1 00:00:04,080 --> 00:00:06,960 Speaker 1: On this episode of newts World, we are continuing our 2 00:00:07,040 --> 00:00:12,560 Speaker 1: series on artificial intelligence as OpenAIS chat GPT four continues 3 00:00:12,600 --> 00:00:15,600 Speaker 1: to develop and gain media attention and public debate. It 4 00:00:15,640 --> 00:00:19,279 Speaker 1: has provoked a conversation in the creative community about when 5 00:00:19,640 --> 00:00:21,840 Speaker 1: and how it should be used and who owns the 6 00:00:21,920 --> 00:00:24,720 Speaker 1: rights to the work. The use of artificial intelligence in 7 00:00:24,760 --> 00:00:27,160 Speaker 1: creative writing is part of what the Writer's Guild of 8 00:00:27,200 --> 00:00:31,120 Speaker 1: America Strike is about. While my guest today has taken 9 00:00:31,160 --> 00:00:34,519 Speaker 1: a new approach. Journalist Stephen Marsh has been immersed in 10 00:00:34,520 --> 00:00:37,519 Speaker 1: the world of AI since twenty seventeen. He is a 11 00:00:37,560 --> 00:00:40,080 Speaker 1: believer in the benefits of AI as a tool for 12 00:00:40,159 --> 00:00:43,720 Speaker 1: creative expression, and his new novella, Death of an Author 13 00:00:44,200 --> 00:00:48,120 Speaker 1: was written by three cutting edge AI writing platforms. The 14 00:00:48,200 --> 00:00:51,160 Speaker 1: result is a gripping mystery that is ninety five percent 15 00:00:51,200 --> 00:00:54,520 Speaker 1: written by AI. His AI author's name is Aiden Marshen 16 00:00:55,000 --> 00:00:58,520 Speaker 1: and five percent by Marsh, who skillfully crafted the story, 17 00:00:58,520 --> 00:01:01,880 Speaker 1: outline and machine prompts. So I'm really pleased to welcome 18 00:01:01,920 --> 00:01:05,440 Speaker 1: my guest, Stephen Marsh. He is a novelist and essayist. 19 00:01:05,680 --> 00:01:09,120 Speaker 1: He wrote his first algorithmically generated story for Wired in 20 00:01:09,160 --> 00:01:13,720 Speaker 1: twenty seventeen. And has written about artificial intelligence The New Yorker, 21 00:01:13,840 --> 00:01:16,680 Speaker 1: The Atlantic, and The New York Times. His books include 22 00:01:17,000 --> 00:01:30,120 Speaker 1: The Next Civil War and On Writing and Failure. Steven welcome, 23 00:01:30,160 --> 00:01:31,640 Speaker 1: and I have to say this is one of the 24 00:01:31,640 --> 00:01:33,560 Speaker 1: more fun things we've done on Newtsworld. 25 00:01:33,680 --> 00:01:35,720 Speaker 2: I mean, you know, is there a more fascinating subject 26 00:01:35,720 --> 00:01:38,080 Speaker 2: in the world right now? I don't actually think there is. 27 00:01:38,600 --> 00:01:41,720 Speaker 1: It's certainly at the absolute cutting edge. But I'm curious, 28 00:01:42,000 --> 00:01:45,119 Speaker 1: when did you start getting interested in artificial intelligence? 29 00:01:45,720 --> 00:01:49,440 Speaker 2: In twenty twelve, I wrote an article for the LA 30 00:01:49,560 --> 00:01:53,000 Speaker 2: Review of Books called against Digital Humanities, because I really 31 00:01:53,000 --> 00:01:56,720 Speaker 2: disapproved of this new technol It was a techy kind 32 00:01:56,760 --> 00:01:59,520 Speaker 2: of esthetic that they were applying to literary criticism, and 33 00:01:59,520 --> 00:02:01,600 Speaker 2: I wrote a but it's one of those things that 34 00:02:01,680 --> 00:02:03,960 Speaker 2: while I was critiquing it, I sort of think, well, 35 00:02:04,200 --> 00:02:07,360 Speaker 2: you know, that part's interesting, and that part's interesting, and 36 00:02:07,400 --> 00:02:09,280 Speaker 2: maybe I should call that guy to talk to him. 37 00:02:09,639 --> 00:02:12,600 Speaker 2: And then I sort of fell in with computer scientists 38 00:02:12,680 --> 00:02:18,000 Speaker 2: in Toronto who were working on linguistic forms of artificial intelligence. 39 00:02:18,000 --> 00:02:20,560 Speaker 2: In fact, they were at the dog park in my neighborhood, 40 00:02:20,880 --> 00:02:23,160 Speaker 2: and my daughter liked to go visit the dog park, 41 00:02:23,240 --> 00:02:26,320 Speaker 2: so I met a bunch of them that way, and 42 00:02:26,360 --> 00:02:29,600 Speaker 2: I became really pretty obsessed with it and pretty fascinated 43 00:02:29,639 --> 00:02:31,200 Speaker 2: with it. And as you say, I started doing it 44 00:02:31,240 --> 00:02:34,520 Speaker 2: in twenty seventeen, which was prehistory, I mean before the 45 00:02:34,520 --> 00:02:38,440 Speaker 2: invention of the Transformer, and the Transformer is the technology 46 00:02:38,440 --> 00:02:41,400 Speaker 2: that has changed everything in my view, with generative AI. 47 00:02:41,760 --> 00:02:44,440 Speaker 1: For those of us who are genuinely ignorant, what is 48 00:02:44,520 --> 00:02:45,320 Speaker 1: the Transformer? 49 00:02:45,680 --> 00:02:48,880 Speaker 2: The Transformer was invented at Google by a bunch of 50 00:02:48,919 --> 00:02:52,880 Speaker 2: Canadians in twenty seventeen. If I were to try to 51 00:02:52,919 --> 00:02:54,960 Speaker 2: describe it to you, I would do a poor job. 52 00:02:55,480 --> 00:02:57,760 Speaker 2: Like I've tried very hard to understand it. I'm not 53 00:02:57,840 --> 00:03:01,760 Speaker 2: an idiot. But it's an complicated piece of technology that 54 00:03:01,880 --> 00:03:06,600 Speaker 2: essentially takes every token of language, which is simple letters 55 00:03:06,600 --> 00:03:10,760 Speaker 2: and symbols, and creates parameters or patterns between them and 56 00:03:11,000 --> 00:03:13,880 Speaker 2: the position of every other token of language. And then 57 00:03:13,919 --> 00:03:17,839 Speaker 2: it just creates this incredibly powerful text prediction software, which 58 00:03:17,880 --> 00:03:20,600 Speaker 2: is all that any of this is right. And the 59 00:03:20,639 --> 00:03:24,800 Speaker 2: text prediction capacities have what they call emergent properties, which 60 00:03:24,880 --> 00:03:27,960 Speaker 2: is they're just really weird. Things start to emerge from 61 00:03:27,960 --> 00:03:30,000 Speaker 2: them pretty quickly. If I wrote a piece for the 62 00:03:30,040 --> 00:03:34,239 Speaker 2: New Yorker, where I asked GPT three the T and 63 00:03:34,400 --> 00:03:38,960 Speaker 2: GPT is transformer, by the way, to continue Coleridge's poem 64 00:03:39,160 --> 00:03:42,240 Speaker 2: Kubla Khan. And it did so in an absolutely coherent 65 00:03:42,280 --> 00:03:44,080 Speaker 2: way that if you told me that Coleridge wrote it, 66 00:03:44,080 --> 00:03:46,960 Speaker 2: I would believe you. It has the strange properties of 67 00:03:46,960 --> 00:03:50,680 Speaker 2: this text prediction software emerge really from the Transformer. 68 00:03:51,040 --> 00:03:55,400 Speaker 1: Back in twenty seventeen, you published a short story, Twinkle Twinkle, 69 00:03:56,000 --> 00:03:59,240 Speaker 1: as kind of an experiment see whether or not algorithms 70 00:03:59,240 --> 00:04:01,880 Speaker 1: could help you rite. That was pretty bold at that point, 71 00:04:02,000 --> 00:04:02,280 Speaker 1: wasn't it. 72 00:04:02,760 --> 00:04:03,000 Speaker 3: Well? 73 00:04:03,080 --> 00:04:05,080 Speaker 2: Yeah, and it took a huge amount of effort because 74 00:04:05,080 --> 00:04:07,000 Speaker 2: I had to get a computer scientist to design it 75 00:04:07,040 --> 00:04:09,960 Speaker 2: for me, right, Like he actually built sci fi q, 76 00:04:10,160 --> 00:04:13,480 Speaker 2: which was a kind of software that created these stylistic 77 00:04:13,640 --> 00:04:15,920 Speaker 2: metrics that we could measure, and then I had to 78 00:04:16,360 --> 00:04:19,440 Speaker 2: coalesce around certain corpuses of science fiction, like I wanted 79 00:04:19,440 --> 00:04:21,640 Speaker 2: to write a science fiction story, so I got it 80 00:04:21,680 --> 00:04:25,200 Speaker 2: to create an algorithm of my favorite fifty science fiction 81 00:04:25,680 --> 00:04:29,039 Speaker 2: short stories and it was pretty good. But the Transformer 82 00:04:29,080 --> 00:04:32,120 Speaker 2: based stuff is of course on a completely different level. 