WEBVTT - Bloomberg Wall Street Week - October 25th, 2024 0:00:15.920 --> 0:00:19.119 This is Wall Street Week. I'm David Weston bringing you 0:00:19.400 --> 0:00:23.479 stories of capitalism. This week an industrial boom south of 0:00:23.520 --> 0:00:26.360 the border. We went to Mexico to see how tariffs 0:00:26.400 --> 0:00:30.120 and transportation costs are reshaping the American supply chain and 0:00:30.520 --> 0:00:33.600 what that could mean for China. But we begin with 0:00:33.680 --> 0:00:37.199 a biography, not one about a person or even really 0:00:37.320 --> 0:00:37.720 a thing. 0:00:38.320 --> 0:00:39.080 How do we. 0:00:39.080 --> 0:00:43.519 Classify something like artificial intelligence, something that has never existed 0:00:43.560 --> 0:00:47.559 before and something that might be beyond our understanding. This 0:00:47.640 --> 0:00:51.080 is a story about identity, who we are as humans, 0:00:51.280 --> 0:00:54.680 whether our identity is being merged with computers, and if so, 0:00:55.280 --> 0:00:56.440 what we need to do about it. 0:00:58.760 --> 0:01:00.440 There's nothing special that people. 0:01:00.720 --> 0:01:04.560 They are very wonderful things, and incredibly we're incredibly complicated, 0:01:05.880 --> 0:01:08.160 and we're very wonderful to other people. 0:01:09.600 --> 0:01:12.199 That's people and what we care about. We've evolved that way. 0:01:13.800 --> 0:01:18.560 But there's nothing that you can't replicate in other material. 0:01:19.240 --> 0:01:22.400 If anyone should know about whether computers can do everything 0:01:22.440 --> 0:01:25.760 that humans can, it would be Jeffrey Hinton, the Computer 0:01:25.880 --> 0:01:29.640 scientists and University of Toronto professor emeritus, is known as 0:01:29.680 --> 0:01:33.360 the godfather of AI. After working for over a decade 0:01:33.360 --> 0:01:35.880 at Google, he quit his job in twenty twenty three 0:01:36.040 --> 0:01:39.440 and became outspoken about the risks of the technology. He 0:01:39.440 --> 0:01:44.000 helped pioneer this, even as he continues to receive accolades 0:01:44.040 --> 0:01:46.920 for his work. We sat down with him after he 0:01:46.959 --> 0:01:50.160 had just received the Nobel Prize in physics. 0:01:50.840 --> 0:01:52.720 I was in a hotel in California and my phone 0:01:52.720 --> 0:01:54.800 went off at one in the morning, and I thought 0:01:54.800 --> 0:01:56.800 about not answering it, and then they said I won 0:01:56.840 --> 0:02:01.160 the Nobel Prize in physics, and I thought, wait a minute, 0:02:01.240 --> 0:02:02.400 I'm not a physicist. 0:02:02.200 --> 0:02:03.240 This can't be right. 0:02:03.680 --> 0:02:06.440 But it was right. And in December it will be 0:02:06.480 --> 0:02:10.240 Professor Hinton, along with John Hopfield, who will be receiving 0:02:10.280 --> 0:02:13.160 the award at a ceremony in Stockholm for their breakthrough 0:02:13.200 --> 0:02:17.480 work on machine learning. That generative AI is reshaping the 0:02:17.520 --> 0:02:22.280 business landscape is beyond doubt. Bloomberg Intelligence expects it to 0:02:22.400 --> 0:02:25.640 drive big tech firms to nearly double their CAPEX spending 0:02:25.680 --> 0:02:29.320 from twenty twenty three to twenty twenty five, with revenue 0:02:29.360 --> 0:02:32.480 from AI set to approach a trillion dollars by the 0:02:32.600 --> 0:02:35.120 end of the decade. But before we get to the 0:02:35.160 --> 0:02:38.520 future of AI, we turned to its past. We asked 0:02:38.560 --> 0:02:41.200 Jeffrey Hinton to go back to the beginning, when this 0:02:41.360 --> 0:02:45.040 new creature that can do what humans can do first arrived. 0:02:46.280 --> 0:02:50.320 I would say it was born in about nineteen eighty 0:02:50.440 --> 0:02:56.160 three with something that Terry Sinowski and I did, which 0:02:56.240 --> 0:02:59.320 was called the Boltsham Machine. It never really panned out, 0:02:59.400 --> 0:03:01.800 but it was a generaty of AI. It was a 0:03:01.800 --> 0:03:04.520 model that would generate images, not very good images, but 0:03:04.520 --> 0:03:05.399 it would generate them. 0:03:05.560 --> 0:03:06.160 Well. 0:03:06.400 --> 0:03:08.639 There could be a debate about the birth of it, 0:03:08.680 --> 0:03:12.519 but I would go all the way back to Alan 0:03:12.600 --> 0:03:16.120 Turing and then John von Neumann, so that would be 0:03:16.520 --> 0:03:19.200 in the thirties, forties and fifties. 