27,400 in a wildfire without further human ado, 877 01:15:30,530 --> 01:15:35,970 but the advantages outweigh the disadvantages by so much 878 01:15:35,970 --> 01:15:39,760 that we are not going to stop its development. 879 01:15:49,540 --> 01:15:51,180 Europe is waking up. 880 01:15:52,160 --> 01:15:58,390 Lots of companies in Europe are realizing that the next wave of AI 881 01:15:58,730 --> 01:16:01,540 will be much bigger than the current wave. 882 01:16:03,310 --> 01:16:08,000 The next wave of AI will be about robots. 883 01:16:09,810 --> 01:16:15,450 All these machines that make things, that produce stuff, 884 01:16:15,450 --> 01:16:19,370 that build other machines, they are going to become smart. 885 01:16:26,660 --> 01:16:29,950 In the not-so-distant future, we will have robots 886 01:16:29,950 --> 01:16:33,290 that we can teach like we teach kids. 887 01:16:35,640 --> 01:16:38,660 For example, I will talk to a little robot and I will say, 888 01:16:39,970 --> 01:16:42,550 "Look here, robot, look here. 889 01:16:42,550 --> 01:16:44,620 "Let's assemble a smartphone. 890 01:16:44,620 --> 01:16:48,720 "We take this slab of plastic like that and we takes a screwdriver like that, 891 01:16:48,720 --> 01:16:52,310 "and now we screw in everything like this. 892 01:16:52,310 --> 01:16:54,310 "No, no, not like this. 893 01:16:54,310 --> 01:16:57,690 "Like this, look, robot, look, like this." 894 01:16:58,820 --> 01:17:01,890 And he will fail a couple of times but rather quickly, 895 01:17:01,890 --> 01:17:05,860 he will learn to do the same thing much better than I could do it. 896 01:17:06,710 --> 01:17:11,400 And then we stop the learning and we make a million copies, and sell it. 897 01:17:32,410 --> 01:17:35,800 Regulation of AI sounds like an attractive idea, 898 01:17:35,800 --> 01:17:38,210 but I don't think it's possible. 899 01:17:40,250 --> 01:17:42,720 One of the reasons why it won't work is 900 01:17:42,720 --> 01:17:46,360 the sheer curiosity of scientists. 901 01:17:47,260 --> 01:17:49,560 They don't give a damn for regulation. 902 01:17:52,640 --> 01:17:56,740 Military powers won't give a damn for regulations, either. 903 01:17:56,740 --> 01:17:59,020 They will say, "If we, the Americans don't do it, 904 01:17:59,020 --> 01:18:00,720 "then the Chinese will do it." 905 01:18:00,720 --> 01:18:03,150 And the Chinese will say, "Oh, if we don't do it, 906 01:18:03,150 --> 01:18:04,810 "then the Russians will do it." 907 01:18:07,700 --> 01:18:10,930 No matter what kind of political regulation is out there, 908 01:18:10,930 --> 01:18:14,660 all these military industrial complexes, 909 01:18:14,660 --> 01:18:17,570 they will almost by definition have to ignore that 910 01:18:18,670 --> 01:18:21,250 because they want to avoid falling behind. 911 01:18:26,030 --> 01:18:27,690 Welcome to Xinhua. 912 01:18:27,690 --> 01:18:30,260 I'm the world's first female AI news anchor developed 913 01:18:30,260 --> 01:18:32,690 jointly by Xinhua and search engine company Sogou. 914 01:18:33,050 --> 01:18:36,310 - A program developed by the company OpenAI can write 915 01:18:36,310 --> 01:18:39,500 coherent and credible stories just like human beings. 916 01:18:39,500 --> 01:18:41,520 - It's one small step for machine, 917 01:18:41,520 --> 01:18:44,330 one giant leap for machine kind. 918 01:18:44,330 --> 01:18:47,170 IBM's newest artificial intelligence system took on 919 01:18:47,170 --> 01:18:51,570 experienced human debaters and won a live debate. 920 01:18:51,570 --> 01:18:54,160 - Computer-generated videos known as deep fakes 921 01:18:54,160 --> 01:18:57,980 are being used to put women's faces on pornographic videos. 