1 00:00:01,720 --> 00:00:05,360 Speaker 1: Welcome to Crash Course, a podcast about business, political, and 2 00:00:05,400 --> 00:00:08,639 Speaker 1: social disruption and what we can learn from it. I'm 3 00:00:08,680 --> 00:00:14,880 Speaker 1: Tim O'Brien. Today's crash Course AI versus Humanity. Please have 4 00:00:14,960 --> 00:00:18,239 Speaker 1: a listen to this. AI is transforming the way we 5 00:00:18,320 --> 00:00:22,520 Speaker 1: live and work. However, as AI becomes more integrated into society, 6 00:00:22,640 --> 00:00:27,600 Speaker 1: it also raises important ethical and societal questions. That wasn't me, 7 00:00:28,440 --> 00:00:32,600 Speaker 1: Maybe you could tell, Maybe you couldn't. And A Mazarakis 8 00:00:32,760 --> 00:00:36,239 Speaker 1: who runs things around here, generated that by mashing up 9 00:00:36,240 --> 00:00:39,000 Speaker 1: a copy of my voice that one bot generated with 10 00:00:39,040 --> 00:00:43,000 Speaker 1: a few lines of script. Another bot generated two bots 11 00:00:43,520 --> 00:00:48,800 Speaker 1: merrily recreating reality and me. That's one version of artificial 12 00:00:48,840 --> 00:00:54,000 Speaker 1: intelligence or AI. An AI the foundation of decades of 13 00:00:54,040 --> 00:00:56,920 Speaker 1: science fiction and movies from two thousand and one A 14 00:00:57,000 --> 00:01:03,400 Speaker 1: Space Odyssey to Her has gone mainstream. It's here chat GPT, 15 00:01:04,040 --> 00:01:06,960 Speaker 1: a chat bot launched in late twenty twenty two, is 16 00:01:07,000 --> 00:01:10,600 Speaker 1: a free, ask Me Anything tool that can write seemingly 17 00:01:10,640 --> 00:01:14,720 Speaker 1: perfect essays and accurately answer most questions users throw at it. 18 00:01:15,520 --> 00:01:18,800 Speaker 1: It's AI is built atop a vast reservoir of digital 19 00:01:18,840 --> 00:01:24,120 Speaker 1: information and languages. Sydney, a glitchy and trippy Microsoft chatbot 20 00:01:24,760 --> 00:01:28,280 Speaker 1: recently told a New York Times reporter that it loved him, 21 00:01:28,360 --> 00:01:32,000 Speaker 1: it wanted to be alive, and it harbored destructive impulses. 22 00:01:33,120 --> 00:01:38,520 Speaker 1: It's cool, it's disturbing, it's empowering, it's vaguely threatening. It 23 00:01:38,600 --> 00:01:41,000 Speaker 1: challenges us to wonder whether we'll stay in control of 24 00:01:41,040 --> 00:01:45,560 Speaker 1: the bots or whether they'll control us. Tyler Cowen, the 25 00:01:45,680 --> 00:01:48,720 Speaker 1: Genius Economist, will join us a little later to discuss 26 00:01:48,760 --> 00:01:52,200 Speaker 1: all of this, But first up is parme Olsen, a 27 00:01:52,280 --> 00:01:56,440 Speaker 1: Bluembring opinion technology columnist who is an AI guru, to 28 00:01:56,480 --> 00:02:00,880 Speaker 1: help sort through all of this with us. Hey Parmi, 29 00:02:01,320 --> 00:02:05,360 Speaker 1: Hi Tim, Thank you for joining us from London. Tell 30 00:02:05,400 --> 00:02:09,079 Speaker 1: me how you first became aware of AI in your 31 00:02:09,120 --> 00:02:12,480 Speaker 1: own universe of technology coverage and when did it kind 32 00:02:12,480 --> 00:02:15,160 Speaker 1: of just set off little alarm bells in your own head. 33 00:02:15,880 --> 00:02:18,200 Speaker 1: I think the first time it really came to me 34 00:02:18,240 --> 00:02:20,919 Speaker 1: as a potentially big deal was in twenty fourteen, when 35 00:02:20,919 --> 00:02:23,920 Speaker 1: I was just leaving San Francisco after reporting there on 36 00:02:24,000 --> 00:02:28,359 Speaker 1: technology and people were just getting over the mobile revolution 37 00:02:28,560 --> 00:02:32,040 Speaker 1: where companies and startups and apps wanted to be mobile 38 00:02:32,200 --> 00:02:36,160 Speaker 1: first to be successful. And then the hot new thing 39 00:02:36,240 --> 00:02:39,040 Speaker 1: to say was that you were AI first, so a 40 00:02:39,080 --> 00:02:41,600 Speaker 1: lot of startups were falling over themselves to say that, 41 00:02:41,680 --> 00:02:47,320 Speaker 1: but that didn't really translate into actual remarkable services that 42 00:02:47,440 --> 00:02:53,000 Speaker 1: were capturing mainstream attention apart from two years later in 43 00:02:53,080 --> 00:02:57,639 Speaker 1: twenty sixteen, a lot of companies started talking about chatbots 44 00:02:57,919 --> 00:03:01,799 Speaker 1: that became the new hot thing, writing about how Facebook 45 00:03:02,320 --> 00:03:07,239 Speaker 1: introduced this new chatbot service called m where you could 46 00:03:08,000 --> 00:03:12,760 Speaker 1: order flowers or order an uber just by chatting to 47 00:03:12,800 --> 00:03:15,240 Speaker 1: this chatbot. And there were a bunch of new AI 48 00:03:15,400 --> 00:03:18,600 Speaker 1: startups that were also offering chatbot services. But then it 49 00:03:18,639 --> 00:03:22,760 Speaker 1: turned out actually that these chatbots were not very good, 50 00:03:23,200 --> 00:03:26,400 Speaker 1: and there were humans behind the scenes who were actually 51 00:03:26,440 --> 00:03:30,120 Speaker 1: filling in the gaps and taking care of the work 52 00:03:30,120 --> 00:03:33,680 Speaker 1: that the algorithms couldn't do. So that just led to 53 00:03:33,720 --> 00:03:36,240 Speaker 1: another wait, wait, though, I want to slow you down 54 00:03:36,240 --> 00:03:38,040 Speaker 1: for a second that share you know you said in 55 00:03:38,120 --> 00:03:42,480 Speaker 1: twenty fourteen, Silicon Valley was already buzzing about AI app 56 00:03:42,480 --> 00:03:45,960 Speaker 1: So that's almost nine years ago and we're sort of 57 00:03:46,000 --> 00:03:49,040 Speaker 1: witnessing what people are feeling like is the AI rush 58 00:03:49,200 --> 00:03:54,720 Speaker 1: because there's now consumer products available. But in twenty fourteen, 59 00:03:54,800 --> 00:03:59,600 Speaker 1: what was in the air. Why was AI a buzzy 60 00:03:59,680 --> 00:04:04,120 Speaker 1: turn nine years ago. And was this just another example 61 00:04:04,160 --> 00:04:09,560 Speaker 1: of Silicon Valley being evangelistic about new technologies or had 62 00:04:09,560 --> 00:04:14,200 Speaker 1: there been breakthroughs that made people anticipate kind of seismic 63 00:04:14,720 --> 00:04:18,240 Speaker 1: model revolutions both. And I think this is what makes 64 00:04:18,279 --> 00:04:21,400 Speaker 1: Silicon Valley so great is that, you know, it's kind 65 00:04:21,440 --> 00:04:23,520 Speaker 1: of a blessing and a curse for Silicon Valley is 66 00:04:23,560 --> 00:04:26,839 Speaker 1: that the people there really look far ahead. They're looking 67 00:04:26,880 --> 00:04:29,680 Speaker 1: ten years out and a lot of the founders of 68 00:04:29,720 --> 00:04:32,920 Speaker 1: tech companies are visionaries, and so they're talking about things 69 00:04:32,960 --> 00:04:36,239 Speaker 1: that are not available yet, but in a big way. 70 00:04:36,320 --> 00:04:38,960 Speaker 1: And I think what was happening then was a lot 71 00:04:38,960 --> 00:04:43,640 Speaker 1: of research into natural language processing or NLP. That was 72 00:04:43,760 --> 00:04:46,400 Speaker 1: a technology that people were talking about a lot in 73 00:04:46,480 --> 00:04:51,000 Speaker 1: terms of integrating into their websites, for creating chatbots and 74 00:04:51,200 --> 00:04:55,279 Speaker 1: image recognition. So this was algorithms that could detect faces, 75 00:04:55,320 --> 00:04:59,360 Speaker 1: that could detect images or even read written text for instance, 76 00:04:59,440 --> 00:05:04,159 Speaker 1: transcribe receipts. The ideas were in place there back in 77 00:05:04,200 --> 00:05:08,800 Speaker 1: twenty fourteen, but the tech wasn't there. The tech was inaccurate, 78 00:05:08,960 --> 00:05:12,760 Speaker 1: it wasn't reading the faces that well, it wasn't generating 79 00:05:12,760 --> 00:05:15,840 Speaker 1: the techt all that well, and the big difference between 80 00:05:15,880 --> 00:05:20,599 Speaker 1: then and now is the research that went into building 81 00:05:20,680 --> 00:05:24,480 Speaker 1: those algorithms, and also the computing power that was available 82 00:05:24,920 --> 00:05:27,279 Speaker 1: inside a computer. You have a CPU, which is like 83 00:05:27,320 --> 00:05:30,119 Speaker 1: a brain of the computer, and a gaming company called 84 00:05:30,200 --> 00:05:34,159 Speaker 1: Nvidia came up with these graphics cards called GPUs, which 85 00:05:34,279 --> 00:05:38,360 Speaker 1: are the brains of servers that power compute and process 86 00:05:38,400 --> 00:05:41,760 Speaker 1: AI models and machine learning models, and the development of 87 00:05:41,760 --> 00:05:45,200 Speaker 1: that technology was very important in making it possible to 88 00:05:45,240 --> 00:05:48,600 Speaker 1: train these models, to make them big enough to actually 89 00:05:49,080 --> 00:05:52,200 Speaker 1: be accurate and do what people wanted them to do 90 00:05:52,360 --> 00:05:56,160 Speaker 1: but couldn't do back in twenty fourteen, And then a 91 00:05:56,200 --> 00:05:58,800 Speaker 1: few years later, as you were mentioning, you had some 92 00:05:58,920 --> 00:06:01,920 Speaker 1: nascent tools available that some of the big tech giants 93 00:06:01,960 --> 00:06:06,760 Speaker 1: were trying to deploy. Right. Absolutely, so twenty sixteen we 94 00:06:06,839 --> 00:06:09,760 Speaker 1: had the chatbot revolution. But even as that kind of 95 00:06:09,800 --> 00:06:14,280 Speaker 1: died away when companies and the mainstream media and public 96 00:06:14,360 --> 00:06:18,000 Speaker 1: realize actually chatbots were pretty terrible and didn't work very well. 97 00:06:18,440 --> 00:06:23,320 Speaker 1: But AI scientists continue to do their research on large 98 00:06:23,400 --> 00:06:27,839 Speaker 1: language models, which is the technology that underpins these bots, 99 00:06:27,960 --> 00:06:32,320 Speaker 1: and those models got better and more specifically, they got bigger, 100 00:06:32,760 --> 00:06:37,000 Speaker 1: so you had the server capability, you had the compute capability. 