0:03:19.720 --> 0:03:22.680 Marty Chavez of Sixth Street has spent his career developing 0:03:22.720 --> 0:03:25.519 AI and he traces its birth back to the very 0:03:25.560 --> 0:03:27.720 foundations of computer science. 0:03:28.480 --> 0:03:32.200 So it's been around a while, right, people thinking about 0:03:32.560 --> 0:03:37.680 what is computation, What is the nature of computation? And 0:03:38.240 --> 0:03:42.720 do human beings do computations? Is everything that's happening in 0:03:42.720 --> 0:03:46.000 our brains or our minds just a form of computation. 0:03:46.120 --> 0:03:50.240 People were asking these questions really almost almost one hundred 0:03:50.320 --> 0:03:52.200 years ago and starting to work on it. 0:03:52.320 --> 0:03:55.040 Whatever the beginnings of AI, most agree that it took 0:03:55.200 --> 0:03:58.400 many people in many decades to get here, But they 0:03:58.400 --> 0:04:01.880 also agree that in just the last few years the 0:04:01.960 --> 0:04:05.480 game has changed. A series of rapid breakthroughs has brought 0:04:05.520 --> 0:04:09.000 what were far out predictions much closer to reality. 0:04:09.720 --> 0:04:14.520 The big milestone is the large language models, which work 0:04:14.600 --> 0:04:18.160 by looking at a context and then generating guesses about 0:04:18.160 --> 0:04:19.560 what they think the next word should be. 0:04:19.960 --> 0:04:21.000 That's why they generative. 0:04:22.000 --> 0:04:24.800 Then that was followed by things like Dali from open 0:04:24.839 --> 0:04:27.680 Ai that could generate images. Before that, there were things 0:04:27.680 --> 0:04:30.880 that could generate captions for images, so they would look 0:04:30.880 --> 0:04:33.360 at an image and then they would generate a string 0:04:33.360 --> 0:04:34.760 of words that describe the image. 0:04:36.560 --> 0:04:38.880 Those were the early versions of generative area. 0:04:39.080 --> 0:04:43.600 Now we're getting multimodal chatbots which can generate everything. They 0:04:43.600 --> 0:04:45.240 can generate images and language. 0:04:45.480 --> 0:04:50.480 The huge milestone that I think of is one that 0:04:50.560 --> 0:04:55.960 we hit twenty sixteen twenty seventeen. This was a milestone 0:04:56.080 --> 0:04:59.279 in the evolution of the of the neural networks. It 0:04:59.360 --> 0:05:03.240 maybe didn't get a lot of public attention, but it 0:05:03.360 --> 0:05:07.240 certainly got the attention of everybody in the field. And 0:05:07.920 --> 0:05:11.839 this milestone came about because of the work of many 0:05:11.880 --> 0:05:15.360 of the researchers. I mentioned a few of them for 0:05:15.520 --> 0:05:19.720 over decades, but another contributor to the milestone was the 0:05:19.760 --> 0:05:25.240 Internet and the a sudden availability of vast amounts of 0:05:25.839 --> 0:05:27.120 labeled data. 0:05:27.640 --> 0:05:30.520 The breakthroughs that Hinton and others worked to achieve have 0:05:30.600 --> 0:05:33.960 developed artificial intelligence to the point that now, in a 0:05:34.000 --> 0:05:38.119 text conversation, it might be indistinguishable from a human being. 0:05:38.720 --> 0:05:41.440 A study this year by cognitive scientists that you see 0:05:41.480 --> 0:05:44.719 San Diego showed that people mistook chat GPT for a 0:05:44.800 --> 0:05:48.039 human more than fifty percent of the time, clearing the 0:05:48.080 --> 0:05:51.920 bar for the so called touring test. Humans themselves only 0:05:51.960 --> 0:05:54.479 passed the test about two thirds of the time. But 0:05:54.560 --> 0:05:58.480 as surprising as that is, it could be just the beginning. 