1 00:00:05,559 --> 00:00:11,240 Speaker 1: Will writers and artists and musicians become unemployed by AI? 2 00:00:11,920 --> 00:00:15,280 Speaker 1: What are the new capabilities that we're seeing all around us, 3 00:00:15,360 --> 00:00:16,120 Speaker 1: and what is this. 4 00:00:16,120 --> 00:00:18,840 Speaker 2: Going to mean for human creativity? 5 00:00:19,200 --> 00:00:22,000 Speaker 1: And what does this have to do with diamonds and 6 00:00:22,079 --> 00:00:27,120 Speaker 1: Westworld and effort and Frankenstein in Beethoven and the Stark 7 00:00:27,200 --> 00:00:33,000 Speaker 1: Family and Game of Thrones. Welcome to Inner Cosmos with 8 00:00:33,080 --> 00:00:37,639 Speaker 1: me David Eagleman. I'm a neuroscientist and an author at 9 00:00:37,680 --> 00:00:41,480 Speaker 1: Stanford University, and in this episode, I get to dive 10 00:00:41,520 --> 00:00:46,520 Speaker 1: into something that's right at the intersection of science and creativity. 11 00:00:51,479 --> 00:00:55,720 Speaker 1: Most of my podcasts are about evergreen topics about our 12 00:00:55,760 --> 00:00:59,920 Speaker 1: brains and our psychology, but there's something so extraordinary happy 13 00:01:00,440 --> 00:01:01,080 Speaker 1: right now. 14 00:01:01,560 --> 00:01:05,600 Speaker 2: We're in the middle of a revolution with AI, and 15 00:01:05,640 --> 00:01:09,640 Speaker 2: what's called generative AI in particular. So I'm going to 16 00:01:09,720 --> 00:01:12,960 Speaker 2: do a two part episode on this. For today, I'm 17 00:01:12,959 --> 00:01:15,880 Speaker 2: going to dig into what generative AI is and what 18 00:01:16,000 --> 00:01:20,080 Speaker 2: it means for human creativity, and then in the next episode, 19 00:01:20,120 --> 00:01:24,360 Speaker 2: I'm going to tackle the question of sentience. Are these 20 00:01:24,400 --> 00:01:28,720 Speaker 2: ais conscious and if not, now, could they be soon? 21 00:01:29,400 --> 00:01:31,760 Speaker 2: And how would we know when we get there? 22 00:01:35,000 --> 00:01:39,240 Speaker 1: So let's start in twenty seventeen when almost no one 23 00:01:39,280 --> 00:01:42,440 Speaker 1: in the world paid attention when a team at Google 24 00:01:42,520 --> 00:01:47,280 Speaker 1: Brain introduced a new way of building an artificial neural network. 25 00:01:47,920 --> 00:01:51,040 Speaker 1: So this was different than the architectures that came before it, 26 00:01:51,480 --> 00:01:55,560 Speaker 1: which were called things like convolutional neural networks and recurrent 27 00:01:55,640 --> 00:01:58,800 Speaker 1: neural networks. Instead, they presented a new model that was 28 00:01:58,840 --> 00:02:03,520 Speaker 1: called a transformer. Now, transformer is not one of those 29 00:02:03,640 --> 00:02:06,760 Speaker 1: robots that shapeshift into trucks and helicopters. 30 00:02:07,280 --> 00:02:10,040 Speaker 2: Instead, a transformer model is. 31 00:02:10,000 --> 00:02:14,160 Speaker 1: A way to tackle sequential data like the words that 32 00:02:14,200 --> 00:02:16,280 Speaker 1: are in a sentence or the frames in a video. 33 00:02:16,800 --> 00:02:20,080 Speaker 1: And a transformer model takes in everything at once, and 34 00:02:20,120 --> 00:02:23,239 Speaker 1: it essentially pays attention to different parts of the data. 35 00:02:23,800 --> 00:02:29,400 Speaker 1: And this allows training on enormous data sets, bigger than 36 00:02:29,440 --> 00:02:33,560 Speaker 1: what was trained on before. Like now it's essentially everything 37 00:02:33,680 --> 00:02:36,960 Speaker 1: that has been written by humans that is on the Internet, 38 00:02:37,160 --> 00:02:41,840 Speaker 1: which is petabytes of data. So these models they digest 39 00:02:41,919 --> 00:02:44,840 Speaker 1: all of that and what do they do. They essentially 40 00:02:44,919 --> 00:02:48,200 Speaker 1: look at a sequence of inputs like the words and 41 00:02:48,240 --> 00:02:52,280 Speaker 1: a sentence, and they ask what word is most likely 42 00:02:52,360 --> 00:02:56,440 Speaker 1: to come next in that sequence. Now we'll come back 43 00:02:56,480 --> 00:02:58,000 Speaker 1: to that in a second, but I just want to 44 00:02:58,040 --> 00:03:04,600 Speaker 1: note that this transformer model is finding uses way beyond text. So, 45 00:03:04,680 --> 00:03:08,160 Speaker 1: for example, a recent Nature paper used this kind of 46 00:03:08,200 --> 00:03:11,720 Speaker 1: model to look at amino acids, which run in a 47 00:03:11,720 --> 00:03:14,800 Speaker 1: sequence to make proteins, and they looked at these chains 48 00:03:14,800 --> 00:03:18,200 Speaker 1: of amino