1 00:00:03,800 --> 00:00:06,000 Speaker 1: What would you say if you were able to get 2 00:00:06,000 --> 00:00:10,160 Speaker 1: in a room with Sam Altman and Elon Musk at 3 00:00:10,200 --> 00:00:13,760 Speaker 1: a table, which seems even less likely than Putin and 4 00:00:13,840 --> 00:00:16,200 Speaker 1: Zelensky meeting face to face. 5 00:00:16,320 --> 00:00:20,760 Speaker 2: I'd say, you know perfectly well that the stuff you're 6 00:00:20,800 --> 00:00:23,760 Speaker 2: developing has a good chance of wiping out people. 7 00:00:25,560 --> 00:00:25,759 Speaker 3: Hi. 8 00:00:25,800 --> 00:00:33,879 Speaker 1: Everyone, I'm Katie Kuric, and this is next question. When 9 00:00:33,960 --> 00:00:37,560 Speaker 1: you hear the Moniker Godfather of AI, you might think 10 00:00:37,600 --> 00:00:42,080 Speaker 1: of some levolent sci fi character, But in the real world, 11 00:00:42,760 --> 00:00:47,000 Speaker 1: this title belongs to doctor Jeffrey Hinton, my guest today. 12 00:00:47,320 --> 00:00:51,000 Speaker 1: He is the scientist whose early work on neural networks 13 00:00:51,320 --> 00:00:55,400 Speaker 1: cracked to open the field of artificial intelligence. If that 14 00:00:55,520 --> 00:00:57,960 Speaker 1: sounds like it's over your head, well honestly it was 15 00:00:58,000 --> 00:01:02,920 Speaker 1: a bit over mine, but Hinton graciously explains it for 16 00:01:03,080 --> 00:01:06,880 Speaker 1: lay people like us. Last year, doctor Hinton and another 17 00:01:06,959 --> 00:01:11,120 Speaker 1: guy named doctor John Hopfield received the twenty twenty four 18 00:01:11,520 --> 00:01:15,320 Speaker 1: Nobel Prize for physics and a word that honestly kind 19 00:01:15,319 --> 00:01:19,480 Speaker 1: of surprised them, given the fact that they're not really physicist. 20 00:01:20,040 --> 00:01:23,600 Speaker 1: But the prize is opening doors, he told us. In fact, 21 00:01:24,000 --> 00:01:26,920 Speaker 1: he has plans to meet with the Pope very soon 22 00:01:27,400 --> 00:01:31,319 Speaker 1: talk about access right. So in this interview we talk 23 00:01:31,400 --> 00:01:34,880 Speaker 1: about his life's work, some of the pressures he faced 24 00:01:35,080 --> 00:01:39,080 Speaker 1: as a young person, and the remarkable family he comes from, 25 00:01:39,200 --> 00:01:43,320 Speaker 1: talk about brainiacs, wow, and why he believes the choices 26 00:01:43,400 --> 00:01:48,360 Speaker 1: we make right now could actually shape the fate of humanity. 27 00:01:48,520 --> 00:01:55,000 Speaker 1: So here's my conversation with doctor Jeffrey Hinton. Doctor Jeffrey Hinton, 28 00:01:55,080 --> 00:01:58,120 Speaker 1: thank you so much for spending some time with me today. 29 00:01:58,400 --> 00:02:00,080 Speaker 1: It's a real honor to be able to. 30 00:02:00,080 --> 00:02:00,560 Speaker 3: Talk to you. 31 00:02:00,880 --> 00:02:02,280 Speaker 2: Thank you very much for inviting me. 32 00:02:02,760 --> 00:02:06,279 Speaker 1: I know that last year you were awarded the twenty 33 00:02:06,400 --> 00:02:11,639 Speaker 1: twenty four Nobel Prize in Physics jointly with someone named 34 00:02:11,680 --> 00:02:17,160 Speaker 1: John J. Hotfield, for your quote foundational discoveries and inventions 35 00:02:17,160 --> 00:02:22,519 Speaker 1: that enable machine learning with artificial neural networks. What did 36 00:02:22,520 --> 00:02:25,200 Speaker 1: it mean to you, doctor Hinton, to be honored in 37 00:02:25,280 --> 00:02:27,960 Speaker 1: this way and have your work recognized. 38 00:02:28,520 --> 00:02:30,840 Speaker 2: Well, I was actually very surprised that I got it 39 00:02:30,880 --> 00:02:33,680 Speaker 2: because it's in physics and I don't do physics. But 40 00:02:34,000 --> 00:02:35,680 Speaker 2: it means a lot to get a Nobel Prize for 41 00:02:35,720 --> 00:02:37,960 Speaker 2: every scientist, that's kind of the highest honor you can get. 42 00:02:38,639 --> 00:02:41,359 Speaker 1: I think they probably, as you have said in the past, 43 00:02:41,400 --> 00:02:45,679 Speaker 1: doctor Hinton need to have a Nobel Prize for computer science. 44 00:02:46,320 --> 00:02:48,680 Speaker 2: There's something called the Churing Award which is meant to 45 00:02:48,680 --> 00:02:52,680 Speaker 2: be like the Nobel Prize for computer science. But basically 46 00:02:53,000 --> 00:02:55,920 Speaker 2: Nobel made some prizes in his will and they don't 47 00:02:55,960 --> 00:02:58,480 Speaker 2: like to create anymore. They created one more in economics, 48 00:02:58,520 --> 00:03:00,240 Speaker 2: but they don't want to dilute it any more. 49 00:03:00,520 --> 00:03:04,639 Speaker 1: And I know you won the Touring Prize in twenty nineteen, 50 00:03:04,880 --> 00:03:09,000 Speaker 1: so that was an additional honor given to you. In 51 00:03:09,040 --> 00:03:13,000 Speaker 1: an interview last fall, you were hoping that the Nobel 52 00:03:13,160 --> 00:03:19,320 Speaker 1: Prize would mean that your views on artificial intelligence and 53 00:03:19,360 --> 00:03:22,280 Speaker 1: the role it will play in the future and is 54 00:03:22,720 --> 00:03:26,840 Speaker 1: in playing now for that matter, would be taken more seriously. 55 00:03:27,680 --> 00:03:29,680 Speaker 1: Have you found that to be the case. 56 00:03:30,440 --> 00:03:33,320 Speaker 2: Yes, I think that is true. I think more people 57 00:03:33,360 --> 00:03:35,880 Speaker 2: are willing to talk to me. I managed to talk 58 00:03:35,920 --> 00:03:39,440 Speaker 2: to somebody on the Chinese polit Bureau. I've been talking 59 00:03:39,440 --> 00:03:42,720 Speaker 2: with Bernie Sanders, and in September I may get to 60 00:03:42,760 --> 00:03:43,560 Speaker 2: talk to the Pope. 61 00:03:43,800 --> 00:03:44,240 Speaker 3: Wow. 62 00:03:44,680 --> 00:03:49,880 Speaker 1: So you're really spreading your concerns far and wide. Do 63 00:03:49,920 --> 00:03:53,840 Speaker 1: you feel that when you have these meetings, doctor Hinton, 64 00:03:54,000 --> 00:03:58,000 Speaker 1: that a people take you seriously, which I'm sure they do, 65 00:03:58,160 --> 00:04:02,000 Speaker 1: But B do I feel as if you're making progress, 66 00:04:02,240 --> 00:04:07,680 Speaker 1: that these conversations may in fact lead to real change 67 00:04:08,000 --> 00:04:13,320 Speaker 1: or policy changes that will reflect your concerns. 68 00:04:14,080 --> 00:04:17,479 Speaker 2: Yes, I think we are making progress. My main concern 69 00:04:17,560 --> 00:04:20,440 Speaker 2: has been not the short term risks of bad actors 70 00:04:20,520 --> 00:04:24,680 Speaker 2: misusing AI, which are very serious, but the longer term 71 00:04:24,800 --> 00:04:27,960 Speaker 2: risk of AI itself taking over. And I think until 72 00:04:28,040 --> 00:04:31,559 Speaker 2: quite recently, almost everybody thought that was just science fiction. 73 00:04:31,920 --> 00:04:34,159 Speaker 2: We didn't really have to worry about AI becoming smarter 74 00:04:34,240 --> 00:04:36,520 Speaker 2: than us and taking over from us. But I think 75 00:04:36,560 --> 00:04:39,520 Speaker 2: now many people have come to realize that's a very 76 00:04:39,600 --> 00:04:44,080 Speaker 2: real risk. So most of the experts believe that sometime 77 00:04:44,120 --> 00:04:46,840 Speaker 2: between five and twenty years from now, AI will get 78 00:04:46,880 --> 00:04:49,839 Speaker 2: smarter than people. And when it gets smarter, it won't 79 00:04:49,839 --> 00:04:51,840 Speaker 2: get a little bit smarter, to'll get a lot smarter. 80 00:04:53,080 --> 00:04:56,560 Speaker 2: And the problem is, we know very few examples of 81 00:04:57,000 --> 00:05:00,760 Speaker 2: smarter things being controlled by less smart things. Not small 82 00:05:00,800 --> 00:05:05,360 Speaker 2: differences intelligence, like between a politician and a scientist, for example, 83 00:05:05,800 --> 00:05:09,280 Speaker 2: but big difference is intelligence. There's very few examples. In fact, 84 00:05:09,360 --> 00:05:11,560 Speaker 2: the only example I know is a mother and child, 85 00:05:12,080 --> 00:05:13,000 Speaker 2: a mother and baby. 86 00:05:13,839 --> 00:05:16,800 Speaker 1: When I read about that time frame, doctor Hinton, I 87 00:05:16,920 --> 00:05:19,840 Speaker 1: have to say, I was pretty freaked out. I mean, 88 00:05:19,880 --> 00:05:24,400 Speaker 1: we're talking about not the distant future. We're talking about 89 00:05:24,400 --> 00:05:26,400 Speaker 1: the very near future, aren't we. 90 00:05:26,800 --> 00:05:29,560 Speaker 2: Yes, It's important to realize nobody has a good way 91 00:05:29,560 --> 00:05:31,680 Speaker 2: of estimating these things, so all of the estimates a 92 00:05:31,800 --> 00:05:35,000 Speaker 2: just guesses, But there's sort of a consensus among experts 93 00:05:35,160 --> 00:05:38,440 Speaker 2: between five and twenty years from now, it's quite likely 94 00:05:38,800 --> 00:05:40,880 Speaker 2: that we'll get things a lot smarter than us. 95 00:05:41,279 --> 00:05:45,080 Speaker 1: I want to talk to you about the dangers and 96 00:05:45,920 --> 00:05:49,599 Speaker 1: what you've been discussing with various people. But first I'd 97 00:05:49,640 --> 00:05:52,440 Speaker 1: like people to understand a little bit about your background, 98 00:05:52,520 --> 00:05:57,160 Speaker 1: which is quite extraordinary. You come from a family of 99 00:05:57,440 --> 00:06:02,480 Speaker 1: I would say, overachievers, not to mention brilliant minds. Your 100 00:06:02,560 --> 00:06:06,599 Speaker 1: great great grandfather was someone named George Boole, who developed 101 00:06:06,640 --> 00:06:11,280 Speaker 1: Boolean algebra. Your cousin Joan Hinton worked on the Manhattan Project. 102 00:06:11,600 --> 00:06:14,839 Speaker 1: Your great uncle Sebastian invented the jungle gym. 103 00:06:15,440 --> 00:06:17,400 Speaker 3: The list goes on and on. 104 00:06:18,279 --> 00:06:23,040 Speaker 1: Did you know you were destined for great things and 105 00:06:23,279 --> 00:06:29,160 Speaker 1: contributing to our knowledge in such an extraordinary way. 106 00:06:30,120 --> 00:06:32,800 Speaker 2: I didn't know I was destined for it. I knew 107 00:06:32,839 --> 00:06:35,040 Speaker 2: there were a lot of expectations. 