83 00:04:32,400 --> 00:04:35,039 Speaker 2: I mean, with the Transformer in twenty twenty, I was 84 00:04:35,080 --> 00:04:38,599 Speaker 2: able to write a seventeen percent computer generated story where 85 00:04:39,000 --> 00:04:41,359 Speaker 2: I would just write blocks of text and then have 86 00:04:41,440 --> 00:04:44,640 Speaker 2: the machine finish it right, or create gaps and then 87 00:04:44,680 --> 00:04:47,000 Speaker 2: have the machine finish it in a very coherent way. 88 00:04:47,360 --> 00:04:49,960 Speaker 2: And then I also trained stylistic boughts with a different 89 00:04:50,040 --> 00:04:53,640 Speaker 2: large language model called cohere out of Canada, where I 90 00:04:53,720 --> 00:04:59,400 Speaker 2: actually trained them to write like Nabokov, Jarles Dickens, Jane 91 00:04:59,440 --> 00:05:01,039 Speaker 2: Austen and and I had each of them write a 92 00:05:01,080 --> 00:05:03,760 Speaker 2: sentence to get a sort of perfected love story, if 93 00:05:03,800 --> 00:05:06,760 Speaker 2: you will, right. And then Jacob Weisberg called me in 94 00:05:06,839 --> 00:05:08,280 Speaker 2: January and said, hey, do you want to write a 95 00:05:08,320 --> 00:05:10,800 Speaker 2: novel using this stuff? And I said yes, please, let's 96 00:05:10,800 --> 00:05:12,880 Speaker 2: figure out what we can do. This project out of 97 00:05:12,880 --> 00:05:15,440 Speaker 2: an author is obviously on a completely different scale and 98 00:05:15,560 --> 00:05:19,200 Speaker 2: requires a completely different methodology than any other creative AI 99 00:05:19,279 --> 00:05:20,360 Speaker 2: that I've worked on before. 100 00:05:20,800 --> 00:05:23,000 Speaker 1: What's the scale of evolution you've seen? 101 00:05:23,600 --> 00:05:27,880 Speaker 2: Oh, I mean from zero to a million? Right from 102 00:05:28,279 --> 00:05:30,320 Speaker 2: sci fi q, which was a little handy thing that 103 00:05:30,360 --> 00:05:32,200 Speaker 2: I had on my own computer and was just for me. 104 00:05:32,640 --> 00:05:35,279 Speaker 2: That was really nice. Now you have Google's about to 105 00:05:35,279 --> 00:05:38,120 Speaker 2: put Palm two into every Google doc, so every Google 106 00:05:38,160 --> 00:05:41,840 Speaker 2: doc that you open you'll be able to continue any 107 00:05:41,920 --> 00:05:42,760 Speaker 2: text with AI. 108 00:05:43,000 --> 00:05:43,200 Speaker 3: Right. 109 00:05:43,400 --> 00:05:46,880 Speaker 2: And they're using PALM two, which is a hugely powerful 110 00:05:47,240 --> 00:05:50,120 Speaker 2: large language model. And when I interviewed the developers of 111 00:05:50,120 --> 00:05:53,320 Speaker 2: that in twenty twenty two, and it was already capable 112 00:05:53,320 --> 00:05:57,039 Speaker 2: of really crazy like low level chain reasoning and able 113 00:05:57,080 --> 00:05:59,680 Speaker 2: to translate into languages that it was not trained on, 114 00:06:00,000 --> 00:06:02,880 Speaker 2: which is just, I mean, really strange. And you know, 115 00:06:03,000 --> 00:06:04,800 Speaker 2: the thing that I think you should remember about these 116 00:06:04,880 --> 00:06:08,520 Speaker 2: leaps forward in AI which happen is that the scientists 117 00:06:08,600 --> 00:06:12,960 Speaker 2: who develop them don't understand them. They don't know why 118 00:06:13,000 --> 00:06:15,840 Speaker 2: they do that, they don't know why these results are 119 00:06:15,839 --> 00:06:17,520 Speaker 2: coming out. They just know that they do it. 120 00:06:17,720 --> 00:06:21,120 Speaker 1: This is an empirical rather than a theoretical evolution. 121 00:06:21,400 --> 00:06:24,159 Speaker 2: The transformer, of course is a theoretical thing, but it's 122 00:06:24,160 --> 00:06:27,760 Speaker 2: actually rather that because of the nature of the size 123 00:06:27,760 --> 00:06:30,480 Speaker 2: of the number of parameters. So like GPT three was 124 00:06:30,480 --> 00:06:33,240 Speaker 2: one hundred and seventy five billion parameters, open Aye is 125 00:06:33,279 --> 00:06:36,320 Speaker 2: not telling us how many parameters there are in GPT 126 00:06:36,520 --> 00:06:39,640 Speaker 2: four Palm the version that I saw was five hundred 127 00:06:39,640 --> 00:06:44,760 Speaker 2: and forty billion parameters. So that is essentially unfathomable by 128 00:06:44,800 --> 00:06:48,080 Speaker 2: a human being. You can't go and understand five hundred 129 00:06:48,080 --> 00:06:52,240 Speaker 2: and forty billion patterns. If you started today and tried 130 00:06:52,279 --> 00:06:54,560 Speaker 2: to do it four one hundred years, you couldn't understand 131 00:06:54,640 --> 00:06:59,320 Speaker 2: it right. So it's a question Rather than empirical or theoretical, 132 00:06:59,640 --> 00:07:02,800 Speaker 2: it is a kind of plunge into the unfathomable, and 133 00:07:02,839 --> 00:07:06,279 Speaker 2: that is different. This is a probabilistic technology. It's not 134 00:07:06,400 --> 00:07:09,679 Speaker 2: a truth based technology. When it generates something, it isn't 135 00:07:09,920 --> 00:07:13,160 Speaker 2: generating out of facts. It is generating strictly on the 136 00:07:13,160 --> 00:07:18,800 Speaker 2: basis of probabilities. It's really neither theoretical nor empirical. It's probabilistic. 137 00:07:19,040 --> 00:07:22,360 Speaker 1: The comedian Buck Henry once said, any technology you can't 138 00:07:22,400 --> 00:07:23,360 Speaker 1: explain is magic. 139 00:07:23,920 --> 00:07:28,760 Speaker 2: If magic is the encounter with powers that are unfathomable, 140 00:07:29,440 --> 00:07:33,360 Speaker 2: then this is literally magic because no one can fathom them. 141 00:07:33,560 --> 00:07:35,200 Speaker 2: The idea that they're going to regulate it and make 142 00:07:35,200 --> 00:07:38,040 Speaker 2: it transparent or whatever like. You use this technology to 143 00:07:38,200 --> 00:07:40,800 Speaker 2: do things that you can't do with your own brain, 144 00:07:41,200 --> 00:07:43,920 Speaker 2: and that's the source of its power. That's what it's 145 00:07:43,960 --> 00:07:46,280 Speaker 2: going to be used to do. And that of course 146 00:07:46,320 --> 00:07:50,440 Speaker 2: has huge implications on intellectual life and so on. I 147 00:07:50,520 --> 00:07:53,560 Speaker 2: wrote a piece for The Atlantic that was called Gods 148 00:07:53,560 --> 00:07:56,920 Speaker 2: and machines, right, Like, these are magical entities and the 149 00:07:57,000 --> 00:07:59,920 Speaker 2: strictly the technical sense of a power that you know 150 00:08:00,120 --> 00:08:02,760 Speaker 2: is powerful but you can't understand why. Right. I mean 151 00:08:02,760 --> 00:08:05,480 Speaker 2: that to me is when I've experienced magic in my life. 152 00:08:05,560 --> 00:08:09,920 Speaker 2: That's what I experience, right, And this technology is magic 153 00:08:10,000 --> 00:08:11,360 Speaker 2: in that technical sense. 