1 00:00:01,735 --> 00:00:03,268 BEN: It's the DNA of the next tech revolution. 2 00:00:03,304 --> 00:00:06,872 This is such a huge data set that there's no way that a human 3 00:00:06,907 --> 00:00:09,374 or even a team of humans can look at all of it. 4 00:00:09,410 --> 00:00:12,144 The race for artificial intelligence is on. 5 00:00:12,179 --> 00:00:14,813 In quantum, you have this thing called a qubit, 6 00:00:14,848 --> 00:00:16,682 which can be zero and one at the same time, 7 00:00:16,717 --> 00:00:18,317 and that's where the power comes from. 8 00:00:18,352 --> 00:00:20,986 AI might make our lives better... 9 00:00:21,021 --> 00:00:23,288 The life expectancy for human civilization 10 00:00:23,324 --> 00:00:25,857 might then easily measure in billions of years. 11 00:00:25,893 --> 00:00:27,893 ...but could it destroy the human race? 12 00:00:27,928 --> 00:00:30,229 You ought to be really concerned about the strong AI 13 00:00:30,264 --> 00:00:31,530 that has guns on it. 14 00:00:31,565 --> 00:00:41,573 ♪ 15 00:00:51,018 --> 00:00:53,151 As the host of Cyberwar, I've travelled the world to talk to 16 00:00:53,187 --> 00:00:56,088 brilliant hackers, scientists and programmers, 17 00:00:56,123 --> 00:00:58,257 and many of them have told me they're thinking about 18 00:00:58,292 --> 00:00:59,891 artificial intelligence. 19 00:01:01,228 --> 00:01:03,562 Today we're in the middle of an AI boom. 20 00:01:03,597 --> 00:01:05,330 Huge advancements in artificial intelligence 21 00:01:05,366 --> 00:01:08,867 have made things like Siri and self-driving cars possible. 22 00:01:08,902 --> 00:01:11,937 AI is in video games, security surveillance systems, 23 00:01:11,972 --> 00:01:15,107 smart home devices, and advanced weapons systems. 24 00:01:17,077 --> 00:01:19,611 But as researchers race to make the next breakthrough, 25 00:01:19,647 --> 00:01:22,014 a lot are warning that we haven't actually 26 00:01:22,049 --> 00:01:23,949 thought this through. 27 00:01:23,984 --> 00:01:26,852 And it's not necessarily the Terminator scenario 28 00:01:26,887 --> 00:01:29,621 they're afraid of, where an AI system becomes self aware 29 00:01:29,657 --> 00:01:31,290 and decides to kill us all. 30 00:01:33,294 --> 00:01:35,694 The real threat could lie in the unintended consequences 31 00:01:35,729 --> 00:01:37,763 no one sees coming. 32 00:01:37,798 --> 00:01:40,866 I'm at Stanford University, where researchers are designing 33 00:01:40,901 --> 00:01:44,436 AI that comes in a very non-threatening package. 34 00:01:44,471 --> 00:01:48,106 Meet Jackrabbot, a cute little machine that's built to navigate 35 00:01:48,142 --> 00:01:50,776 and move through crowds of human buddies. 36 00:01:50,811 --> 00:01:52,744 This is a robot that's programmed to learn on its own, 37 00:01:52,780 --> 00:01:54,379 through example. 38 00:01:56,183 --> 00:01:58,250 This technique, called "deep learning", 39 00:01:58,285 --> 00:02:00,952 has been fuelling a lot of recent AI breakthroughs. 40 00:02:02,189 --> 00:02:05,123 Alexandre Alahi and Alexandre Robicquet are part of a team 41 00:02:05,159 --> 00:02:08,260 that built Jackrabbot, and they say their little creation 42 00:02:08,295 --> 00:02:10,595 is more advanced than a self-driving car. 43 00:02:12,366 --> 00:02:14,499 So the robot enter the scene, analyze it, and then 44 00:02:14,535 --> 00:02:17,336 navigate through it according to all the data that we gather. 45 00:02:17,371 --> 00:02:20,505 So it's just watching people, and looking at people, 46 00:02:20,541 --> 00:02:23,108 and what they're doing, and judging, you know, 47 00:02:23,143 --> 00:02:25,077 the space and the time? 48 00:02:25,112 --> 00:02:27,145 Yeah, and also trying to understand how they interact 49 00:02:27,181 --> 00:02:28,980 between each other. 50 00:02:29,016 --> 00:02:31,683 The main aspect would be to understand well, how... 51 00:02:31,719 --> 00:02:33,752 what is a safe distance that I keep with someone else, 52 00:02:33,787 --> 00:02:36,254 or how do I accelerate or decelerate when I get close 53 00:02:36,290 --> 00:02:38,623 to someone and react accordingly? 54 00:02:38,659 --> 00:02:41,493 So how do you gather that information? 55 00:02:41,528 --> 00:02:43,528 You just like... this just looks like 56 00:02:43,564 --> 00:02:46,398 some sort of surveillance footage. 57 00:02:46,433 --> 00:02:51,937 So last summer, our team spent like 2 months gathering daily 58 00:02:51,972 --> 00:02:56,641 at rush hour, top view data from the Stanford crowd, 59 00:02:56,677 --> 00:02:59,211 so we can actually understand how they avoid, 60 00:02:59,246 --> 00:03:02,414 how they behave and how they navigate in such a crowd. 