0:05:59.200 --> 0:06:02.160 How far along is AI in its life? Again, if 0:06:02.160 --> 0:06:06.160 we're anthropomorphizing, here, is it a child? Is an adolescent? 0:06:06.200 --> 0:06:06.640 Where is it? 0:06:06.560 --> 0:06:12.880 It's more like a toddler, maybe like a three year old. 0:06:13.240 --> 0:06:17.000 Let's imagine that we're in that gangly, awkward teenager phase 0:06:17.440 --> 0:06:23.440 of AI. What happens next? There's a huge debate. 0:06:23.680 --> 0:06:26.919 Whether AI is a toddler or an adolescent. Today. The 0:06:27.000 --> 0:06:29.920 question is how far can it go? How much can 0:06:29.960 --> 0:06:33.120 it learn? And perhaps most important, will it catch up 0:06:33.160 --> 0:06:36.599 to us humans? For Jeffrey Hinton. It's a matter of when, 0:06:37.120 --> 0:06:42.360 not if. Does that mean that general of artificial intelligence 0:06:43.160 --> 0:06:46.200 will be able to do whatever humans can do? Is 0:06:46.240 --> 0:06:49.800 there anything that it cannot do that we humans can do? 0:06:50.200 --> 0:06:53.799 Yes, most people have the belief there's something very special 0:06:53.839 --> 0:06:58.039 at people, which is consciousness or sentience or subjective experience, 0:06:58.520 --> 0:06:59.599 and we have this in a theater. 0:07:00.120 --> 0:07:01.080 Probably all nonsense. 0:07:01.440 --> 0:07:04.279 The history of it is, these large language models predicting 0:07:04.279 --> 0:07:07.640 the next words using backpropagation were initially designed as a 0:07:07.680 --> 0:07:10.840 model of how people work. Now there's a whole bunch 0:07:10.880 --> 0:07:13.000 of people who believe in symbolic logic and that's how 0:07:13.000 --> 0:07:15.720 people must think. They think it must be true, who 0:07:15.720 --> 0:07:18.200 think people work a completely different way, and they're wrong. 0:07:18.240 --> 0:07:20.200 They never managed to produce stuff that works nearly as 0:07:20.240 --> 0:07:24.120 well as these big language models. So they'll tell you 0:07:24.160 --> 0:07:26.320 these big language models are quite unlike people. They're just 0:07:26.440 --> 0:07:28.680 using correlations and so on, and it's all just not 0:07:28.880 --> 0:07:32.800 really understanding. No, our best model of how people understand 0:07:33.160 --> 0:07:35.160 is exactly the same model as how these big language 0:07:35.160 --> 0:07:38.360 models understand. They're just like us, and they're much more 0:07:38.400 --> 0:07:41.160 like us than they are like standard computer software. So 0:07:41.360 --> 0:07:44.320 understanding is all about converting words into features, having the 0:07:44.320 --> 0:07:48.800 features interact, and having those predict the next word. We 0:07:48.840 --> 0:07:50.600 do it pretty much the same way they do it. 0:07:50.640 --> 0:07:54.960 What about creativity? Can general AI be as creative as 0:07:55.000 --> 0:07:55.520 humans are? 0:07:56.200 --> 0:07:58.840 As someone else might have said, There you go again, 0:08:00.240 --> 0:08:03.920 you want people to be special. If you take standard 0:08:03.920 --> 0:08:07.160 tests of creativity, AIS already do better than most people. 0:08:07.520 --> 0:08:11.680 What about in a large corporation, you have worker bees 0:08:11.760 --> 0:08:13.840 we often have heard of them. Then you have middle management, 0:08:14.200 --> 0:08:17.080 upper management and CEOs? Could they replace CEOs? 0:08:17.800 --> 0:08:20.440 Okay, so there's a good scenario for the future which 0:08:20.480 --> 0:08:21.040 goes like this. 0:08:21.600 --> 0:08:22.880 Suppose you have a big. 0:08:22.600 --> 0:08:26.440 Company with a rather dumb CEO. Maybe he's the son 0:08:26.480 --> 0:08:30.720 of the previous CEO, but he has an executive assistant. 0:08:30.280 --> 0:08:32.800 Who's very good, almost certainly a woman. 