154 00:08:11,960 --> 00:08:14,360 Speaker 1: I've been looking at a variety of impacts and it 155 00:08:14,440 --> 00:08:17,640 Speaker 1: strikes me that what you can do is regulate the 156 00:08:17,760 --> 00:08:21,440 Speaker 1: human but not regulate the system. If I use AI 157 00:08:21,600 --> 00:08:24,960 Speaker 1: to cheat you, then you can come after me for cheating, 158 00:08:25,040 --> 00:08:27,000 Speaker 1: but you can't possibly regulate the AI. 159 00:08:27,600 --> 00:08:31,880 Speaker 2: The key here is to regulate outcomes rather than processes, 160 00:08:32,160 --> 00:08:35,320 Speaker 2: because the processes are unfathomable, but the outcomes are not. 161 00:08:35,360 --> 00:08:38,000 Speaker 2: I'd also point out that like most of the harms 162 00:08:38,040 --> 00:08:41,280 Speaker 2: that people think about when I talk about AI are 163 00:08:41,760 --> 00:08:45,720 Speaker 2: really rooted in other things, right, Like, I mean, Facebook 164 00:08:45,760 --> 00:08:48,280 Speaker 2: had to take down two point two billion fake profiles 165 00:08:48,320 --> 00:08:50,920 Speaker 2: in twenty nineteen. That has nothing to do with generative AI. 166 00:08:51,200 --> 00:08:55,160 Speaker 2: That's cut and paste. That's enough to do massive quantities 167 00:08:55,160 --> 00:08:59,320 Speaker 2: of misinformation. We know that Instagram is associated with higher 168 00:08:59,360 --> 00:09:01,600 Speaker 2: levels of depress for teenage girls and has all kinds 169 00:09:01,600 --> 00:09:04,760 Speaker 2: of consequences on teenage girls. We know that, and also 170 00:09:05,040 --> 00:09:08,480 Speaker 2: that algorithm. You can regulate it. You absolutely can regulate it, 171 00:09:08,600 --> 00:09:13,760 Speaker 2: whereas AI is essentially unfathomable. The only way to regulate 172 00:09:13,800 --> 00:09:16,960 Speaker 2: it is essentially to stop it happening, right, And I 173 00:09:16,960 --> 00:09:19,679 Speaker 2: think that would be a huge wasted opportunity really, Like 174 00:09:19,720 --> 00:09:21,560 Speaker 2: it's a zero one kind of proposition. 175 00:09:21,760 --> 00:09:23,440 Speaker 1: Look, even if the United States wanted to do it, 176 00:09:23,440 --> 00:09:24,920 Speaker 1: there are lots of countries that are not going to 177 00:09:24,960 --> 00:09:28,320 Speaker 1: do it, So it's going to develop, It's going to happen. 178 00:09:28,880 --> 00:09:31,320 Speaker 2: The truth is, the United States is the least likely 179 00:09:31,360 --> 00:09:35,320 Speaker 2: to regulate Canada. We regulate everything like for Canada, like 180 00:09:35,480 --> 00:09:37,959 Speaker 2: Europe is going to regulate. Europe is definitely going to 181 00:09:38,040 --> 00:09:40,800 Speaker 2: come up with some regulation. They could quite possibly kill 182 00:09:40,800 --> 00:09:45,040 Speaker 2: the industry there. China, I don't think they're in this conversation, right, Like, 183 00:09:45,080 --> 00:09:49,040 Speaker 2: they use AI in very set ways. Also, we never 184 00:09:49,080 --> 00:09:51,760 Speaker 2: see what they do, so it's very impossible for us 185 00:09:51,800 --> 00:09:53,719 Speaker 2: to judge what they're actually producing. 186 00:09:54,000 --> 00:09:57,160 Speaker 1: I mean, isn't it actually a potentially huge threat to 187 00:09:57,160 --> 00:10:01,360 Speaker 1: any titulo jarian system because it enables you potentially to 188 00:10:01,400 --> 00:10:04,480 Speaker 1: have access to ways of thinking and researching and doing 189 00:10:05,200 --> 00:10:07,400 Speaker 1: that are virtually beyond control. 190 00:10:07,800 --> 00:10:10,480 Speaker 2: No, no, the opposite. It is the dream tool of 191 00:10:10,559 --> 00:10:11,840 Speaker 2: totalitarian governments. 192 00:10:12,520 --> 00:10:13,319 Speaker 1: Why is that. 193 00:10:13,880 --> 00:10:18,040 Speaker 2: It's capable of the most systematic thinking imaginable. If you 194 00:10:18,080 --> 00:10:21,120 Speaker 2: ask an AI to register every face in a city, 195 00:10:21,400 --> 00:10:24,720 Speaker 2: that's easy. I mean, we're there, that's done. If you 196 00:10:24,800 --> 00:10:29,439 Speaker 2: want AI to do elaborate and I mean the real 197 00:10:29,559 --> 00:10:32,720 Speaker 2: danger here is allowing AI to be involved in any 198 00:10:32,760 --> 00:10:35,920 Speaker 2: decision making at all. I am very optimistic about AI. 199 00:10:36,080 --> 00:10:39,000 Speaker 2: I think it's a hugely beneficial power. I mean, I 200 00:10:39,080 --> 00:10:41,880 Speaker 2: really don't like AI doomers, but you know, let's be clear, 201 00:10:42,000 --> 00:10:45,560 Speaker 2: Like you cannot have an AI make decisions because you 202 00:10:45,600 --> 00:10:49,079 Speaker 2: can't fire an AI. No AI is accountable to anyone. 203 00:10:49,400 --> 00:10:53,000 Speaker 2: That's been true since computation, since the invention of computation too, 204 00:10:53,160 --> 00:10:56,319 Speaker 2: Like you cannot have a computer make legal decisions because 205 00:10:56,320 --> 00:11:01,160 Speaker 2: there's no recourse. Right. It is absolutely the tool of 206 00:11:01,240 --> 00:11:05,120 Speaker 2: every authoritarian government. It can impose mass control and mass 207 00:11:05,160 --> 00:11:08,719 Speaker 2: surveillance on a level that would make Big Brother look 208 00:11:08,800 --> 00:11:10,200 Speaker 2: like nothing, I mean nothing. 209 00:11:10,520 --> 00:11:12,439 Speaker 1: So this is the automator Willian dream. 210 00:11:12,720 --> 00:11:16,240 Speaker 2: Well, it's a tool, and just like every other tool, 211 00:11:16,320 --> 00:11:19,720 Speaker 2: like film, is also the dream of authoritarian governments. Any 212 00:11:19,880 --> 00:11:23,440 Speaker 2: mass technology can be used in a different way. I 213 00:11:23,440 --> 00:11:27,280 Speaker 2: think people don't quite understand the probabilistic nature of AI. 214 00:11:27,800 --> 00:11:30,040 Speaker 2: They don't quite understand that we're not dealing here with 215 00:11:30,080 --> 00:11:32,439 Speaker 2: a truth engine at all. We're dealing here with a 216 00:11:32,520 --> 00:11:36,880 Speaker 2: language engine, and the distinction between that is actually extremely vital, 217 00:11:37,160 --> 00:11:41,360 Speaker 2: like extremely vital to understand. Absolutely, this could be a 218 00:11:41,400 --> 00:11:45,000 Speaker 2: tool for abuse as much as for anything else. But 219 00:11:45,200 --> 00:11:48,120 Speaker 2: I really believe that one needs to be regulated or 220 00:11:48,160 --> 00:11:51,680 Speaker 2: outcomes rather than processes, because I think if you try 221 00:11:51,679 --> 00:11:54,800 Speaker 2: and regulate processes, you're going to fail. Sort of the 222 00:11:54,800 --> 00:11:58,000 Speaker 2: AI domerism that I encounter. I think they want people 223 00:11:58,040 --> 00:11:59,880 Speaker 2: to be scared of it because there is no solution. 224 00:12:00,360 --> 00:12:04,679 Speaker 2: Whereas there is absolutely a solution to instagram algorithms one 225 00:12:04,760 --> 00:12:07,679 Speaker 2: hundred percent, there is absolutely a way to regulate that 226 00:12:07,720 --> 00:12:10,199 Speaker 2: in a meaningful way. The outcomes are totally understood. All 227 00:12:10,200 --> 00:12:12,480 Speaker 2: it takes is the political will. But they don't want 228 00:12:12,520 --> 00:12:16,560 Speaker 2: that because that could actually change technology. Whereas regulating AI, 229 00:12:17,120 --> 00:12:20,160 Speaker 2: I've never met anyone who can explain to me a 230 00:12:20,200 --> 00:12:24,280 Speaker 2: meaningful way to regulate it without destroying it. Because the 231 00:12:24,360 --> 00:12:26,920 Speaker 2: main claim in the EU was who are the leaders 232 00:12:26,920 --> 00:12:32,559 Speaker 2: in this? Was transparency? Well, is this inherently an obscure technology. 