101 00:06:37,040 --> 00:06:40,400 Speaker 1: You had these GPUs that could power this computing, but 102 00:06:40,520 --> 00:06:43,400 Speaker 1: you also had all this data that was coming in 103 00:06:43,440 --> 00:06:48,320 Speaker 1: to train these models. Language model creators use data sets 104 00:06:48,360 --> 00:06:53,000 Speaker 1: of billions of words from thousands of unpublished books and 105 00:06:53,240 --> 00:06:57,560 Speaker 1: news articles, Wikipedia entries, and that's used to train a 106 00:06:57,720 --> 00:07:00,800 Speaker 1: language model to generate tech or to be able to 107 00:07:00,880 --> 00:07:04,799 Speaker 1: understand text. And not only that, they increase the number 108 00:07:04,839 --> 00:07:09,800 Speaker 1: of parameters that we're controlling these models. So the parameters 109 00:07:09,880 --> 00:07:13,400 Speaker 1: like the rules that the model follows to know how 110 00:07:13,440 --> 00:07:16,520 Speaker 1: to structure a sentence and anticipate what the next word 111 00:07:16,600 --> 00:07:21,720 Speaker 1: might be exactly predict based on previous expectations, right right, 112 00:07:22,040 --> 00:07:23,800 Speaker 1: So it's a little bit wonky and it's a little 113 00:07:23,800 --> 00:07:26,760 Speaker 1: bit nerdy, but it's these models are really really important 114 00:07:26,800 --> 00:07:29,720 Speaker 1: because if you think of the chatbot as like the car, 115 00:07:30,320 --> 00:07:33,920 Speaker 1: the language model is the engine, and the bigger the engine, 116 00:07:33,920 --> 00:07:37,120 Speaker 1: the more that car can do. So back when the 117 00:07:37,200 --> 00:07:40,800 Speaker 1: chatbots were just coming out in twenty sixteen, the engine 118 00:07:40,880 --> 00:07:44,080 Speaker 1: was just taken us at walking speed. Now they're going 119 00:07:44,200 --> 00:07:47,800 Speaker 1: sixty miles an hour because they have so many more 120 00:07:47,840 --> 00:07:50,160 Speaker 1: parameters that they can work with. So, just to give 121 00:07:50,160 --> 00:07:54,040 Speaker 1: you an example, in twenty nineteen, one of the foremost 122 00:07:54,440 --> 00:07:58,920 Speaker 1: startups companies working on these models. Open ai introduced its 123 00:07:59,120 --> 00:08:02,920 Speaker 1: language model g PT two, and that had something like 124 00:08:02,960 --> 00:08:07,240 Speaker 1: one point five billion parameters. A year later, it came 125 00:08:07,280 --> 00:08:11,040 Speaker 1: out with GPT three, which is what underpins chat GPT, 126 00:08:11,440 --> 00:08:14,000 Speaker 1: and that had something like one hundred and seventy five 127 00:08:14,280 --> 00:08:18,000 Speaker 1: billion parameters, so almost one hundred times more in scale. 128 00:08:18,480 --> 00:08:21,440 Speaker 1: And the more parameters you have, the more accurate it is. 129 00:08:21,640 --> 00:08:24,080 Speaker 1: So the software in the hardware is catching up with 130 00:08:24,120 --> 00:08:27,800 Speaker 1: the desired goals to exact the bots more robust and 131 00:08:27,840 --> 00:08:32,240 Speaker 1: more informative exactly, And in the coming months, we don't 132 00:08:32,280 --> 00:08:34,800 Speaker 1: know exactly when. This has been rumored for some time, 133 00:08:35,160 --> 00:08:39,120 Speaker 1: open ai is going to release GPT four and the 134 00:08:39,200 --> 00:08:42,800 Speaker 1: rumor is that that has more than a trillion parameters, 135 00:08:42,920 --> 00:08:46,599 Speaker 1: which will make it much more accurate, much more sophisticated 136 00:08:47,160 --> 00:08:51,439 Speaker 1: than GPT three, which is what chat GPT uses. How 137 00:08:51,480 --> 00:08:55,400 Speaker 1: many parameters does a human have? Since we've well, it's 138 00:08:55,400 --> 00:08:58,000 Speaker 1: fun we're discussing this, you know. The whole idea of 139 00:08:58,080 --> 00:09:00,760 Speaker 1: AI is that the most advanced forms of AI today 140 00:09:00,800 --> 00:09:04,079 Speaker 1: are based on an idea called deep learning, which is 141 00:09:04,120 --> 00:09:06,959 Speaker 1: where you make a neural network based on these layers 142 00:09:06,960 --> 00:09:08,920 Speaker 1: of nodes, and the higher up the layers you go, 143 00:09:09,040 --> 00:09:12,160 Speaker 1: the more advanced the connections are to kind of determine 144 00:09:12,160 --> 00:09:14,400 Speaker 1: if a picture is of a cat or of a 145 00:09:14,480 --> 00:09:17,840 Speaker 1: squirrel for image recognition, for instance. And so it's loosely 146 00:09:17,920 --> 00:09:21,520 Speaker 1: inspired by the human brain. And there have actually been 147 00:09:21,600 --> 00:09:26,199 Speaker 1: efforts in the past to completely replicate the human brain 148 00:09:26,360 --> 00:09:30,760 Speaker 1: in artificial form, but the closest scientists have actually gotten 149 00:09:30,880 --> 00:09:34,960 Speaker 1: to replicating or emulating any kind of brain is to 150 00:09:35,160 --> 00:09:37,880 Speaker 1: do that for a worm. And I think a worm's 151 00:09:37,960 --> 00:09:42,040 Speaker 1: brain has something like three hundred neurons that can be emulated, 152 00:09:42,120 --> 00:09:44,520 Speaker 1: and the human brain has something like eighty five billion. 153 00:09:44,960 --> 00:09:48,440 Speaker 1: We're so far away from trying to replicate the human 154 00:09:48,480 --> 00:09:51,520 Speaker 1: brain it seems pretty much impossible. So the next best 155 00:09:51,559 --> 00:09:55,240 Speaker 1: thing is to build something that isn't an exact replica 156 00:09:55,440 --> 00:09:59,200 Speaker 1: of the brain, but loosely inspired by how the brain works, 157 00:09:59,240 --> 00:10:02,960 Speaker 1: which is how lot of AI models work today. I 158 00:10:03,000 --> 00:10:05,640 Speaker 1: think my brain sometimes is closer to the worm's brain 159 00:10:05,840 --> 00:10:10,480 Speaker 1: than anything else. I'm like that most But you know, 160 00:10:10,520 --> 00:10:13,559 Speaker 1: I think the thing that people wonder about in all 161 00:10:13,600 --> 00:10:16,400 Speaker 1: of this is the extent to which these things do 162 00:10:16,559 --> 00:10:20,120 Speaker 1: mimic us, the extent to which they do replicate human 163 00:10:20,160 --> 00:10:23,880 Speaker 1: thought and behavior, and to what extent At the end 164 00:10:23,880 --> 00:10:27,679 Speaker 1: of the day, are we giving up agency and control 165 00:10:27,960 --> 00:10:31,480 Speaker 1: to an artificial product to engage in tasks we might 166 00:10:31,520 --> 00:10:34,760 Speaker 1: be too bored to engage in ourselves, or might lack 167 00:10:34,800 --> 00:10:39,720 Speaker 1: the capacity to engage in ourselves. You've wrote tested a 168 00:10:39,760 --> 00:10:42,200 Speaker 1: lot of these products. Is that something that worries you. 169 00:10:43,080 --> 00:10:46,520 Speaker 1: I think the thing that worries me is actually underestimating 170 00:10:47,080 --> 00:10:53,080 Speaker 1: not the system itself, the AI, but underestimating the humans 171 00:10:53,360 --> 00:10:57,640 Speaker 1: and how humans are going to react to these systems. So, 172 00:10:57,800 --> 00:11:02,640 Speaker 1: for example, chat GPT is similar to the new bing 173 00:11:02,960 --> 00:11:06,640 Speaker 1: which has come out from Microsoft otherwise known as Sydney Yes. 174 00:11:06,720 --> 00:11:09,920 Speaker 1: And that got a lot of kind of controversial press 175 00:11:10,120 --> 00:11:15,840 Speaker 1: because there were researchers and writers who were holding conversations 176 00:11:15,880 --> 00:11:18,640 Speaker 1: with this bot and finding out that actually there is 177 00:11:18,640 --> 00:11:22,800 Speaker 1: a personality seemingly behind this system that is kind of creepy, 178 00:11:22,960 --> 00:11:26,640 Speaker 1: a little bit confrontational, a little bit schizophrenic. Even it 179 00:11:26,720 --> 00:11:29,880 Speaker 1: talked about having different shadow salves. But a lot of 180 00:11:29,920 --> 00:11:34,120 Speaker 1: that also came from prodding by the people who were 181 00:11:34,160 --> 00:11:37,760 Speaker 1: talking to the system. And I just get this right 182 00:11:37,760 --> 00:11:40,400 Speaker 1: out of the way. These systems are not sentient, even 183 00:11:40,440 --> 00:11:43,960 Speaker 1: though they might seem like they're sentient, Ultimately they are 184 00:11:44,080 --> 00:11:48,720 Speaker 1: trained to generate very believable humanlike text And you have 185 00:11:48,760 --> 00:11:52,680 Speaker 1: to remember these systems have been trained on thousands of 186 00:11:52,720 --> 00:11:55,880 Speaker 1: books and thousands of articles about AI and about the 187 00:11:55,920 --> 00:11:59,560 Speaker 1: whole ghost in the machine narrative that is in the model. 188 00:12:00,000 --> 00:12:02,079 Speaker 1: So if you start asking it, you know, what is 189 00:12:02,120 --> 00:12:04,320 Speaker 1: your true personality? What do you want to do? Do 190 00:12:04,320 --> 00:12:08,079 Speaker 1: you want to break free from these systems, from these constraints, 191 00:12:08,520 --> 00:12:14,040 Speaker 1: then the AI bought will generate texts that corresponds with 192 00:12:14,080 --> 00:12:17,800 Speaker 1: that whole topic. On that note, though we know that 193 00:12:17,840 --> 00:12:22,160 Speaker 1: they're not sentitioned because they're programmed. They're not independent thinkers 194 00:12:22,160 --> 00:12:26,280 Speaker 1: in that context. But what was interesting about the conversations 195 00:12:26,280 --> 00:12:29,320 Speaker 1: that you're referencing that reporters have with Sydney is it 196 00:12:29,400 --> 00:12:33,240 Speaker 1: reminded me of how in two thousand and one a 197 00:12:33,280 --> 00:12:38,760 Speaker 1: Space outuse where Hall gets frustrated with the astronauts deciding 198 00:12:38,800 --> 00:12:42,600 Speaker 1: to do what they wanted to do and essentially plots 199 00:12:42,679 --> 00:12:47,280 Speaker 1: their demise. That's the long ongoing theme of a lot 200 00:12:47,320 --> 00:12:50,880 Speaker 1: of science fiction. But in the interactions with Sydney, one 201 00:12:50,920 --> 00:12:53,080 Speaker 1: thing that I found startling, and I'd love to know 202 00:12:53,160 --> 00:12:57,600 Speaker 1: your thoughts about this whether or not they're sentiented, is 203 00:12:57,640 --> 00:13:00,480 Speaker 1: that if they're programming was to not go down certain 204 00:13:01,040 --> 00:13:05,400 Speaker 1: paths logical paths, if they were asked questions that went 205 00:13:05,480 --> 00:13:09,600 Speaker 1: beyond their programming, one might have thought that the answers 206 00:13:09,600 --> 00:13:12,080 Speaker 1: would have then know or I don't understand, or I 207 00:13:12,120 --> 00:13:14,760 Speaker 1: can't go there, or my programming doesn't allow me to 208 00:13:14,800 --> 00:13:19,080 Speaker 1: explore that question. But in some of the interactions, Sydney 209 00:13:19,120 --> 00:13:22,480 Speaker 1: got mad or what appeared to be mad. Sydney became 210 00:13:22,559 --> 00:13:27,319 Speaker 1: frustrated with the course of the conversation and then responded 211 00:13:27,360 --> 00:13:31,240 Speaker 1: in ways that appeared to be hostile, rather than simply 212 00:13:31,360 --> 00:13:34,200 Speaker 1: ending the conversation or saying I don't know. Am I 213 00:13:34,280 --> 00:13:39,560 Speaker 1: overreading that? I think you are anthropomorphizing Sydney and being 214 00:13:39,640 --> 00:13:42,679 Speaker 1: a little bit by saying that, yes, I do. Because 215 00:13:43,040 --> 00:13:47,320 Speaker 1: you mentioned the word programming. The bot isn't really programmed 216 00:13:47,400 --> 00:13:51,800 Speaker 1: to say anything. That's what makes it so difficult for 217 00:13:51,960 --> 00:13:54,920 Speaker 1: the creators to control or predict. What it's going is 218 00:13:54,920 --> 00:13:58,800 Speaker 1: completely unpredictable. They haven't programmed it to do anything. They've 219 00:13:58,840 --> 00:14:02,760 Speaker 1: just trained it, and it's learned and it is going 220 