922 01:19:02,510 --> 01:19:06,450 - Artificial intelligence evolves at a very crazy pace. 923 01:19:07,970 --> 01:19:10,180 You know, it's like progressing so fast. 924 01:19:10,180 --> 01:19:12,940 In some ways, we're only at the beginning right now. 925 01:19:14,490 --> 01:19:17,620 You have so many potential applications, it's a gold mine. 926 01:19:19,590 --> 01:19:23,450 Since 2012, when deep learning became a big game changer 927 01:19:23,450 --> 01:19:25,370 in the computer vision community, 928 01:19:25,370 --> 01:19:28,730 we were one of the first to actually adopt deep learning 929 01:19:28,730 --> 01:19:31,080 and apply it in the field of computer graphics. 930 01:19:34,640 --> 01:19:39,800 A lot of our research is funded by government, military intelligence agencies. 931 01:19:43,950 --> 01:19:47,840 The way we create these photoreal mappings, 932 01:19:47,840 --> 01:19:50,300 usually the way it works is that we need two subjects, 933 01:19:50,300 --> 01:19:53,740 a source and a target, and I can do a face replacement. 934 01:19:58,880 --> 01:20:03,700 One of the applications is, for example, I want to manipulate someone's face 935 01:20:03,700 --> 01:20:05,450 saying things that he did not. 936 01:20:08,920 --> 01:20:12,140 It can be used for creative things, for funny contents, 937 01:20:12,140 --> 01:20:16,110 but obviously, it can also be used for just simply 938 01:20:16,110 --> 01:20:18,660 manipulate videos and generate fake news. 939 01:20:21,050 --> 01:20:23,340 This can be very dangerous. 940 01:20:24,790 --> 01:20:29,300 If it gets into the wrong hands, it can get out of control very quickly. 941 01:20:33,370 --> 01:20:35,750 - We're entering an era in which our enemies can 942 01:20:35,750 --> 01:20:39,590 make it look like anyone is saying anything at any point in time, 943 01:20:39,590 --> 01:20:41,830 even if they would never say those things. 944 01:20:42,450 --> 01:20:46,970 Moving forward, we need to be more vigilant with what we trust from the Internet. 945 01:20:46,970 --> 01:20:50,710 It may sound basic, but how we move forward 946 01:20:50,710 --> 01:20:55,180 in the age of information is going to be the difference 947 01:20:55,180 --> 01:20:58,160 between whether we survive or whether we become 948 01:20:58,160 --> 01:21:00,270 some kind of fucked up dystopia. 949 01:22:31,070 --> 01:22:37,337 - One criticism that is frequently raised against my work is saying that, 950 01:22:37,337 --> 01:22:43,610 "Hey, you know there were stupid ideas in the past like phrenology or physiognomy.- 951 01:22:45,110 --> 01:22:49,210 - "There were people claiming that you can read a character 952 01:22:49,210 --> 01:22:52,010 "of a person just based on their face." 953 01:22:53,820 --> 01:22:56,050 People would say, "This is rubbish. 954 01:22:56,050 --> 01:23:01,050 "We know it was just thinly veiled racism and superstition." 955 01:23:05,300 --> 01:23:09,650 But the fact that someone made a claim in the past 956 01:23:09,650 --> 01:23:14,650 and tried to support this claim with invalid reasoning, 957 01:23:14,960 --> 01:23:18,970 doesn't automatically invalidate the claim. 958 01:23:23,790 --> 01:23:25,710 Of course, people should have rights 959 01:23:25,710 --> 01:23:31,020 to their privacy when it comes to sexual orientation or political views, 960 01:23:32,310 --> 01:23:35,600 but I'm also afraid that in the current technological environment, 961 01:23:35,600 --> 01:23:38,220 this is essentially impossible. 962 01:23:42,640 --> 01:23:44,910 People should realize there's no going back. 963 01:23:44,910 --> 01:23:48,330 There's no running away from the algorithms. 964 01:23:51,120 --> 01:23:55,670 The sooner we accept the inevitable and inconvenient truth 965 01:23:56,540 --> 01:23:59,430 that privacy is gone, 966 01:24:01,770 --> 01:24:05,380 the sooner we can actually start thinking about 967 01:24:05,380 --> 01:24:07,440 how to make sure that our societies 968 01:24:07,440 --> 01:24:11,660 are ready for the Post-Privacy Age. 969 01:24:35,080 --> 01:24:37,610 - While speaking about facial recognition, 970 01:24:37,610 --> 01:24:43,700 in my deep thoughts, I sometimes get to the very dark era of our history. 