233 00:12:32,679 --> 00:12:36,240 Speaker 2: It's inherently an unfathomable technology. That's the nature of the transformer. 234 00:12:36,760 --> 00:12:40,360 Speaker 2: I just think numerism is really a mistaken approach to this. 235 00:12:40,640 --> 00:12:44,480 Speaker 2: I think going straight to the apocalypse and human extinction 236 00:12:44,600 --> 00:12:46,520 Speaker 2: events as a way of thinking through this is a 237 00:12:46,520 --> 00:12:49,800 Speaker 2: big mistake. But also I think, like, well, why not 238 00:12:49,920 --> 00:12:52,719 Speaker 2: start with regulating some really basic aspects of the new 239 00:12:52,720 --> 00:12:55,480 Speaker 2: era that we've entered, because we know about those and 240 00:12:55,600 --> 00:12:57,400 Speaker 2: we're really really very early here. 241 00:13:08,400 --> 00:13:10,559 Speaker 1: March the Majority will come out on June sixth and 242 00:13:10,679 --> 00:13:13,000 Speaker 1: is available for pre order now. But what struck me 243 00:13:13,200 --> 00:13:16,840 Speaker 1: when we were writing it is how relevant it is today. 244 00:13:16,840 --> 00:13:20,120 Speaker 1: It's really more like a cookbook or a roadmap for 245 00:13:20,200 --> 00:13:23,839 Speaker 1: how to both create a majority and how to use 246 00:13:23,880 --> 00:13:28,280 Speaker 1: that majority to get a democratic president to sign conservative reforms. 247 00:13:28,280 --> 00:13:31,199 Speaker 1: But Joe, what was your sense of its relevance to today? 248 00:13:31,360 --> 00:13:34,920 Speaker 3: Well, I think it's really important today because it provides 249 00:13:34,960 --> 00:13:38,320 Speaker 3: again our roadmap for how you get things done and 250 00:13:38,320 --> 00:13:41,720 Speaker 3: how you can get reforms through Congress, and how you 251 00:13:41,800 --> 00:13:45,080 Speaker 3: win elections, because you can't do anything if you don't 252 00:13:45,080 --> 00:13:48,440 Speaker 3: win elections. So I think it's important for those two reasons. 253 00:13:48,559 --> 00:13:51,040 Speaker 1: Then I felt that it really did provide a whole 254 00:13:51,080 --> 00:13:54,520 Speaker 1: set of principled ways for people to think, and that 255 00:13:54,600 --> 00:13:57,600 Speaker 1: if they do that, they're going to be dramatically more effective, 256 00:13:57,679 --> 00:14:00,480 Speaker 1: again in the Reagan tradition. So I just don't mind 257 00:14:00,480 --> 00:14:03,280 Speaker 1: everyone that March The Majority will come out on June sixth, 258 00:14:03,600 --> 00:14:12,880 Speaker 1: and it is available for pre order now. So you 259 00:14:13,040 --> 00:14:15,520 Speaker 1: decided to take a sort of different approach, and instead 260 00:14:15,559 --> 00:14:19,520 Speaker 1: of being afraid, you decided you would see whether or 261 00:14:19,560 --> 00:14:23,080 Speaker 1: not Ai could write a book which I think is fascinating, 262 00:14:23,280 --> 00:14:25,120 Speaker 1: which I assume is available for sale. 263 00:14:25,600 --> 00:14:28,040 Speaker 2: Oh it is. It's an audiobook and it's an ebook 264 00:14:28,080 --> 00:14:31,200 Speaker 2: on pushkin dot FM. We have Eduardo Ballerini reading it, 265 00:14:31,240 --> 00:14:33,400 Speaker 2: who The New Yorker called the Voice of God. We 266 00:14:33,520 --> 00:14:36,400 Speaker 2: decided the quality of the book was really important, because 267 00:14:36,440 --> 00:14:39,520 Speaker 2: of course, you can get Ai to write a garbage book, like, 268 00:14:39,600 --> 00:14:42,080 Speaker 2: that's no problem at all. The question is can you 269 00:14:42,280 --> 00:14:45,840 Speaker 2: use it to make something that people actually want to consume? Right, 270 00:14:46,160 --> 00:14:49,680 Speaker 2: I couldn't find a voice Ai that I thought people 271 00:14:49,680 --> 00:14:52,360 Speaker 2: would listen to for two hours, So we used Eduardo 272 00:14:52,400 --> 00:14:55,200 Speaker 2: Ballerini instead, who's like the best of the best, and 273 00:14:55,280 --> 00:14:56,920 Speaker 2: didn't worry about it. You know, if it's like if 274 00:14:56,960 --> 00:14:58,800 Speaker 2: Ai isn't there yet. We just won't use it, but 275 00:14:58,840 --> 00:15:03,120 Speaker 2: we will use the AI that we can, if that 276 00:15:03,160 --> 00:15:03,640 Speaker 2: makes sense. 277 00:15:03,960 --> 00:15:05,840 Speaker 1: You came out as an audiobook, but you didn't bring 278 00:15:05,880 --> 00:15:06,840 Speaker 1: it out in print. 279 00:15:07,080 --> 00:15:09,520 Speaker 2: No, because it was the speed The idea of this 280 00:15:09,600 --> 00:15:13,960 Speaker 2: started January twenty eighth this year. You're interviewing me about it, 281 00:15:14,000 --> 00:15:17,040 Speaker 2: and it's available for purchase in May. So I actually 282 00:15:17,080 --> 00:15:19,360 Speaker 2: don't know. I don't think any book has been released 283 00:15:19,360 --> 00:15:22,080 Speaker 2: that fast since like the eighteenth century, you know, and 284 00:15:22,120 --> 00:15:25,440 Speaker 2: like pamphlet culture, Like that's the last time someone put 285 00:15:25,440 --> 00:15:27,200 Speaker 2: it something impressed this fast? 286 00:15:27,400 --> 00:15:33,480 Speaker 1: Right, As somebody who writes books regularly, that's an amazing turnaround, right. 287 00:15:33,480 --> 00:15:36,000 Speaker 2: I mean you know, like it's only because Jacob has 288 00:15:36,000 --> 00:15:40,240 Speaker 2: complete control over Pushkin that that could even be conceived of. 289 00:15:40,600 --> 00:15:43,200 Speaker 2: There's no way it could be in print. You literally 290 00:15:43,200 --> 00:15:45,880 Speaker 2: just can't find the presses to do it that quickly. Right, 291 00:15:45,960 --> 00:15:48,080 Speaker 2: So it's an ebook and it's audio. 292 00:15:48,680 --> 00:15:50,840 Speaker 1: How did you conceptize this when somebody says to you, 293 00:15:51,200 --> 00:15:53,320 Speaker 1: why do you try it out as a novel? What 294 00:15:53,480 --> 00:15:54,600 Speaker 1: is your thought press? 295 00:15:55,080 --> 00:15:57,840 Speaker 2: I'd done these earlier experiments, so I knew some of 296 00:15:57,880 --> 00:16:01,000 Speaker 2: the limitations and weaknesses, and I'd use a bunch of 297 00:16:01,000 --> 00:16:04,160 Speaker 2: different tech use a bunch of different technology to write 298 00:16:04,200 --> 00:16:07,960 Speaker 2: things before. So I wasn't starting from nothing, but I knew, 299 00:16:08,000 --> 00:16:11,560 Speaker 2: for example, that getting any AI to write a plot 300 00:16:11,960 --> 00:16:15,240 Speaker 2: was very poor. They're just very lousy at plot. Like 301 00:16:15,400 --> 00:16:19,120 Speaker 2: what they are incredibly good at is really elaborate descriptions 302 00:16:19,240 --> 00:16:23,000 Speaker 2: and dialogue. They can do that stuff just an incredibly 303 00:16:23,000 --> 00:16:25,960 Speaker 2: good way. And what they're particularly good at is imitating 304 00:16:26,040 --> 00:16:29,080 Speaker 2: formulaic speech, like if you ask it to write something 305 00:16:29,160 --> 00:16:34,640 Speaker 2: in the style of a Lacinian professor of linguistics in 306 00:16:34,720 --> 00:16:38,040 Speaker 2: nineteen seventy three, Like, it can do that to a t, 307 00:16:38,440 --> 00:16:40,680 Speaker 2: I mean, better than any human writer, just better than 308 00:16:40,720 --> 00:16:44,760 Speaker 2: any human writer. So the question became we obviously we 309 00:16:44,800 --> 00:16:46,760 Speaker 2: wanted to make it a detective book. We wanted to 310 00:16:46,800 --> 00:16:48,760 Speaker 2: make it a thriller. We want the question here is 311 00:16:48,800 --> 00:16:52,400 Speaker 2: can you read it? Can it be fun to read? Right? 