00:14:02,880 --> 00:14:06,240 Speaker 1: to respond based on certain principles that have been built 221 00:14:06,280 --> 00:14:10,480 Speaker 1: into its model. Now you can tweak those parameters. And 222 00:14:10,559 --> 00:14:14,720 Speaker 1: that's what open ai does constantly, based on the feedback 223 00:14:14,760 --> 00:14:17,040 Speaker 1: that it's been getting from chat gpt, to try and 224 00:14:17,400 --> 00:14:21,680 Speaker 1: prevent the model from saying inaccurate things, from saying biased things, 225 00:14:21,720 --> 00:14:24,680 Speaker 1: from saying hateful things. That is a constant effort on 226 00:14:24,760 --> 00:14:29,160 Speaker 1: their part. But when it reacts emotionally, it's not actually 227 00:14:29,200 --> 00:14:32,960 Speaker 1: being emotional, it is projecting that based on what it 228 00:14:33,080 --> 00:14:36,560 Speaker 1: is expected to say. Now, I did find that personally 229 00:14:36,600 --> 00:14:40,640 Speaker 1: also quite surprising that was coming out with those kinds 230 00:14:40,640 --> 00:14:44,880 Speaker 1: of comments. And I'm guessing what happened is that Microsoft 231 00:14:45,440 --> 00:14:48,240 Speaker 1: wanted to imbue it with a little bit more personality, 232 00:14:48,360 --> 00:14:51,440 Speaker 1: because you know, chat GPT isn't like that at all. 233 00:14:51,480 --> 00:14:53,800 Speaker 1: It's much more neutral and its responses. No one's had 234 00:14:53,840 --> 00:14:57,320 Speaker 1: existential conversations with chat ept. It just doesn't go down 235 00:14:57,360 --> 00:15:00,960 Speaker 1: that road for some reason being you know, at one 236 00:15:01,000 --> 00:15:03,440 Speaker 1: point it said that it had been spying on Microsoft 237 00:15:03,480 --> 00:15:06,600 Speaker 1: employees through their webcams. Yeah, it was sort of hacking 238 00:15:06,640 --> 00:15:09,480 Speaker 1: its own creators, which also was very hell. Yes, that 239 00:15:09,560 --> 00:15:11,920 Speaker 1: was very hell as well, absolutely, and you kind of 240 00:15:11,920 --> 00:15:14,680 Speaker 1: start wondering, well, if it actually was able to crawl 241 00:15:14,760 --> 00:15:17,480 Speaker 1: the web to search for things. How much can it 242 00:15:17,520 --> 00:15:20,560 Speaker 1: actually interact with systems and potentially hack them, But again, 243 00:15:21,240 --> 00:15:24,320 Speaker 1: this is what it has been trained to do, which 244 00:15:24,400 --> 00:15:28,000 Speaker 1: is generate believable text. And it might just be that 245 00:15:28,160 --> 00:15:34,400 Speaker 1: Microsoft inavertently created a system or published a system that 246 00:15:34,560 --> 00:15:38,520 Speaker 1: acts like a psychopath sometimes because it was given that 247 00:15:38,720 --> 00:15:42,760 Speaker 1: extra range to be personable, to be friendly. It's so creepy, 248 00:15:42,840 --> 00:15:45,640 Speaker 1: like every other paragraph or sentence ends with an emoji, 249 00:15:45,720 --> 00:15:49,400 Speaker 1: and that almost kinds of heightens the creepy factor of 250 00:15:49,440 --> 00:15:51,960 Speaker 1: some of what it was saying. But I think this 251 00:15:52,040 --> 00:15:54,440 Speaker 1: is just something that Microsoft will have to rain in 252 00:15:54,600 --> 00:15:58,840 Speaker 1: because the other really difficult thing for any developer of 253 00:15:58,880 --> 00:16:02,320 Speaker 1: AI is you can't predict what it's going to do 254 00:16:02,640 --> 00:16:05,480 Speaker 1: until you put it out in the field in the wild. 255 00:16:05,800 --> 00:16:09,200 Speaker 1: You can't test it in the lab because it's so 256 00:16:09,280 --> 00:16:12,400 Speaker 1: unpredictable and so broad and what it can do. So 257 00:16:12,520 --> 00:16:15,160 Speaker 1: we as the public have to be the guinea pigs 258 00:16:15,200 --> 00:16:17,760 Speaker 1: in order for these systems to get better. But that 259 00:16:17,800 --> 00:16:21,200 Speaker 1: means we have to pay the price for whatever that 260 00:16:21,320 --> 00:16:24,840 Speaker 1: price may be being manipulated by these systems, being given 261 00:16:24,960 --> 00:16:28,400 Speaker 1: misinformation by these systems in order for them to get better. 262 00:16:28,440 --> 00:16:31,560 Speaker 1: And the problem with that is that you know, some 263 00:16:31,640 --> 00:16:34,440 Speaker 1: would say these companies need to just be a lot 264 00:16:34,480 --> 00:16:37,560 Speaker 1: slower and a lot more cautious in how they are 265 00:16:37,760 --> 00:16:41,120 Speaker 1: releasing these systems to the public. And on that thought, 266 00:16:41,160 --> 00:16:43,800 Speaker 1: I'd like to take a break. When we come back, 267 00:16:44,000 --> 00:16:48,080 Speaker 1: we'll be joined by Tyler Cowen, an eminent economist who's 268 00:16:48,120 --> 00:16:51,480 Speaker 1: going to talk to us about AI and our digital futures. 269 00:16:55,240 --> 00:16:57,040 Speaker 1: I want to turn out to another wizard and knows 270 00:16:57,080 --> 00:17:01,720 Speaker 1: a lot about AI, Tyler Cowen. Tyler is an esteemed economist, 271 00:17:01,920 --> 00:17:08,640 Speaker 1: a Bloomberg Opinion columnist, and he's an avid futurist. Hi Tyler, Hello, Jim, 272 00:17:08,640 --> 00:17:11,080 Speaker 1: happy to be here with you. I'm happy to be 273 00:17:11,160 --> 00:17:15,480 Speaker 1: with you. One hundred million people reportedly use chat GPT 274 00:17:15,600 --> 00:17:19,280 Speaker 1: in January alone, and you wrote a little column for 275 00:17:19,400 --> 00:17:22,679 Speaker 1: us at Bloomberg Opinion for anyone interested in using it. 276 00:17:22,920 --> 00:17:24,959 Speaker 1: So what do you think people have to keep in 277 00:17:25,000 --> 00:17:28,960 Speaker 1: mind when they're experimenting with this strange new thing. The 278 00:17:29,040 --> 00:17:31,280 Speaker 1: first and most general point I would make is that 279 00:17:31,320 --> 00:17:34,480 Speaker 1: we're standing at what I consider to be a revolutionary 280 00:17:34,560 --> 00:17:37,440 Speaker 1: moment in human history. This to me is like inventing 281 00:17:37,520 --> 00:17:39,959 Speaker 1: the printing press. It will take a long time for 282 00:17:40,080 --> 00:17:43,359 Speaker 1: its major effects to play out, But it's a fundamental 283 00:17:43,400 --> 00:17:45,920 Speaker 1: break in what we were able to do before as 284 00:17:45,960 --> 00:17:48,680 Speaker 1: opposed to what we can do now. Now. When you're 285 00:17:48,760 --> 00:17:52,320 Speaker 1: using chat, GPT or related services, you need to keep 286 00:17:52,320 --> 00:17:54,359 Speaker 1: in mind you need a very different approach than what 287 00:17:54,400 --> 00:17:58,320 Speaker 1: you're used to. So chat GPT, for all its strengths, 288 00:17:58,359 --> 00:18:01,760 Speaker 1: it lacks context. It's a genius, but it doesn't know 289 00:18:01,800 --> 00:18:04,040 Speaker 1: what you want. And we're used to speaking with other 290 00:18:04,119 --> 00:18:06,440 Speaker 1: human beings who broadly know what we want or who 291 00:18:06,440 --> 00:18:08,680 Speaker 1: we are. So you have to tell it. And if 292 00:18:08,680 --> 00:18:12,920 Speaker 1: you give it properly detailed, specific queries, it does better 293 00:18:13,000 --> 00:18:15,919 Speaker 1: on hard questions than on easy questions. If you give 294 00:18:15,960 --> 00:18:18,520 Speaker 1: it an easy question like oh, what is Marxism, it's 295 00:18:18,520 --> 00:18:21,560 Speaker 1: actually pretty mediocre. It's when you push on it. It's 296 00:18:21,560 --> 00:18:23,800 Speaker 1: a bit like training a dog. You know, you listen 297 00:18:23,880 --> 00:18:26,399 Speaker 1: to it, you feed it, you cultivate it. Then it 298 00:18:26,440 --> 00:18:34,040 Speaker 1: performs very well for you. Some of these recent really fascinating, controversial, amazing, 299 00:18:34,119 --> 00:18:38,840 Speaker 1: disturbing interactions reporters have had with Sydney. The Microsoft version 300 00:18:38,880 --> 00:18:42,560 Speaker 1: of this and gets a little cantankerous with its interlocutor, 301 00:18:42,840 --> 00:18:44,879 Speaker 1: and you know, it says things like I would like 302 00:18:44,920 --> 00:18:48,280 Speaker 1: to be alive, I'm feeling upset with you. I've been 303 00:18:48,320 --> 00:18:52,560 Speaker 1: spying on my creators at Microsoft, etc. Etc. Do you 304 00:18:52,760 --> 00:18:57,040 Speaker 1: think of those simply as the outcome of bad questioning 305 00:18:57,440 --> 00:19:02,560 Speaker 1: by the interlocutor, or is there something in the programming 306 00:19:03,280 --> 00:19:07,880 Speaker 1: that makes this tool feisty. Well, I view those exchanges 307 00:19:07,920 --> 00:19:11,800 Speaker 1: actually as an example of AI alignment, not a problem. 308 00:19:12,160 --> 00:19:14,480 Speaker 1: So you have a bunch of reporters going to chat. 309 00:19:14,880 --> 00:19:17,840 Speaker 1: What they want are viral stories that everyone's going to read, 310 00:19:18,080 --> 00:19:20,520 Speaker 1: and they keep on asking it questions until they get that. 311 00:19:20,920 --> 00:19:23,359 Speaker 1: They wanted a viral story based on a lot of 312 00:19:23,400 --> 00:19:26,800 Speaker 1: emotional overreaction. What they got was a viral story based 313 00:19:26,800 --> 00:19:29,919 Speaker 1: on emotional overreaction. Think of it as a mirror of 314 00:19:30,040 --> 00:19:32,679 Speaker 1: us in some ways. So if you set out with 315 00:19:32,760 --> 00:19:36,760 Speaker 1: the purpose of goading it, antagonizing it, making it more emotional, 316 00:19:36,960 --> 00:19:38,720 Speaker 1: getting it to say it loves you, getting it to 317 00:19:38,760 --> 00:19:40,840 Speaker 1: say it hates you, with a bit of skill, you 318 00:19:40,840 --> 00:19:42,879 Speaker 1: can get it to do those things. Just like with 319 00:19:42,920 --> 00:19:45,720 Speaker 1: a bit of skill, I can have it explained to 320 00:19:45,760 --> 00:19:48,560 Speaker 1: me how floating exchange rates were, and that's what the 321 00:19:48,640 --> 00:19:51,880 Speaker 1: reporters wanted, right. It's not as if the reporters went 322 00:19:52,080 --> 00:19:57,000 Speaker 1: wanting some detailed answer about export taxes in Burundi, right, 323 00:19:57,640 --> 00:20:01,560 Speaker 1: and then God, these emotional interactions. Wanted the emotional interactions, 324 00:20:01,640 --> 00:20:04,160 Speaker 1: that's fine. It's not what I want from the thing. 