971 01:24:45,490 --> 01:24:49,020 When the people had to live in the system, 972 01:24:49,020 --> 01:24:52,660 where some part of the society was accepted 973 01:24:53,680 --> 01:24:56,900 and some part of the society was accused to death. 974 01:25:01,670 --> 01:25:06,180 What would Mengele do to have such an instrument in his hands? 975 01:25:10,740 --> 01:25:14,390 It would be very quick and efficient for selection 976 01:25:18,390 --> 01:25:22,100 and this is the apocalyptic vision. 977 01:26:15,880 --> 01:26:19,870 - So in the near future, the entire story of you 978 01:26:19,870 --> 01:26:25,340 will exist in a vast array of connected databases of faces, 979 01:26:25,750 --> 01:26:28,390 genomes, behaviors and emotion. 980 01:26:30,950 --> 01:26:35,320 So, you will have a digital avatar of yourself online, 981 01:26:35,320 --> 01:26:39,510 which records how well you are doing as a citizen, 982 01:26:39,510 --> 01:26:41,910 what kind of a relationship do you have, 983 01:26:41,910 --> 01:26:45,890 what kind of political orientation and sexual orientation. 984 01:26:50,020 --> 01:26:54,120 Based on all of those data, those algorithms will be able to 985 01:26:54,120 --> 01:26:58,390 manipulate your behavior with an extreme precision, 986 01:26:58,390 --> 01:27:03,890 changing how we think and probably in the future, how we feel. 987 01:27:25,660 --> 01:27:30,510 - The beliefs and desires of the first AGIs will be extremely important. 988 01:27:32,730 --> 01:27:35,180 So, it's important to program them correctly. 989 01:27:36,200 --> 01:27:37,850 I think that if this is not done, 990 01:27:38,810 --> 01:27:44,190 then the nature of evolution of natural selection will favor 991 01:27:44,190 --> 01:27:48,340 those systems, prioritize their own survival above all else. 992 01:27:51,860 --> 01:27:56,660 It's not that it's going to actively hate humans and want to harm them, 993 01:27:58,780 --> 01:28:01,100 but it's just going to be too powerful 994 01:28:01,400 --> 01:28:05,470 and I think a good analogy would be the way humans treat animals. 995 01:28:06,970 --> 01:28:08,200 It's not that we hate animals. 996 01:28:08,260 --> 01:28:11,930 I think humans love animals and have a lot of affection for them, 997 01:28:12,340 --> 01:28:17,610 but when the time comes to build a highway between two cities, 998 01:28:17,610 --> 01:28:20,130 we are not asking the animals for permission. 999 01:28:20,130 --> 01:28:23,340 We just do it because it's important for us. 1000 01:28:24,250 --> 01:28:29,010 And I think by default, that's the kind of relationship that's going to be between us 1001 01:28:29,010 --> 01:28:34,700 and AGIs which are truly autonomous and operating on their own behalf. 1002 01:28:47,190 --> 01:28:51,400 If you have an arms-race dynamics between multiple kings 1003 01:28:51,400 --> 01:28:54,140 trying to build the AGI first, 1004 01:28:54,140 --> 01:28:58,000 they will have less time to make sure that the AGI that they build 1005 01:28:58,970 --> 01:29:00,520 will care deeply for humans. 1006 01:29:03,530 --> 01:29:06,540 Because the way I imagine it is that there is an avalanche, 1007 01:29:06,540 --> 01:29:09,220 there is an avalanche of AGI development. 1008 01:29:09,220 --> 01:29:12,300 Imagine it's a huge unstoppable force. 1009 01:29:15,780 --> 01:29:18,740 And I think it's pretty likely the entire surface of the 1010 01:29:18,740 --> 01:29:22,070 earth would be covered with solar panels and data centers. 1011 01:29:26,230 --> 01:29:28,700 Given these kinds of concerns, 1012 01:29:28,700 --> 01:29:32,830 it will be important that the AGI is somehow built 1013 01:29:32,830 --> 01:29:35,570 as a cooperation with multiple countries. 