61 00:03:02,449 --> 00:03:04,349 Let's go, robot. 62 00:03:09,022 --> 00:03:10,389 Free! 63 00:03:11,859 --> 00:03:13,692 So wait, he's being controlled by you? 64 00:03:13,727 --> 00:03:15,227 For right now, yeah. 65 00:03:15,262 --> 00:03:17,295 So does he have a fully autonomous mode? 66 00:03:17,331 --> 00:03:18,530 Yeah, he could. 67 00:03:18,565 --> 00:03:20,198 And what does he do? He just follows? 68 00:03:20,234 --> 00:03:22,501 No, we actually can give him-- like if you map the place 69 00:03:22,536 --> 00:03:24,703 and the area, he can go from point A to point B, 70 00:03:24,738 --> 00:03:27,072 but we're just having a simple collision avoidance right now. 71 00:03:27,107 --> 00:03:29,508 But like you train to avoid every single obstacle 72 00:03:29,543 --> 00:03:31,710 and go where you want him to go. 73 00:03:31,745 --> 00:03:32,911 Can we drive him around a little bit? 74 00:03:32,946 --> 00:03:34,513 Yeah, absolutely. 75 00:03:34,548 --> 00:03:37,516 And those-- that spinning thing, that's just like a 3D sensor? 76 00:03:37,551 --> 00:03:39,351 Right, right. 77 00:03:39,386 --> 00:03:42,354 So we're capturing depth data, visual data, 78 00:03:42,389 --> 00:03:46,491 and we're combining all these sensors to detect humans, 79 00:03:46,527 --> 00:03:49,728 understand its surrounding to look at itself, 80 00:03:49,763 --> 00:03:52,030 and predict also where people will go. 81 00:03:53,467 --> 00:03:54,900 Do people freak out when they see it a lot? 82 00:03:54,935 --> 00:03:55,901 No, they like it a lot. 83 00:03:55,936 --> 00:03:57,736 They come hug it, they talk to it. 84 00:03:57,771 --> 00:03:59,738 He doesn't... He cannot talk yet. 85 00:03:59,773 --> 00:04:01,873 Is that what you hope the rise of the machines is? 86 00:04:01,909 --> 00:04:03,542 It's a friendly rise of the machines? 87 00:04:03,577 --> 00:04:04,676 For that one, yes. 88 00:04:04,711 --> 00:04:05,877 - Yeah. - Absolutely. 89 00:04:07,548 --> 00:04:09,915 Jackrabbot's creators hope that as AI advances, 90 00:04:09,950 --> 00:04:12,250 machines like this will be built to carry luggage 91 00:04:12,286 --> 00:04:14,252 through airports, or help the blind navigate 92 00:04:14,288 --> 00:04:16,421 through pedestrian traffic. 93 00:04:16,457 --> 00:04:18,557 The Stanford team is working to get AI to the point 94 00:04:18,592 --> 00:04:20,225 where Jackrabbots can one day do 95 00:04:20,260 --> 00:04:22,928 what most people can just do by nature. 96 00:04:22,963 --> 00:04:25,931 But today's computers are already performing tasks 97 00:04:25,966 --> 00:04:28,200 no human is capable of performing, 98 00:04:28,235 --> 00:04:31,069 even if it's not quite AI. 99 00:04:31,104 --> 00:04:33,371 That technology continues to evolve, 100 00:04:33,407 --> 00:04:36,241 and the so-called supercomputer is at the cutting edge. 101 00:04:39,112 --> 00:04:42,047 Bryan Biegel works in NASA's Advanced Supercomputing 102 00:04:42,082 --> 00:04:44,783 Division, where they're working with machines that can actually 103 00:04:44,818 --> 00:04:46,318 help us see into the future... 104 00:04:46,353 --> 00:04:48,553 Let's go to the vislab, and I'll show you some of the things 105 00:04:48,589 --> 00:04:50,922 that our users do with our supercomputers. 106 00:04:50,958 --> 00:04:54,593 ...using code so advanced it took 15 years to write. 107 00:04:56,163 --> 00:04:58,763 Yeah, so this is actually a simulation of Earth's oceans. 108 00:04:58,799 --> 00:05:02,167 This is... It uses measured data from NASA's satellites 109 00:05:02,202 --> 00:05:04,636 and puts it into a really huge computer model 110 00:05:04,671 --> 00:05:08,273 on the supercomputers, and predicts the ocean behaviour. 111 00:05:08,308 --> 00:05:12,277 So this single simulation actually used up to 70,000 112 00:05:12,312 --> 00:05:16,014 processors, and generated over 3 petabytes of data. 113 00:05:16,049 --> 00:05:18,750 That's like 10,000 times what you would have 114 00:05:18,785 --> 00:05:20,285 on your entire laptop. 115 00:05:20,320 --> 00:05:21,853 Let me show you this in full scale. 116 00:05:21,889 --> 00:05:24,322 This is actually scaled down so you can see it on the monitors, 117 00:05:24,358 --> 00:05:25,924 but if I go to the next one, 118 00:05:25,959 --> 00:05:29,494 you can see what the full-scale version of this simulation is. 119 00:05:29,530 --> 00:05:33,298 For example, this is the ice over Antarctic Ocean, 120 00:05:33,333 --> 00:05:35,133 and you can see where it cracks, 121 00:05:35,168 --> 00:05:36,968 and it sends ripples out into the ocean. 