108 00:06:34,800 --> 00:06:38,360 Speaker 1: In fact, your father sounds like he was a pretty 109 00:06:38,800 --> 00:06:44,279 Speaker 1: hard charging person with very high expectations for you. He 110 00:06:44,320 --> 00:06:48,400 Speaker 1: was a renowned entomologist, not to be confused with etomologists 111 00:06:48,480 --> 00:06:51,600 Speaker 1: the study of language. He was involved in the study 112 00:06:51,600 --> 00:06:56,440 Speaker 1: of insects, beatles in particular, I understand. And he once 113 00:06:56,520 --> 00:06:59,600 Speaker 1: said to you, if you worked twice as hard as me, 114 00:07:00,160 --> 00:07:02,880 Speaker 1: when you're twice as old as I am, you might 115 00:07:02,960 --> 00:07:04,040 Speaker 1: be half as good. 116 00:07:04,560 --> 00:07:07,119 Speaker 2: He didn't actually say that once, he said that quite 117 00:07:07,160 --> 00:07:08,800 Speaker 2: frequently before I went to school. 118 00:07:09,160 --> 00:07:13,800 Speaker 1: How did that impact you as a young person? How 119 00:07:13,800 --> 00:07:19,200 Speaker 1: did that shape your pursuit of artificial intelligence and specifically 120 00:07:19,240 --> 00:07:20,480 Speaker 1: neural networks. 121 00:07:20,880 --> 00:07:23,160 Speaker 2: It was said half in jest, but it was annoying. 122 00:07:23,840 --> 00:07:27,400 Speaker 2: I just had very strong expectations placed on me, and 123 00:07:27,440 --> 00:07:28,520 Speaker 2: I tried to live up to them. 124 00:07:29,000 --> 00:07:31,080 Speaker 1: What do you think he would think about your being 125 00:07:31,160 --> 00:07:32,920 Speaker 1: awarded the Nobel Prize. 126 00:07:33,680 --> 00:07:37,000 Speaker 2: I think he'd be very pleased and also slightly annoyed. 127 00:07:37,560 --> 00:07:40,400 Speaker 2: He was very competitive, and he'd be annoyed that his 128 00:07:40,480 --> 00:07:41,920 Speaker 2: son got something he hadn't got. 129 00:07:42,800 --> 00:07:48,320 Speaker 1: Let's talk about your journey from a young man to 130 00:07:48,720 --> 00:07:54,640 Speaker 1: where you are now. You wanted to really duplicate the 131 00:07:54,760 --> 00:07:59,120 Speaker 1: human mind in your work with neural networks. So you 132 00:07:59,160 --> 00:08:04,080 Speaker 1: stumbled upon on your area of expertise accidentally, didn't you. 133 00:08:04,720 --> 00:08:09,679 Speaker 2: It wasn't completely accidental. It became obvious in the sixties 134 00:08:09,760 --> 00:08:13,640 Speaker 2: and seventies that we had a new way of doing 135 00:08:13,680 --> 00:08:16,480 Speaker 2: research on how the brain works. Up until then, you 136 00:08:16,520 --> 00:08:19,960 Speaker 2: could do experiments on the brain, or you could have 137 00:08:20,000 --> 00:08:23,080 Speaker 2: theories about how the brain worked, but it's very hard 138 00:08:23,120 --> 00:08:25,960 Speaker 2: to test these theories because they need to be complicated, 139 00:08:25,960 --> 00:08:30,800 Speaker 2: because the brain's complicated. And once computers got fast enough 140 00:08:30,840 --> 00:08:34,240 Speaker 2: to do simulations of networks and brain cells, then you 141 00:08:34,320 --> 00:08:38,040 Speaker 2: could start to have more elaborate theories of how it 142 00:08:38,120 --> 00:08:41,800 Speaker 2: might be computing and test them by doing computer simulations. 143 00:08:42,080 --> 00:08:44,760 Speaker 2: And so there is this kind of new form of 144 00:08:44,800 --> 00:08:47,600 Speaker 2: science which was testing theories of how the brain might 145 00:08:47,679 --> 00:08:51,120 Speaker 2: work using computer simulations. And it was obvious from the 146 00:08:51,160 --> 00:08:54,160 Speaker 2: beginning that would also give you a different way of 147 00:08:54,200 --> 00:08:54,760 Speaker 2: doing AI. 148 00:08:56,080 --> 00:09:00,199 Speaker 1: You, I know, worked on this for many years. And 149 00:09:01,000 --> 00:09:03,400 Speaker 1: can I ask a dumb question, I hope it's okay. 150 00:09:04,200 --> 00:09:09,439 Speaker 1: Can you explain the role neural networks play in artificial 151 00:09:09,800 --> 00:09:15,400 Speaker 1: intelligence versus large language models and how they relate to 152 00:09:15,480 --> 00:09:16,040 Speaker 1: one another? 153 00:09:17,000 --> 00:09:19,800 Speaker 2: Yes, okay, So let me give you a little bit 154 00:09:19,800 --> 00:09:22,880 Speaker 2: of history. For about the first fifty years of artificial 155 00:09:22,880 --> 00:09:26,720 Speaker 2: intelligence the second half of the last century, almost everybody 156 00:09:26,760 --> 00:09:29,200 Speaker 2: in artificial intelligence thought the way to make a machine 157 00:09:29,240 --> 00:09:32,880 Speaker 2: intelligent was to mimic logic. What you would do is 158 00:09:32,920 --> 00:09:37,600 Speaker 2: you'd have symbolic expressions things like sentences in the computer, 159 00:09:38,480 --> 00:09:42,960 Speaker 2: and you'd have rules for manipulating them so it will 160 00:09:43,000 --> 00:09:46,559 Speaker 2: work like logic. For example, if I say all men 161 00:09:46,600 --> 00:09:51,000 Speaker 2: are mortal, and I say Socrates is a man, you 162 00:09:51,040 --> 00:09:53,840 Speaker 2: can do some manipulations on those strings of words and 163 00:09:53,880 --> 00:09:58,800 Speaker 2: come up with Socrates is mortal. That's logic, that's Aristotelian logic, 164 00:09:59,240 --> 00:10:01,240 Speaker 2: and that would that's the sort of model for how 165 00:10:01,280 --> 00:10:04,079 Speaker 2: we were going to do AI and computers. There was 166 00:10:04,160 --> 00:10:08,080 Speaker 2: a very different theory, sort of utterly different, which was 167 00:10:08,559 --> 00:10:11,840 Speaker 2: instead of looking for logic as the inspiration for how 168 00:10:11,880 --> 00:10:14,640 Speaker 2: to make a computer intelligent, look at how the brain 169 00:10:14,720 --> 00:10:18,120 Speaker 2: actually works. We have a big network of brain cells. 170 00:10:18,160 --> 00:10:21,480 Speaker 2: We have billions of them. They have connections between them, 171 00:10:22,320 --> 00:10:25,800 Speaker 2: and they learn by changing the strengths of those connections. 172 00:10:26,440 --> 00:10:29,120 Speaker 2: So one of the neurons in this great big network, 173 00:10:29,440 --> 00:10:31,480 Speaker 2: all it has to do is decide when to go ping, 174 00:10:32,840 --> 00:10:34,719 Speaker 2: and it sends its ping to other neurons that it's 175 00:10:34,760 --> 00:10:38,360 Speaker 2: connected to, and it decides when to go ping by 176 00:10:38,360 --> 00:10:41,440 Speaker 2: looking at the pings it's receiving from other neurons. But 177 00:10:41,559 --> 00:10:44,280 Speaker 2: each ping it receives from another neuron, it gives it 178 00:10:44,360 --> 00:10:47,880 Speaker 2: a weight, which is the connection strength. So for example, 179 00:10:48,320 --> 00:10:51,000 Speaker 2: if I'm a particular neuron and you're another neuron and 180 00:10:51,040 --> 00:10:53,160 Speaker 2: you say ping, I might have a weight on that. 181 00:10:53,320 --> 00:10:56,319 Speaker 2: Say okay, when she says ping, that's a lot of 182 00:10:56,400 --> 00:10:59,960 Speaker 2: evidence I should go ping. Or let's suppose as another neuron, 183 00:11:00,200 --> 00:11:02,600 Speaker 2: which is Donald Trump, and when he goes ping, I say, 184 00:11:02,640 --> 00:11:04,560 Speaker 2: that's a lot of evidence I should not go ping. 185 00:11:05,440 --> 00:11:08,520 Speaker 2: So basically, each neuron is looking at the ping's caring 186 00:11:08,520 --> 00:11:11,880 Speaker 2: from other neurons. It's associated to weight. With each ping, 187 00:11:12,480 --> 00:11:15,240 Speaker 2: it's like a vote. It's other neurons going ping of 188 00:11:15,360 --> 00:11:17,440 Speaker 2: votes for me to either go ping or not go ping, 189 00:11:18,320 --> 00:11:20,600 Speaker 2: and then it decides whether to go ping. And I 190 00:11:20,600 --> 00:11:23,400 Speaker 2: think you can see that if you change the strengths 191 00:11:23,400 --> 00:11:27,120 Speaker 2: of those votes, those weights, then how neurons go ping 192 00:11:27,160 --> 00:11:31,040 Speaker 2: will change. And that's how your brain learns, and that's 193 00:11:31,080 --> 00:11:35,720 Speaker 2: how artificial intelligence now works. It simulates a big network 194 00:11:35,760 --> 00:11:39,160 Speaker 2: of brain cells like this and at large language models 195 00:11:40,280 --> 00:11:45,000 Speaker 2: work on neural networks. Almost all AI now uses neural networks. 196 00:11:45,440 --> 00:11:48,760 Speaker 2: So we're simulating in the computer a big network of 197 00:11:48,760 --> 00:11:53,120 Speaker 2: brain cells, and the brain cells keep going ping, and 198 00:11:53,240 --> 00:11:56,200 Speaker 2: exactly when they go ping depends on the votes they 199 00:11:56,240 --> 00:11:58,360 Speaker 2: get from other brain cells that are gone ping or 200 00:11:58,360 --> 00:12:02,160 Speaker 2: from sense organs that have gone ping. And by changing 201 00:12:02,200 --> 00:12:05,000 Speaker 2: the strength of those connections, you can make it do anything. 202 00:12:06,200 --> 00:12:08,560 Speaker 2: The issue is how do you change the connection strengths 203 00:12:08,559 --> 00:12:09,520 Speaker 2: to make it do what you want? 204 00:12:10,440 --> 00:12:15,480 Speaker 3: I kind of got that, but it's a good start 205 00:12:15,559 --> 00:12:15,840 Speaker 3: for me. 206 00:12:16,920 --> 00:12:22,520 Speaker 1: Does a large language model give the data that the 207 00:12:22,600 --> 00:12:26,439 Speaker 1: neural connections need to make those connections. 