312 00:16:52,480 --> 00:16:52,520 Speaker 1: Like? 313 00:16:52,520 --> 00:16:54,840 Speaker 2: Because I thought the other experiences were a bit more lyrical, 314 00:16:55,240 --> 00:16:57,200 Speaker 2: and you know, I knew that the tech could do that, 315 00:16:57,240 --> 00:17:00,680 Speaker 2: but the question was like can it actually be gripping? Right? 316 00:17:01,080 --> 00:17:04,560 Speaker 2: And so originally I wanted to write it like Jim Thompson. 317 00:17:04,880 --> 00:17:07,480 Speaker 2: He's my favorite. I mean, they call him the Dimestore Dostoevsky, 318 00:17:07,600 --> 00:17:11,159 Speaker 2: and it's like completely accurate thing. And in his books, 319 00:17:11,160 --> 00:17:14,639 Speaker 2: like every sentence has like multiple levels of meaning, like 320 00:17:15,240 --> 00:17:17,040 Speaker 2: what the guy is saying is different from what he 321 00:17:17,080 --> 00:17:19,760 Speaker 2: means is also different from what the reader understands by it. 322 00:17:20,200 --> 00:17:23,840 Speaker 2: And the computers were just absolutely incapable of doing that. 323 00:17:24,040 --> 00:17:27,440 Speaker 2: Any AI that I worked on was just absolutely incapable 324 00:17:27,520 --> 00:17:30,840 Speaker 2: of that kind of ironic double meaning. So I was like, okay, 325 00:17:31,800 --> 00:17:34,480 Speaker 2: I was fooling around with a bunch of different language models, 326 00:17:34,480 --> 00:17:36,760 Speaker 2: and I was like, Okay, this isn't going to work. 327 00:17:36,880 --> 00:17:38,720 Speaker 2: What can we do here? So I thought I could 328 00:17:38,720 --> 00:17:41,639 Speaker 2: do something closer to Raymond Chandler right where it's like 329 00:17:42,240 --> 00:17:48,040 Speaker 2: really strong functional electric pros that moves, that keeps you moving. 330 00:17:48,680 --> 00:17:51,520 Speaker 2: And so basically what I did is I would go 331 00:17:51,600 --> 00:17:54,959 Speaker 2: to chat GPT and I would say write a scene, 332 00:17:55,040 --> 00:17:58,080 Speaker 2: and then I would give it the very close specifications 333 00:17:58,119 --> 00:18:00,000 Speaker 2: of what I wanted. And then I would say, can 334 00:18:00,080 --> 00:18:02,439 Speaker 2: taining the following information? And I would tell all the 335 00:18:02,480 --> 00:18:05,640 Speaker 2: information that I wanted. It would generate it, it would 336 00:18:05,720 --> 00:18:08,320 Speaker 2: generate a text. I would then take that into a 337 00:18:08,359 --> 00:18:12,000 Speaker 2: program called pseudo rite, which is a stochastic writing instrument 338 00:18:12,280 --> 00:18:16,280 Speaker 2: It's also GPT based, so you can select the text 339 00:18:16,520 --> 00:18:20,399 Speaker 2: and then you can add, you can shorten, you can rephrase. 340 00:18:20,720 --> 00:18:23,160 Speaker 2: It also has a customized button so you can put 341 00:18:23,359 --> 00:18:25,520 Speaker 2: you can put in the customized button, like make more active, 342 00:18:25,640 --> 00:18:27,479 Speaker 2: which I did with almost everything in the book. As 343 00:18:27,560 --> 00:18:30,200 Speaker 2: a writer, you know like you need to make everything 344 00:18:30,240 --> 00:18:32,639 Speaker 2: more active all the time. So make it more active, 345 00:18:32,760 --> 00:18:36,240 Speaker 2: make it more conversational, and then make it more like Hemingway, 346 00:18:36,640 --> 00:18:40,520 Speaker 2: or make it more like Chinese landscape poetry, or make 347 00:18:40,560 --> 00:18:44,960 Speaker 2: it more like whatever, and then whatever the occasion required. 348 00:18:45,320 --> 00:18:47,000 Speaker 2: And then that would give me a text that I 349 00:18:47,000 --> 00:18:48,919 Speaker 2: could work with, and then I would just cut and 350 00:18:48,920 --> 00:18:51,439 Speaker 2: paste that into the manuscript. And then I also wanted 351 00:18:51,440 --> 00:18:55,200 Speaker 2: good lines, because Chandler has lines like it was a blonde, 352 00:18:55,840 --> 00:18:57,640 Speaker 2: a blonde to make a bishop pick a hole through 353 00:18:57,680 --> 00:19:01,960 Speaker 2: a stained glass window. And I wanted really good lines 354 00:19:02,000 --> 00:19:04,760 Speaker 2: like that. So I used a program called cohere for that, 355 00:19:05,320 --> 00:19:08,720 Speaker 2: and that was a different process that involved creating prompts 356 00:19:08,760 --> 00:19:12,000 Speaker 2: and then training the prompts on examples, and then running 357 00:19:12,000 --> 00:19:15,159 Speaker 2: the prompts until I got images and metaphors that I wanted, 358 00:19:15,480 --> 00:19:18,080 Speaker 2: and a lot like a very conscious process. I would say, 359 00:19:18,160 --> 00:19:20,480 Speaker 2: like it wasn't like going to chat GPT and saying like, hey, 360 00:19:20,800 --> 00:19:23,400 Speaker 2: write me a novel. I was very much a creator here. 361 00:19:23,640 --> 00:19:26,000 Speaker 1: Could it master both Hemingway and Faulkner. 362 00:19:26,560 --> 00:19:31,560 Speaker 2: It's really good at Hemingway. The more involved the style, 363 00:19:31,680 --> 00:19:34,199 Speaker 2: the more clear the style is, the better it is 364 00:19:34,200 --> 00:19:37,280 Speaker 2: at imitating it. Coleridge was incredibly good at it because 365 00:19:37,320 --> 00:19:40,160 Speaker 2: Coleridge had this crazy and Xanda dude to Kubla Khan 366 00:19:40,240 --> 00:19:43,680 Speaker 2: a stately pleasure dome decree where Alph the Sacred River ran, 367 00:19:43,880 --> 00:19:47,760 Speaker 2: you know, like crazy vocabulary and stuff. That's what it's 368 00:19:47,840 --> 00:19:51,480 Speaker 2: best at. What it's weaker on is the more basic things. 369 00:19:52,440 --> 00:19:55,600 Speaker 2: Formach's paradox, it says that one of the features of AI, 370 00:19:56,119 --> 00:19:59,080 Speaker 2: the more complicated the task is, the better it is 371 00:19:59,640 --> 00:20:02,480 Speaker 2: at it. The simpler the task is, the worse it 372 00:20:02,560 --> 00:20:05,719 Speaker 2: is at it. So getting an AI to understand the 373 00:20:05,840 --> 00:20:09,400 Speaker 2: faces of every single person in a city relatively simple. 374 00:20:09,680 --> 00:20:12,640 Speaker 2: Getting it to understand cow like what a child does 375 00:20:12,680 --> 00:20:14,959 Speaker 2: when they see a cow and like recognize it as 376 00:20:14,960 --> 00:20:18,199 Speaker 2: a cow, incredibly difficult. And you know, I found that 377 00:20:18,280 --> 00:20:23,000 Speaker 2: applied absolutely to writing, Like if you ask it to 378 00:20:23,040 --> 00:20:28,480 Speaker 2: write something like James Jrace's Ulysses, huge complicated stylistic systems 379 00:20:29,000 --> 00:20:32,600 Speaker 2: that it's amazing at getting it to tell you an 380 00:20:32,640 --> 00:20:36,000 Speaker 2: interesting story that might be an interest to you. Almost impossible. 