122 00:05:37,004 --> 00:05:38,803 And they're working on... 123 00:05:38,839 --> 00:05:41,139 they're still continuing to advance the accuracy of this 124 00:05:41,174 --> 00:05:43,508 model so they can predict decades into the future. 125 00:05:43,544 --> 00:05:46,945 Would you say this is artificial intelligence? 126 00:05:46,980 --> 00:05:48,713 It seems like it's getting there, right? 127 00:05:48,749 --> 00:05:52,350 It's so complex, but we have programmed in this case 128 00:05:52,386 --> 00:05:54,352 every single line of this code. 129 00:05:54,388 --> 00:05:56,788 It's a little different than artificial intelligence, 130 00:05:56,823 --> 00:05:58,890 where it's more of a-- even more of a black box, 131 00:05:58,926 --> 00:06:01,359 where we don't know exactly how it's gonna work 132 00:06:01,395 --> 00:06:03,962 and how that artificial brain is gonna evolve. 133 00:06:03,997 --> 00:06:07,299 Why would NASA be interested in artificial intelligence? 134 00:06:07,334 --> 00:06:09,534 Well, there are a few reasons. 135 00:06:09,570 --> 00:06:12,537 Of course NASA's trying to send artificial probes further and 136 00:06:12,573 --> 00:06:16,041 further out into the solar system, and eventually beyond. 137 00:06:16,076 --> 00:06:20,078 We can't program those for all of the things that 138 00:06:20,113 --> 00:06:23,982 it's gonna encounter, so we need them to be able to do their job 139 00:06:24,017 --> 00:06:26,718 without us constantly telling them everything to do. 140 00:06:26,753 --> 00:06:28,887 Even in interpreting data like this, 141 00:06:28,922 --> 00:06:31,923 this is such a huge data set that there's no way that 142 00:06:31,959 --> 00:06:34,593 a human or even a team of humans can look at all of it. 143 00:06:34,628 --> 00:06:38,997 Computing is gonna continue to expand dramatically over even 144 00:06:39,032 --> 00:06:42,734 the next 10 years, and so we'll be able to do an even better job 145 00:06:42,769 --> 00:06:45,236 of modeling the history of the universe, 146 00:06:45,272 --> 00:06:47,806 the future of the universe, the history and future 147 00:06:47,841 --> 00:06:51,209 of our planet, and go further in exploring our universe. 148 00:06:56,950 --> 00:06:59,351 to artificial intelligence could be creating 149 00:06:59,386 --> 00:07:02,721 systems that take cues from the way our own brains work, 150 00:07:02,756 --> 00:07:05,023 using a process called "deep learning". 151 00:07:07,060 --> 00:07:09,194 I'm here at Berkeley to meet one of the leading minds 152 00:07:09,229 --> 00:07:11,029 in the development of AI. 153 00:07:11,064 --> 00:07:13,064 Stuart Russell has been at this for more than 20 years. 154 00:07:14,901 --> 00:07:17,769 He co-wrote THE definitive textbook on AI, 155 00:07:17,804 --> 00:07:21,072 and he understands both its promise and its dangers. 156 00:07:22,676 --> 00:07:26,144 So the brain has an enormous network of neurons. 157 00:07:26,179 --> 00:07:28,813 So a neuron is a cell that has these long, 158 00:07:28,849 --> 00:07:31,282 thin connections to other neurons, 159 00:07:31,318 --> 00:07:34,619 so it kind of looks like a big tangle of electrical spaghetti. 160 00:07:34,655 --> 00:07:39,090 And we have tens of billions of neurons, 161 00:07:39,126 --> 00:07:43,995 and deep learning networks are much, much, much simpler. 162 00:07:44,031 --> 00:07:48,466 What's similar about them is that both the brain and these 163 00:07:48,502 --> 00:07:52,103 deep learning networks can learn how to perform a given function, 164 00:07:52,139 --> 00:07:55,073 and they learn by being given lots of examples. 165 00:07:55,108 --> 00:07:57,242 But deep learning networks, for example, 166 00:07:57,277 --> 00:08:00,111 if you want to learn to recognize cats and dogs 167 00:08:00,147 --> 00:08:04,115 and candles and... staplers and things like that, 168 00:08:04,151 --> 00:08:07,118 you can show it millions of photographs with these things, 169 00:08:07,154 --> 00:08:08,853 with labeled things is what they are. 170 00:08:08,889 --> 00:08:13,425 So on tasks like recognizing a wide range of categories, 171 00:08:13,460 --> 00:08:15,927 the computations they run now have a thousand different 172 00:08:15,962 --> 00:08:17,429 categories of objects. 173 00:08:17,464 --> 00:08:20,865 And in the last five years, we've gone from systems that 174 00:08:20,901 --> 00:08:24,602 might get 5% accuracy on that thousand category task 175 00:08:24,638 --> 00:08:27,439 to systems that are getting 98% accuracy. 176 00:08:27,474 --> 00:08:28,940 So you've been at this for a while. 