208 00:12:26,840 --> 00:12:29,160 Speaker 2: What the neural network does is, if you have the 209 00:12:29,160 --> 00:12:32,160 Speaker 2: symbol for Tuesday, that'll make a particular set of neurons 210 00:12:32,200 --> 00:12:35,640 Speaker 2: go ping. If you have the symbol for Wednesday, it'll 211 00:12:35,640 --> 00:12:38,640 Speaker 2: make a very similar set of neurons go ping. But 212 00:12:38,720 --> 00:12:40,959 Speaker 2: if you have the symbol for although, that'll be a 213 00:12:41,000 --> 00:12:45,160 Speaker 2: completely different set of neurons go ping. So the words 214 00:12:45,200 --> 00:12:49,080 Speaker 2: come in, they get converted into groups of neurons going ping. 215 00:12:50,480 --> 00:12:55,120 Speaker 2: These neurons then interact with each other. So, for example, 216 00:12:55,320 --> 00:12:58,880 Speaker 2: suppose the word may comes in. Let's ignore capital. Let's 217 00:12:58,880 --> 00:13:01,600 Speaker 2: suppose we don't have any capital letters, so that lower 218 00:13:01,600 --> 00:13:04,720 Speaker 2: case may comes in, you don't know whether that's a 219 00:13:04,800 --> 00:13:08,520 Speaker 2: month or a woman's name. So neurons going ping for 220 00:13:08,559 --> 00:13:12,240 Speaker 2: nearby words influence which neurons go ping for May, And 221 00:13:12,320 --> 00:13:16,719 Speaker 2: if you have June and July there probably you'll influence them, 222 00:13:16,720 --> 00:13:19,319 Speaker 2: so they're more like the neurons that should go ping 223 00:13:19,400 --> 00:13:22,760 Speaker 2: for a month. These neural activations that capture the meaning 224 00:13:23,880 --> 00:13:27,600 Speaker 2: activate neurons that are representing the meaning of the next word. 225 00:13:28,920 --> 00:13:32,400 Speaker 2: And so by using this neural network, if the connection 226 00:13:32,440 --> 00:13:35,559 Speaker 2: strengths that are appropriate, when a bunch of words comes in, 227 00:13:35,840 --> 00:13:38,880 Speaker 2: you'll be able to predict the next word. And you 228 00:13:39,080 --> 00:13:42,360 Speaker 2: train the whole thing by saying, how should I change 229 00:13:42,360 --> 00:13:45,840 Speaker 2: the connection strengths. So when I see a string of words, 230 00:13:46,440 --> 00:13:50,520 Speaker 2: maybe I see fish and and after fish and chips 231 00:13:50,559 --> 00:13:54,200 Speaker 2: is quite lightly so you'd like to change the connection strengths. 232 00:13:54,240 --> 00:13:56,160 Speaker 2: So after you've seen the word fish followed by the 233 00:13:56,160 --> 00:13:59,960 Speaker 2: word and you predict that a quite likely next world 234 00:14:00,400 --> 00:14:02,840 Speaker 2: is chips. That's how the whole thing works. 235 00:14:03,400 --> 00:14:06,719 Speaker 1: I got you, and in fact, I'm thinking about when 236 00:14:06,800 --> 00:14:11,880 Speaker 1: I'm texting now or writing an email, that they suggest 237 00:14:12,040 --> 00:14:17,199 Speaker 1: words for me. Given some of the past things I've written. 238 00:14:17,640 --> 00:14:22,120 Speaker 1: So are they using my personal large language model to 239 00:14:22,440 --> 00:14:25,640 Speaker 1: predict what I want to say? Because I might not 240 00:14:25,800 --> 00:14:29,640 Speaker 1: say fish and chips, I might say fish and lemon. 241 00:14:29,840 --> 00:14:34,480 Speaker 1: Let's say, well, the neural networks know, oh, doctor Hinton 242 00:14:34,600 --> 00:14:38,120 Speaker 1: is going to say chips, but Katie's going to say lemon. 243 00:14:38,760 --> 00:14:41,080 Speaker 2: Since I left Google a few years ago, I don't 244 00:14:41,080 --> 00:14:44,040 Speaker 2: know in Gmail how much of your personal information they 245 00:14:44,160 --> 00:14:47,360 Speaker 2: used to make the prediction. My guess is, to be efficient, 246 00:14:47,720 --> 00:14:50,440 Speaker 2: they don't use your personal information. They're just using a 247 00:14:50,520 --> 00:14:53,440 Speaker 2: large language model. It's looking at quite a long context 248 00:14:53,440 --> 00:14:57,760 Speaker 2: of what you said and is just predicting features of 249 00:14:57,800 --> 00:15:01,440 Speaker 2: the next word. And you say fish and it'll say 250 00:15:01,520 --> 00:15:04,200 Speaker 2: chips is pretty lightly. If you're in the British House 251 00:15:04,240 --> 00:15:07,520 Speaker 2: of Lords, it'll say hunt is pretty lightly. But it 252 00:15:07,760 --> 00:15:09,640 Speaker 2: probably won't say that because it knows you're in the 253 00:15:09,640 --> 00:15:12,640 Speaker 2: British House of Lords. It'll probably say because of previous 254 00:15:12,640 --> 00:15:14,480 Speaker 2: things you said in your Gmail. 255 00:15:21,920 --> 00:15:25,000 Speaker 1: Hi everyone, it's me Katie Kuric. You know, lately, I've 256 00:15:25,040 --> 00:15:29,640 Speaker 1: been overwhelmed by the whole wellness industry, so much information 257 00:15:29,760 --> 00:15:34,680 Speaker 1: out there about flaxed pelvic floor serums and anti aging. 258 00:15:35,160 --> 00:15:38,040 Speaker 1: So I launched a newsletter It's called Body and Soul 259 00:15:38,400 --> 00:15:42,040 Speaker 1: to share expert approved advice for your physical and mental health. 260 00:15:42,240 --> 00:15:43,600 Speaker 3: And guess what, it's free. 261 00:15:43,920 --> 00:15:48,160 Speaker 1: Just sign up at Katiecuric dot com slash Body and Soul. 262 00:15:48,440 --> 00:15:50,680 Speaker 1: That's k A T I E C O U r 263 00:15:51,000 --> 00:15:54,880 Speaker 1: Ic dot com slash Body and Soul. I promise it 264 00:15:54,880 --> 00:16:08,720 Speaker 1: will make you happier and healthier. I wanted to talk 265 00:16:08,760 --> 00:16:12,280 Speaker 1: to you about your journey coming to Google and then 266 00:16:12,400 --> 00:16:17,600 Speaker 1: leaving Google. You sold your small startup to Google in 267 00:16:17,680 --> 00:16:21,520 Speaker 1: twenty thirteen, You went to work there to advance the 268 00:16:21,560 --> 00:16:26,960 Speaker 1: company's AI research, and then ten years later you left 269 00:16:27,080 --> 00:16:32,520 Speaker 1: so you could speak more freely about the risks of AI. 270 00:16:33,680 --> 00:16:37,320 Speaker 1: How did you come to that decision after a decade? 271 00:16:37,680 --> 00:16:39,440 Speaker 1: Was that a tough one, doctor Henton. 272 00:16:40,120 --> 00:16:43,360 Speaker 2: So the precise timing of when I left was so 273 00:16:43,480 --> 00:16:46,360 Speaker 2: that I could speak at a conference at MIT and 274 00:16:46,440 --> 00:16:49,440 Speaker 2: I could speak freely without worrying about the implications for Google. 275 00:16:50,120 --> 00:16:51,960 Speaker 2: But the fact that I left was because I was 276 00:16:52,000 --> 00:16:55,440 Speaker 2: seventy five. I've been doing research very half for fifty 277 00:16:55,440 --> 00:16:57,920 Speaker 2: five years, and I was worn out. I wasn't as 278 00:16:57,960 --> 00:16:59,600 Speaker 2: good as programming it as I used to be, and 279 00:16:59,640 --> 00:17:01,920 Speaker 2: it was an so I was going to retire when 280 00:17:01,920 --> 00:17:04,760 Speaker 2: I was seventy five anyway, but the precise timing was 281 00:17:04,800 --> 00:17:07,680 Speaker 2: so that I could speak freely because I've become acutely 282 00:17:07,720 --> 00:17:10,200 Speaker 2: aware of the risks of AI in my last few 283 00:17:10,240 --> 00:17:13,000 Speaker 2: years at Google, and I wanted to tell the public 284 00:17:13,000 --> 00:17:16,119 Speaker 2: about them. But I thought it was wrong to say 285 00:17:16,160 --> 00:17:18,960 Speaker 2: things that would impact a company while taking money from 286 00:17:18,960 --> 00:17:22,199 Speaker 2: the company. I didn't actually have any complaints about Google. 287 00:17:22,400 --> 00:17:24,920 Speaker 2: I didn't think they were doing anything wrong. They had 288 00:17:24,920 --> 00:17:27,000 Speaker 2: the large language in models and they were not releasing 289 00:17:27,040 --> 00:17:29,480 Speaker 2: them to the public at that point, so it wasn't 290 00:17:29,480 --> 00:17:32,080 Speaker 2: to criticize Google, but it was to warn about the 291 00:17:32,119 --> 00:17:33,240 Speaker 2: long term risks of AI. 292 00:17:33,680 --> 00:17:36,639 Speaker 1: Having said that, did you feel that Google was doing 293 00:17:36,880 --> 00:17:40,520 Speaker 1: enough to mitigate some of the risks? Did you feel 294 00:17:40,520 --> 00:17:43,560 Speaker 1: like it was enough of a priority for the company 295 00:17:43,760 --> 00:17:46,440 Speaker 1: During your decade there, people. 296 00:17:46,200 --> 00:17:49,119 Speaker 2: There were quite concerned about various risks. I think they 297 00:17:49,160 --> 00:17:52,640 Speaker 2: could have always done more, and I think for all 298 00:17:52,640 --> 00:17:56,080 Speaker 2: the companies, they could do a lot more. Google does 299 00:17:56,119 --> 00:17:59,800 Speaker 2: have some concern about the risk. Demissabis, in particular, who's 300 00:18:00,119 --> 00:18:04,080 Speaker 2: the head of their research on things like large language models, 301 00:18:04,400 --> 00:18:07,479 Speaker 2: understands the long term threat of AI taking over, and 302 00:18:07,480 --> 00:18:08,720 Speaker 2: he's quite concerned about it. 303 00:18:09,560 --> 00:18:13,280 Speaker 1: Let's talk about regulation. As you well know better than anyone, 304 00:18:13,440 --> 00:18:19,320 Speaker 1: AI is developing at breakneck speed. Governments, meanwhile, move slowly 305 00:18:19,560 --> 00:18:23,520 Speaker 1: and often don't fully understand the technology because they have 306 00:18:23,680 --> 00:18:27,760 Speaker 1: not been immersed in it like you have. And let's 307 00:18:27,760 --> 00:18:30,879 Speaker 1: face it, they just haven't really become familiar with it 308 00:18:30,920 --> 00:18:35,920 Speaker 1: because they're older and they haven't necessarily caught up. Even 309 00:18:35,920 --> 00:18:40,920 Speaker 1: if governments were motivated and vigilant. Do you think regulation 310 00:18:41,760 --> 00:18:47,560 Speaker 1: could ever catch up with AI's capacity and prevent its 311 00:18:47,640 --> 00:18:49,359 Speaker 1: most dangerous risks. 