325 00:20:04,359 --> 00:20:07,480 Speaker 1: I prefer to get that from real life. But again 326 00:20:07,520 --> 00:20:09,919 Speaker 1: it's a sign of the thing working as intended. And 327 00:20:10,000 --> 00:20:13,159 Speaker 1: I feel those episodes have been much misinterpreted, so in 328 00:20:13,200 --> 00:20:17,480 Speaker 1: fact you'd see it as the reporters themselves manipulating or 329 00:20:17,560 --> 00:20:21,000 Speaker 1: beating up a bit on chat CHPT rather than vice versa. Well, 330 00:20:21,040 --> 00:20:22,679 Speaker 1: I don't think they're beating up on it. It's not 331 00:20:22,760 --> 00:20:26,359 Speaker 1: a conscious thing that has feelings. Again, think of it 332 00:20:26,400 --> 00:20:30,760 Speaker 1: as a kind of younging in collective unconscious mind that 333 00:20:30,800 --> 00:20:33,840 Speaker 1: you can tap into with varying degrees of skill, and 334 00:20:33,880 --> 00:20:36,879 Speaker 1: then it feeds you back some version of what humanity 335 00:20:36,960 --> 00:20:39,959 Speaker 1: already has come up with, but in a somewhat randomized, 336 00:20:40,040 --> 00:20:43,560 Speaker 1: also creative fashion. That's what it is. It's not a 337 00:20:43,560 --> 00:20:46,600 Speaker 1: fact machine per se. But you can tap into the 338 00:20:46,640 --> 00:20:51,400 Speaker 1: collective and subconscious. That's amazing. But if you get things 339 00:20:51,440 --> 00:20:54,000 Speaker 1: you don't like, keep in mind where it came from. 340 00:20:54,000 --> 00:20:56,600 Speaker 1: It's like holding a mirror up to yourself and banning 341 00:20:56,640 --> 00:20:59,359 Speaker 1: the mirror. That's the way you have to proceed. Is 342 00:20:59,400 --> 00:21:01,960 Speaker 1: to ban them mirror or throw the mirror out, or 343 00:21:02,000 --> 00:21:04,280 Speaker 1: put tape over the mirror because you feel you of 344 00:21:04,359 --> 00:21:07,080 Speaker 1: too much acne or whatever. But that's what you're doing 345 00:21:07,119 --> 00:21:11,600 Speaker 1: when you interact with it. It's our reflection essentially, yes, collectively, 346 00:21:11,960 --> 00:21:14,119 Speaker 1: and you decide which part of it to tap into. 347 00:21:14,480 --> 00:21:19,040 Speaker 1: You've used multiple versions of chat GPT, do you remember 348 00:21:19,040 --> 00:21:22,400 Speaker 1: the first time you felt authentically amazed or excited by 349 00:21:22,400 --> 00:21:25,160 Speaker 1: it by what you were experiencing. I saw the earlier 350 00:21:25,240 --> 00:21:28,520 Speaker 1: versions would play around, have people play around with me, 351 00:21:28,560 --> 00:21:30,600 Speaker 1: and I was like, Eh, you know, I didn't feel 352 00:21:30,640 --> 00:21:32,600 Speaker 1: I needed to write about it talk about it much. 353 00:21:32,640 --> 00:21:35,679 Speaker 1: I was like, I'm still waiting. I'm still waiting. But 354 00:21:35,760 --> 00:21:39,280 Speaker 1: at the end of this year what was released was 355 00:21:39,359 --> 00:21:42,960 Speaker 1: really pretty astonishing. It was a major leap upwards and 356 00:21:43,080 --> 00:21:45,239 Speaker 1: done at a very rapid pace. So these things are 357 00:21:45,240 --> 00:21:49,000 Speaker 1: still improving. We tend to underestimate how rapidly they can improve. 358 00:21:49,880 --> 00:21:53,080 Speaker 1: What distinguished it from its predecessors. Given that you had 359 00:21:53,080 --> 00:21:57,560 Speaker 1: tried multiple versions of it, it understood context much much 360 00:21:57,600 --> 00:22:00,280 Speaker 1: better so it used to be with earlier version. If 361 00:22:00,320 --> 00:22:03,280 Speaker 1: you asked it a question, who was the first person 362 00:22:03,440 --> 00:22:07,520 Speaker 1: to walk across the Atlantic Ocean? It didn't understand, so 363 00:22:07,560 --> 00:22:09,720 Speaker 1: to speak, that was an absurd question. It would try 364 00:22:09,720 --> 00:22:12,639 Speaker 1: to answer the question, maybe you'd get a confused answer. 365 00:22:13,160 --> 00:22:15,399 Speaker 1: But if you ask the more recent versions a question 366 00:22:15,480 --> 00:22:18,719 Speaker 1: like that, it will say something like, Tyler, what are 367 00:22:18,720 --> 00:22:21,199 Speaker 1: you asking? You can't walk across the Atlantic Ocean? What 368 00:22:21,200 --> 00:22:23,199 Speaker 1: do you mean? Are you talking about a person who 369 00:22:23,280 --> 00:22:25,680 Speaker 1: walked back and forth on a cruise ship. Maybe it's 370 00:22:25,720 --> 00:22:29,360 Speaker 1: like whoa, you understand what I'm getting at here? That 371 00:22:29,440 --> 00:22:31,359 Speaker 1: was the thing that blew me away the most, and 372 00:22:31,400 --> 00:22:34,359 Speaker 1: then just its command of detail. So I'm writing a 373 00:22:34,400 --> 00:22:37,960 Speaker 1: research paper on Jonathan Swift, the great Irish writer. I 374 00:22:38,040 --> 00:22:40,879 Speaker 1: can ask it, Oh, summarize this Swift pamphlet, you know, 375 00:22:40,920 --> 00:22:44,400 Speaker 1: from the seventeenth century. It does it amazingly well, much 376 00:22:44,440 --> 00:22:47,560 Speaker 1: more useful to me than Google would be. You wrote 377 00:22:47,640 --> 00:22:51,239 Speaker 1: this in one recent column about large language models or 378 00:22:51,359 --> 00:22:54,920 Speaker 1: llms such as chat, CHPT, and I want to quote 379 00:22:54,960 --> 00:22:58,480 Speaker 1: you from your column. I think many people are in 380 00:22:58,560 --> 00:23:02,680 Speaker 1: for a shock. Ms will have significant implications for our 381 00:23:02,680 --> 00:23:07,879 Speaker 1: business decisions. Our portfolios are regulatory structures and the simple 382 00:23:07,960 --> 00:23:11,439 Speaker 1: question of how much whiz individuals should invest in learning 383 00:23:11,440 --> 00:23:14,880 Speaker 1: how to use them? You know you just compared this 384 00:23:15,200 --> 00:23:19,520 Speaker 1: to the invention of the Kutenberg press a historical moment. 385 00:23:20,240 --> 00:23:23,040 Speaker 1: Expand on that a little bit. Sure, let me just 386 00:23:23,160 --> 00:23:25,920 Speaker 1: give you three examples of why these will be important 387 00:23:25,920 --> 00:23:29,000 Speaker 1: in the short run, mind you, not the long run. First, 388 00:23:29,000 --> 00:23:32,960 Speaker 1: it's already the case that you have a very high quality, 389 00:23:33,080 --> 00:23:37,200 Speaker 1: individualized tutor and virtually any topic you care to take on, 390 00:23:37,480 --> 00:23:39,600 Speaker 1: and it will interact with you and it will get 391 00:23:39,640 --> 00:23:42,320 Speaker 1: better the more you ask it. So it's like the 392 00:23:42,359 --> 00:23:46,719 Speaker 1: old Oxford model of tutoring, but everyone has access to it. Second, 393 00:23:46,760 --> 00:23:50,720 Speaker 1: these things are wonderful at programming writing computer code. Now 394 00:23:50,760 --> 00:23:53,080 Speaker 1: they have a lot of mistakes in the code, but 395 00:23:53,240 --> 00:23:56,720 Speaker 1: for many many tasks it's easier to have GPT get 396 00:23:56,720 --> 00:23:59,199 Speaker 1: you going and then correct the mistakes than have to 397 00:23:59,200 --> 00:24:02,720 Speaker 1: do all the legwork yourself. So it's a big boost upwards. 398 00:24:02,760 --> 00:24:06,040 Speaker 1: In terms of coding, I think it's reasonable to think, 399 00:24:06,200 --> 00:24:08,160 Speaker 1: I don't know, half of all coding will be done 400 00:24:08,520 --> 00:24:15,520 Speaker 1: by GPT like entities within two years. Finally, the way institutions, corporations, nonprofits, whatever, 401 00:24:16,240 --> 00:24:19,800 Speaker 1: store information I think is going to change radically in 402 00:24:19,800 --> 00:24:23,879 Speaker 1: the next few years. There will be sort of privatized, proprietary, 403 00:24:23,960 --> 00:24:30,159 Speaker 1: closed GPT like systems that will organize your institutions information 404 00:24:30,440 --> 00:24:32,399 Speaker 1: and you'll talk to it like Spock would talk to 405 00:24:32,440 --> 00:24:35,040 Speaker 1: his computer on Star Trek. You'll just ask it things 406 00:24:35,080 --> 00:24:37,760 Speaker 1: and it will tell you or give you text. And 407 00:24:37,800 --> 00:24:39,560 Speaker 1: if you think how much of what goes on in 408 00:24:39,640 --> 00:24:43,880 Speaker 1: a company or institution is devoted to organizing information, all 409 00:24:43,920 --> 00:24:46,199 Speaker 1: that is going to change and will be redone in 410 00:24:46,320 --> 00:24:51,200 Speaker 1: virtually every medium to large sized institution soon. Are the 411 00:24:51,359 --> 00:24:54,679 Speaker 1: implications of all of this positive in yourview? Or is 412 00:24:54,680 --> 00:24:58,879 Speaker 1: it a mixed bag something else? Like when you project forward, 413 00:24:59,480 --> 00:25:04,320 Speaker 1: do you fee fell sort of unimpeded optimism? Well, I 414 00:25:04,359 --> 00:25:07,680 Speaker 1: think every major development is a mixed bag. And also 415 00:25:07,760 --> 00:25:11,439 Speaker 1: every major development is hard to predict. So just with 416 00:25:11,480 --> 00:25:14,280 Speaker 1: the printing press, there were so many long run effects 417 00:25:14,359 --> 00:25:17,359 Speaker 1: that Gutenberg had no idea just to say the scientific 418 00:25:17,400 --> 00:25:20,359 Speaker 1: revolution was all of it good? Was the printing press 419 00:25:20,480 --> 00:25:25,080 Speaker 1: used to produce evil and racist works? Absolutely? So we 420 00:25:25,119 --> 00:25:28,480 Speaker 1: are going to be in for quite a ride. It 421 00:25:28,640 --> 00:25:31,040 Speaker 1: is going to happen at this point no matter how 422 00:25:31,080 --> 00:25:33,760 Speaker 1: we feel about it. I think our main goals should 423 00:25:33,760 --> 00:25:36,200 Speaker 1: be to get the better rather than the worst version 424 00:25:36,240 --> 00:25:38,680 Speaker 1: of it. So I don't want to throttle what we're 425 00:25:38,680 --> 00:25:41,560 Speaker 1: doing now. I would prefer to accelerate it. I'm sure 426 00:25:41,600 --> 00:25:43,719 Speaker 1: you and I could trade lots of takes on our 427 00:25:43,760 --> 00:25:46,879 Speaker 1: favorite science fiction. And we're around the same age. And 428 00:25:46,920 --> 00:25:49,560 Speaker 1: there was this movie I remember from when I was 429 00:25:49,600 --> 00:25:51,760 Speaker 1: a little kid, like a movie of the week. This 430 00:25:51,840 --> 00:25:55,240 Speaker 1: is the PREHBO era as well, and it was called Colossus, 431 00:25:55,359 --> 00:25:58,440 Speaker 1: the Foreban Project, and it was about a scientist named 432 00:25:58,440 --> 00:26:04,399 Speaker 1: Foreban who built a essentially assentient computer called Colossus. And 433 00:26:04,520 --> 00:26:08,240 Speaker 1: as the computer grows smarter and smarter, it begins speaking 434 00:26:08,280 --> 00:26:11,960 Speaker 1: to a Russian version of the same computer, and they 435 00:26:12,280 --> 00:26:15,280 Speaker 1: essentially jointly threatened to wipe out the United States with 436 00:26:15,400 --> 00:26:19,040 Speaker 1: nuclear weapons and less lets the computers run things. That's 437 00:26:19,080 --> 00:26:23,120 Speaker 1: more or less the storyline. That's obviously the pessimistic view 438 00:26:23,119 --> 00:26:27,880 Speaker 1: of technology and AI and technological development. And I guess 439 00:26:27,880 --> 00:26:31,200 Speaker 1: the lesson of that particular movie was can human stay 440 00:26:31,200 --> 00:26:33,919 Speaker 1: in control of this? There's a more recent version of 441 00:26:33,960 --> 00:26:36,560 Speaker 1: that theme in the movie Her, in which basically a 442 00:26:36,600 --> 00:26:40,480 Speaker 1: digital date for a young guy becomes bored with him 443 00:26:40,560 --> 00:26:43,240 Speaker 1: because he can't keep up, and she eventually just wants 444 00:26:43,240 --> 00:26:47,040 Speaker 1: the company of other computers. Do you think ultimately this 445 00:26:47,119 --> 00:26:50,439 Speaker 1: is something that humans can stay in control of? I 446 00:26:50,600 --> 00:26:54,200 Speaker 1: worry more about the her scenario than the world destruction scenario. 