177 00:08:28,975 --> 00:08:31,276 I mean, how does it make you feel, that the progress now 178 00:08:31,311 --> 00:08:34,612 is becoming exponential? 179 00:08:34,648 --> 00:08:36,548 The exponential word is a dangerous one, 'cause... 180 00:08:36,583 --> 00:08:38,016 (Laughing) 181 00:08:38,051 --> 00:08:40,952 'cause it tends to suggest that things will continue to 182 00:08:40,987 --> 00:08:43,455 accelerate without end, but it may turn out that it will 183 00:08:43,490 --> 00:08:50,061 plateau, and that for other tasks we need new breakthroughs. 184 00:08:50,097 --> 00:08:51,863 Quantum computing? 185 00:08:51,898 --> 00:08:54,632 Quantum computing possibly, 186 00:08:54,668 --> 00:08:58,369 but that's kind of a cheat in some sense. 187 00:08:58,405 --> 00:09:01,740 That says that rather than really understand 188 00:09:01,775 --> 00:09:04,743 how the brain manages to do these amazing tasks 189 00:09:04,778 --> 00:09:08,012 with what is really not that much hardware, 190 00:09:08,048 --> 00:09:11,015 and so I almost hope that quantum computing... 191 00:09:11,051 --> 00:09:12,550 Doesn't happen. 192 00:09:12,586 --> 00:09:15,153 Doesn't happen, or happens a long time in the future where-- 193 00:09:15,188 --> 00:09:19,190 and you know, give us a chance to keep working on finding 194 00:09:19,226 --> 00:09:22,527 the secrets of intelligence, the things that make us smart, 195 00:09:22,562 --> 00:09:24,829 and gain real understanding. 196 00:09:24,865 --> 00:09:27,432 'Cause just brute forcing it isn't really understanding 197 00:09:27,467 --> 00:09:29,033 what's going on. 198 00:09:30,804 --> 00:09:32,437 Quantum computing may be the next giant leap 199 00:09:32,472 --> 00:09:35,607 in building the artificial intelligence of the future. 200 00:09:35,642 --> 00:09:38,209 It's a technology so powerful that we really could use it 201 00:09:38,245 --> 00:09:40,879 to develop AI without having to understand 202 00:09:40,914 --> 00:09:43,581 how human learning works. 203 00:09:43,617 --> 00:09:45,850 And though Stuart Russell hopes that quantum computing 204 00:09:45,886 --> 00:09:47,752 won't happen for a long time, 205 00:09:47,788 --> 00:09:51,890 Rupak Biswas can't wait - and he isn't. 206 00:09:51,925 --> 00:09:55,293 He runs the Quantum Artificial Intelligence Lab at NASA's 207 00:09:55,328 --> 00:09:59,063 Ames Research Center, where scientists are working on 208 00:09:59,099 --> 00:10:03,635 a powerful experimental computer called the D-Wave. 209 00:10:03,670 --> 00:10:06,237 It's built on the principles of quantum mechanics, 210 00:10:06,273 --> 00:10:09,440 which I was hoping Rupak could explain to me. 211 00:10:09,476 --> 00:10:11,376 If someone tells you that they'll explain to you 212 00:10:11,411 --> 00:10:13,778 what quantum mechanics is, you should run away from them 213 00:10:13,814 --> 00:10:15,880 because no one really understands this field. 214 00:10:15,916 --> 00:10:17,482 - (Laughing) - It's a very complex field. 215 00:10:17,517 --> 00:10:21,286 Rupak agreed to show me NASA's quantum computer. 216 00:10:21,321 --> 00:10:24,556 This is the D-Wave quantum annealing system. 217 00:10:24,591 --> 00:10:28,493 So what you'll see here is basically a black box, 218 00:10:28,528 --> 00:10:31,629 and that is where the D-Wave processor is. 219 00:10:31,665 --> 00:10:35,333 This is the Star Trek computer that's more powerful-- 220 00:10:35,368 --> 00:10:37,802 million of times more powerful than anything Microsoft's got. 221 00:10:37,838 --> 00:10:40,672 Well sure, yeah, for certain classes of problems, yes. 222 00:10:40,707 --> 00:10:42,640 This is something that is really more powerful, 223 00:10:42,676 --> 00:10:44,742 and we expect to get a lot of research done on this. 224 00:10:46,379 --> 00:10:48,613 NASA is hoping to use the quantum computer to develop 225 00:10:48,648 --> 00:10:51,916 quantum algorithms, and algorithms are a key component 226 00:10:51,952 --> 00:10:54,085 of the code running AI. 227 00:10:54,120 --> 00:10:57,355 In layman's terms, an algorithm is a kind of recipe 228 00:10:57,390 --> 00:10:59,424 for solving a problem. 229 00:10:59,459 --> 00:11:02,260 Basically, it's the set of step-by-step instructions 230 00:11:02,295 --> 00:11:05,697 given to a computer to help it accomplish a task. 231 00:11:05,732 --> 00:11:09,834 The question here though is that the algorithm could be different 232 00:11:09,870 --> 00:11:12,170 on a supercomputer than on a quantum computer, 233 00:11:12,205 --> 00:11:13,805 even though you are trying to solve the same problem. 