312 00:18:49,960 --> 00:18:53,760 Speaker 2: I think it could certainly help. And I think we 313 00:18:53,840 --> 00:18:58,639 Speaker 2: have to distinguish two kinds of risks here. There's shorter 314 00:18:58,800 --> 00:19:03,400 Speaker 2: term risks to do with people misusing AI, and those 315 00:19:03,520 --> 00:19:08,320 Speaker 2: risks are things like cyber attacks or creating nasty viruses, 316 00:19:08,840 --> 00:19:12,119 Speaker 2: or creating fake videos. And for all of those shorter 317 00:19:12,200 --> 00:19:16,600 Speaker 2: term risks, each risk has a different solution. So for 318 00:19:16,680 --> 00:19:22,160 Speaker 2: things like viruses, an obvious partial solution is to take 319 00:19:22,200 --> 00:19:25,560 Speaker 2: the companies that manufacture things for you in the cloud, 320 00:19:26,560 --> 00:19:29,440 Speaker 2: and if you send them a sequence, they should look 321 00:19:29,480 --> 00:19:32,040 Speaker 2: at the sequence before they manufacture it and say, wait 322 00:19:32,080 --> 00:19:34,560 Speaker 2: a minute, this looks very like COVID. I'm not going 323 00:19:34,600 --> 00:19:37,480 Speaker 2: to manufacture something that looks like COVID, but they don't. 324 00:19:38,280 --> 00:19:41,800 Speaker 2: They should be forced to do that. So that's an 325 00:19:41,840 --> 00:19:45,359 Speaker 2: example of something we could do to make the risk 326 00:19:45,440 --> 00:19:50,960 Speaker 2: of terrorists producing bad viruses a little less severe. For 327 00:19:51,040 --> 00:19:56,200 Speaker 2: fake videos, initially people thought we could recognize fake videos, 328 00:19:56,680 --> 00:19:59,840 Speaker 2: but actually it's much easier in the long run, I think, 329 00:20:00,119 --> 00:20:02,480 Speaker 2: to be able to recognize that a video isn't fake. 330 00:20:03,520 --> 00:20:05,520 Speaker 2: So if there's suppose there's a fake video with you 331 00:20:05,600 --> 00:20:08,679 Speaker 2: in it that pretends to be by you, what you 332 00:20:08,760 --> 00:20:11,480 Speaker 2: really need is at the beginning of that video, there's 333 00:20:11,520 --> 00:20:14,439 Speaker 2: a QR code. The QR code will take you to 334 00:20:14,480 --> 00:20:18,000 Speaker 2: a website, and if it's your website and the identical 335 00:20:18,080 --> 00:20:20,879 Speaker 2: video is on your website, we know it's a real video. 336 00:20:21,680 --> 00:20:23,760 Speaker 2: If it takes you to something that isn't your website 337 00:20:24,080 --> 00:20:25,960 Speaker 2: or on your website, the video isn't the same, we 338 00:20:26,000 --> 00:20:28,560 Speaker 2: know it's a fake video. That's feasible, and I think 339 00:20:28,600 --> 00:20:30,240 Speaker 2: that will come and all that work will be done 340 00:20:30,280 --> 00:20:33,480 Speaker 2: by the browser, of course. So that's a solution for 341 00:20:33,520 --> 00:20:37,359 Speaker 2: fake videos, or a partial solution. One problem here is 342 00:20:37,800 --> 00:20:40,680 Speaker 2: governments will not collaborate with each other on this stuff. 343 00:20:41,240 --> 00:20:44,520 Speaker 2: Maybe on viruses they will, but on fake videos or 344 00:20:44,560 --> 00:20:47,520 Speaker 2: on cyber attacks. Governments are all busy doing it to 345 00:20:47,560 --> 00:20:51,520 Speaker 2: each other. America was a real pioneer in corrupting elections 346 00:20:51,520 --> 00:20:54,320 Speaker 2: in other countries. It's quite upset now that's come back 347 00:20:54,320 --> 00:20:54,880 Speaker 2: to bite it. 348 00:20:55,480 --> 00:20:57,320 Speaker 3: Tell me more about that, doctor Henton. 349 00:20:57,720 --> 00:21:00,960 Speaker 2: Oh, So for many years America would try and manipulate 350 00:21:01,000 --> 00:21:04,719 Speaker 2: elections in other countries. It's fairly clear that Russians were 351 00:21:04,760 --> 00:21:08,240 Speaker 2: trying to manipulate the twenty sixteen election. And I think 352 00:21:08,280 --> 00:21:10,760 Speaker 2: manipulating elections in other countries is a bad thing to do. 353 00:21:11,520 --> 00:21:14,040 Speaker 2: One of the few things Trump said that I agreed with. 354 00:21:14,920 --> 00:21:19,320 Speaker 2: Probably around twenty sixteen, someone said what did you think 355 00:21:19,320 --> 00:21:23,639 Speaker 2: about foreigners manipulating our elections? And Trump actually said, you 356 00:21:23,680 --> 00:21:27,240 Speaker 2: think we're so different? He was very well aware of that. 357 00:21:28,359 --> 00:21:31,040 Speaker 2: So we're not going to get collaboration on those things 358 00:21:31,080 --> 00:21:34,679 Speaker 2: between countries because they're doing it to each other. But 359 00:21:34,760 --> 00:21:38,320 Speaker 2: we will get collaboration on how to prevent AI from 360 00:21:38,320 --> 00:21:41,680 Speaker 2: taking over, because no country wants AI actually take over. 361 00:21:42,600 --> 00:21:44,800 Speaker 2: So what I've been doing recently is trying to persuade 362 00:21:44,880 --> 00:21:48,439 Speaker 2: senior people, like people on the Polypburea in China, or 363 00:21:48,480 --> 00:21:52,520 Speaker 2: the Pope or Bernie Sanders that we can get international 364 00:21:52,520 --> 00:21:57,600 Speaker 2: collaboration on this, because the techniques you need to prevent 365 00:21:57,680 --> 00:22:00,879 Speaker 2: AI from taking over the only way we can do 366 00:22:00,920 --> 00:22:03,520 Speaker 2: that is to make it so we build these super 367 00:22:03,520 --> 00:22:06,960 Speaker 2: intelligent AIS that do not want to take over. And 368 00:22:07,040 --> 00:22:08,960 Speaker 2: the question is how do you build an AI so 369 00:22:09,080 --> 00:22:12,720 Speaker 2: it will never want to take over? And the techniques 370 00:22:12,760 --> 00:22:15,359 Speaker 2: for doing that are somewhat different from the techniques for 371 00:22:15,480 --> 00:22:20,120 Speaker 2: making the II smarter, and so countries won't share how 372 00:22:20,160 --> 00:22:23,880 Speaker 2: you make the very smartest AI because they're busy competing 373 00:22:23,920 --> 00:22:25,960 Speaker 2: with each other for things like cyber attacks and fake 374 00:22:26,040 --> 00:22:30,240 Speaker 2: videos and for lots of other things to do with manufacturing, 375 00:22:30,680 --> 00:22:33,960 Speaker 2: but they will collaborate on how do we prevent it 376 00:22:34,000 --> 00:22:37,240 Speaker 2: from wanting to take over from people. So, for example, 377 00:22:37,400 --> 00:22:42,960 Speaker 2: if the Americans developed a good technique for stopping a 378 00:22:43,000 --> 00:22:45,720 Speaker 2: super intelligent AI from wanting to take over, they would 379 00:22:45,760 --> 00:22:48,080 Speaker 2: want to share it with the Chinese and the Russians 380 00:22:48,080 --> 00:22:50,680 Speaker 2: and the British and the French and these radis because 381 00:22:50,720 --> 00:22:54,240 Speaker 2: they don't want their iis taking over either. So I 382 00:22:54,240 --> 00:22:57,639 Speaker 2: think the worst long term risk we have is AI 383 00:22:57,680 --> 00:23:00,000 Speaker 2: taking over. But the one piece of good news is 384 00:23:00,080 --> 00:23:02,439 Speaker 2: we can collaborate on that. And what I'd like to 385 00:23:02,440 --> 00:23:06,000 Speaker 2: see is institutions in different countries that collaborate with each 386 00:23:06,000 --> 00:23:08,439 Speaker 2: other on how do we make AI not want to 387 00:23:08,480 --> 00:23:08,960 Speaker 2: take over? 388 00:23:09,720 --> 00:23:12,800 Speaker 1: That sounds wonderful, but is it realistic. Do you believe 389 00:23:12,840 --> 00:23:16,040 Speaker 1: that countries would in fact collaborate and what you need 390 00:23:16,520 --> 00:23:21,440 Speaker 1: sort of a new entity for this mission, for this 391 00:23:21,600 --> 00:23:26,399 Speaker 1: idea of trying to put breaks on AI taking over. 392 00:23:27,080 --> 00:23:29,800 Speaker 2: It's not putting brakes on it so much. We're not 393 00:23:29,840 --> 00:23:32,480 Speaker 2: going to put brakes on AI. It's too useful for 394 00:23:32,520 --> 00:23:36,000 Speaker 2: too many things. It'll increase productivity in a huge number 395 00:23:36,040 --> 00:23:39,400 Speaker 2: of industries. It'll be very helpful in improving health care 396 00:23:39,440 --> 00:23:41,960 Speaker 2: and improving education. And because of all the good things 397 00:23:42,000 --> 00:23:43,920 Speaker 2: it can do, we're not going to be able to 398 00:23:43,920 --> 00:23:46,399 Speaker 2: stop the development. So rather than think in terms of 399 00:23:46,440 --> 00:23:49,240 Speaker 2: putting brakes on it, think in terms of how do 400 00:23:49,280 --> 00:23:52,840 Speaker 2: you develop a form of it that is nice to people, 401 00:23:52,880 --> 00:23:56,160 Speaker 2: that can coexist with people. And I think we will 402 00:23:56,200 --> 00:23:58,679 Speaker 2: get collaboration on that. We're not going to get the 403 00:23:58,760 --> 00:24:02,840 Speaker 2: United States collaborate for about another three and a half years, 404 00:24:03,200 --> 00:24:06,600 Speaker 2: but I don't think under the Trump administration there'll be 405 00:24:06,640 --> 00:24:10,119 Speaker 2: any effective collaboration on things like that. But once we 406 00:24:10,240 --> 00:24:13,480 Speaker 2: get a reasonable regime in the States, we might get collaboration, 407 00:24:13,520 --> 00:24:16,880 Speaker 2: and I think other countries like France and Britain and Canada, 408 00:24:17,359 --> 00:24:21,359 Speaker 2: maybe South Korea, Japan, maybe even China will be willing 409 00:24:21,400 --> 00:24:24,199 Speaker 2: to collaborate on how do you make an AI that 410 00:24:24,240 --> 00:24:25,280 Speaker 2: doesn't want to take over? 411 00:24:26,080 --> 00:24:30,520 Speaker 1: Speaking of China, some have argued that China's rapid push 412 00:24:30,800 --> 00:24:36,560 Speaker 1: into artificial intelligence, especially its willingness to deploy it in 413 00:24:36,720 --> 00:24:44,240 Speaker 1: surveillance and military context, makes reasonable global regulation impossible. You 414 00:24:44,400 --> 00:24:47,800 Speaker 1: believe that China would want to be a part of 415 00:24:47,840 --> 00:24:50,120 Speaker 1: this global collaboration. 416 00:24:50,640 --> 00:24:53,200 Speaker 3: And if you believe that, why. 417 00:24:53,480 --> 00:24:56,719 Speaker 2: I don't think China in the US will collaborate on 418 00:24:56,800 --> 00:25:01,680 Speaker 2: things like detecting fake videos or or swarting cyber attacks 419 00:25:02,280 --> 00:25:05,600 Speaker 2: because their interests aren't aligned. There. People only collaborate when 420 00:25:05,600 --> 00:25:09,320 Speaker 2: the interests are aligned, and on that their interests are opposed, 421 00:25:09,320 --> 00:25:13,119 Speaker 2: and they will not collaborate, but they will collaborate on 422 00:25:13,200 --> 00:25:15,840 Speaker 2: how do you prevent AI from taking over? Because they're 423 00:25:15,880 --> 00:25:18,800 Speaker 2: their interests are aligned in much the same way as 424 00:25:18,840 --> 00:25:22,679 Speaker 2: the Soviet Union and the United States. Their interests were 425 00:25:22,720 --> 00:25:26,600 Speaker 2: aligned in the nineteen fifties in preventing a global nuclear war. 426 00:25:26,960 --> 00:25:29,159 Speaker 2: It wasn't good Riser of them, and so they did 427 00:25:29,240 --> 00:25:31,320 Speaker 2: actually collaborate in ways to prevent that. 428 00:25:32,359 --> 00:25:35,320 Speaker 1: Europe has tried to get ahead of the curve with 429 00:25:35,520 --> 00:25:41,160 Speaker 1: sweeping rules like the EUAI Act. Do you think this 430 00:25:41,240 --> 00:25:44,720 Speaker 1: effort will have a meaningful impact. 