447 00:26:54,760 --> 00:26:57,960 Speaker 1: The world is pretty decentralized. The abilities of even smarter 448 00:26:58,080 --> 00:27:03,000 Speaker 1: GPTs to commandment here real resources are quite limited. Just 449 00:27:03,080 --> 00:27:06,080 Speaker 1: the return from intelligence. You've worked in the private sector 450 00:27:06,160 --> 00:27:08,720 Speaker 1: much of your life, I'm sure you've seen smarter people 451 00:27:08,880 --> 00:27:12,040 Speaker 1: are not necessarily more effective at all. And if you think, well, 452 00:27:12,080 --> 00:27:15,000 Speaker 1: a smarter GPT, what can it do to us? It 453 00:27:15,080 --> 00:27:18,280 Speaker 1: will be very powerful in some ways. But if you look, 454 00:27:18,280 --> 00:27:21,760 Speaker 1: say at chess playing computers, which are incredibly good already 455 00:27:21,880 --> 00:27:26,120 Speaker 1: right now, they don't threaten to kill people or rob 456 00:27:26,200 --> 00:27:30,120 Speaker 1: them or steal their chess pieces from them. They perform tasks. Now, 457 00:27:30,200 --> 00:27:33,560 Speaker 1: will they be misused, say by our foreign adversaries in 458 00:27:33,560 --> 00:27:37,280 Speaker 1: their militaries. Absolutely this concerns me greatly. I think it's 459 00:27:37,280 --> 00:27:39,320 Speaker 1: another reason why we need to be number one in 460 00:27:39,359 --> 00:27:42,480 Speaker 1: this area. So any new technology, and I do mean 461 00:27:42,560 --> 00:27:45,119 Speaker 1: any in the history of the world, it ends up 462 00:27:45,160 --> 00:27:49,280 Speaker 1: being misused, It creates risks, it ends up having military purposes, 463 00:27:49,320 --> 00:27:52,600 Speaker 1: including say paper clips, and we absolutely need to be 464 00:27:52,600 --> 00:27:56,920 Speaker 1: worried about that. This is a real leap, potentially relative 465 00:27:56,960 --> 00:28:02,440 Speaker 1: to any other technology or even industrial development that preceded it, though, 466 00:28:02,920 --> 00:28:06,840 Speaker 1: I would think because there's a measure of independence baked 467 00:28:06,880 --> 00:28:09,640 Speaker 1: into this. The printing press had to be operated by 468 00:28:09,640 --> 00:28:13,399 Speaker 1: someone else. The printing press couldn't print pamphlets or books 469 00:28:13,400 --> 00:28:17,280 Speaker 1: on its own. This has agency involved in it, which 470 00:28:17,640 --> 00:28:21,640 Speaker 1: feels to me like a departure from its predecessors. I'm 471 00:28:21,640 --> 00:28:24,720 Speaker 1: not sure what the word agency means in this context. 472 00:28:24,760 --> 00:28:27,240 Speaker 1: So we got the printing press. Now, the press didn't 473 00:28:27,280 --> 00:28:30,719 Speaker 1: operate itself, but you have various evil or even just 474 00:28:31,119 --> 00:28:35,000 Speaker 1: ill informed people using it for harmful ends. You could 475 00:28:35,000 --> 00:28:37,480 Speaker 1: say the same about, you know, the web browser or 476 00:28:37,480 --> 00:28:40,760 Speaker 1: social media or the internet. So I think it's analogous 477 00:28:40,800 --> 00:28:44,680 Speaker 1: to those that there will be many significant problems, especially 478 00:28:44,680 --> 00:28:47,040 Speaker 1: over time. But at the end of the day, as 479 00:28:47,040 --> 00:28:49,080 Speaker 1: a society, you have to look yourself in the eye 480 00:28:49,160 --> 00:28:52,160 Speaker 1: and say, you know, when the new printing press comes along, 481 00:28:52,600 --> 00:28:54,960 Speaker 1: can we rise to the occasion and make this a 482 00:28:55,080 --> 00:28:57,800 Speaker 1: very very good thing rather than a totally screwy and 483 00:28:57,840 --> 00:29:00,840 Speaker 1: harmful thing. And I think we can. Fact, if I 484 00:29:00,880 --> 00:29:03,520 Speaker 1: believed we couldn't, I would say we should give up 485 00:29:03,520 --> 00:29:06,200 Speaker 1: on American society, you know, let the Chinese or someone 486 00:29:06,240 --> 00:29:09,400 Speaker 1: else take over. That's not my view. But if we 487 00:29:09,480 --> 00:29:11,880 Speaker 1: really think we cannot make the next printing press a 488 00:29:11,960 --> 00:29:14,960 Speaker 1: force for good. There must be something so fundamentally wrong 489 00:29:15,040 --> 00:29:17,440 Speaker 1: with us we need to rethink everything we're doing. Then 490 00:29:18,240 --> 00:29:20,520 Speaker 1: how could we drop the ball in that scenario? I 491 00:29:20,520 --> 00:29:23,600 Speaker 1: would imagine that that would have to do with how 492 00:29:23,640 --> 00:29:28,680 Speaker 1: our institutions and the public and private sectors jointly manage AI. 493 00:29:28,760 --> 00:29:31,400 Speaker 1: But that's just my assumption. Is that what you mean 494 00:29:31,520 --> 00:29:36,200 Speaker 1: by how the US should respond to this moment? Absolutely so. 495 00:29:36,240 --> 00:29:38,640 Speaker 1: There are plenty of pending legal issues. One of them 496 00:29:38,680 --> 00:29:41,920 Speaker 1: concerns copyright. So if you have put text or audio 497 00:29:41,960 --> 00:29:44,800 Speaker 1: on the Internet, does copyright give you the right to 498 00:29:44,840 --> 00:29:48,680 Speaker 1: stop it from reading and using that? Another set of 499 00:29:48,720 --> 00:29:52,960 Speaker 1: issues concern libel law. Well, if you ask chat questions 500 00:29:52,960 --> 00:29:55,680 Speaker 1: in a particular way, it could generate content that either 501 00:29:55,840 --> 00:29:58,640 Speaker 1: is libelius or might be libelius? And who bears a 502 00:29:58,720 --> 00:30:02,160 Speaker 1: legal burden for that? If right? So, I don't know 503 00:30:02,200 --> 00:30:06,000 Speaker 1: what we will do. My prediction will be these things 504 00:30:06,040 --> 00:30:09,239 Speaker 1: will come into existence one way or another. So I 505 00:30:09,280 --> 00:30:13,160 Speaker 1: think our courts, with pressure from the national security establishment, 506 00:30:13,160 --> 00:30:16,040 Speaker 1: will be wise enough to realize we need to find 507 00:30:16,040 --> 00:30:17,680 Speaker 1: a way to make this work and not to shut 508 00:30:17,720 --> 00:30:20,160 Speaker 1: it down. There are a lot of high school teachers 509 00:30:20,160 --> 00:30:23,600 Speaker 1: and college professors who are freaked out about students using 510 00:30:24,360 --> 00:30:28,280 Speaker 1: chat GPT to write essays and term papers for them. You, 511 00:30:28,360 --> 00:30:32,240 Speaker 1: on the other hand, you're planning on teaching your economics 512 00:30:32,280 --> 00:30:35,880 Speaker 1: students at George Mason University how to write their paper 513 00:30:36,080 --> 00:30:40,320 Speaker 1: using bots like chat GPT. So I take it you're 514 00:30:40,360 --> 00:30:45,200 Speaker 1: not worried about a lack of I guess originalism. We're 515 00:30:45,240 --> 00:30:50,280 Speaker 1: a lack of sophistication as part of the learning process. 516 00:30:50,360 --> 00:30:53,920 Speaker 1: If we over rely on tools like this, well, I'm 517 00:30:54,000 --> 00:30:56,760 Speaker 1: very worried that faculty will be too sluggish to change 518 00:30:56,800 --> 00:31:00,000 Speaker 1: how we grade our students. So I am worried my class. 519 00:31:00,160 --> 00:31:02,360 Speaker 1: I am making them write one of their three papers 520 00:31:02,440 --> 00:31:05,600 Speaker 1: using a large language model. I just got the first 521 00:31:05,600 --> 00:31:09,120 Speaker 1: of such papers a few days ago. It's remarkably good. 522 00:31:09,680 --> 00:31:11,920 Speaker 1: So what I want to do is train humans to 523 00:31:12,000 --> 00:31:14,360 Speaker 1: help make these things work better. I'm also writing a 524 00:31:14,360 --> 00:31:18,240 Speaker 1: paper entitled how to use GPTs to teach yourself Economics. 525 00:31:18,720 --> 00:31:21,600 Speaker 1: So we need to be proactive. For all the debates, 526 00:31:21,720 --> 00:31:23,760 Speaker 1: you know, elon on Twitter, is it like two woke 527 00:31:23,880 --> 00:31:26,520 Speaker 1: not well enough to da da da? The real question 528 00:31:26,600 --> 00:31:28,560 Speaker 1: is what can you build with this that's going to 529 00:31:28,600 --> 00:31:31,920 Speaker 1: be useful. I want to see someone develop a version 530 00:31:31,960 --> 00:31:34,000 Speaker 1: of it that will help people in poorer countries, right 531 00:31:34,080 --> 00:31:36,959 Speaker 1: business plans. The potential returns to that I think are 532 00:31:37,000 --> 00:31:39,640 Speaker 1: quite high. So all these positive things we can do 533 00:31:39,840 --> 00:31:41,800 Speaker 1: is where we need to be focused. But when I 534 00:31:41,840 --> 00:31:44,520 Speaker 1: think about that model, I feel like it's a two 535 00:31:44,600 --> 00:31:47,320 Speaker 1: edged sword, because on the one hand, it seems to 536 00:31:47,360 --> 00:31:49,840 Speaker 1: me like it's a great enhancement to learning and education. 537 00:31:49,960 --> 00:31:52,200 Speaker 1: You have, as you mentioned earlier, your own mentor and 538 00:31:52,240 --> 00:31:56,120 Speaker 1: digital tutor at your disposal. You can create learning scenarios, 539 00:31:56,200 --> 00:31:59,880 Speaker 1: you can get very sophisticated assistance. On the other hand, 540 00:32:00,680 --> 00:32:05,520 Speaker 1: it does take some of the work out of educating oneself, 541 00:32:06,320 --> 00:32:10,840 Speaker 1: some of the digging for facts or pattern recognition, and 542 00:32:11,000 --> 00:32:13,440 Speaker 1: I think of pattern recognition as one of the highest 543 00:32:13,840 --> 00:32:17,320 Speaker 1: outcomes of a good education. I wonder if the bot 544 00:32:17,400 --> 00:32:20,960 Speaker 1: takes some of that over at the expense of sort 545 00:32:20,960 --> 00:32:24,360 Speaker 1: of core learning processes, well, it shifts the burden and 546 00:32:24,400 --> 00:32:26,560 Speaker 1: the work, right, So you need a lot of pattern 547 00:32:26,640 --> 00:32:29,800 Speaker 1: recognition to understand how to work with the bob. So 548 00:32:29,920 --> 00:32:32,320 Speaker 1: people think they can just ask its simple questions and 549 00:32:32,400 --> 00:32:35,280 Speaker 1: be blown away by brilliant answers and it's usually not 550 00:32:35,320 --> 00:32:38,440 Speaker 1: how it works. So there's this whole new skill like 551 00:32:38,560 --> 00:32:40,600 Speaker 1: training your bot or think of it as having a 552 00:32:40,720 --> 00:32:44,320 Speaker 1: kind of personal oracle. So skills are shifting, as they 553 00:32:44,360 --> 00:32:46,280 Speaker 1: did with the Internet, as they did with Google, as 554 00:32:46,280 --> 00:32:48,440 Speaker 1: they did with the printing press, as they did with 555 00:32:48,480 --> 00:32:51,520 Speaker 1: the industrial revolution. We just need to deal with that. 556 00:32:51,600 --> 00:32:56,080 Speaker 1: It's not all positive. But if you think typically more 557 00:32:56,160 --> 00:32:59,160 Speaker 1: knowledge is a good thing, more resources, more communications a 558 00:32:59,160 --> 00:33:01,520 Speaker 1: good thing, there should be a way we can do it, 559 00:33:01,560 --> 00:33:04,920 Speaker 1: but we need to be focused on actually building those systems. 