0:08:33.040 --> 0:08:36.719 This executive assistant is asked to achieve things by the 0:08:36.800 --> 0:08:39.319 dumb CEO, and she achieves them in ways he doesn't 0:08:39.360 --> 0:08:42.000 understand and would never have thought of, so he becomes 0:08:42.040 --> 0:08:45.320 much more effective. He gets what he wants, he thinks 0:08:45.360 --> 0:08:49.240 he's doing it. It all happens behind the scenes and 0:08:49.360 --> 0:08:53.080 very efficiently. That's how it could be if we all 0:08:53.200 --> 0:08:56.719 had our own AI assistants who were smarter than us, 0:08:56.760 --> 0:08:58.680 and they were thoroughly benevolent, and they never wanted to 0:08:58.720 --> 0:09:02.720 take over because somehow achieve that that's the good scenario. 0:09:03.160 --> 0:09:05.760 I had a boss once in business who said, a 0:09:05.880 --> 0:09:09.960 danger is wishing yourselves to success. Do we wish ourselves 0:09:10.080 --> 0:09:14.280 to a safer world by hoping that General AI cannot 0:09:14.280 --> 0:09:16.360 do everything we can do, because it might be very 0:09:16.400 --> 0:09:18.600 threatening to many people if we believed it can do 0:09:18.720 --> 0:09:19.679 everything we can do. 0:09:20.880 --> 0:09:23.720 Yes, it is very threatening and I'm very worried about it, 0:09:24.679 --> 0:09:29.920 and we should definitely put huge effort right now into 0:09:29.960 --> 0:09:31.760 figuring out whether we can keep control of it. 0:09:32.080 --> 0:09:34.319 What effort should we put forward, what should we do? 0:09:34.840 --> 0:09:38.560 Okay, so for something like climate change, everybody knows what 0:09:38.600 --> 0:09:41.600 we should do. There's a very simple solution, just don't 0:09:41.600 --> 0:09:45.400 burn carbon. It's just we don't have the political will. 0:09:45.640 --> 0:09:47.240 We still give big subsidies to. 0:09:47.200 --> 0:09:51.839 Oil companies, which is lunatic. So we know what to do, 0:09:51.880 --> 0:09:56.120 we just won't do it with AI. To keep AI 0:09:56.280 --> 0:10:00.160 safe from the long term danger of more INtime and 0:10:00.240 --> 0:10:02.960 things than us just taking over from us, we don't 0:10:03.000 --> 0:10:05.000 even know what to do. What we need to do 0:10:05.080 --> 0:10:09.360 now is understand that there really is this danger. These 0:10:09.400 --> 0:10:11.600 things really do understand and they really are going to 0:10:11.600 --> 0:10:14.600 get smarter than us. But also how are we going 0:10:14.679 --> 0:10:17.280 to keep control of them? We have some chance because 0:10:17.400 --> 0:10:18.439 we're creating. 0:10:18.080 --> 0:10:22.600 Them, and that's where we turn next in the biography 0:10:22.600 --> 0:10:25.840 of AI. How much good can it do? And what 0:10:25.880 --> 0:10:28.120 do we need to do to make sure it becomes 0:10:28.120 --> 0:10:31.320 a responsible citizen instead of a threat to us. 0:10:31.320 --> 0:10:35.640 All many people say, you know, I'll create more jobs 0:10:36.320 --> 0:10:39.840 for this particular thing. I'm not convinced of that. What 0:10:39.880 --> 0:10:45.199 we're doing in the Industrial Revolution we make human strength irrelevant. 0:10:46.200 --> 0:10:50.440 Now we're making human intelligence irrelevant, and that's very scary. 0:10:54.000 --> 0:10:57.400 You're listening to Bloomberg Wall Street Week with David Weston 0:10:57.559 --> 0:11:06.240 from Bloomberg Radio on November fifth, America designs. 0:11:05.760 --> 0:11:10.600 A remarkable seismic shifts in American politics, and Bloomberg is covering. 0:11:10.280 --> 0:11:14.000 Every angle now through election day live from our nation's capital. 0:11:14.040 --> 0:11:16.720 Our polling shows more enthusiasm among black voters. 0:11:16.720 --> 0:11:19.480 Spell and s of Pop with Jim Mathew and Kayley Lyne. 0:11:19.720 --> 0:11:22.800 Where voters see Trump and Harris on the issues. 0:11:22.920 --> 0:11:24.720 How important is higher turnouts? 0:11:24.760 --> 0:11:27.840 Listen live weekdays at noon and five pm Eastern, or 0:11:27.880 --> 0:11:31.839 on demand wherever you get your podcasts. Bloomberg Radio. Context 0:11:31.920 --> 0:11:32.880 changes everything. 