234 00:11:13,840 --> 00:11:15,139 So then what is the difference? 235 00:11:15,175 --> 00:11:17,141 Because, you know, to a lot of people when you say 236 00:11:17,177 --> 00:11:19,010 supercomputer and quantum computer, 237 00:11:19,045 --> 00:11:20,511 aren't they all supercomputers? 238 00:11:20,547 --> 00:11:21,613 What's the difference? 239 00:11:21,648 --> 00:11:23,715 Right, so supercomputers are basically-- 240 00:11:23,750 --> 00:11:27,785 it's based on this transistor, and the transistor is... 241 00:11:27,821 --> 00:11:32,223 in layman terms, is basically a very small switch. 242 00:11:32,259 --> 00:11:34,125 So it's either zero or one. 243 00:11:34,160 --> 00:11:37,028 So a bit - and in a traditional computer, it's called a bit - 244 00:11:37,063 --> 00:11:39,964 it's either zero or one, whereas in quantum you have this thing 245 00:11:40,000 --> 00:11:43,134 called a qubit, which can be zero and one at the same time. 246 00:11:43,169 --> 00:11:46,504 And that's essentially the difference between a computer or 247 00:11:46,539 --> 00:11:49,974 a supercomputer or what we would consider classical computing 248 00:11:50,010 --> 00:11:53,344 versus quantum computing, where the quantum computer allows you 249 00:11:53,380 --> 00:11:55,647 to be in two stages at the same time. 250 00:11:55,682 --> 00:11:57,415 And that's where the power comes from. 251 00:11:57,450 --> 00:12:00,051 A quantum computer could help solve problems 252 00:12:00,086 --> 00:12:02,420 that no technology can solve today; 253 00:12:02,455 --> 00:12:05,890 finding cures for diseases or designing space colonies. 254 00:12:05,926 --> 00:12:09,027 But what if we choose to use that awesome computing power 255 00:12:09,062 --> 00:12:11,529 for destructive purposes? 256 00:12:11,564 --> 00:12:13,564 You know, if you think about nuclear energy, 257 00:12:13,600 --> 00:12:16,901 you can use nuclear energy to solve world's energy problems, 258 00:12:16,937 --> 00:12:18,903 but you can also use it for bad things. 259 00:12:18,939 --> 00:12:20,772 You know, so all of these things, 260 00:12:20,807 --> 00:12:23,408 if they are used improperly and get in the wrong hands, 261 00:12:23,443 --> 00:12:24,542 it could lead to trouble. 262 00:12:32,218 --> 00:12:35,153 BEN: I'm in Oxford to meet the founding engineer of Skype. 263 00:12:35,188 --> 00:12:37,522 Since leaving the company, Jaan Tallinn has become 264 00:12:37,557 --> 00:12:39,290 one of the most prominent voices 265 00:12:39,326 --> 00:12:41,693 in the field of artificial intelligence. 266 00:12:41,728 --> 00:12:44,495 He sees great advantages in developing AI, 267 00:12:44,531 --> 00:12:47,498 but he also warns of the risks of making the technology more 268 00:12:47,534 --> 00:12:52,236 powerful without understanding its potential harms. 269 00:12:52,272 --> 00:12:56,874 Once you have systems that are basically smarter than humans 270 00:12:56,910 --> 00:13:01,145 when it comes to developing further intelligent systems, 271 00:13:01,181 --> 00:13:03,181 then you have intelligent systems developing 272 00:13:03,216 --> 00:13:05,249 intelligent systems that in turn go on to develop 273 00:13:05,285 --> 00:13:07,385 even more intelligent systems. 274 00:13:07,420 --> 00:13:09,053 And you have this intelligence explosion. 275 00:13:10,523 --> 00:13:12,256 He thinks the next step in AI 276 00:13:12,292 --> 00:13:14,826 is the creation of so-called "general intelligence". 277 00:13:16,229 --> 00:13:19,697 What's the difference between what we're producing now and 278 00:13:19,733 --> 00:13:22,266 "general intelligence" of the future? 279 00:13:22,302 --> 00:13:24,435 If you think about a chess-playing computer, 280 00:13:24,471 --> 00:13:28,506 it's just modelling the chessboard in its memory 281 00:13:28,541 --> 00:13:32,243 and looking at scenarios how this game can play out, 282 00:13:32,278 --> 00:13:34,912 and what are the actions that it can do on the chessboard. 283 00:13:34,948 --> 00:13:38,850 So it actually would be good from a chess-playing perspective 284 00:13:38,885 --> 00:13:41,219 to not only model the chessboard, 285 00:13:41,254 --> 00:13:44,522 but also what's going on in the brain of your opponent. 286 00:13:44,557 --> 00:13:46,791 So it's one thing for a computer to understand 287 00:13:46,826 --> 00:13:49,761 the game of chess, but if it can look up from the board 288 00:13:49,796 --> 00:13:53,297 and understand me, that's a whole new level of scary. 289 00:13:53,333 --> 00:13:56,234 But I'm not the only one that's worried. 