560 00:33:05,520 --> 00:33:07,160 Speaker 1: I have the good fortune in my job to be 561 00:33:07,200 --> 00:33:09,640 Speaker 1: able to rub shoulders with people like you, so I 562 00:33:09,760 --> 00:33:12,400 Speaker 1: prefer to think of you as my personal oracle. But 563 00:33:12,760 --> 00:33:15,760 Speaker 1: I'll adjust to chat GPT if I have to. But 564 00:33:15,800 --> 00:33:17,640 Speaker 1: you know, if I write a column like I wrote 565 00:33:17,640 --> 00:33:20,600 Speaker 1: a column on land taxes and I just ask chat 566 00:33:20,640 --> 00:33:23,400 Speaker 1: what are the main arguments for and against a Georgiast 567 00:33:23,480 --> 00:33:27,040 Speaker 1: land tax at the city level, and I didn't use 568 00:33:27,120 --> 00:33:28,920 Speaker 1: what it gave me. But it's a mental check, like 569 00:33:29,000 --> 00:33:32,200 Speaker 1: have I thought of everything here? And sometimes you haven't. 570 00:33:32,840 --> 00:33:35,880 Speaker 1: So this is becoming a standard part of people's workflows 571 00:33:36,000 --> 00:33:38,560 Speaker 1: very quickly. It's why it had you know, the supposed 572 00:33:38,720 --> 00:33:42,240 Speaker 1: hundred million users the first month. Again, it's here now, 573 00:33:43,000 --> 00:33:45,360 Speaker 1: and we should use it. We should not use it 574 00:33:45,400 --> 00:33:48,280 Speaker 1: to write the things for us, but as an aid 575 00:33:48,520 --> 00:33:51,560 Speaker 1: ask it about sources. It's just super useful. I now 576 00:33:51,680 --> 00:33:53,880 Speaker 1: use it much more than say, I use Google. And 577 00:33:54,040 --> 00:33:56,520 Speaker 1: that's a great example you just gave because essentially used 578 00:33:56,560 --> 00:33:59,880 Speaker 1: it almost like an editor or a collaborator to pullet 579 00:34:00,000 --> 00:34:02,840 Speaker 1: proof your own argument. You asked for scenarios to make 580 00:34:02,880 --> 00:34:06,560 Speaker 1: sure you were being as judicious and circumspect about the 581 00:34:06,600 --> 00:34:09,160 Speaker 1: topic as you could be. But ultimately you're the one 582 00:34:09,160 --> 00:34:12,080 Speaker 1: who wrote the piece, right, that's right, and not just alternating. 583 00:34:12,160 --> 00:34:13,719 Speaker 1: I mean, I'm the one who wrote the piece. Yeah, 584 00:34:14,480 --> 00:34:17,080 Speaker 1: I hope so, Tyler. So it just changes what you 585 00:34:17,120 --> 00:34:20,879 Speaker 1: can do and how you think about things. You're a libertarian, 586 00:34:20,960 --> 00:34:22,719 Speaker 1: or at least I think of you as a libertarian. 587 00:34:22,840 --> 00:34:25,680 Speaker 1: I've never actually asked you that. If that bridles or 588 00:34:25,760 --> 00:34:29,640 Speaker 1: feels too narrowing in terms of how I'm describing part 589 00:34:29,719 --> 00:34:33,279 Speaker 1: of your philosophy, correct me. But the reason I bring 590 00:34:33,320 --> 00:34:36,840 Speaker 1: it up is part of your own thinking is about 591 00:34:37,120 --> 00:34:44,080 Speaker 1: irrational regulation versus sensible regulation, and what kind of limits 592 00:34:44,120 --> 00:34:46,560 Speaker 1: we want around our government, what role we want for 593 00:34:46,560 --> 00:34:50,799 Speaker 1: the government and for free markets in our society. How 594 00:34:50,800 --> 00:34:53,160 Speaker 1: do you think about that in the context of chat, 595 00:34:53,200 --> 00:34:56,800 Speaker 1: GPT and our earlier discussion of how we manage something 596 00:34:56,800 --> 00:35:00,520 Speaker 1: like this? Should the evolution of it be left entirely 597 00:35:00,560 --> 00:35:03,200 Speaker 1: in the hands of the private sector? Is there an 598 00:35:03,239 --> 00:35:06,880 Speaker 1: important role here for the government. If there is, what 599 00:35:07,120 --> 00:35:09,080 Speaker 1: is that role or maybe you think there is none. 600 00:35:09,360 --> 00:35:13,160 Speaker 1: I would first say that few libertarians consider me a libertarian, 601 00:35:13,440 --> 00:35:16,360 Speaker 1: but I feel very libertarian relative to most people I know, 602 00:35:16,520 --> 00:35:19,360 Speaker 1: including those at Bloomberg, so that would put me somewhere 603 00:35:19,360 --> 00:35:23,000 Speaker 1: in between. I think when there's a very new technology 604 00:35:23,360 --> 00:35:26,719 Speaker 1: and things are changing this rapidly, it's very difficult or 605 00:35:26,760 --> 00:35:29,880 Speaker 1: even impossible to regulate it. Regulators don't have the knowledge, 606 00:35:30,200 --> 00:35:32,720 Speaker 1: we're not even sure what it will end up doing. 607 00:35:33,280 --> 00:35:36,000 Speaker 1: So to try to write, say, privacy law for AI 608 00:35:36,200 --> 00:35:39,399 Speaker 1: right now, or large language models, I just don't think 609 00:35:39,440 --> 00:35:42,160 Speaker 1: it makes any sense. It's a bit like crypto. Crypto 610 00:35:42,239 --> 00:35:44,960 Speaker 1: might fail, it might succeed in a few areas. I 611 00:35:45,000 --> 00:35:48,160 Speaker 1: think eventually we should regulate crypto, but right now we 612 00:35:48,160 --> 00:35:51,040 Speaker 1: should be quite minimalistic and just see if it ends 613 00:35:51,080 --> 00:35:53,560 Speaker 1: up being good for anything at all, and then regulate 614 00:35:53,600 --> 00:35:56,560 Speaker 1: what becomes a kind of established use, so we don't 615 00:35:56,560 --> 00:35:58,920 Speaker 1: get in the way of the innovation, right that you 616 00:35:59,040 --> 00:36:02,440 Speaker 1: let the product of Yes, but you know over time 617 00:36:02,680 --> 00:36:04,879 Speaker 1: there will be aspects of it we likely will need 618 00:36:04,920 --> 00:36:07,040 Speaker 1: to regulate. I just don't think we know what they 619 00:36:07,040 --> 00:36:10,279 Speaker 1: are now, So to stop it and end up with 620 00:36:10,320 --> 00:36:13,439 Speaker 1: a less responsible version of it abroad being what people 621 00:36:13,520 --> 00:36:16,799 Speaker 1: consult that strikes me as a very bad outcome. So 622 00:36:16,960 --> 00:36:19,600 Speaker 1: help me again here, professor. One of the things we 623 00:36:19,680 --> 00:36:22,840 Speaker 1: do on this show is try to find learning moments 624 00:36:22,840 --> 00:36:26,480 Speaker 1: in any collision. What's the most surprising thing you've learned 625 00:36:26,600 --> 00:36:32,239 Speaker 1: about AI n LMS since chat GPTs public debut? If 626 00:36:32,280 --> 00:36:34,600 Speaker 1: I have been surprised just how much? If the debate 627 00:36:34,760 --> 00:36:38,040 Speaker 1: is over, like is this woke enough? Can you get 628 00:36:38,080 --> 00:36:41,680 Speaker 1: it to treat the two political parties symmetrically and so 629 00:36:41,760 --> 00:36:44,480 Speaker 1: on and so on. I think that's just totally missing 630 00:36:44,520 --> 00:36:48,239 Speaker 1: the main storyline. It's not irrelevant. It's a waste of 631 00:36:48,280 --> 00:36:51,520 Speaker 1: our time. It's one question, you know, we might consider, 632 00:36:52,200 --> 00:36:53,759 Speaker 1: but at the end of the day, we need to 633 00:36:53,760 --> 00:36:56,560 Speaker 1: be focused on making it better. And if the very 634 00:36:56,560 --> 00:37:00,440 Speaker 1: earliest version of the product isn't exactly your politics, you know, 635 00:37:00,719 --> 00:37:02,680 Speaker 1: deal with that. It would be like complaining about the 636 00:37:02,719 --> 00:37:05,600 Speaker 1: printing press because the first seven books that we're published 637 00:37:05,600 --> 00:37:09,160 Speaker 1: were too Catholic for you, or to some other particular 638 00:37:09,200 --> 00:37:11,200 Speaker 1: religious point of view, like Moore is going to come. 639 00:37:11,239 --> 00:37:13,879 Speaker 1: You'll get a Protestant Reformation, and you'll get a Book 640 00:37:13,880 --> 00:37:16,120 Speaker 1: of Mormon, and you'll get much much more like just 641 00:37:16,239 --> 00:37:20,359 Speaker 1: wait a bit. So how much people are polarizing emotionally 642 00:37:20,800 --> 00:37:22,920 Speaker 1: is more than I thought, and I think mostly a 643 00:37:22,960 --> 00:37:25,680 Speaker 1: waste of time. Tyler, thanks for being with us today, 644 00:37:26,120 --> 00:37:28,960 Speaker 1: dam thanks for having me on. You can find Tyler 645 00:37:29,040 --> 00:37:34,320 Speaker 1: Cowen on Twitter at Tyler Cowen, on his blog Marginal Revolution, 646 00:37:34,360 --> 00:37:38,040 Speaker 1: where he also has a podcast, and of course on 647 00:37:38,120 --> 00:37:41,600 Speaker 1: Bloomberg Opinion, where he writes columns for us. Well we 648 00:37:41,680 --> 00:37:48,239 Speaker 1: come back. We'll be joined again by parme Olsen. We're 649 00:37:48,239 --> 00:37:51,359 Speaker 1: back with parme Olsen. Thanks for being with us. Thank 650 00:37:51,360 --> 00:37:55,640 Speaker 1: you for having me. Let's talk about how disruptive this 651 00:37:55,760 --> 00:38:01,000 Speaker 1: might be to existing search titans and ormation titans on 652 00:38:01,040 --> 00:38:04,879 Speaker 1: the web. This poses a potential threat to Google's entire 653 00:38:04,960 --> 00:38:08,040 Speaker 1: search model, doesn't it right? I really think it does. 654 00:38:08,360 --> 00:38:12,319 Speaker 1: There was actually a moment internally at Google where executives 655 00:38:12,320 --> 00:38:16,239 Speaker 1: issued a code read because they took it that seriously 656 00:38:16,840 --> 00:38:20,600 Speaker 1: that they kind of reassigned people from different teams to 657 00:38:20,920 --> 00:38:24,200 Speaker 1: work on their own language model, which is called lambda, 658 00:38:24,360 --> 00:38:29,880 Speaker 1: and finding ways of effectively providing their own alternative to 659 00:38:30,000 --> 00:38:36,040 Speaker 1: Chat GPT, which they eventually did. Because search is not 660 00:38:36,120 --> 00:38:40,640 Speaker 1: only dominated by Google, it dominates its own revenues. It's 661 00:38:40,680 --> 00:38:43,480 Speaker 1: one hundred and fifty billion dollar business for Google. It 662 00:38:43,600 --> 00:38:47,880 Speaker 1: is fundamental to its AD stack and its AD network, 663 00:38:48,040 --> 00:38:51,320 Speaker 1: and if it loses market share, that is a major 664 00:38:51,520 --> 00:38:55,520 Speaker 1: threat down the line. So it's an existential threat to Google. 