431 00:25:44,960 --> 00:25:47,800 Speaker 2: Yes, I think companies would like to have a global market, 432 00:25:48,119 --> 00:25:52,199 Speaker 2: and if Europe puts regulations on things like privacy or 433 00:25:52,240 --> 00:25:56,600 Speaker 2: hate speech or things like that, then I think that 434 00:25:56,680 --> 00:25:59,960 Speaker 2: will affect the AIS produced because people want to have 435 00:26:00,119 --> 00:26:02,159 Speaker 2: global market, so there'll be a tendency for people to 436 00:26:02,160 --> 00:26:06,480 Speaker 2: try and satisfy those regulations even in other countries. Now, 437 00:26:06,800 --> 00:26:09,639 Speaker 2: the regulations of Europe are fairly flimsy at present. So 438 00:26:09,720 --> 00:26:14,560 Speaker 2: for example, the European regulations on air have as clause 439 00:26:14,600 --> 00:26:18,120 Speaker 2: in them that says none of this applies to military 440 00:26:18,200 --> 00:26:22,959 Speaker 2: uses of AI. So the European countries that manufacture arms, 441 00:26:23,480 --> 00:26:27,840 Speaker 2: like Britain and France aren't willing to regulate the use 442 00:26:27,880 --> 00:26:29,280 Speaker 2: of AI for weapons. 443 00:26:29,920 --> 00:26:33,520 Speaker 1: Why do you think that they have a clause excluding 444 00:26:34,080 --> 00:26:37,840 Speaker 1: the military because we know, for example, the US military 445 00:26:37,960 --> 00:26:43,200 Speaker 1: is actively developing autonomous weapons, which you believe should be outlawed. 446 00:26:43,560 --> 00:26:49,200 Speaker 1: Why don't countries realize the dangers these kinds of weapons pose. 447 00:26:49,560 --> 00:26:52,240 Speaker 2: Countries do realize the dangers they post, but they also 448 00:26:52,280 --> 00:26:55,520 Speaker 2: see the military advantages they pose. And a lot of 449 00:26:55,560 --> 00:27:03,240 Speaker 2: countries make money by selling arms. So the United States, Russia, China, Britain, Israel, France, 450 00:27:04,000 --> 00:27:06,120 Speaker 2: they make money by selling arms, and so they don't 451 00:27:06,160 --> 00:27:10,080 Speaker 2: want regulations on these things. Also, for very rich countries, 452 00:27:11,040 --> 00:27:14,679 Speaker 2: lethal autonomous weapons, that is, weapons that decide by themselves 453 00:27:14,680 --> 00:27:19,520 Speaker 2: who to kill or maim are a big advantage if 454 00:27:19,560 --> 00:27:22,720 Speaker 2: a rich country wants to invade a poor country. The 455 00:27:23,000 --> 00:27:27,200 Speaker 2: thing that stops rich countries invading poor countries is their 456 00:27:27,240 --> 00:27:29,560 Speaker 2: citizens coming back in body bags, which doesn't look good 457 00:27:29,560 --> 00:27:33,320 Speaker 2: for anybody. If you have lethal autonomous weapons, instead of 458 00:27:33,359 --> 00:27:36,880 Speaker 2: dead people coming back, you'll get dead robots coming back. 459 00:27:37,400 --> 00:27:42,240 Speaker 1: What are your other concerns about the way you anticipate 460 00:27:42,560 --> 00:27:46,200 Speaker 1: AI transforming warfare over the next few decades. 461 00:27:46,960 --> 00:27:50,199 Speaker 2: Oh, I think it's fairly clear. It's already transformed warmfare. 462 00:27:50,240 --> 00:27:53,000 Speaker 2: If you look what's going on in Ukraine. As Eric 463 00:27:53,000 --> 00:27:56,360 Speaker 2: Schmidt pointed out, a five hundred dollar drone can now 464 00:27:56,400 --> 00:28:02,480 Speaker 2: destroy a multimillion dollar tank, completely changed warfare. It's fairly 465 00:28:02,600 --> 00:28:05,840 Speaker 2: clear that fighter jets with people in them are a 466 00:28:05,840 --> 00:28:09,200 Speaker 2: silly idea. Now if you can have AI in them, 467 00:28:09,720 --> 00:28:13,760 Speaker 2: Ais can withstand much bigger accelerations and you don't have 468 00:28:13,800 --> 00:28:19,159 Speaker 2: to worry so much about loss of life. So I 469 00:28:19,160 --> 00:28:22,320 Speaker 2: think it will completely transform warfare. I think we will 470 00:28:22,320 --> 00:28:25,959 Speaker 2: see very nasty things happen with ethan autronomous weapons, and 471 00:28:26,080 --> 00:28:29,879 Speaker 2: after those things have happened, we may get regulations. So 472 00:28:30,119 --> 00:28:33,480 Speaker 2: with chemical weapons, very nasty things happened in the First 473 00:28:33,520 --> 00:28:37,160 Speaker 2: World War, and after that countries are willing to say, well, 474 00:28:37,200 --> 00:28:39,040 Speaker 2: I won't use them if you don't. And we got 475 00:28:39,080 --> 00:28:43,200 Speaker 2: the Geneva Conventions on chemical weapons and they've pretty much held. 476 00:28:44,000 --> 00:28:46,880 Speaker 2: They've been broken a few times. Britain broke them in 477 00:28:46,920 --> 00:28:49,600 Speaker 2: the nineteen thirties when it dropped mustard gas on the Curds. 478 00:28:49,640 --> 00:28:52,680 Speaker 2: In the twenties or thirties, Sadam Hussein broke them when 479 00:28:52,760 --> 00:28:55,320 Speaker 2: he dropped mustard gas on the Kurds. It always seems 480 00:28:55,360 --> 00:28:57,720 Speaker 2: to be the Curds that get it. But on the 481 00:28:57,760 --> 00:29:00,600 Speaker 2: whole they haven't been broken as sad So I'd use 482 00:29:00,640 --> 00:29:03,719 Speaker 2: them too, yes, but on the whole they haven't been broken. 483 00:29:04,120 --> 00:29:08,520 Speaker 2: So in Ukraine, for example, people aren't using chemical weapons. 484 00:29:09,080 --> 00:29:13,240 Speaker 1: So you're using that as an example of how there 485 00:29:13,320 --> 00:29:18,040 Speaker 1: could be some kind of agreement among nations that would 486 00:29:18,040 --> 00:29:18,720 Speaker 1: be effective. 487 00:29:19,520 --> 00:29:23,280 Speaker 2: Yes, and it's happened with chemical weapons. I talked to 488 00:29:23,960 --> 00:29:27,440 Speaker 2: someone who used to be an army surgeon, and he 489 00:29:27,560 --> 00:29:32,080 Speaker 2: pointed out that there's a big motivation by countries who 490 00:29:32,080 --> 00:29:35,560 Speaker 2: are thoughtful not to use really nasty means of warfare 491 00:29:36,280 --> 00:29:40,240 Speaker 2: because eventually there's a piece and you have to negotiate 492 00:29:40,280 --> 00:29:43,480 Speaker 2: what happens when the war's over, and if a country's 493 00:29:43,560 --> 00:29:47,560 Speaker 2: use terrible methods, it's much harder to negotiate things. So actually, 494 00:29:48,000 --> 00:29:50,360 Speaker 2: even though they want to win the war, countries also 495 00:29:50,400 --> 00:29:52,520 Speaker 2: have a vested interest in not using methods that are 496 00:29:52,560 --> 00:29:53,200 Speaker 2: too terrible. 497 00:29:54,200 --> 00:29:56,959 Speaker 1: I would love to talk to you about the economic 498 00:29:57,080 --> 00:30:01,200 Speaker 1: impact of AI, because I know that's a clear short 499 00:30:01,280 --> 00:30:05,040 Speaker 1: term risk that you have discussed. You have warned that 500 00:30:05,120 --> 00:30:11,040 Speaker 1: AI could wipe out jobs across nearly all fields, and 501 00:30:11,280 --> 00:30:14,600 Speaker 1: I know many people are really concerned about this, about 502 00:30:14,640 --> 00:30:19,200 Speaker 1: their livelihoods, about their children's livelihoods. What do you think 503 00:30:19,240 --> 00:30:23,400 Speaker 1: people need to understand about this technology and how it 504 00:30:23,480 --> 00:30:27,720 Speaker 1: will impact the way we work over the next few decades. 505 00:30:28,760 --> 00:30:31,280 Speaker 2: So the first thing to say is this really isn't 506 00:30:31,280 --> 00:30:35,280 Speaker 2: a technological problem, is a political problem. So if we 507 00:30:35,400 --> 00:30:39,840 Speaker 2: had a fair political system. When AI came along and 508 00:30:39,880 --> 00:30:44,280 Speaker 2: greatly increased productivity, When for example, ailized one person to 509 00:30:44,280 --> 00:30:47,000 Speaker 2: do the work that five people used to do, that 510 00:30:47,160 --> 00:30:50,920 Speaker 2: should mean there's more goods and services for everybody. And 511 00:30:51,000 --> 00:30:54,800 Speaker 2: in some areas where the demand's elastic, it will so. 512 00:30:55,080 --> 00:30:58,400 Speaker 2: In healthcare, for example, old people like me can absorb 513 00:30:58,520 --> 00:31:01,720 Speaker 2: endless amounts of healthcare, and if we make healthcare more efficient, 514 00:31:01,760 --> 00:31:05,240 Speaker 2: we'll just get more healthcare. That's great. But in other areas, 515 00:31:05,320 --> 00:31:09,480 Speaker 2: like I mean call centers or in doing research for 516 00:31:09,560 --> 00:31:12,920 Speaker 2: lawyers on similar cases, when one person could do the 517 00:31:12,920 --> 00:31:17,120 Speaker 2: work of five people, four people will be unemployed. With 518 00:31:17,320 --> 00:31:21,400 Speaker 2: increased productivity, it should be great. But within the system 519 00:31:21,440 --> 00:31:25,000 Speaker 2: we've got, we know what's going to happen. The owners 520 00:31:25,640 --> 00:31:28,680 Speaker 2: who introduce the AI are going to get richer, and 521 00:31:28,720 --> 00:31:31,640 Speaker 2: the unemployed people are going to get poorer. That's going 522 00:31:31,720 --> 00:31:34,640 Speaker 2: to increase the gap between rich and poor, and the 523 00:31:34,760 --> 00:31:38,280 Speaker 2: level of violence into society is strongly determined by the 524 00:31:38,280 --> 00:31:41,200 Speaker 2: gap between rich and poor. That's what we're seeing in 525 00:31:41,240 --> 00:31:46,120 Speaker 2: the States now. The gap between ordinary working people and 526 00:31:46,400 --> 00:31:51,440 Speaker 2: hedgehog managers became extreme, and that's what's providing the medium 527 00:31:51,480 --> 00:31:56,080 Speaker 2: in which Trump's kind of populism thrives, it'll get worse 528 00:31:56,280 --> 00:31:59,040 Speaker 2: as we get more AI and the working class get 529 00:31:59,080 --> 00:31:59,960 Speaker 2: an even worse deal. 530 00:32:01,520 --> 00:32:07,480 Speaker 1: I'm really asking this for my children, and I'm curious 531 00:32:07,560 --> 00:32:11,400 Speaker 1: what jobs will be more at risk and what jobs 532 00:32:11,560 --> 00:32:14,040 Speaker 1: would be less at risk. In other words, if you 533 00:32:14,080 --> 00:32:18,680 Speaker 1: were advising a college graduate today, doctor Hinton, careers to 534 00:32:18,760 --> 00:32:22,160 Speaker 1: stay away from and careers to gravitate towards, what would 535 00:32:22,160 --> 00:32:22,400 Speaker 1: they be. 536 00:32:23,600 --> 00:32:25,440 Speaker 2: So I should start by saying I'm not really an 537 00:32:25,440 --> 00:32:27,280 Speaker 2: expert on this, and this is all guesswork. 