290 00:13:56,269 --> 00:14:00,104 In 2015, a group of prominent thinkers signed an open letter, 291 00:14:00,140 --> 00:14:03,041 warning researchers of the risks of making AI more powerful 292 00:14:03,076 --> 00:14:05,109 without understanding its potential harms. 293 00:14:07,013 --> 00:14:09,380 The signatories include Stuart Russell, 294 00:14:09,416 --> 00:14:12,750 Jaan Tallinn, Stephen Hawking, and Elon Musk. 295 00:14:14,154 --> 00:14:16,554 Nick Bostrom also signed the letter. 296 00:14:16,589 --> 00:14:18,923 He's a Swedish philosopher who's concerned that 297 00:14:18,958 --> 00:14:21,459 a mega-powerful AI that is capable of fulfilling the goals 298 00:14:21,494 --> 00:14:25,063 we give it could cause our extinction. 299 00:14:25,098 --> 00:14:27,131 What if we don't understand the full consequences 300 00:14:27,167 --> 00:14:29,233 of what we're asking it to do? 301 00:14:29,269 --> 00:14:31,669 What if we leave a crucial detail out? 302 00:14:33,106 --> 00:14:36,474 Take the myth of King Midas. 303 00:14:36,509 --> 00:14:37,842 You know, he asked... 304 00:14:37,877 --> 00:14:40,144 everything he touches should be turned into gold, 305 00:14:40,180 --> 00:14:42,346 which sounds like a great idea because you'll be very wealthy 306 00:14:42,382 --> 00:14:45,450 if you can turn coffee mugs into gold. 307 00:14:45,485 --> 00:14:47,752 Then he touches his food, it turns into gold, 308 00:14:47,787 --> 00:14:50,455 he touches his daughter, turns into a gold sculpture, 309 00:14:50,490 --> 00:14:52,457 so not such a cool idea. 310 00:14:52,492 --> 00:14:55,126 Turns out that it's actually quite difficult to write down 311 00:14:55,161 --> 00:14:58,930 some objective functions such that it would actually be good 312 00:14:58,965 --> 00:15:02,600 if that objective function were maximally realized. 313 00:15:02,635 --> 00:15:06,337 It seemed to me that this could be the most important thing 314 00:15:06,372 --> 00:15:08,506 in all of human history. 315 00:15:08,541 --> 00:15:12,343 What happens if AI succeeds at its original ambition? 316 00:15:12,378 --> 00:15:14,378 Which has all along been to achieve full general 317 00:15:14,414 --> 00:15:18,549 intelligence, where we have potentially artificial agents 318 00:15:18,585 --> 00:15:22,887 that can strategize, can deceive, that can form 319 00:15:22,922 --> 00:15:26,491 long goals and find creative ways of achieving them. 320 00:15:26,526 --> 00:15:30,328 And at that point, that kind of AI is not necessarily 321 00:15:30,363 --> 00:15:33,865 best thought of as merely a tool, merely another gadget. 322 00:15:33,900 --> 00:15:36,067 At that point, really talking about creating 323 00:15:36,102 --> 00:15:37,902 another intelligent life form. 324 00:15:37,937 --> 00:15:42,440 So could this potentially be - AI - the last human invention? 325 00:15:42,475 --> 00:15:48,312 Yeah, so once you have general intelligence at the human or 326 00:15:48,348 --> 00:15:51,682 superhuman level, then it's not just that you have made 327 00:15:51,718 --> 00:15:54,552 a breakthrough in AI, but you have made indirectly 328 00:15:54,587 --> 00:15:56,854 a breakthrough in every other area as well. 329 00:15:56,890 --> 00:15:59,390 So the AI can do research, science, 330 00:15:59,425 --> 00:16:01,659 development, all the other things that humans do. 331 00:16:01,694 --> 00:16:04,529 So potentially what you have is a kind of telescoping of 332 00:16:04,564 --> 00:16:07,698 the future, where all those possible technologies that, 333 00:16:07,734 --> 00:16:10,701 you know, maybe we would have developed given 40,000 years 334 00:16:10,737 --> 00:16:13,671 to work on it - you know, space colonization, 335 00:16:13,706 --> 00:16:16,407 cures for aging, all these other things. 336 00:16:16,442 --> 00:16:20,711 So in other words, we could be this hyper-invincible, 337 00:16:20,747 --> 00:16:23,748 space-hopping species with AI, 338 00:16:23,783 --> 00:16:25,349 but we also could go extinct by it? 339 00:16:25,385 --> 00:16:26,684 Yeah. 340 00:16:26,719 --> 00:16:31,122 I think... that machine superintelligence is - 341 00:16:31,157 --> 00:16:34,559 depending on how optimistic you feel on a given day - 342 00:16:34,594 --> 00:16:37,061 either the keyhole through which 343 00:16:37,096 --> 00:16:40,131 Earth-originating intelligent life has to pass, 344 00:16:40,166 --> 00:16:41,465 and we could crash into the wall 345 00:16:41,501 --> 00:16:43,100 instead of actually going through this. 346 00:16:43,136 --> 00:16:44,635 But if you make it through there, 347 00:16:44,671 --> 00:16:47,672 then the life expectancy for human civilization might then 348 00:16:47,707 --> 00:16:49,807 easily measure in billions of years. 