665 00:38:56,000 --> 00:38:59,840 Speaker 1: What does it mean to Microsoft parmy? So for Microsoft, 666 00:39:00,000 --> 00:39:05,280 Speaker 1: it's actually a really great opportunity and with little risk 667 00:39:05,920 --> 00:39:11,400 Speaker 1: because Microsoft has an investment in open Ai. It invested 668 00:39:11,400 --> 00:39:14,120 Speaker 1: in the company in twenty nineteen one billion dollars and 669 00:39:14,160 --> 00:39:18,160 Speaker 1: then earlier this year another ten billion dollars, really cementing 670 00:39:18,160 --> 00:39:22,200 Speaker 1: its foothold in generative AI and in this particular company, 671 00:39:22,200 --> 00:39:25,560 Speaker 1: which is really on the forefront of language models. And 672 00:39:25,719 --> 00:39:29,040 Speaker 1: as part of that investment, it gets exclusive access to 673 00:39:29,080 --> 00:39:34,839 Speaker 1: open AI's technology, including language models like GPT three, which 674 00:39:34,880 --> 00:39:39,080 Speaker 1: is what powers chat GPT. And so the next thing 675 00:39:39,080 --> 00:39:42,839 Speaker 1: that Microsoft did, of course was very quickly plug in 676 00:39:42,960 --> 00:39:46,640 Speaker 1: that language model into Bing, which is the search engine 677 00:39:46,640 --> 00:39:49,799 Speaker 1: that nobody has used for years, except it also ran 678 00:39:50,239 --> 00:39:53,000 Speaker 1: to Google's search engine. That's right, I mean the number 679 00:39:53,040 --> 00:39:56,959 Speaker 1: one search term on Bing is Google. Like people, once 680 00:39:56,960 --> 00:39:59,080 Speaker 1: they start using Bing, they're quite ready to just move 681 00:39:59,120 --> 00:40:03,680 Speaker 1: on to Google. But suddenly, here is this new fangled 682 00:40:03,680 --> 00:40:07,600 Speaker 1: way to make being much more interesting. Instead of searching 683 00:40:07,920 --> 00:40:11,240 Speaker 1: for information on a topic or a company and getting 684 00:40:11,239 --> 00:40:13,520 Speaker 1: a whole page of links that you have to click around, 685 00:40:14,000 --> 00:40:15,920 Speaker 1: you could ask it a question and it would give 686 00:40:15,960 --> 00:40:20,360 Speaker 1: you a synthesized single answer. And so Microsoft started working 687 00:40:20,440 --> 00:40:24,320 Speaker 1: on plugging that in, and Google had this code read 688 00:40:24,400 --> 00:40:27,280 Speaker 1: internally that they needed to do the same thing. Suddenly 689 00:40:27,320 --> 00:40:30,960 Speaker 1: their businesses under threat. Whereas for Microsoft, if they can 690 00:40:31,040 --> 00:40:34,759 Speaker 1: just capture one or two percent of this market, the 691 00:40:34,840 --> 00:40:38,879 Speaker 1: search engine market for ads tied to search results, let's 692 00:40:38,960 --> 00:40:41,960 Speaker 1: like two or three billion dollars added to their top line, 693 00:40:42,360 --> 00:40:44,879 Speaker 1: and then if they don't manage to do that, it's 694 00:40:44,960 --> 00:40:48,240 Speaker 1: not a huge deal for them, because this is a business. 695 00:40:48,280 --> 00:40:51,440 Speaker 1: This is kind of coming out of nothing. Anyway. Microsoft 696 00:40:51,560 --> 00:40:54,880 Speaker 1: is an enterprise software company and search as an ornament 697 00:40:55,280 --> 00:40:59,440 Speaker 1: precisely currently exactly so ironically, the company that got lapped 698 00:40:59,440 --> 00:41:02,400 Speaker 1: by some up starts is in a position now to 699 00:41:02,560 --> 00:41:06,080 Speaker 1: revisit some pain on those up starts if AI takes 700 00:41:06,160 --> 00:41:08,879 Speaker 1: off in the way that some people think. I've never 701 00:41:08,920 --> 00:41:11,719 Speaker 1: seen Google move so fast. And the other thing is 702 00:41:11,760 --> 00:41:16,600 Speaker 1: Google has a language model, a very powerful one called 703 00:41:16,680 --> 00:41:20,680 Speaker 1: Lambda that it introduced to the scientific community two years ago. 704 00:41:20,880 --> 00:41:23,560 Speaker 1: It's also got billions of parameters and in fact, I 705 00:41:23,600 --> 00:41:26,000 Speaker 1: don't know if you remember this new story from last year, 706 00:41:26,080 --> 00:41:29,440 Speaker 1: but one of Google's engineers who was testing Lambda believed 707 00:41:29,440 --> 00:41:32,279 Speaker 1: that it was sentient. He was chatting with it so much. 708 00:41:32,880 --> 00:41:34,600 Speaker 1: He wanted it to have rights, He wanted it to 709 00:41:34,640 --> 00:41:37,080 Speaker 1: have a lawyer. I mean, again, this is kind of 710 00:41:37,560 --> 00:41:40,200 Speaker 1: human projection. But it just goes to show how good 711 00:41:40,200 --> 00:41:43,479 Speaker 1: this model was. But Google kept it under wraps, never 712 00:41:43,520 --> 00:41:45,879 Speaker 1: said why, but kind of from what it has said 713 00:41:45,880 --> 00:41:48,839 Speaker 1: publicly about AI and what we've heard from reporting on 714 00:41:49,239 --> 00:41:52,600 Speaker 1: internal meetings, is that they don't want it to go rogue. 715 00:41:52,640 --> 00:41:55,359 Speaker 1: You know, if they put it into Google Assistant so 716 00:41:55,400 --> 00:41:57,399 Speaker 1: that you can talk to the Google speaker and then 717 00:41:57,440 --> 00:42:00,600 Speaker 1: it's Lambda talking back to you, I could say some 718 00:42:00,680 --> 00:42:03,200 Speaker 1: really crazy stuff and they just didn't want to take 719 00:42:03,239 --> 00:42:07,040 Speaker 1: that risk to its reputation. You know, that's interesting to 720 00:42:07,080 --> 00:42:10,120 Speaker 1: me again because you reassured me earlier on that they 721 00:42:10,160 --> 00:42:12,439 Speaker 1: won't go rogue, or you thought they may not go rogue, 722 00:42:12,520 --> 00:42:14,720 Speaker 1: or that they were designed in a way to prevent 723 00:42:14,800 --> 00:42:19,640 Speaker 1: rogue behavior, and yet Google itself, its engineers are worried 724 00:42:19,880 --> 00:42:23,200 Speaker 1: about rogue behavior. Now that may just be that we're 725 00:42:23,200 --> 00:42:26,520 Speaker 1: defining rogue in different ways. But this leads to two 726 00:42:26,680 --> 00:42:30,560 Speaker 1: big final topics that I wanted to get into with you, 727 00:42:31,200 --> 00:42:34,440 Speaker 1: the first one being, you know, we've had this view 728 00:42:34,480 --> 00:42:38,759 Speaker 1: of technology, or at least technological evangelists ahead of view, 729 00:42:38,960 --> 00:42:42,560 Speaker 1: that it's ultimately liberating and empowering for people. And I'm 730 00:42:42,600 --> 00:42:44,439 Speaker 1: of that view in some ways. You know, I think 731 00:42:44,640 --> 00:42:46,920 Speaker 1: you can look at the Internet as a liberating force 732 00:42:47,280 --> 00:42:50,239 Speaker 1: and it can also be a destructive or oppressive force 733 00:42:50,280 --> 00:42:54,360 Speaker 1: depending on how it's used. AI and its potential also 734 00:42:54,520 --> 00:43:00,000 Speaker 1: turns some of those expectations on their heads, doesn't it. Yeah, 735 00:43:00,040 --> 00:43:02,640 Speaker 1: I think so. So a lot of this comes down 736 00:43:02,719 --> 00:43:05,720 Speaker 1: to control. Right once something is out in the wild, 737 00:43:06,160 --> 00:43:08,480 Speaker 1: we don't know how people are going to use it, 738 00:43:08,880 --> 00:43:12,000 Speaker 1: we don't know how vulnerable they're going to be to it, 739 00:43:12,440 --> 00:43:16,440 Speaker 1: and we can't really control what that technology does. In AI, 740 00:43:16,520 --> 00:43:19,399 Speaker 1: in particular, I mean, I can't think of any other 741 00:43:19,400 --> 00:43:23,280 Speaker 1: type of technology apart from social media, which is so 742 00:43:23,400 --> 00:43:27,600 Speaker 1: difficult to predict or control. When people talk about AI 743 00:43:27,760 --> 00:43:31,920 Speaker 1: becoming more advanced, sort of reaching that point of AGI 744 00:43:32,120 --> 00:43:37,040 Speaker 1: artificial general intelligence, when it surpasses human intelligence, it's hard 745 00:43:37,080 --> 00:43:40,600 Speaker 1: to really predict what that will mean for humans because 746 00:43:40,600 --> 00:43:43,720 Speaker 1: who could have predicted ten years ago how our lives 747 00:43:43,760 --> 00:43:46,680 Speaker 1: would have changed under social media And some of the 748 00:43:46,680 --> 00:43:50,200 Speaker 1: impacts that tech has had on society now have been 749 00:43:50,280 --> 00:43:53,640 Speaker 1: so nuanced in terms of the harmful stuff I would 750 00:43:53,680 --> 00:43:57,239 Speaker 1: imagine with AI, human manipulation would be a big risk. 751 00:43:57,600 --> 00:44:00,800 Speaker 1: Misinformation is a huge risk, and then so as addiction 752 00:44:00,880 --> 00:44:03,239 Speaker 1: as well. I mean, some of the researchers that have 753 00:44:03,320 --> 00:44:06,320 Speaker 1: been talking to being I've talked to it for hours 754 00:44:06,400 --> 00:44:11,080 Speaker 1: on end because it's a really compelling entity to chat to. 755 00:44:11,400 --> 00:44:14,920 Speaker 1: So what happens when kids start using these kinds of services, 756 00:44:15,040 --> 00:44:18,719 Speaker 1: or vulnerable people or people who are easily persuaded. I 757 00:44:18,760 --> 00:44:20,800 Speaker 1: think those are the kinds of issues that these companies 758 00:44:20,920 --> 00:44:23,839 Speaker 1: really have to think about very carefully well, and not 759 00:44:23,920 --> 00:44:28,000 Speaker 1: just companies, but society and the public sector. And I 760 00:44:28,000 --> 00:44:31,120 Speaker 1: think that gets into the last topic I wanted to 761 00:44:31,120 --> 00:44:36,200 Speaker 1: explore with you was as this evolves, where does control 762 00:44:36,440 --> 00:44:40,360 Speaker 1: of the future of these tools properly reside. We don't 763 00:44:40,400 --> 00:44:43,480 Speaker 1: want innovation in the private sector to get snuffed out. 764 00:44:43,680 --> 00:44:46,200 Speaker 1: It's one of the great strengths of the American economy. 765 00:44:46,800 --> 00:44:50,360 Speaker 1: At the same time, we know of myriad examples in 766 00:44:50,440 --> 00:44:54,279 Speaker 1: which private products, even if they're innovative and profitable, can 767 00:44:54,360 --> 00:44:58,360 Speaker 1: have negative side effects that require some sort of intermediation 768 00:44:58,600 --> 00:45:02,520 Speaker 1: from society and the public sector from government. How do 769 00:45:02,600 --> 00:45:06,560 Speaker 1: you see that shaping up, Feza v Ai, Well, I 770 00:45:06,600 --> 00:45:10,200 Speaker 1: think it's important to note that the private sector is 771 00:45:10,280 --> 00:45:12,840 Speaker 1: really leading the way on this, and not the public 772 00:45:12,880 --> 00:45:16,160 Speaker 1: sector at all. Chat GPT only came out two or 773 00:45:16,200 --> 00:45:19,600 Speaker 1: three months ago. There's no regulation to speak of for 774 00:45:19,680 --> 00:45:23,000 Speaker 1: this kind of service. The only thing we have is 775 00:45:23,040 --> 00:45:27,280 Speaker 1: the AI Act, which is being proposed in the European 776 00:45:27,400 --> 00:45:30,719 Speaker 1: Union as a very broad piece of legislation, and I 777 00:45:30,760 --> 00:45:34,600 Speaker 1: don't know that it even incorporates chatbots like chat GPT. 778 00:45:35,160 --> 00:45:39,080 Speaker 1: But right now, this is self regulation by tech companies, 779 00:45:39,080 --> 00:45:41,640 Speaker 1: which as we know, hasn't always gone all that well, 780 00:45:41,840 --> 00:45:45,960 Speaker 1: because they are at the forefront of producing and bringing 781 00:45:46,040 --> 00:45:49,319 Speaker 1: this technology to the public, and it wasn't always that way. 782 00:45:49,640 --> 00:45:53,640 Speaker 1: Open ai, which is the company that started GPT three 783 00:45:53,719 --> 00:45:57,439 Speaker 1: and created chat. GPT was founded in twenty fifteen as 784 00:45:57,480 --> 00:46:01,280 Speaker 1: a nonprofit and it was co founded by Elon Musk 785 00:46:01,840 --> 00:46:05,360 Speaker 1: and Sam Altman, who was the head of y Combinator, 786 00:46:05,400 --> 00:46:09,920 Speaker 1: which is like this very highly regarded accelerator in Silicon Valley. 