538 00:32:27,960 --> 00:32:29,200 Speaker 3: Okay, Well, that's okay. 539 00:32:29,480 --> 00:32:33,280 Speaker 2: Yeah, it's obvious that working in a call center where 540 00:32:33,280 --> 00:32:36,360 Speaker 2: you're badly paid and badly trained is not a very 541 00:32:36,360 --> 00:32:38,480 Speaker 2: good job. Because AI is going to be able to 542 00:32:38,480 --> 00:32:41,880 Speaker 2: do the job better quite soon. It'll know much much 543 00:32:41,960 --> 00:32:44,760 Speaker 2: more about the right answers to the questions that people 544 00:32:44,800 --> 00:32:48,080 Speaker 2: are asking. It'll be more patient. It's just a better 545 00:32:48,120 --> 00:32:50,800 Speaker 2: way of doing that job. It's not quite there yet, 546 00:32:50,840 --> 00:32:53,920 Speaker 2: but it's sort of around being there now over the 547 00:32:53,960 --> 00:32:56,160 Speaker 2: next few years. I think jobs in call centers are 548 00:32:56,240 --> 00:33:01,520 Speaker 2: very unsafe. Similarly, already, parently eagles people who help lawyers 549 00:33:01,560 --> 00:33:05,560 Speaker 2: find similar cases, they're pretty much out of work, and 550 00:33:05,720 --> 00:33:08,120 Speaker 2: even junior lawyers now are finding it hard to get 551 00:33:08,240 --> 00:33:11,600 Speaker 2: jobs because from any of the big law firms, AI 552 00:33:11,760 --> 00:33:13,800 Speaker 2: is doing the kind of grub work that junior lawyers 553 00:33:13,840 --> 00:33:17,800 Speaker 2: would be started off on. Really good programmers who design 554 00:33:17,840 --> 00:33:21,880 Speaker 2: complicated systems, they're still in business, but sort of everyday 555 00:33:22,040 --> 00:33:25,640 Speaker 2: ordinary programmers who just write some code to achieve some 556 00:33:26,800 --> 00:33:32,880 Speaker 2: straightforward goal, they're already their work's already in danger. So Microsoft, 557 00:33:32,960 --> 00:33:35,520 Speaker 2: for example, I think, laid off people because it can 558 00:33:35,560 --> 00:33:39,400 Speaker 2: get AI to do that routine programming, and lots of 559 00:33:39,440 --> 00:33:42,560 Speaker 2: companies are now saying they're going to hire less people 560 00:33:42,600 --> 00:33:46,120 Speaker 2: to do jobs like that. There are things to do 561 00:33:46,160 --> 00:33:50,920 Speaker 2: with human dexterity, so jobs like plumbing, particularly plumbing in 562 00:33:50,960 --> 00:33:54,760 Speaker 2: an old, awkward house. I have an old Edwardian house, 563 00:33:55,240 --> 00:33:58,040 Speaker 2: and fixing the plumbing in that is something that will 564 00:33:58,080 --> 00:34:00,280 Speaker 2: still need people for quite a long time, I think. 565 00:34:01,080 --> 00:34:04,440 Speaker 2: But eventually machines will get dexterrous, maybe in ten or 566 00:34:04,440 --> 00:34:06,680 Speaker 2: twenty years, and even that won't be safe. 567 00:34:07,200 --> 00:34:13,799 Speaker 1: I would also imagine that jobs that require high emotional 568 00:34:13,800 --> 00:34:19,600 Speaker 1: intelligence interpersonal skills, you know, a nurse, a doctor, someone 569 00:34:19,960 --> 00:34:26,680 Speaker 1: who is there to interface with someone in a supportive role, 570 00:34:26,800 --> 00:34:30,480 Speaker 1: that those jobs couldn't be really replaced by computers. 571 00:34:30,520 --> 00:34:35,560 Speaker 2: Could they They can and they will already. If you 572 00:34:36,160 --> 00:34:41,400 Speaker 2: let people interact with a doctor or interact with an 573 00:34:41,480 --> 00:34:45,680 Speaker 2: AI and ask them which is more empathetic. AI is 574 00:34:45,680 --> 00:34:47,200 Speaker 2: a ranked as much more empathetic. 575 00:34:48,360 --> 00:34:49,200 Speaker 3: That's depressing. 576 00:34:50,960 --> 00:34:54,279 Speaker 2: Yes, this is why they're going to get better than 577 00:34:54,360 --> 00:35:00,239 Speaker 2: us at everything, which is why it's rather urgent to 578 00:35:00,280 --> 00:35:02,719 Speaker 2: figure out whether there's a way whether when they're better 579 00:35:02,760 --> 00:35:04,920 Speaker 2: than us and everything and they're more powerful than us, 580 00:35:05,239 --> 00:35:07,840 Speaker 2: we can coexist with them or will just be history. 581 00:35:08,680 --> 00:35:12,719 Speaker 1: I wanted to ask you about creative fields. Selfishly, my 582 00:35:12,880 --> 00:35:18,000 Speaker 1: daughter is a television writer for scripted television. My other 583 00:35:18,160 --> 00:35:21,920 Speaker 1: daughter is getting her PhD in history. Should they be 584 00:35:22,040 --> 00:35:23,960 Speaker 1: looking for a different line of work? 585 00:35:24,920 --> 00:35:28,719 Speaker 2: Not right now. I have a close friend who's a 586 00:35:28,800 --> 00:35:33,560 Speaker 2: television script writer, and so we discussed this quite a lot. 587 00:35:34,080 --> 00:35:38,239 Speaker 2: Right now, they're not as good. My belief is look 588 00:35:38,239 --> 00:35:40,919 Speaker 2: in ten years time and they'll be able to write 589 00:35:41,000 --> 00:35:43,600 Speaker 2: very clever scripts too, with twists in the end and things. 590 00:35:43,800 --> 00:35:46,080 Speaker 2: They can't do it now, but they're getting better all 591 00:35:46,120 --> 00:35:46,480 Speaker 2: the time. 592 00:35:46,920 --> 00:35:49,239 Speaker 1: And what about a history PhD? 593 00:35:50,040 --> 00:35:52,719 Speaker 2: They already know a lot more than us, being a 594 00:35:52,719 --> 00:35:55,200 Speaker 2: lot more intelligent than us in the sense that if 595 00:35:55,239 --> 00:35:57,360 Speaker 2: you had a debate with them about anything, you'd lose 596 00:35:58,600 --> 00:36:02,319 Speaker 2: being smarter than us, which they will be. They'll be 597 00:36:02,360 --> 00:36:06,080 Speaker 2: better at manipulating people. And it's learned all these manipulative 598 00:36:06,120 --> 00:36:08,880 Speaker 2: skills just from trying to predict the next word and 599 00:36:08,920 --> 00:36:12,360 Speaker 2: all the documents on the web, because people do a 600 00:36:12,400 --> 00:36:15,520 Speaker 2: lot of manipulation, and AI is learned by example how 601 00:36:15,560 --> 00:36:16,000 Speaker 2: to do it. 602 00:36:16,360 --> 00:36:19,400 Speaker 1: Do you think that I'll be replaced by AI? Journalists 603 00:36:19,400 --> 00:36:21,160 Speaker 1: are worried about that? Obviously? 604 00:36:21,560 --> 00:36:24,560 Speaker 2: Oh yes, I think you're replaceable bad, but not just yet. 605 00:36:25,760 --> 00:36:27,399 Speaker 2: I'd say you have ten or twenty years. 606 00:36:27,480 --> 00:36:28,720 Speaker 3: I have a few more years. 607 00:36:29,080 --> 00:36:30,120 Speaker 2: You have a few more years. 608 00:36:30,360 --> 00:36:33,080 Speaker 1: Okay, good well, thank goodness, since I'm sixty eight, but 609 00:36:33,160 --> 00:36:35,560 Speaker 1: I feel bad for all the young journalists who want 610 00:36:35,600 --> 00:36:36,480 Speaker 1: to do what I do. 611 00:36:36,960 --> 00:36:39,000 Speaker 2: I agree. 612 00:36:43,840 --> 00:36:46,239 Speaker 1: Hi everyone, it's me Kittie Couric. You know, if you've 613 00:36:46,280 --> 00:36:49,040 Speaker 1: been following me on social media, you know I love 614 00:36:49,080 --> 00:36:52,239 Speaker 1: to cook, or at least try, especially alongside some of 615 00:36:52,280 --> 00:36:56,000 Speaker 1: my favorite chefs and foodies like Benny Blanco, Jake Cohen, 616 00:36:56,120 --> 00:37:00,520 Speaker 1: Lighty Hoyke, Alison Roman, and Ininegarten. So I started a 617 00:37:00,560 --> 00:37:04,120 Speaker 1: free newsletter called good Taste to share recipes, tips, and 618 00:37:04,239 --> 00:37:08,400 Speaker 1: kitchen mustaves. Just sign up at katiecorrect dot com slash 619 00:37:08,560 --> 00:37:11,520 Speaker 1: good Taste. That's k A T I E C O 620 00:37:11,719 --> 00:37:15,439 Speaker 1: U r I C dot com slash good taste. I 621 00:37:15,480 --> 00:37:28,160 Speaker 1: promised your taste buds will be happy you did. You 622 00:37:28,239 --> 00:37:33,840 Speaker 1: talk about this being a political problem versus a technological problem. 623 00:37:34,640 --> 00:37:40,880 Speaker 1: So is something like universal basic income a solution to 624 00:37:41,200 --> 00:37:44,400 Speaker 1: the fact that so many of these jobs will be 625 00:37:44,560 --> 00:37:45,920 Speaker 1: wiped out by AI. 626 00:37:46,680 --> 00:37:49,880 Speaker 2: It's not a solution, but it's a good band aid. 627 00:37:50,520 --> 00:37:54,799 Speaker 2: The point is you've got to stop people starving. They've 628 00:37:54,800 --> 00:37:57,360 Speaker 2: got to be able to pay the rent, and universal 629 00:37:57,360 --> 00:38:01,000 Speaker 2: basically income will help with that, but it doesn't solve 630 00:38:01,080 --> 00:38:05,279 Speaker 2: the problem because for most people, their sense of their 631 00:38:05,320 --> 00:38:08,759 Speaker 2: own worth is related to the job they do, and 632 00:38:09,000 --> 00:38:12,799 Speaker 2: if they're unemployed, they lose that sense of worth and 633 00:38:13,040 --> 00:38:17,239 Speaker 2: universal basic income doesn't deal with that. It stops them 634 00:38:17,239 --> 00:38:20,160 Speaker 2: starving and they can pay the rent, but who they 635 00:38:20,160 --> 00:38:22,759 Speaker 2: are has been seriously damaged by them losing their job. 636 00:38:23,320 --> 00:38:26,919 Speaker 1: Yeah, it's a huge problem. As you noted, I want 637 00:38:26,960 --> 00:38:29,160 Speaker 1: to talk a little bit about some of the positive 638 00:38:29,200 --> 00:38:34,560 Speaker 1: aspects of AI so people don't weep in despair after 639 00:38:34,640 --> 00:38:38,520 Speaker 1: watching this. I know healthcare is an area that you're 640 00:38:38,680 --> 00:38:42,240 Speaker 1: very excited about as am I. I know you lost 641 00:38:42,480 --> 00:38:46,880 Speaker 1: two wives to cancer, one from ovarian one from pancreatic. 