349 00:16:57,317 --> 00:16:59,116 BEN: The scientists and philosophers working on 350 00:16:59,152 --> 00:17:02,453 artificial intelligence are sure that AI can improve our lives, 351 00:17:02,488 --> 00:17:05,423 yet the same people worry that it could destroy us. 352 00:17:05,458 --> 00:17:06,791 But how would that happen? 353 00:17:06,826 --> 00:17:08,259 - Hi, I'm Ben. - Hi! 354 00:17:08,294 --> 00:17:10,261 - Nice to meet you, Ben. - Nice to meet you. 355 00:17:10,296 --> 00:17:13,464 What if the AI we build is actually designed to kill? 356 00:17:15,335 --> 00:17:18,269 Heather Roff has testified before the UN as an expert 357 00:17:18,304 --> 00:17:19,737 on autonomous weapons. 358 00:17:21,107 --> 00:17:24,742 I once sat next to a grad student on a plane, 359 00:17:24,777 --> 00:17:27,144 and he told me that he was an AI researcher. 360 00:17:27,180 --> 00:17:28,679 And so I was very interested, saying, 361 00:17:28,715 --> 00:17:30,348 "What is it that you work on?" 362 00:17:30,383 --> 00:17:33,284 And he said, "Well, I work on image recognition." 363 00:17:33,319 --> 00:17:35,152 And I was like, "Really? Tell me more about that." 364 00:17:35,188 --> 00:17:38,256 And he said, "Well, I'm looking at how we identify 365 00:17:38,291 --> 00:17:40,424 different birds, and how we identify 366 00:17:40,460 --> 00:17:42,660 different types of coral and starfish." 367 00:17:42,695 --> 00:17:44,128 And I said, "Who funds your research?" 368 00:17:44,163 --> 00:17:46,597 And he said, "The Office of Naval Research." 369 00:17:46,633 --> 00:17:49,367 And I said, "Do you really think that your research 370 00:17:49,402 --> 00:17:51,502 is going to be confined to birds and starfish?" 371 00:17:51,537 --> 00:17:53,537 And he didn't actually have an answer for me. 372 00:17:53,573 --> 00:17:55,373 He didn't really think through the next step. 373 00:17:55,408 --> 00:17:58,509 How bad would it be if superintelligence is developed 374 00:17:58,544 --> 00:18:00,211 by the military? 375 00:18:00,246 --> 00:18:03,614 So I think if a superintelligence emerges, 376 00:18:03,650 --> 00:18:07,385 we're all in trouble for a variety of reasons. 377 00:18:07,420 --> 00:18:09,954 One of the things that we know about military applications 378 00:18:09,989 --> 00:18:12,356 is that they're not for the benefit of humanity, right? 379 00:18:12,392 --> 00:18:15,159 They're directed towards offensive harming. 380 00:18:15,194 --> 00:18:17,695 We're not talking about creating an AI that's gonna be 381 00:18:17,730 --> 00:18:21,032 trying to solve... climate change 382 00:18:21,067 --> 00:18:23,567 or solve-- or create poetry. 383 00:18:23,603 --> 00:18:27,138 We're not worried about those types of AI applications. 384 00:18:27,173 --> 00:18:29,907 We could be worried about those ones becoming superintelligent, 385 00:18:29,942 --> 00:18:33,244 but you ought to be really concerned about the strong AI 386 00:18:33,279 --> 00:18:34,812 that has guns on it. 387 00:18:34,847 --> 00:18:37,381 What does a superintelligent weapon look like? 388 00:18:37,417 --> 00:18:40,251 A superintelligence could be connected to everything. 389 00:18:40,286 --> 00:18:43,821 If it has a network, it has a capability of being connected 390 00:18:43,856 --> 00:18:46,691 through Wi-Fi, or it could figure out new ways of 391 00:18:46,726 --> 00:18:48,826 connecting itself if you shut off that Wi-Fi. 392 00:18:48,861 --> 00:18:50,828 It could propagate itself and its software 393 00:18:50,863 --> 00:18:52,997 on different servers and different things 394 00:18:53,032 --> 00:18:54,899 so you could never really truly get rid of it. 395 00:18:54,934 --> 00:18:58,035 It could hook itself into missile defense systems 396 00:18:58,071 --> 00:19:00,905 and nuclear arsenals, and it could do whatever it liked. 397 00:19:00,940 --> 00:19:03,040 I mean, that's the whole thing about being a superintelligence, 398 00:19:03,076 --> 00:19:05,443 is you're everywhere. 399 00:19:05,478 --> 00:19:07,244 You're, you know... 400 00:19:07,280 --> 00:19:09,347 Think Skynet but scarier. 401 00:19:09,382 --> 00:19:11,749 (Chuckling) 402 00:19:11,784 --> 00:19:13,084 Wow, can't wait. 403 00:19:13,119 --> 00:19:14,518 (Laughing) 404 00:19:14,554 --> 00:19:16,620 Thanks to all those AI researchers for getting that. 405 00:19:16,656 --> 00:19:19,890 Well, I mean, the AI researchers... 406 00:19:19,926 --> 00:19:22,093 I think it's not fair to blame them. 407 00:19:22,128 --> 00:19:25,463 I think they're trying to do things that are good with AI. 408 00:19:25,498 --> 00:19:27,765 It's the moment someone takes a scalpel from the surgeon 409 00:19:27,800 --> 00:19:29,200 and makes it a knife for killing. 