381 00:20:36,280 --> 00:20:38,600 Speaker 2: It was a question of running into the strengths and 382 00:20:38,640 --> 00:20:40,800 Speaker 2: away from the limitations, if you see what I mean. 383 00:20:41,200 --> 00:20:45,320 Speaker 1: So, did you have any notion when you started of 384 00:20:45,359 --> 00:20:46,639 Speaker 1: what the books should look like? 385 00:20:46,880 --> 00:20:49,040 Speaker 2: Oh? Yeah, I knew that the machines were not good 386 00:20:49,040 --> 00:20:52,159 Speaker 2: at the plot. I took copious notes about the plot 387 00:20:52,600 --> 00:20:55,639 Speaker 2: and the division. Now, as I started to create this 388 00:20:55,720 --> 00:20:59,760 Speaker 2: text with the machine, it would add details. The metaphors 389 00:20:59,760 --> 00:21:04,040 Speaker 2: are completely alien like, they're really foreign. Ultimately, I feel 390 00:21:04,040 --> 00:21:05,639 Speaker 2: like what I did here is get an alien to 391 00:21:05,680 --> 00:21:08,720 Speaker 2: write a short novel, an alien who happens to be 392 00:21:09,240 --> 00:21:11,800 Speaker 2: the sum total of our collective language. But there were 393 00:21:11,840 --> 00:21:14,760 Speaker 2: these moments where obviously I am in control of this. 394 00:21:14,760 --> 00:21:18,000 Speaker 2: This is a highly controlled aesthetic experiment. But there are 395 00:21:18,000 --> 00:21:21,840 Speaker 2: other moments when the machines got quite wild and that 396 00:21:22,000 --> 00:21:24,119 Speaker 2: you could feel that sort of I don't even know 397 00:21:24,119 --> 00:21:27,199 Speaker 2: how to describe it, like a burst of literary energy, 398 00:21:27,280 --> 00:21:29,800 Speaker 2: I guess. And then when I was able to get 399 00:21:29,840 --> 00:21:34,800 Speaker 2: the machines literary energy within my own control, those were 400 00:21:34,880 --> 00:21:37,879 Speaker 2: rare moments. I would say that maybe thirty or forty 401 00:21:37,920 --> 00:21:40,159 Speaker 2: moments during the writing of the book. But that really 402 00:21:40,200 --> 00:21:44,320 Speaker 2: felt powerful, like awesome in the old sense of the word, 403 00:21:44,600 --> 00:21:47,919 Speaker 2: like I was encountering something sublime, and that was a 404 00:21:47,920 --> 00:21:50,440 Speaker 2: different aesthetic experience, I think than anything I've had before 405 00:21:50,480 --> 00:21:52,560 Speaker 2: as a writer. There were creative moments that I've never 406 00:21:52,560 --> 00:21:53,160 Speaker 2: had before. 407 00:21:53,280 --> 00:21:56,320 Speaker 1: Does it take massive computing power to create the AI? 408 00:21:57,000 --> 00:22:00,640 Speaker 1: But you can use the application with a relatively simple computer. 409 00:22:01,040 --> 00:22:04,960 Speaker 2: I'm on an Mattbook Air. This requires a twenty dollars 410 00:22:05,040 --> 00:22:09,280 Speaker 2: a month subscription to GPT plus. I think I pay 411 00:22:09,320 --> 00:22:12,000 Speaker 2: pseudo rite like twelve dollars a month. I don't quote 412 00:22:12,040 --> 00:22:13,480 Speaker 2: me on that. I'm not sure how much I pay them. 413 00:22:13,600 --> 00:22:17,679 Speaker 2: These are like Netflix style subscriptions. We'll see what Google 414 00:22:17,720 --> 00:22:21,080 Speaker 2: does with bart on Google Docs. Google is an incredibly 415 00:22:21,280 --> 00:22:24,320 Speaker 2: powerful AI that has not really been accessed by the 416 00:22:24,320 --> 00:22:28,119 Speaker 2: public yet, and when it solves that public component, it 417 00:22:28,160 --> 00:22:29,720 Speaker 2: could just dwarf everyone else. 418 00:22:30,280 --> 00:22:33,800 Speaker 1: You end up with a remarkably inexpensive co author. 419 00:22:34,200 --> 00:22:37,040 Speaker 2: Absolutely, you end up with a twenty dollars a month writer. 420 00:22:37,440 --> 00:22:39,760 Speaker 1: How did you hook up with Malcolm Gladwell, who's been 421 00:22:39,800 --> 00:22:41,560 Speaker 1: astonishingly successful. 422 00:22:41,880 --> 00:22:44,720 Speaker 2: I mean, it was just Jacob I've never met Malcolmside. Well, 423 00:22:44,960 --> 00:22:47,080 Speaker 2: actually I did meet him once. It's a party here, 424 00:22:47,280 --> 00:22:49,880 Speaker 2: but I have no relationship with Malcolm Gladwell. We did 425 00:22:49,920 --> 00:22:52,719 Speaker 2: get the machine to write his blurbs for this book, 426 00:22:52,960 --> 00:22:54,879 Speaker 2: but I did have the machine, right. I think it 427 00:22:54,920 --> 00:22:56,280 Speaker 2: was three or four of them where it was like 428 00:22:56,520 --> 00:22:59,920 Speaker 2: one was effusive and one was if you're feeling obligated 429 00:23:00,040 --> 00:23:02,640 Speaker 2: than the other was like just medium, and he chose 430 00:23:02,640 --> 00:23:04,720 Speaker 2: the effusive, So you know, I feel like that's the 431 00:23:04,840 --> 00:23:05,760 Speaker 2: choice of some kind. 432 00:23:18,720 --> 00:23:20,960 Speaker 1: March the Majority will come out on June sixth and 433 00:23:21,040 --> 00:23:23,320 Speaker 1: is available for pre order now. But what struck me 434 00:23:23,560 --> 00:23:27,200 Speaker 1: when we were writing it is how relevant it is today. 435 00:23:27,200 --> 00:23:30,480 Speaker 1: It's really more like a cookbook or a roadmap for 436 00:23:30,560 --> 00:23:34,200 Speaker 1: how to both create a majority and how to use 437 00:23:34,200 --> 00:23:38,600 Speaker 1: that majority to get a democratic president to sign conservative reforms. 438 00:23:38,600 --> 00:23:41,560 Speaker 1: But Joe, what was your sense of its relevance to today? 439 00:23:41,680 --> 00:23:45,280 Speaker 3: Well, I think it's really important today because it provides 440 00:23:45,320 --> 00:23:48,639 Speaker 3: again our roadmap for how you get things done and 441 00:23:48,680 --> 00:23:52,080 Speaker 3: how you can get reforms through Congress, and how you 442 00:23:52,119 --> 00:23:55,400 Speaker 3: win elections, because you can't do anything if you don't 443 00:23:55,440 --> 00:23:58,800 Speaker 3: win elections. So I think it's important for those two reasons. 444 00:23:58,920 --> 00:24:01,359 Speaker 1: Then I felt that it really to provide a whole 445 00:24:01,400 --> 00:24:04,840 Speaker 1: set of principled ways for people to think, and that 446 00:24:04,920 --> 00:24:07,960 Speaker 1: if they do that, they're going to be dramatically more effective, 447 00:24:08,040 --> 00:24:10,520 Speaker 1: again in the Reagan tradition. So I just want to 448 00:24:10,520 --> 00:24:12,919 Speaker 1: mind everyone that March the Majority will come out on 449 00:24:13,040 --> 00:24:16,760 Speaker 1: June sixth, and it is available for pre order. Now, 450 00:24:22,240 --> 00:24:26,800 Speaker 1: you wrote the book under the pen name Aiden Marshene. Yeah, 451 00:24:26,840 --> 00:24:28,560 Speaker 1: why did you decide to use a pen name? 452 00:24:28,920 --> 00:24:33,560 Speaker 2: Well, it's an extremely complicated moment because, like, we really 453 00:24:33,600 --> 00:24:36,520 Speaker 2: don't have the words. One of the things that's ironic 454 00:24:36,600 --> 00:24:40,439 Speaker 2: about this incredible moment is that it's about language, it's 455 00:24:40,440 --> 00:24:42,480 Speaker 2: about the transformation of language, but we don't really have 456 00:24:42,560 --> 00:24:46,040 Speaker 2: the words to describe what we're dealing with here. So 457 00:24:46,800 --> 00:24:49,159 Speaker 2: you know, like, on the one hand, I am absolutely 458 00:24:49,200 --> 00:24:52,600 Speaker 2: the creator of this literary work, right, like one hundred percent. 