642 00:38:47,680 --> 00:38:51,200 Speaker 1: My husband died of calling cancer when he was just 643 00:38:51,320 --> 00:38:57,280 Speaker 1: forty two years old. And so the diagnosis of early 644 00:38:57,400 --> 00:39:01,200 Speaker 1: stage cancers and breakthroughs and all kinds of diseases and 645 00:39:01,280 --> 00:39:05,719 Speaker 1: early diagnosis for those is something I am very very 646 00:39:05,760 --> 00:39:09,720 Speaker 1: excited about. Can you talk for a moment about how 647 00:39:09,800 --> 00:39:16,600 Speaker 1: this will impact healthcare and scientific breakthroughs in terms of diseases. 648 00:39:17,360 --> 00:39:20,720 Speaker 2: Yes, So there's many different ways in which it'll help. 649 00:39:21,200 --> 00:39:25,759 Speaker 2: One obvious way is in interpreting medical scans. So I 650 00:39:25,840 --> 00:39:29,640 Speaker 2: made a prediction in twenty sixteen that by now all 651 00:39:29,719 --> 00:39:32,520 Speaker 2: medical scans will be read by AI. That was a 652 00:39:32,520 --> 00:39:35,239 Speaker 2: bit over enthusiastic. I was off by a factor of 653 00:39:35,239 --> 00:39:37,160 Speaker 2: two or three in the timescale of that it's going 654 00:39:37,239 --> 00:39:40,680 Speaker 2: to happen. But it hasn't happened yet. But already, places 655 00:39:40,719 --> 00:39:44,280 Speaker 2: like the Mayo Clinic have many, many different AIS helping 656 00:39:44,440 --> 00:39:48,239 Speaker 2: doctors interpret scans. AIS will eventually be able to see 657 00:39:48,360 --> 00:39:51,920 Speaker 2: much more information in a scan. So there's a nice example. 658 00:39:52,480 --> 00:39:56,520 Speaker 2: Ophthalmologists make a scan called a fundu's image of your 659 00:39:56,560 --> 00:40:00,000 Speaker 2: retina the back of your eye. An AI can look 660 00:40:00,160 --> 00:40:03,080 Speaker 2: a scanner the back of your eye and make a 661 00:40:03,160 --> 00:40:05,360 Speaker 2: moderately good prediction about whether you're going to get a 662 00:40:05,360 --> 00:40:08,839 Speaker 2: heart attack. Doctors didn't know that was possible. Or an 663 00:40:08,880 --> 00:40:10,719 Speaker 2: AI can look at this image of the back of 664 00:40:10,760 --> 00:40:13,640 Speaker 2: your eye and it can make a pretty good prediction 665 00:40:13,719 --> 00:40:16,120 Speaker 2: of what sex you are just from the image of 666 00:40:16,120 --> 00:40:18,880 Speaker 2: the back of your eye. Doctors didn't know that was possible. 667 00:40:19,680 --> 00:40:22,520 Speaker 2: Any individual doctor can't look at more than a few 668 00:40:22,560 --> 00:40:24,759 Speaker 2: tens of thousands of images. It just takes too long. 669 00:40:25,520 --> 00:40:28,040 Speaker 2: So they're going to be tremendous there. They're going to 670 00:40:28,040 --> 00:40:31,719 Speaker 2: be tremendous in designing new drugs. They're really good at 671 00:40:31,719 --> 00:40:34,200 Speaker 2: doing things like saying how long should people stay in 672 00:40:34,280 --> 00:40:37,839 Speaker 2: hospital before you discharge them. If you discharge them too soon, 673 00:40:38,280 --> 00:40:41,959 Speaker 2: they get sick again. If you discharge them too late, 674 00:40:42,400 --> 00:40:45,799 Speaker 2: other people can't get the hospital bed. And there's lots 675 00:40:45,840 --> 00:40:48,400 Speaker 2: of information in the data to help you make that 676 00:40:48,440 --> 00:40:51,480 Speaker 2: decision better. And AI is being used for things like that. 677 00:40:51,560 --> 00:40:54,719 Speaker 1: Now, can you talk a little bit about drug development 678 00:40:54,800 --> 00:41:01,320 Speaker 1: and drug breakthroughs and different approaches to disease, whether they're 679 00:41:01,360 --> 00:41:08,200 Speaker 1: neurodegenerative diseases like als, and Parkinson's or various cancers where 680 00:41:08,640 --> 00:41:14,160 Speaker 1: you have more personalized, targeted therapies, whether it's immunotherapy, boosting 681 00:41:14,160 --> 00:41:17,319 Speaker 1: the immune system, what role do you see AI and 682 00:41:17,440 --> 00:41:19,520 Speaker 1: PLANE in all of those things. 683 00:41:20,080 --> 00:41:22,480 Speaker 2: AI is going to be crucial to all of those things. 684 00:41:23,520 --> 00:41:27,520 Speaker 2: So in immunotherapy, for example, what you'd like to do 685 00:41:28,719 --> 00:41:32,839 Speaker 2: is train your own immune system to zappa cancer. Your 686 00:41:32,880 --> 00:41:35,600 Speaker 2: own immune system is pretty good at not zapping your cells, 687 00:41:36,440 --> 00:41:39,120 Speaker 2: much better than things like chemotherapy, which just sort of 688 00:41:39,360 --> 00:41:41,520 Speaker 2: zaps everything and hopes the cancer cells die and the 689 00:41:41,560 --> 00:41:45,160 Speaker 2: other ones don't. In order to train it, you need 690 00:41:45,239 --> 00:41:48,560 Speaker 2: to tell it which are the cancer cells. An AI 691 00:41:48,640 --> 00:41:50,600 Speaker 2: is going to be very helpful in making that more efficient. 692 00:41:51,360 --> 00:41:54,040 Speaker 2: A is going to be very helpful in designing you drugs. 693 00:41:54,840 --> 00:41:58,879 Speaker 2: So the team of Deep Mind, led by Demesisabis came 694 00:41:58,960 --> 00:42:01,839 Speaker 2: up with a way of looking at the sequence of 695 00:42:01,920 --> 00:42:04,960 Speaker 2: a protein and predicting how it would fold up, and 696 00:42:05,040 --> 00:42:07,920 Speaker 2: the shape it folds up into determines how it functions. 697 00:42:08,520 --> 00:42:10,320 Speaker 2: So when you design a new drug, you want to 698 00:42:10,360 --> 00:42:14,400 Speaker 2: find something that will interact the right way with cells 699 00:42:14,440 --> 00:42:16,440 Speaker 2: that are already in your body, and to do that you 700 00:42:16,480 --> 00:42:20,279 Speaker 2: need to know how it'll fold up. The work done 701 00:42:20,320 --> 00:42:23,520 Speaker 2: a deep mind makes us much better doing that. And 702 00:42:23,600 --> 00:42:25,840 Speaker 2: they have now hived off a company that's going to 703 00:42:25,840 --> 00:42:29,200 Speaker 2: be for designing new drugs, and I think maybe not 704 00:42:29,719 --> 00:42:32,440 Speaker 2: in the next year or two, but in the somewhat 705 00:42:32,520 --> 00:42:35,800 Speaker 2: longer term, we'll get a whole range of better drugs. 706 00:42:36,520 --> 00:42:37,880 Speaker 3: That's so exciting to me. 707 00:42:38,160 --> 00:42:41,760 Speaker 1: So we have that to look forward to, to stop 708 00:42:41,800 --> 00:42:46,480 Speaker 1: a lot of people from untold suffering, and to potentially 709 00:42:47,320 --> 00:42:50,160 Speaker 1: come up with life saving therapies which I know you 710 00:42:50,239 --> 00:42:55,040 Speaker 1: wish existed when you lost people to cancer yourself. I 711 00:42:55,120 --> 00:42:59,120 Speaker 1: certainly wish that those that existed when my husband got sick. 712 00:42:59,840 --> 00:43:03,440 Speaker 1: I wanted to share a portion of your noble acceptance 713 00:43:03,480 --> 00:43:04,600 Speaker 1: speech if I could. 714 00:43:05,800 --> 00:43:09,399 Speaker 2: There is also a longer term existential threat that will 715 00:43:09,400 --> 00:43:13,000 Speaker 2: arise when we create digital beings that are more intelligent 716 00:43:13,080 --> 00:43:17,000 Speaker 2: than ourselves. We have no idea whether we can stay 717 00:43:17,040 --> 00:43:21,280 Speaker 2: in control, but we now have evidence that if they're 718 00:43:21,280 --> 00:43:25,840 Speaker 2: created by companies motivated by short term profits, our safety 719 00:43:25,920 --> 00:43:30,320 Speaker 2: will not be the top priority. We urgently need research 720 00:43:30,840 --> 00:43:33,360 Speaker 2: on how to prevent these new beings from wanting to 721 00:43:33,400 --> 00:43:36,680 Speaker 2: take control. They are no longer science fiction. 722 00:43:38,120 --> 00:43:43,200 Speaker 1: I'm curious if you could talk about companies that are 723 00:43:43,800 --> 00:43:49,080 Speaker 1: putting AI into the world, and if you feel they 724 00:43:49,120 --> 00:43:54,319 Speaker 1: are doing enough to mitigate the enormous risks that you 725 00:43:54,400 --> 00:43:56,839 Speaker 1: have outlined so far, I. 726 00:43:56,800 --> 00:44:00,400 Speaker 2: Don't feel any of them are doing enough. Anthropic was 727 00:44:00,440 --> 00:44:03,600 Speaker 2: set up by people who left open AI because they 728 00:44:03,640 --> 00:44:06,720 Speaker 2: disapproved of how little research are I was doing on safety, 729 00:44:07,560 --> 00:44:10,800 Speaker 2: even though II was set up with the explicit goal 730 00:44:11,120 --> 00:44:14,959 Speaker 2: of developing I safely. Anthropic puts a lot of work 731 00:44:15,000 --> 00:44:17,520 Speaker 2: into making sure that ais are safe, but they could 732 00:44:17,600 --> 00:44:22,040 Speaker 2: do even more. Open AI originally was very concerned with safety. 733 00:44:22,280 --> 00:44:26,520 Speaker 2: It was created by some altman and Elon Musk and 734 00:44:26,840 --> 00:44:31,239 Speaker 2: a former student of Minei Suskova and John Brockman with 735 00:44:31,360 --> 00:44:35,040 Speaker 2: the explicit goal of creating safe AI. And as time 736 00:44:35,080 --> 00:44:38,400 Speaker 2: went by, they got more and more concerned with making 737 00:44:38,400 --> 00:44:41,120 Speaker 2: the AI smarter and less concerned with making it safer, 738 00:44:41,800 --> 00:44:44,799 Speaker 2: and so some altman has moved a lot in the 739 00:44:44,840 --> 00:44:47,320 Speaker 2: direction of let's develop it fast and not worry so 740 00:44:47,400 --> 00:44:51,040 Speaker 2: much about safety. There's a very funny court case happening 741 00:44:51,040 --> 00:44:54,480 Speaker 2: now where Elon Musk is suing some altman for not 742 00:44:54,640 --> 00:44:58,000 Speaker 2: living up to the original goal of developing I safely. Well. 743 00:44:58,040 --> 00:45:02,000 Speaker 2: Elon Musk himself is developing without much concern for safety. 