1014 01:29:38,060 --> 01:29:41,180 The future is going to be good for the AIs, regardless. 1015 01:29:42,130 --> 01:29:45,020 It would be nice if it would be good for humans as well. 1016 01:30:06,750 --> 01:30:10,580 - Is there a lot of responsibility weighing on my shoulders? 1017 01:30:10,580 --> 01:30:11,800 Not really. 1018 01:30:13,420 --> 01:30:16,740 Was there a lot of responsibility on the shoulders 1019 01:30:16,740 --> 01:30:19,100 of the parents of Einstein? 1020 01:30:20,190 --> 01:30:21,780 The parents somehow made him, 1021 01:30:21,780 --> 01:30:25,450 but they had no way of predicting what he would do, 1022 01:30:25,450 --> 01:30:27,300 and how he would change the world. 1023 01:30:28,580 --> 01:30:32,540 And so, you can't really hold them responsible for that. 1024 01:30:57,640 --> 01:31:00,240 So, I'm not a very human-centric person. 1025 01:31:01,840 --> 01:31:05,500 I think I'm a little stepping stone in the evolution 1026 01:31:05,500 --> 01:31:08,060 of the Universe towards higher complexity. 1027 01:31:10,820 --> 01:31:14,860 But it's also clear to me that I'm not the crown of creation 1028 01:31:14,860 --> 01:31:19,320 and that humankind as a whole is not the crown of creation, 1029 01:31:21,140 --> 01:31:26,290 but we are setting the stage for something that is bigger than us that transcends us. 1030 01:31:28,980 --> 01:31:32,780 And then will go out there in a way where humans cannot follow 1031 01:31:32,780 --> 01:31:37,940 and transform the entire Universe, or at least, the reachable Universe. 1032 01:31:41,520 --> 01:31:46,520 So, I find beauty and awe in seeing myself 1033 01:31:46,640 --> 01:31:50,260 as part of this much grander theme. 1034 01:32:14,070 --> 01:32:16,030 - AI is inevitable. 1035 01:32:17,560 --> 01:32:23,370 We need to make sure we have the necessary human regulation 1036 01:32:23,370 --> 01:32:27,940 to prevent the weaponization of artificial intelligence. 1037 01:32:28,690 --> 01:32:32,000 We don't need any more weaponization 1038 01:32:32,000 --> 01:32:33,960 of such a powerful tool. 1039 01:32:37,100 --> 01:32:41,960 - One of the most critical things, I think, is the need for international governance. 1040 01:32:43,920 --> 01:32:46,310 We have an imbalance of power here because now 1041 01:32:46,310 --> 01:32:49,140 we have corporations with more power, might and ability, 1042 01:32:49,140 --> 01:32:51,100 than entire countries. 1043 01:32:51,100 --> 01:32:54,050 How do we make sure that people's voices are getting heard? 1044 01:32:58,030 --> 01:32:59,930 - It can't be a law-free zone. 1045 01:32:59,930 --> 01:33:02,100 It can't be a rights-free zone. 1046 01:33:02,100 --> 01:33:05,870 We can't embrace all of these wonderful new technologies 1047 01:33:05,870 --> 01:33:10,380 for the 21st century without trying to bring with us 1048 01:33:10,380 --> 01:33:15,380 the package of human rights that we fought so hard 1049 01:33:15,640 --> 01:33:19,190 to achieve, and that remains so fragile. 1050 01:33:28,680 --> 01:33:32,180 - AI isn't good and it isn't evil, either. 1051 01:33:32,180 --> 01:33:36,890 It's just going to amplify the desires and goals of whoever controls it. 1052 01:33:36,890 --> 01:33:41,240 And AI today is under the control of a very, very small group of people. 1053 01:33:44,430 --> 01:33:49,260 The most important question that we humans have to ask ourselves at this point in history 1054 01:33:49,260 --> 01:33:51,380 requires no technical knowledge. 1055 01:33:51,590 --> 01:33:57,540 It's the question of what sort of future society do we want to create 1056 01:33:57,540 --> 01:33:59,740 with all this technology we're making? 1057 01:34:01,410 --> 01:34:05,110 What do we want the role of humans to be in this world?