459 00:24:52,600 --> 00:24:55,720 Speaker 2: It is in my control, So traditionally we would just 460 00:24:55,760 --> 00:24:58,800 Speaker 2: call that an author. On the other hand, I didn't 461 00:24:58,960 --> 00:25:02,560 Speaker 2: write ninety five percent of the words. So if you 462 00:25:02,680 --> 00:25:05,520 Speaker 2: don't write the words, can you call yourself the author 463 00:25:05,600 --> 00:25:09,040 Speaker 2: of a text? So essentially, Aidan Marchen was a way 464 00:25:09,160 --> 00:25:14,160 Speaker 2: of understanding or just recognizing the collaborative nature of it, right, 465 00:25:14,320 --> 00:25:16,800 Speaker 2: Like I'm just saying, like, this is me plus these 466 00:25:16,840 --> 00:25:20,560 Speaker 2: three technologies that have built it. And also there are 467 00:25:20,560 --> 00:25:23,000 Speaker 2: things in this book that I would not write. I 468 00:25:23,040 --> 00:25:25,960 Speaker 2: wanted to get those sorts of things right in there 469 00:25:26,000 --> 00:25:28,880 Speaker 2: because I think the experience of this book should be 470 00:25:29,600 --> 00:25:32,040 Speaker 2: what the machines can do more than just me. You 471 00:25:32,040 --> 00:25:34,640 Speaker 2: can go buy a dozen of my books, but this 472 00:25:34,680 --> 00:25:37,480 Speaker 2: is something a little different. The problem here is the 473 00:25:37,480 --> 00:25:41,800 Speaker 2: word author, Like does it fit for any of these conditions. 474 00:25:41,840 --> 00:25:45,560 Speaker 2: It's very hard. I remember when I was interviewing I 475 00:25:45,600 --> 00:25:47,639 Speaker 2: think it was the VP of Google Deep Mind. In 476 00:25:47,680 --> 00:25:50,639 Speaker 2: their promotional materials for Palm, they had said it was 477 00:25:50,680 --> 00:25:54,120 Speaker 2: capable of understanding, and I challenged him about this because 478 00:25:54,119 --> 00:25:56,520 Speaker 2: I was like, you know that this is just text prediction, 479 00:25:57,200 --> 00:25:59,560 Speaker 2: and that text prediction is not understanding the way that 480 00:25:59,640 --> 00:26:04,639 Speaker 2: we understand understanding. Like I know what Samuel Taylor Coleridge is. 481 00:26:04,960 --> 00:26:07,320 Speaker 2: Your machine does not know what Samuel Taylor Coleridge is. 482 00:26:07,359 --> 00:26:10,320 Speaker 2: And he said, look, you're right. But on the other hand, 483 00:26:10,640 --> 00:26:13,760 Speaker 2: when we say to this machine, write this in Bengali, 484 00:26:14,160 --> 00:26:17,280 Speaker 2: it knows what Bengali means. What word are we supposed 485 00:26:17,280 --> 00:26:19,720 Speaker 2: to use other than understanding, and I was like, you know, 486 00:26:19,760 --> 00:26:23,400 Speaker 2: you're probably right. We need two words for understanding, one 487 00:26:23,400 --> 00:26:27,560 Speaker 2: which is registering meaning and the other which is having 488 00:26:27,600 --> 00:26:30,479 Speaker 2: a grounding and meaning. But you know, we've never faced 489 00:26:30,480 --> 00:26:33,760 Speaker 2: a condition where those two things have been separate, and 490 00:26:33,800 --> 00:26:36,040 Speaker 2: now we are similar. And it's the same thing with 491 00:26:36,080 --> 00:26:39,000 Speaker 2: the authorship question. We don't actually have the language for 492 00:26:39,040 --> 00:26:41,680 Speaker 2: what this is yet. I mean, I'm thinking of calling 493 00:26:41,720 --> 00:26:44,240 Speaker 2: myself AI art director or something like that. 494 00:26:44,440 --> 00:26:48,800 Speaker 1: Are you under current rules? Are you able to copyright 495 00:26:48,920 --> 00:26:49,439 Speaker 1: your book? 496 00:26:49,640 --> 00:26:53,080 Speaker 2: It's actually a gray area. No one quite knows yet. 497 00:26:53,400 --> 00:26:56,879 Speaker 2: I mean, obviously I have five percent of the book 498 00:26:57,160 --> 00:26:59,280 Speaker 2: and I have the essay at the end, so there 499 00:26:59,320 --> 00:27:02,359 Speaker 2: is a copyright. Now are those aspects of it? There's 500 00:27:02,359 --> 00:27:04,040 Speaker 2: no question that I have the moral right to it. 501 00:27:04,080 --> 00:27:07,320 Speaker 2: I am the creator of it. But the actual copyright 502 00:27:07,400 --> 00:27:11,399 Speaker 2: law is actually incredibly behind the times with this. But 503 00:27:11,480 --> 00:27:13,080 Speaker 2: you know it's going to catch up in two years 504 00:27:13,080 --> 00:27:15,320 Speaker 2: and I'll probably get the copyright then, right, it doesn't 505 00:27:15,320 --> 00:27:17,920 Speaker 2: technically fit the definition of copyrightable work at the moment, 506 00:27:18,000 --> 00:27:19,760 Speaker 2: which is totally absurd. Really. 507 00:27:20,160 --> 00:27:23,520 Speaker 1: It's interesting to watch the Writers Guild with their strike 508 00:27:23,640 --> 00:27:27,800 Speaker 1: and their concerns since they represent people whose entire livelihood 509 00:27:27,840 --> 00:27:30,439 Speaker 1: is writing. One of their concerns has to be that 510 00:27:30,480 --> 00:27:33,879 Speaker 1: somebody's going to come along who can, in effect train 511 00:27:34,000 --> 00:27:36,800 Speaker 1: AI to write as though they were the screenplay writer 512 00:27:38,000 --> 00:27:40,520 Speaker 1: in a way that eliminates the screenplay litter. 513 00:27:40,920 --> 00:27:43,920 Speaker 2: Well, I got to say, from my experience of doing this, 514 00:27:44,720 --> 00:27:47,440 Speaker 2: good luck. If you think some executives going to sit 515 00:27:47,440 --> 00:27:50,560 Speaker 2: down at chat GPT and say write me John Wick 516 00:27:50,640 --> 00:27:54,000 Speaker 2: five and it's going to do it. No, I don't 517 00:27:54,240 --> 00:27:57,560 Speaker 2: see it happening anytime soon. I would put that the 518 00:27:57,640 --> 00:28:02,480 Speaker 2: replacement of screenwriters very low on my list of possible outcomes. 519 00:28:02,680 --> 00:28:08,480 Speaker 4: Copywriters, AD copywriters, Google AD copywriters, low level legal clerks, 520 00:28:08,960 --> 00:28:12,960 Speaker 4: even people who write ker notices huge challenge. 521 00:28:13,240 --> 00:28:16,359 Speaker 2: But the more formulaic the writing, the easier it is 522 00:28:16,600 --> 00:28:19,080 Speaker 2: for a machine to imitate. That's just a key thing 523 00:28:19,119 --> 00:28:22,200 Speaker 2: to understand. And if you want any kind of originality 524 00:28:22,240 --> 00:28:25,000 Speaker 2: at all, I'm not sure that Hollywood does. You're going 525 00:28:25,080 --> 00:28:26,280 Speaker 2: to need a person for that. 526 00:28:26,720 --> 00:28:30,880 Speaker 1: You wrote an Atlantic article in December called the college 527 00:28:31,080 --> 00:28:34,680 Speaker 1: essay is Dead. As a former college teacher, I found 528 00:28:34,680 --> 00:28:37,600 Speaker 1: that interesting. Do you think this is a function of AI. 529 00:28:37,440 --> 00:28:42,480 Speaker 2: Or what the formulaic writing is? That's what AI does 530 00:28:42,720 --> 00:28:47,600 Speaker 2: to a SUPERB degree that is basically unmatchable by human beings. 531 00:28:48,080 --> 00:28:50,840 Speaker 2: I'm a professional writer. From the time I've been about 532 00:28:50,880 --> 00:28:54,520 Speaker 2: ten years old, I've been trained in a very specific 533 00:28:54,600 --> 00:28:58,360 Speaker 2: skill of how to write an eight hundred word essay 534 00:28:58,640 --> 00:29:01,760 Speaker 2: where you have three sentences that bring you into the issue, 535 00:29:01,840 --> 00:29:05,880 Speaker 2: leading to a thesis statement followed by three arguments, followed 536 00:29:05,880 --> 00:29:09,280 Speaker 2: by a conclusive statement leader. And I learned that skill 537 00:29:09,360 --> 00:29:12,120 Speaker 2: at great cost, and it took me a long time 538 00:29:12,160 --> 00:29:14,840 Speaker 2: to master it. A computer has that like that. Now, 539 00:29:15,240 --> 00:29:19,120 Speaker 2: if you say, write a fifteen hundred word undergraduate essay 540 00:29:19,240 --> 00:29:22,200 Speaker 2: on this subject, it will just do something that is 541 00:29:22,560 --> 00:29:25,360 Speaker 2: at least a B plus right, and with GPT four 542 00:29:25,400 --> 00:29:27,160 Speaker 2: are probably an A right. 