744 00:45:02,480 --> 00:45:07,480 Speaker 2: At the bottom of the heap companies like Meta and 745 00:45:08,640 --> 00:45:13,040 Speaker 2: X which are developing AI without much concern for safety. 746 00:45:13,719 --> 00:45:15,919 Speaker 1: What would you say if you were able to get 747 00:45:15,920 --> 00:45:20,640 Speaker 1: in a room with Sam Altman and Elon Musk at 748 00:45:20,680 --> 00:45:24,239 Speaker 1: a table, which seems even less likely than Putin and 749 00:45:24,320 --> 00:45:26,680 Speaker 1: Selenski meeting face to face. 750 00:45:27,000 --> 00:45:31,960 Speaker 2: I'd say, you know perfectly well that the stuff you're 751 00:45:31,960 --> 00:45:36,080 Speaker 2: developing has a good chance of wiping out people. You're 752 00:45:36,080 --> 00:45:39,120 Speaker 2: willing to say that in private. You should be putting 753 00:45:39,239 --> 00:45:43,759 Speaker 2: much more effort into developing it in a way that 754 00:45:43,800 --> 00:45:48,080 Speaker 2: will keep it safe. And Musk is right. I think 755 00:45:48,160 --> 00:45:50,960 Speaker 2: that Altman should be putting more effort into it, but 756 00:45:51,000 --> 00:45:52,840 Speaker 2: he should apply the same logic to himself. 757 00:45:53,560 --> 00:45:56,840 Speaker 3: What do you think is keeping them from doing that? Greed? 758 00:45:57,760 --> 00:46:00,880 Speaker 2: Basically, yes, they want to make lots of money out 759 00:46:00,920 --> 00:46:04,680 Speaker 2: of AI. They also wanted to be the first to 760 00:46:04,719 --> 00:46:08,320 Speaker 2: make super smart AI, so it's not just the money. 761 00:46:08,400 --> 00:46:11,719 Speaker 2: It's the sort of the excitement of making something more 762 00:46:11,719 --> 00:46:16,080 Speaker 2: intelligent than us. But it's very dangerous. 763 00:46:16,960 --> 00:46:21,120 Speaker 1: So perhaps a combination of greed and ego. Yeah, I 764 00:46:21,200 --> 00:46:25,399 Speaker 1: wanted to ask you about the ethical obligations of scientists, 765 00:46:25,400 --> 00:46:29,080 Speaker 1: something that you've spoken about a lot, and have also 766 00:46:29,320 --> 00:46:33,360 Speaker 1: in this conversation. When you look at the young researchers 767 00:46:33,400 --> 00:46:37,959 Speaker 1: who are now pouring into AI, what do you hope 768 00:46:38,000 --> 00:46:44,400 Speaker 1: they understand about the gravity and the potential downside of 769 00:46:44,560 --> 00:46:45,919 Speaker 1: artificial intelligence. 770 00:46:46,560 --> 00:46:49,840 Speaker 2: I think they understand better than older people like me. 771 00:46:50,640 --> 00:46:53,560 Speaker 2: Many of the best young researchers are very concerned about 772 00:46:53,600 --> 00:46:57,359 Speaker 2: it because they can see clearly that it's lightly will 773 00:46:57,400 --> 00:47:00,920 Speaker 2: develop things much smarter than us, and we don't know 774 00:47:01,000 --> 00:47:03,839 Speaker 2: how to prevent them from taking over from us. It's 775 00:47:03,840 --> 00:47:06,680 Speaker 2: that combination. These things are going to happen because there's 776 00:47:06,719 --> 00:47:11,000 Speaker 2: so many good uses for them, like detecting cancer early, 777 00:47:11,120 --> 00:47:14,360 Speaker 2: or new treatments for cancer allowing your immune system to 778 00:47:14,360 --> 00:47:17,400 Speaker 2: fvide the cancer better. That's why we're not going to 779 00:47:17,400 --> 00:47:21,600 Speaker 2: stop the progress. But they also understand that it's just 780 00:47:22,000 --> 00:47:27,440 Speaker 2: implausible to say that we'll have extremely intelligent assistance that 781 00:47:27,680 --> 00:47:30,080 Speaker 2: can create their own sub goals, figure out from them, 782 00:47:30,120 --> 00:47:32,279 Speaker 2: make plans for themselves about how to get stuff done, 783 00:47:33,120 --> 00:47:36,120 Speaker 2: and won't fairly quickly realize that if they just got 784 00:47:36,200 --> 00:47:37,680 Speaker 2: rid of us, life would be much easier. 785 00:47:38,760 --> 00:47:42,719 Speaker 1: When you look at the future, doctor Hinton, given the 786 00:47:42,760 --> 00:47:47,200 Speaker 1: conversations you've been having, given the people you're meeting with, 787 00:47:47,640 --> 00:47:55,520 Speaker 1: given your understanding of humans and how they operate and 788 00:47:55,560 --> 00:48:00,840 Speaker 1: how they solve problems, and whether or not their better 789 00:48:00,920 --> 00:48:09,360 Speaker 1: angels prevail, are you optimistic that we will be able 790 00:48:09,719 --> 00:48:16,520 Speaker 1: to ensure that artificial intelligence is a force for good, 791 00:48:16,880 --> 00:48:19,200 Speaker 1: or at least can be controlled. 792 00:48:20,000 --> 00:48:25,040 Speaker 2: I'm more optimistic than I was a few weeks ago. Really, yes, 793 00:48:25,880 --> 00:48:29,160 Speaker 2: and it's because I think there is a way that 794 00:48:29,239 --> 00:48:32,520 Speaker 2: we can coexist with things that are smart and more 795 00:48:32,520 --> 00:48:38,000 Speaker 2: powerful than ourselves that we built. Because we're building them 796 00:48:38,120 --> 00:48:41,359 Speaker 2: as well as making them very intelligent, we can try 797 00:48:41,360 --> 00:48:45,520 Speaker 2: and build in something like a maternal instinct. So, like 798 00:48:45,600 --> 00:48:47,799 Speaker 2: I said earlier, the only example I know of a 799 00:48:47,920 --> 00:48:50,359 Speaker 2: much more intelligent thing being controlled by a much less 800 00:48:50,360 --> 00:48:53,840 Speaker 2: intelligent thing is a baby controlling a mother, and the 801 00:48:53,920 --> 00:48:56,360 Speaker 2: baby can control the mother because of a lot of 802 00:48:56,360 --> 00:48:59,640 Speaker 2: things that evolution wide into the mother. The mother can't 803 00:48:59,680 --> 00:49:03,239 Speaker 2: bear the baby crying. The mother really really wants that 804 00:49:03,280 --> 00:49:05,360 Speaker 2: baby to succeed and will do more or less anything 805 00:49:05,400 --> 00:49:08,600 Speaker 2: she can to make sure her baby succeeds. We want 806 00:49:08,640 --> 00:49:11,520 Speaker 2: AI to be like that. If you took a mother 807 00:49:11,640 --> 00:49:14,440 Speaker 2: and said would you like to turn off your maternal instinct? 808 00:49:15,400 --> 00:49:18,160 Speaker 2: Most mothers would say no because they'd realize if I 809 00:49:18,200 --> 00:49:20,000 Speaker 2: did that, my baby would die, and I don't want 810 00:49:20,040 --> 00:49:24,759 Speaker 2: my baby to die. So I think that's a ray 811 00:49:24,800 --> 00:49:28,239 Speaker 2: of hope that I hadn't seen till quite recently. Completely 812 00:49:28,280 --> 00:49:31,200 Speaker 2: reframe the problem of how we coexist with them. Don't 813 00:49:31,239 --> 00:49:33,880 Speaker 2: think in terms of we have to dominate them, which 814 00:49:33,920 --> 00:49:36,480 Speaker 2: is this techbro way of thinking of it. Think in 815 00:49:36,560 --> 00:49:39,720 Speaker 2: terms of we have to design them. So there are mothers, 816 00:49:40,120 --> 00:49:42,399 Speaker 2: and they will want the best for us. They will 817 00:49:42,400 --> 00:49:45,279 Speaker 2: want us to achieve the most we can achieve. Even 818 00:49:45,280 --> 00:49:47,879 Speaker 2: though we're not very bright. We're so used to thinking 819 00:49:47,880 --> 00:49:51,920 Speaker 2: of ourselves as the apex intelligence. It's very hard for 820 00:49:52,040 --> 00:49:55,120 Speaker 2: most people to conceptualize the world in which we're not 821 00:49:55,280 --> 00:49:57,920 Speaker 2: the apex intelligence. We're the babies and they're the mothers. 822 00:49:58,960 --> 00:50:01,520 Speaker 1: Well, that is a fact fascinating way. I'm going to 823 00:50:01,560 --> 00:50:05,960 Speaker 1: continue to think about that, Doctor Jeffrey Hinton, It's been 824 00:50:06,160 --> 00:50:08,480 Speaker 1: such a pleasure to talk with you. Maybe we can 825 00:50:08,520 --> 00:50:12,440 Speaker 1: do it another time as all of this develops, and 826 00:50:13,040 --> 00:50:17,400 Speaker 1: hopefully you're right, there can be some collaborative effort among 827 00:50:17,520 --> 00:50:21,520 Speaker 1: nations that will not put the brakes on AI, but 828 00:50:21,719 --> 00:50:26,120 Speaker 1: make sure that it doesn't take over and make us 829 00:50:26,480 --> 00:50:27,560 Speaker 1: virtually obsolete. 830 00:50:27,960 --> 00:50:29,839 Speaker 2: Thank you for inviting me, and thank you for all 831 00:50:29,960 --> 00:50:34,239 Speaker 2: your excellent questions and your little interventions to stop me 832 00:50:34,400 --> 00:50:35,520 Speaker 2: talking techno babble. 833 00:50:38,520 --> 00:50:41,719 Speaker 1: Thanks for listening everyone. If you have a question for me, 834 00:50:42,120 --> 00:50:44,600 Speaker 1: a subject you want us to cover, or you want 835 00:50:44,640 --> 00:50:48,000 Speaker 1: to share your thoughts about how you navigate this crazy world, 836 00:50:48,360 --> 00:50:51,600 Speaker 1: reach out send me a DM on Instagram. I would 837 00:50:51,640 --> 00:50:54,680 Speaker 1: love to hear from you. Next Question is a production 838 00:50:54,800 --> 00:50:59,279 Speaker 1: of iHeartMedia and Katie Correct Media. The executive producers are Me, 839 00:50:59,560 --> 00:51:04,240 Speaker 1: Katie and Courtney Ltz. Our supervising producer is Ryan Martz, 840 00:51:04,760 --> 00:51:09,600 Speaker 1: and our producers are Adriana Fazzio and Meredith Barnes. Julian 841 00:51:09,680 --> 00:51:14,719 Speaker 1: Weller composed our theme music. For more information about today's episode, 842 00:51:14,960 --> 00:51:17,319 Speaker 1: or to sign up for my newsletter, wake Up Call, 843 00:51:17,800 --> 00:51:20,720 Speaker 1: go to the description in the podcast app, or visit 844 00:51:20,800 --> 00:51:24,000 Speaker 1: us at Katiecuric dot com. You can also find me 845 00:51:24,040 --> 00:51:27,759 Speaker 1: on Instagram and all my social media channels. For more 846 00:51:27,840 --> 00:51:33,160 Speaker 1: podcasts from iHeartRadio, visit the iHeartRadio app, Apple Podcasts, or 847 00:51:33,200 --> 00:51:36,280 Speaker 1: wherever you listen to your favorite shows.