0:11:33.840 --> 0:11:37.400 This is the Bloomberg Green Report. A company called form 0:11:37.640 --> 0:11:39.960 Energy has developed what could turn out to be the 0:11:40.040 --> 0:11:43.160 right product at just the right time. It's a utility 0:11:43.240 --> 0:11:46.160 sized battery that can feed the electricity to the power 0:11:46.200 --> 0:11:49.760 grid for one hundred hours straight. That's twenty five times 0:11:49.800 --> 0:11:52.680 longer than most of the grid tide batteries in use today. 0:11:53.080 --> 0:11:56.160 The batteries will start arriving at power plants from Colorado 0:11:56.240 --> 0:11:59.079 to Virginia next year. They're coming at a time when 0:11:59.160 --> 0:12:01.640 electricity did de manned, is being driven up by new 0:12:01.679 --> 0:12:05.600 factories and data centers, and extreme weather events like Hurricanes 0:12:05.640 --> 0:12:09.640 Helene and Milton have triggered massive blackouts. Those are potent 0:12:09.720 --> 0:12:13.200 selling points. Form Energy is based near Boston. It has 0:12:13.280 --> 0:12:16.120 raised one point two billion dollars from investors who are 0:12:16.160 --> 0:12:20.000 betting big on the company's timing and technology. Bloomberg NEF 0:12:20.040 --> 0:12:23.079 says that's especially remarkable in light of a decline in 0:12:23.160 --> 0:12:27.400 overall climate tech financing since last year. Venture capital has 0:12:27.440 --> 0:12:32.319 been strained by high interest rates and macroeconomic headwinds. Jeff Bellinger, 0:12:32.480 --> 0:12:33.440 Bloomberg Radio. 0:12:33.800 --> 0:12:35.959 Out Here in the middle of these acres, it can 0:12:35.960 --> 0:12:38.480 feel like you're the only person on earth. That's how 0:12:38.520 --> 0:12:40.800 it feels when you're struggling with your mental health. But 0:12:40.920 --> 0:12:43.800 you don't have to feel alone. Find more information at 0:12:43.800 --> 0:12:46.280 Love youormind Today dot org, brought to you by the 0:12:46.360 --> 0:12:48.360 Huntsman Mentuealth Institute in the AD Council. 0:13:03.720 --> 0:13:07.960 This is Bloomberg Wall Street Week with David Weston from 0:13:08.080 --> 0:13:09.640 Bloomberg Radio. 0:13:12.200 --> 0:13:15.360 The life of artificial intelligence is connected to the lives 0:13:15.400 --> 0:13:19.679 of people and also to their livelihoods. AI is already 0:13:19.720 --> 0:13:22.880 being used to make processes more efficient. What could this 0:13:23.000 --> 0:13:27.480 mean for productivity and therefore economic growth and who will 0:13:27.480 --> 0:13:31.240 benefit from it? We pose those questions to Jeffrey Hinton. 0:13:33.080 --> 0:13:36.400 It will be a wonderful thing for productivity, that's true. 0:13:36.679 --> 0:13:39.360 Whether it be a wonderful thing for society is something else. Altogether. 0:13:40.040 --> 0:13:44.320 In a decent society, if you increase productivity a lot, 0:13:44.520 --> 0:13:48.360 everybody's better off. But here, what's going to happen if 0:13:48.400 --> 0:13:53.160 you increase productivity a lot. The rich and the big 0:13:53.200 --> 0:13:57.400 companies are going to get much richer, and ordinary people 0:13:57.440 --> 0:13:59.400 are probably going to be worse off because they lose 0:13:59.440 --> 0:13:59.960 their jobs. 0:14:01.080 --> 0:14:04.800 AI's ability to reshape productivity could also give new life 0:14:04.840 --> 0:14:08.960 to developed economies whose productivity has stalled. US productivity growth 0:14:08.960 --> 0:14:11.600 has been significantly lower in the twenty first century than 0:14:11.640 --> 0:14:15.200 it was between nineteen thirty and two thousand. But for 0:14:15.320 --> 0:14:18.520 all the talk of the sweeping effects of AI, economic 0:14:18.600 --> 0:14:22.520 change has been slow to materialize. So far. Goldwyn Sachs 0:14:22.560 --> 0:14:25.560 reports that only five percent of companies claim to use 0:14:25.680 --> 0:14:30.000 generative AI in regular production, with tech and information businesses 0:14:30.080 --> 0:14:34.080 leading the way. When wider adoption does come, Hinton anticipates 0:14:34.080 --> 0:14:37.040 it will have very different impacts on different parts of 0:14:37.080 --> 0:14:37.760 the workforce. 