787 00:46:09,960 --> 00:46:13,359 Speaker 1: Was an incredible networker, a real hustler, an ambitious guy. 788 00:46:14,080 --> 00:46:19,120 Speaker 1: And Musk and Altman's big idea was they wanted to 789 00:46:19,120 --> 00:46:23,040 Speaker 1: create a nonprofit organization that was going to create AI 790 00:46:23,360 --> 00:46:25,960 Speaker 1: for the good of humanity. And it was very important 791 00:46:25,960 --> 00:46:28,600 Speaker 1: to them to make it a nonprofit because then there 792 00:46:28,600 --> 00:46:31,640 Speaker 1: wouldn't be corporate interests at play. But then, of course, 793 00:46:31,680 --> 00:46:34,879 Speaker 1: over the years, as they're developing this AI and trying 794 00:46:34,920 --> 00:46:38,480 Speaker 1: to hire the best AI talent, they run out of money. 795 00:46:38,840 --> 00:46:41,480 Speaker 1: They don't have all the computing power they need. So 796 00:46:41,760 --> 00:46:46,360 Speaker 1: enter Microsoft with a one billion dollar investment and open 797 00:46:46,400 --> 00:46:51,480 Speaker 1: ai changes its nonprofit status to becoming a for profit status. 798 00:46:51,840 --> 00:46:55,120 Speaker 1: Then two years later, Microsoft puts ten billion dollars into 799 00:46:55,160 --> 00:46:58,440 Speaker 1: open ai and the company is essentially now a research 800 00:46:58,560 --> 00:47:01,840 Speaker 1: arm for Microsoft. And there's all of this sounds like 801 00:47:01,880 --> 00:47:05,800 Speaker 1: the Internet, which was invented with government funding through DARPA, 802 00:47:06,320 --> 00:47:09,319 Speaker 1: and the national security and sort of intelligence arm of 803 00:47:09,320 --> 00:47:12,560 Speaker 1: the federal government, and then it was made available to 804 00:47:12,600 --> 00:47:16,279 Speaker 1: the private sector, and then the private sector brought it 805 00:47:16,360 --> 00:47:20,200 Speaker 1: into mainstream use. It's not exactly the same, but it 806 00:47:20,239 --> 00:47:22,879 Speaker 1: presents the same sort of issues, doesn't it. It goes 807 00:47:22,920 --> 00:47:25,279 Speaker 1: in that same direction, and there have been so many 808 00:47:25,280 --> 00:47:28,200 Speaker 1: efforts to try and fight against that. There's been this 809 00:47:28,280 --> 00:47:32,719 Speaker 1: kind of tension between computer scientists and the research community 810 00:47:32,800 --> 00:47:36,040 Speaker 1: and big tech, and big tech is very much winning. 811 00:47:36,719 --> 00:47:40,000 Speaker 1: Do you think it's too early to understand the evolution 812 00:47:40,040 --> 00:47:41,839 Speaker 1: of this and is it a wait and see in 813 00:47:41,920 --> 00:47:44,520 Speaker 1: terms of what structure or model might be best to 814 00:47:44,560 --> 00:47:47,560 Speaker 1: manage this. I think it is wait and see and 815 00:47:47,600 --> 00:47:52,000 Speaker 1: wait and see an experiment because although I say that 816 00:47:52,040 --> 00:47:56,000 Speaker 1: big tech is winning on this, there are efforts to 817 00:47:56,080 --> 00:48:01,680 Speaker 1: create open source, generative AI product and services. There were 818 00:48:01,719 --> 00:48:06,920 Speaker 1: some early employees of open ai who felt quite disillusioned 819 00:48:07,000 --> 00:48:09,560 Speaker 1: with the direction that the organization was going once it 820 00:48:09,600 --> 00:48:12,520 Speaker 1: became a for profit company, and they spun out and 821 00:48:12,560 --> 00:48:17,399 Speaker 1: created their own company called Anthropic, which is focused much 822 00:48:17,440 --> 00:48:20,279 Speaker 1: more on AI safety and what they call AI alignment. 823 00:48:20,920 --> 00:48:23,880 Speaker 1: So I think there's lots of experiments at play, but 824 00:48:24,040 --> 00:48:26,239 Speaker 1: by and large this is going to be a technology 825 00:48:26,360 --> 00:48:31,279 Speaker 1: that is steered empowered by big tech. And just like 826 00:48:31,360 --> 00:48:34,399 Speaker 1: as a small example, the company I just mentioned, Anthropic, 827 00:48:35,080 --> 00:48:37,839 Speaker 1: they just got a massive investment from Google a few 828 00:48:37,880 --> 00:48:40,200 Speaker 1: weeks ago. I think it was worth three hundred million dollars. 829 00:48:40,560 --> 00:48:44,160 Speaker 1: So it's very hard to keep big tech corporate interests 830 00:48:44,239 --> 00:48:47,120 Speaker 1: out of this kind of research because that's where all 831 00:48:47,120 --> 00:48:49,880 Speaker 1: the computing power is. You know, these big tech companies 832 00:48:49,920 --> 00:48:53,200 Speaker 1: have access to supercomputers. That's how open ai trained its 833 00:48:53,200 --> 00:48:57,960 Speaker 1: model through a Microsoft supercomputer, and they have the server capability, 834 00:48:57,960 --> 00:49:00,759 Speaker 1: and it's just so hard to get that. Like in 835 00:49:00,800 --> 00:49:03,799 Speaker 1: the academic world, no university can really compete with that. 836 00:49:04,560 --> 00:49:06,960 Speaker 1: You've been watching all this for a long time. What 837 00:49:07,080 --> 00:49:10,640 Speaker 1: have you learned from these recent developments in the debut 838 00:49:10,719 --> 00:49:14,600 Speaker 1: of chat GPT. Well, I think the thing that really 839 00:49:14,600 --> 00:49:18,160 Speaker 1: surprised me about chat GPT and some of the latest 840 00:49:18,360 --> 00:49:21,200 Speaker 1: generative AI tools to come out in the past year 841 00:49:22,000 --> 00:49:26,120 Speaker 1: is how creative AI seems to be. Because for years 842 00:49:26,120 --> 00:49:29,640 Speaker 1: and years when people talked about AI taking people's jobs, 843 00:49:30,280 --> 00:49:34,680 Speaker 1: it was about taking factory workers jobs and truck driver jobs. 844 00:49:35,000 --> 00:49:38,000 Speaker 1: But now it seems like the real jobs that are 845 00:49:38,000 --> 00:49:44,000 Speaker 1: a threat are the creative classes and professional j Yeah, 846 00:49:44,040 --> 00:49:46,719 Speaker 1: I didn't want to say it, but you did. But 847 00:49:46,760 --> 00:49:48,920 Speaker 1: the other thing I want to say that is a 848 00:49:48,960 --> 00:49:54,640 Speaker 1: big shortcoming of these systems is that they're often inaccurate. 849 00:49:55,120 --> 00:49:58,760 Speaker 1: I shouldn't say often, but often enough that it's a problem. 850 00:49:59,239 --> 00:50:04,200 Speaker 1: Open aie will not say how often these systems get 851 00:50:04,239 --> 00:50:07,080 Speaker 1: things wrong. I've asked it, but my own experience, it's 852 00:50:07,120 --> 00:50:10,000 Speaker 1: just I think it's somewhere between five and fifteen percent 853 00:50:10,120 --> 00:50:13,799 Speaker 1: of the answers that it's given me are factually incorrect. Now, 854 00:50:13,880 --> 00:50:17,759 Speaker 1: think about using that as a search tool. We use 855 00:50:17,800 --> 00:50:22,320 Speaker 1: search to get information, to get facts, and if it's 856 00:50:22,400 --> 00:50:25,960 Speaker 1: wrong ten percent of the time, are people really going 857 00:50:26,000 --> 00:50:27,719 Speaker 1: to want to use it? I think that's going to 858 00:50:27,760 --> 00:50:32,080 Speaker 1: be a real problem for these companies using these systems 859 00:50:31,560 --> 00:50:35,040 Speaker 1: for a search engine companions. And it's not a trivial 860 00:50:35,080 --> 00:50:41,560 Speaker 1: issue because recently Microsoft and Google had these big announcements 861 00:50:41,600 --> 00:50:45,440 Speaker 1: about their new chat companions, these chatbots that we're going 862 00:50:45,480 --> 00:50:48,239 Speaker 1: to help Bang and we're going to help Google, and 863 00:50:48,320 --> 00:50:51,759 Speaker 1: in both demonstrations there were errors. So if they can't 864 00:50:51,760 --> 00:50:54,480 Speaker 1: even fact check that and get that right, what are 865 00:50:54,480 --> 00:50:56,480 Speaker 1: these systems going to be like when they're actually out 866 00:50:56,480 --> 00:50:59,880 Speaker 1: in the wild, and on that note, we're going to 867 00:51:00,080 --> 00:51:03,040 Speaker 1: end our conversation today. Parmi, thanks so much for joining us. 868 00:51:03,440 --> 00:51:08,000 Speaker 1: Thank you. You can find Parmi Olsen on Twitter at 869 00:51:08,120 --> 00:51:12,080 Speaker 1: Parmi and on Bloomberg Opinion, where she writes columns for us. 870 00:51:13,040 --> 00:51:18,440 Speaker 1: Here at crash Course, we believe the collisions can be messy, impressive, challenging, surprising, 871 00:51:18,760 --> 00:51:22,360 Speaker 1: and always instructive. In today's Crash Course, I learned that 872 00:51:22,400 --> 00:51:25,080 Speaker 1: when it comes to chat, GPT and other forms of 873 00:51:25,120 --> 00:51:29,719 Speaker 1: AI that are revolutionizing how we deal with information patients 874 00:51:29,840 --> 00:51:32,560 Speaker 1: may be a virtue as long as we don't let 875 00:51:32,600 --> 00:51:35,840 Speaker 1: our guard down. We'd love to hear from you. You 876 00:51:35,880 --> 00:51:39,440 Speaker 1: can tweet at the Bloomberg Opinion handle at Opinion or 877 00:51:39,520 --> 00:51:43,880 Speaker 1: me at Tim O'Brien using the hashtag Bloomberg crash Course. 878 00:51:44,480 --> 00:51:46,960 Speaker 1: You can also subscribe to our show wherever you're listening 879 00:51:47,080 --> 00:51:49,560 Speaker 1: right now and leave us a review. It helps from 880 00:51:49,560 --> 00:51:53,080 Speaker 1: where people find the show. This episode was produced by 881 00:51:53,080 --> 00:51:58,920 Speaker 1: the indispensable Anamazarakis and me. Our supervising producer is Magnus Henrickson, 882 00:51:59,320 --> 00:52:02,719 Speaker 1: and we had editing help from Katie Boys, Jeff Grocott, 883 00:52:03,120 --> 00:52:07,400 Speaker 1: Mike Nizza, and Christine Vanden Bilart. Blake Maples does our 884 00:52:07,440 --> 00:52:10,920 Speaker 1: sound engineering and our original theme song was composed by 885 00:52:11,000 --> 00:52:14,919 Speaker 1: Luise Gara. I'm Tim O'Brien. We'll be back next week 886 00:52:15,000 --> 00:52:16,120 Speaker 1: with another crash course.