543 00:29:27,760 --> 00:29:31,520 Speaker 1: So AI may do to brain work what the invention 544 00:29:31,680 --> 00:29:34,920 Speaker 1: of the internal combustion engine did to physical labor, and 545 00:29:34,960 --> 00:29:36,600 Speaker 1: then you suddenly have a replacement. 546 00:29:37,160 --> 00:29:41,240 Speaker 2: Here's what I think. It's going to replace a lot 547 00:29:41,280 --> 00:29:45,560 Speaker 2: of boring formulaic writing things which are a lot of 548 00:29:45,600 --> 00:29:48,440 Speaker 2: our creative life. But it will not replace creative life. 549 00:29:48,600 --> 00:29:52,080 Speaker 2: The value of creativity will actually, in my opinion, increase 550 00:29:52,520 --> 00:29:57,280 Speaker 2: because the mere skill at composition, mere skill at elegant 551 00:29:57,480 --> 00:30:00,280 Speaker 2: hitting the marks of a formula is about to be 552 00:30:00,280 --> 00:30:02,680 Speaker 2: devalued enormously because you can just do it with a 553 00:30:02,720 --> 00:30:07,280 Speaker 2: twenty dollars a month subscription. What will matter is being 554 00:30:07,320 --> 00:30:10,440 Speaker 2: able to express your intention, and not only that, but 555 00:30:10,480 --> 00:30:12,960 Speaker 2: being able to people to feel your intention and your 556 00:30:13,000 --> 00:30:14,480 Speaker 2: will as a writer. 557 00:30:14,560 --> 00:30:17,640 Speaker 1: Well, and being able to have an intention. I always 558 00:30:17,680 --> 00:30:19,800 Speaker 1: tell people the writing isn't hard. 559 00:30:19,840 --> 00:30:23,400 Speaker 2: It's the thinking, exactly, it's the willing. One thing that 560 00:30:23,440 --> 00:30:25,840 Speaker 2: these machines don't do at all, and it's really important 561 00:30:25,840 --> 00:30:29,280 Speaker 2: to remember, is they don't ask questions and they don't 562 00:30:29,480 --> 00:30:34,000 Speaker 2: recognize meaning. Like this novel that I wrote. The reason 563 00:30:34,040 --> 00:30:37,880 Speaker 2: that it's better than anyone else's AI novel is that 564 00:30:38,000 --> 00:30:40,720 Speaker 2: I know what a paragraph looks like. I know what 565 00:30:40,760 --> 00:30:42,600 Speaker 2: a good paragraph looks like. I know what a good 566 00:30:42,640 --> 00:30:44,959 Speaker 2: sentence looks like. I know what good dialogue looks like. 567 00:30:45,200 --> 00:30:47,400 Speaker 2: So when I see it, I know it, and I 568 00:30:47,400 --> 00:30:50,200 Speaker 2: can cut it and move it into a document, right. 569 00:30:50,400 --> 00:30:53,240 Speaker 2: And it does not do recognition, and it does not 570 00:30:53,320 --> 00:30:58,040 Speaker 2: do questioning. It's a tool, I think, much like design. 571 00:30:58,680 --> 00:31:01,840 Speaker 2: When Photoshop arrives, like it seemed like it was the 572 00:31:01,920 --> 00:31:04,440 Speaker 2: end of design, right. It seemed like, oh my god, 573 00:31:04,480 --> 00:31:06,360 Speaker 2: we're not going to need any designers anymore. Why do 574 00:31:06,440 --> 00:31:09,160 Speaker 2: we need people who know how to manipulate photographs when 575 00:31:09,160 --> 00:31:11,240 Speaker 2: you just go on your computer and turn it around 576 00:31:11,240 --> 00:31:14,320 Speaker 2: and no, the opposite happens. Suddenly you need a lot 577 00:31:14,400 --> 00:31:16,760 Speaker 2: of designers, and you need a lot of designers who 578 00:31:16,800 --> 00:31:19,760 Speaker 2: know a lot about how this technology works. Suddenly we 579 00:31:19,800 --> 00:31:22,360 Speaker 2: also live in a world just crammed with design, where 580 00:31:22,360 --> 00:31:26,120 Speaker 2: everything has to be designed. From these headphones, this bottle, 581 00:31:26,520 --> 00:31:29,719 Speaker 2: everything must be well designed. And I think something similar 582 00:31:29,720 --> 00:31:31,880 Speaker 2: could happen with this technology. I don't think it's going 583 00:31:31,960 --> 00:31:35,760 Speaker 2: to replace will at all. It's just going to make 584 00:31:35,800 --> 00:31:37,800 Speaker 2: certain tasks easier to do. 585 00:31:37,800 --> 00:31:40,720 Speaker 1: Do you think your next book will be AI enhanced? 586 00:31:40,880 --> 00:31:43,240 Speaker 2: We did this as an experiment, but the truth is 587 00:31:43,280 --> 00:31:45,240 Speaker 2: the next time, I'm going to use it for what 588 00:31:45,840 --> 00:31:49,120 Speaker 2: it's good at only and do everything else myself. And 589 00:31:49,160 --> 00:31:51,920 Speaker 2: I'm not going to worry about what the percentages of 590 00:31:52,240 --> 00:31:55,600 Speaker 2: are of what, or the percentages are of that or whatever. 591 00:31:55,920 --> 00:31:57,720 Speaker 2: And I'm also going to use it for things that 592 00:31:57,960 --> 00:32:00,560 Speaker 2: only it can do, which is really where we're going 593 00:32:00,600 --> 00:32:02,240 Speaker 2: to get to the really interesting stuff. 594 00:32:02,280 --> 00:32:04,840 Speaker 1: I think that's well. I hope to have a chance 595 00:32:04,920 --> 00:32:07,959 Speaker 1: to pursue this again as the system involves and as 596 00:32:08,000 --> 00:32:10,920 Speaker 1: your own career evolves. This has been really fun. I 597 00:32:10,960 --> 00:32:14,040 Speaker 1: hope that you will give my best to Aiden Marshen, 598 00:32:14,560 --> 00:32:17,640 Speaker 1: and I want to congratulate you both on your new novella, 599 00:32:18,000 --> 00:32:20,479 Speaker 1: Death of an Author, which is now available as an 600 00:32:20,480 --> 00:32:23,280 Speaker 1: audiobook and an ebook, and I encourage our listeners to 601 00:32:23,360 --> 00:32:26,600 Speaker 1: check it out. This is a truly groundbreaking moment and 602 00:32:26,760 --> 00:32:29,040 Speaker 1: something you've done as a pioneer that's says fabulous. 603 00:32:29,280 --> 00:32:30,040 Speaker 2: Thank you very much. 604 00:32:33,760 --> 00:32:35,960 Speaker 1: Thank you to my guest Stephen Marsh. You can get 605 00:32:35,960 --> 00:32:37,800 Speaker 1: a link to order his new book, Death of an 606 00:32:37,840 --> 00:32:41,480 Speaker 1: Author on our show page at newtsworld dot com. Newsworld 607 00:32:41,520 --> 00:32:44,960 Speaker 1: is produced by Ginglish three sixty and iHeartMedia. Our executive 608 00:32:44,960 --> 00:32:48,480 Speaker 1: producer is Guernsey Sloan and our researcher is Rachel Peterson. 609 00:32:48,880 --> 00:32:51,720 Speaker 1: The artwork for the show was created by Steve Penley. 610 00:32:52,160 --> 00:32:54,960 Speaker 1: Special thanks to the team at Gingriish three sixty. If 611 00:32:54,960 --> 00:32:57,480 Speaker 1: you've been enjoying Newtsworld, I hope you'll go to Apple 612 00:32:57,520 --> 00:33:00,760 Speaker 1: Podcast and both rate us with five stars and give 613 00:33:00,840 --> 00:33:03,440 Speaker 1: us a review so others can learn what it's all about. 614 00:33:03,960 --> 00:33:07,040 Speaker 1: Right now, listeners of Newtsworld consign up for my three 615 00:33:07,320 --> 00:33:11,640 Speaker 1: freeweekly columns at gingrichwie sixty dot com slash newsletter, I'm 616 00:33:11,720 --> 00:33:13,720 Speaker 1: nute Gingrich. This is newtsworld.