0:14:38.320 --> 0:14:41.680 Many people say, you know, illegrate more jobs for this 0:14:41.760 --> 0:14:42.560 particular thing. 0:14:43.120 --> 0:14:44.160 I'm not convinced of that. 0:14:44.680 --> 0:14:48.920 What we're doing in the Industrial Revolution we make human 0:14:49.000 --> 0:14:54.520 strength irrelevant. Now we're making human intelligence irrelevant, and that's 0:14:54.640 --> 0:14:59.480 very scary. So there's some areas where demand is very elastic. 0:15:00.040 --> 0:15:03.360 An example would be healthcare. If I could get ten 0:15:03.400 --> 0:15:06.480 hours a week talking to my doctor, I'm over seventy 0:15:07.720 --> 0:15:11.720 I'd be very happy. So if you take someone and 0:15:11.800 --> 0:15:14.360 make them much more efficient by having them work with 0:15:14.440 --> 0:15:18.640 a very intelligent AI, they're not gonna become unemployed. It's 0:15:18.640 --> 0:15:20.360 not that you're now only gonna need a few of them. 0:15:20.560 --> 0:15:22.240 You're just gonna get much more healthcare. 0:15:22.760 --> 0:15:23.080 Great. 0:15:23.880 --> 0:15:26.560 So in elastic areas, it's great. The some areas that 0:15:26.640 --> 0:15:31.920 are less elastic, like I have a niece who answers 0:15:32.280 --> 0:15:35.280 letters a complaint to a health service. She used to 0:15:35.280 --> 0:15:37.760 take twenty five minutes to answer a letter. Now she 0:15:37.840 --> 0:15:41.760 can just scan the letter into chat GBT, it'll give 0:15:41.760 --> 0:15:43.440 an answer. She'll look at it, check it is okay, 0:15:43.480 --> 0:15:47.840 that's five minutes now. I suspect they'll need less people 0:15:47.920 --> 0:15:50.920 like that. It may be they can just everybody can 0:15:50.920 --> 0:15:53.240 complain a lot more, but I suspect they'll need less. 0:15:53.080 --> 0:15:53.640 People like that. 0:15:53.760 --> 0:15:56.400 So some jobs are Some jobs are elastic, others aren't. 0:15:56.640 --> 0:15:59.640 The non elastic ones, I think people will lose their jobs. 0:16:00.320 --> 0:16:04.680 And what's going to happen is the extra wealth created 0:16:04.720 --> 0:16:07.440 by the increasing productivity is not going to go to them. 0:16:08.080 --> 0:16:11.200 Hinton's prize in physics wasn't the only Nobel awarded this 0:16:11.280 --> 0:16:15.880 year for artificial intelligence. Two Google DeepMind scientists, a professor 0:16:15.920 --> 0:16:19.360 at the University of Washington receive Nobels in chemistry for 0:16:19.520 --> 0:16:22.600 using AI to predict the structure of millions of proteins 0:16:22.840 --> 0:16:26.560 and even to invent a whole new protein. Marty Chavez 0:16:26.600 --> 0:16:30.360 A sixth Street says that tool called alpha fold shows 0:16:30.400 --> 0:16:31.920 the power of machine learning. 0:16:32.400 --> 0:16:37.040 As important a breakthrough as this Alpha fold work is 0:16:37.480 --> 0:16:40.880 if you think about the problems in biology, the problems 0:16:40.920 --> 0:16:44.440 in biology are so much more complicated. 0:16:44.600 --> 0:16:44.840 Right. 0:16:45.040 --> 0:16:48.520 Proteins are the basic building block of life. But then 0:16:48.560 --> 0:16:53.400 those proteins organize themselves into organ eels of a cell, 0:16:53.560 --> 0:16:57.880 and then cells, which organize themselves into tissues, which organize 0:16:57.920 --> 0:17:02.720 themselves into organs which go to organ systems human beings, 0:17:03.120 --> 0:17:04.960 populations of human beings. 0:17:05.440 --> 0:17:09.040 Beyond biology and health sciences, generative AI could also have 0:17:09.160 --> 0:17:12.640 profound effects on the climate. For one thing, it will 0:17:12.640 --> 0:17:16.399 require a lot more electricity to drive it. Wells Fargo 0:17:16.440 --> 0:17:19.159 projects a five hundred and fifty percent surge in AI 0:17:19.280 --> 0:17:22.720 power demand by twenty twenty six. Ireland is a poster 0:17:22.840 --> 0:17:25.919 child for the extent of that demand. Today, it's growing 0:17:26.000 --> 0:17:28.720 number of data centers used up twenty one percent of 0:17:28.760 --> 0:17:31.080 the country's electricity in twenty twenty three. 