410 00:19:31,170 --> 00:19:33,070 But others think the greatest dangers will come from 411 00:19:33,106 --> 00:19:36,774 unintended consequences, as Jaan Tallinn explained to me. 412 00:19:38,945 --> 00:19:41,412 I don't think it's correct to say that AI is a technology 413 00:19:41,447 --> 00:19:43,881 just like any other technology, or tool like any other tool. 414 00:19:43,916 --> 00:19:46,550 No, it's a technology that can potentially create 415 00:19:46,586 --> 00:19:48,552 new technologies itself. 416 00:19:48,588 --> 00:19:52,123 Now, there are a lot of smart people, as we know, 417 00:19:52,158 --> 00:19:54,959 looking into this problem and this issue, and... 418 00:19:54,994 --> 00:19:56,093 Not enough though. 419 00:19:56,129 --> 00:19:57,495 - Not enough? - No. 420 00:19:57,530 --> 00:20:00,464 Like imagine that we are building a spaceship that's able 421 00:20:00,500 --> 00:20:02,900 to carry the entire humanity. 422 00:20:04,003 --> 00:20:07,171 And like the boarding has already begun, 423 00:20:07,206 --> 00:20:09,740 and children are already onboard. 424 00:20:09,776 --> 00:20:12,276 And then there's like a small group of people who used to be 425 00:20:12,311 --> 00:20:13,911 completely ignored, who are saying that, 426 00:20:13,946 --> 00:20:17,481 "Look, we're gonna need some steering on this ship." 427 00:20:17,517 --> 00:20:20,317 And now people are going like, "Oh, wait a minute. 428 00:20:20,353 --> 00:20:22,520 Yeah, steering might become handy." 429 00:20:22,555 --> 00:20:23,788 (Laughing) 430 00:20:23,823 --> 00:20:25,322 But in this case, steering is, what, 431 00:20:25,358 --> 00:20:27,992 programming the AI to make sure that it doesn't kill us all? 432 00:20:28,027 --> 00:20:30,928 I think the more general point is that whenever you're 433 00:20:30,963 --> 00:20:34,932 a technology developer, you have the responsibility of... 434 00:20:36,869 --> 00:20:38,803 thinking through the consequences of your actions. 435 00:20:38,838 --> 00:20:45,843 ♪ 436 00:20:45,878 --> 00:20:51,649 It might be an incredibly subtle process which eventually ends up 437 00:20:51,684 --> 00:20:55,453 with the human race becoming sort of enfeebled and dependent 438 00:20:55,488 --> 00:20:58,956 on machines in ways that leave us vulnerable 439 00:20:58,991 --> 00:21:00,858 to any kind of unexpected event. 440 00:21:00,893 --> 00:21:03,627 Is it possible right now we're almost tripping over ourselves, 441 00:21:03,663 --> 00:21:07,531 in that we're coming up with discoveries about AI 442 00:21:07,567 --> 00:21:10,401 and we don't really realize the full implications of 443 00:21:10,436 --> 00:21:13,170 what we'd just discovered, and we just use it? 444 00:21:13,206 --> 00:21:15,172 You know, for centuries or millennia or even-- 445 00:21:15,208 --> 00:21:17,208 "It'll never happen," you hear that, you know. 446 00:21:17,243 --> 00:21:19,043 "We don't have to worry, it's just impossible." 447 00:21:19,078 --> 00:21:23,080 The history of nuclear physics, there was a... 448 00:21:23,115 --> 00:21:25,316 a speech by Rutherford. 449 00:21:25,351 --> 00:21:27,751 He's the guy who split the atom. 450 00:21:27,787 --> 00:21:31,689 In 1933, September 11th, he said, "There is no possibility 451 00:21:31,724 --> 00:21:34,992 that we'll ever be able to extract energy from atoms." 452 00:21:35,027 --> 00:21:39,763 Less than 24 hours later, Szilard invented 453 00:21:39,799 --> 00:21:42,032 the neutron-based nuclear chain reaction, 454 00:21:42,068 --> 00:21:45,169 and instantly realized what it would mean in terms of 455 00:21:45,204 --> 00:21:47,905 the ability to create nuclear explosions. 456 00:21:47,940 --> 00:21:49,740 So it went from never to 24 hours. 457 00:21:51,911 --> 00:21:54,011 The scientist who split the atom didn't understand 458 00:21:54,046 --> 00:21:56,614 that his discovery would so quickly lead to the atom bomb. 459 00:21:59,051 --> 00:22:01,519 But the same discovery later gave us the energy that still 460 00:22:01,554 --> 00:22:04,388 provides electricity to millions around the world. 461 00:22:05,591 --> 00:22:07,791 When it comes to AI, it seems like we're about to split 462 00:22:07,827 --> 00:22:09,793 the proverbial atom again. 463 00:22:09,829 --> 00:22:12,897 The future of humanity could be at stake. 464 00:22:12,932 --> 00:22:16,734 And how we build AI now - what we design it to do - 465 00:22:16,769 --> 00:22:18,369 could make the difference.