0:17:31.680 --> 0:17:33.280 The good news is it's not new in our history. 0:17:33.320 --> 0:17:35.600 We've been through many periods as a country where we 0:17:35.680 --> 0:17:39.159 have had significant energy demand. But what this is going 0:17:39.200 --> 0:17:41.080 to mean is AI is now adding to this as 0:17:41.080 --> 0:17:44.080 we electrify transportation, as we electrify our buildings, we now 0:17:44.119 --> 0:17:47.399 have this big near term demand for energy to power 0:17:47.480 --> 0:17:48.320 these data centers. 0:17:48.880 --> 0:17:52.000 Brian Deese served as the Director of the National Economic 0:17:52.040 --> 0:17:55.360 Council under President Biden and is now focused on studying 0:17:55.359 --> 0:17:58.000 the effects of AI on energy at MIT. 0:17:58.640 --> 0:17:58.720 SO. 0:17:58.840 --> 0:18:01.040 In the near term, it creates this tension of how 0:18:01.040 --> 0:18:04.280 do we bring that additional electricity online? And are we 0:18:04.320 --> 0:18:07.600 bringing cleaner sources online or dirtier sources online, which obviously 0:18:07.600 --> 0:18:10.520 affects emissions. But the longer term, the bigger picture way 0:18:10.560 --> 0:18:13.360 these interact is AI is a technology which will then 0:18:13.359 --> 0:18:16.520 get deployed in lots of aspects of our economy, including 0:18:16.560 --> 0:18:19.520 our energy system. And so there are also opportunities where 0:18:19.640 --> 0:18:22.320 these two technological developments coming at the same time could 0:18:22.320 --> 0:18:25.800 actually be one plus one equals three help us find 0:18:25.880 --> 0:18:29.439 more efficient ways to innovate in the energy space. 0:18:30.240 --> 0:18:33.119 For all the good that generative AI could potentially do 0:18:33.240 --> 0:18:36.200 for the human race, it also poses some real dangers, 0:18:36.640 --> 0:18:40.199 perhaps none so apparent as its application to armed warfare. 0:18:40.680 --> 0:18:45.040 Anya Manual's Aspens Strategy Group publisher report on this very subject. 0:18:45.560 --> 0:18:49.119 AI by its nature is dual use because it is 0:18:49.160 --> 0:18:52.440 a multipurpose technology. It's like electricity or the steam engine. 0:18:52.480 --> 0:18:56.639 It's going to power everything, from the most amazing positive 0:18:57.000 --> 0:18:59.320 use cases to the really dangerous ones. 0:19:00.160 --> 0:19:03.760 All of my lifetime, the biggest geopolitical risk terror even 0:19:04.000 --> 0:19:08.520 has been the threat of thermonuclear destruction. Recently was a 0:19:08.520 --> 0:19:11.000 meeting over in South Korea with an attempt to have 0:19:11.040 --> 0:19:14.199 everybody agreed that if there were to be nuclear weapons used, 0:19:14.280 --> 0:19:16.960 a human had to be in that chain of command, 0:19:17.440 --> 0:19:21.199 and the countries there agreed to it except for one China. 0:19:21.400 --> 0:19:22.320 What do we make of that? 0:19:22.440 --> 0:19:26.439 We are really on day one of this technology and 0:19:26.480 --> 0:19:29.640 how it can be used visa v National security. There 0:19:29.640 --> 0:19:32.000 have been a number of attempts in the last couple 0:19:32.040 --> 0:19:34.240 of years, a lot of them thankfully led by the 0:19:34.320 --> 0:19:39.720 United States to think carefully about AI and nuclear weapons 0:19:40.000 --> 0:19:46.080 use one two. How much you would allow lethal autonomous weapons? 0:19:46.119 --> 0:19:48.439 And this means really, you know, if you have a 0:19:48.480 --> 0:19:52.040 predator drone, there's still a human directing it and deciding 0:19:52.080 --> 0:19:54.280 who to target and who to shoot at. In the 0:19:54.440 --> 0:19:57.840 Ukraine War, You're getting very close now to drone on 0:19:57.920 --> 0:20:02.160 drone dog fights, drones doing their own targeting and shooting, 0:20:02.200 --> 0:20:06.200 and very soon you have a situation that's very dangerous 0:20:06.560 --> 0:20:10.080 where if the AI weapon is always going to be faster, 0:20:11.359 --> 0:20:15.879 the incentive for everyone is to use lethal autonomous weapons 0:20:15.920 --> 0:20:18.920 because they'll always win, and then you'll have huge escalation, 0:20:19.440 --> 0:20:22.520 super dangerous. Lots of different people are trying to get 0:20:22.560 --> 0:20:25.200 at this in different ways. The Korea Conference was one 0:20:25.240 --> 0:20:28.280 of them. I'm not surprised that China hasn't signed on. 0:20:28.680 --> 0:20:31.560 I would just give you one example. You use nuclear weapons. 0:20:32.080 --> 0:20:35.160 You know, when the US came first in the nuclear race, 0:20:35.400 --> 0:20:38.439 we proposed some limited limits on nuclear weapons, I think 0:20:38.480 --> 0:20:42.119 as early as nineteen forty six. Then no real arms 0:20:42.119 --> 0:20:46.480 controlled treaties were negotiated until after you had the Cuban 0:20:46.480 --> 0:20:48.520 missile crisis and we got to the brink, and then 0:20:48.680 --> 0:20:51.040 we had some pretty good ones. But what was going 0:20:51.080 --> 0:20:53.320 on all the time in the background, and what is 0:20:53.359 --> 0:20:56.640 going on now, and it's super important and bears emphasizing, 0:20:57.200 --> 0:20:59.960 is quietly behind the scenes. There is some Chinese sign 0:21:00.000 --> 0:21:04.200 scientists and American and Western scientists talking about these issues. 0:21:04.600 --> 0:21:06.560 There are other track twos that are more on the 0:21:06.600 --> 0:21:10.920 policy area. I'm part of one of them, quiet conversations 0:21:10.920 --> 0:21:13.760 with the Chinese about the dangers of AI and the 0:21:13.840 --> 0:21:18.200 Chinese interlock unders I've talked to are equally worried about 0:21:18.400 --> 0:21:20.800 very similar things than the ones you and I are 0:21:20.840 --> 0:21:21.760 talking about here. 0:21:22.200 --> 0:21:24.920 And for a long time people have realized there's going 0:21:24.920 --> 0:21:26.480 to be an arms race who can get the best 0:21:26.520 --> 0:21:30.119 lethal autonomous weapons fastest. All of the defense departments of 0:21:30.160 --> 0:21:35.280 the people who sell arms, like the US, China, Britain, Israel, Russia, 0:21:35.400 --> 0:21:39.320 all those defense departments are busy working on autonomous lethal weapons. 0:21:40.520 --> 0:21:42.679 They're not going to stop. If one of them stopped, 0:21:42.680 --> 0:21:46.439 the others wouldn't. What we need for that is not 0:21:46.520 --> 0:21:50.200 to stop working on it, but to have a Geneva convention. Now, 0:21:50.480 --> 0:21:53.400 you don't get Geneva conventions things like the Geneva Conventions 0:21:53.400 --> 0:21:56.320 for chemical weapons until after something very nasty has happened. 0:21:56.880 --> 0:21:58.919 So I think realistically, very. 0:21:58.840 --> 0:22:01.160 Nasty things are going to happen with ethan autonomous weapons, 0:22:01.320 --> 0:22:03.840 and then maybe we'll be able to get conventions. So 0:22:03.920 --> 0:22:07.440 for chemical weapons, the conventions have basically worked. Putin isn't 0:22:07.520 --> 0:22:10.240 using them in the Ukraine and they haven't been used much, 0:22:12.119 --> 0:22:15.920 so basically the conventions worked. I'm hoping they would work 0:22:15.920 --> 0:22:19.040 for lethal autonomous weapons, although I'm less confident but nothing's 0:22:19.080 --> 0:22:21.639 going to happen till after some very nasty things have happened. 0:22:22.359 --> 0:22:25.520 Whether AI works for good or for ill, a fundamental 0:22:25.600 --> 0:22:28.440 question is whether humans can maintain control or at least 0:22:28.480 --> 0:22:32.280 influence over it. Professor Hinton says we shouldn't assume that 0:22:32.320 --> 0:22:34.359 we can, at least for long. 0:22:34.840 --> 0:22:36.800 As soon as you make agents which people are busy 0:22:36.800 --> 0:22:39.399 doing now, things that can act in the world. To 0:22:39.440 --> 0:22:41.919 make an effective agent, you have to give it the 0:22:41.960 --> 0:22:44.239 ability to create subgoals. So if you want to get 0:22:44.280 --> 0:22:45.680