acids like techt strings, and they set a 49 00:03:18,360 --> 00:03:21,840 Speaker 1: major new water mark in determining how proteins fold, which 50 00:03:21,880 --> 00:03:25,400 Speaker 1: is a very difficult problem. And people are using transformers 51 00:03:25,400 --> 00:03:30,800 Speaker 1: for everything from making music to reading giant reams of 52 00:03:30,840 --> 00:03:35,040 Speaker 1: medical records and so on. These transformer models are built 53 00:03:35,080 --> 00:03:37,440 Speaker 1: into search already, and soon they're going to be in 54 00:03:37,520 --> 00:03:40,120 Speaker 1: your phone and in your car, and in your bank 55 00:03:40,160 --> 00:03:48,080 Speaker 1: and in your doctor's office. So what everyone in Silicon 56 00:03:48,160 --> 00:03:50,800 Speaker 1: Valley is talking about is how this new kind of 57 00:03:50,880 --> 00:03:55,120 Speaker 1: AI is going to disrupt the workforce. And a lot 58 00:03:55,160 --> 00:03:58,320 Speaker 1: of people are thinking about white collar jobs that have 59 00:03:58,440 --> 00:04:04,200 Speaker 1: traditionally required memorization of long textbooks, and these jobs, whether 60 00:04:04,280 --> 00:04:08,440 Speaker 1: they're legal or medical, suddenly seem to be kind of outmoded. 61 00:04:09,000 --> 00:04:11,880 Speaker 1: And so we're all thinking about what this means for 62 00:04:11,960 --> 00:04:15,120 Speaker 1: the economy because so many jobs are going to be 63 00:04:15,160 --> 00:04:19,920 Speaker 1: displaced by this new technology. Now, there's nothing totally new 64 00:04:19,960 --> 00:04:24,039 Speaker 1: about this kind of worry, because every generation sees new 65 00:04:24,080 --> 00:04:28,240 Speaker 1: technologies take over old jobs. That's natural, and we don't 66 00:04:28,320 --> 00:04:32,840 Speaker 1: lament the fact that we don't have elevator operators anymore, 67 00:04:33,120 --> 00:04:38,000 Speaker 1: or switchboard operators at telephone companies, or factories that make 68 00:04:38,240 --> 00:04:43,800 Speaker 1: VCRs or eight track tape players, because new technologies continuously 69 00:04:43,880 --> 00:04:48,560 Speaker 1: replace the old, and industries change and people adapt. But 70 00:04:48,680 --> 00:04:52,120 Speaker 1: the concern that we're seeing with the AI revolution is 71 00:04:52,200 --> 00:04:55,559 Speaker 1: the speed of it. It's probably the case that we've 72 00:04:56,040 --> 00:05:00,479 Speaker 1: never before had a move forward in technology that's so 73 00:05:00,960 --> 00:05:07,080 Speaker 1: unbelievably rapid. So this is why everyone's talking about this 74 00:05:07,240 --> 00:05:09,919 Speaker 1: with a different point of view than we did with 75 00:05:10,080 --> 00:05:13,080 Speaker 1: previous innovations. But I want to zoom in on something 76 00:05:13,120 --> 00:05:15,360 Speaker 1: a little different for this episode. I want to know 77 00:05:15,400 --> 00:05:19,719 Speaker 1: what this all means for human creativity, because the thing 78 00:05:19,800 --> 00:05:23,080 Speaker 1: to note is these models have been trained up not 79 00:05:23,240 --> 00:05:26,960 Speaker 1: just on the handful of novels and conversations and schoolwork 80 00:05:27,040 --> 00:05:31,599 Speaker 1: that you have experienced on your thin trajectory through space 81 00:05:31,640 --> 00:05:34,880 Speaker 1: and time, but they have been trained with everything that's 82 00:05:34,960 --> 00:05:40,479 Speaker 1: ever been written by humans. Every textbook, every article, every poem, 83 00:05:40,520 --> 00:05:46,800 Speaker 1: every blog post, every novel. We're talking seventy one billion 84 00:05:46,880 --> 00:05:52,719 Speaker 1: web pages and hundreds of trillions of words, It's something 85 00:05:52,760 --> 00:05:57,680 Speaker 1: that's so far beyond any human's capacity to consume even 86 00:05:57,720 --> 00:06:01,360 Speaker 1: a fraction of it, or to really imagine a corpus 87 00:06:01,480 --> 00:06:04,839 Speaker 1: of text that large. Oh and by the way, it 88 00:06:04,880 --> 00:06:08,400 Speaker 1: has a perfect memory for every word that it's read. 89 00:06:08,520 --> 00:06:12,080 Speaker 1: So now you're talking about a system that's not the 90 00:06:12,120 --> 00:06:17,560 Speaker 1: same as a brain, but is incredibly powerful at generating 91 00:06:17,720 --> 00:06:22,159 Speaker 1: text or visual art or music and soon video. And 92 00:06:22,200 --> 00:06:25,000 Speaker 1: so while we'll talk about sentience next week, this week, 93 00:06:25,080 --> 00:06:28,080 Speaker 1: I want to address a social point that has quickly 94 00:06:28,160 --> 00:06:30,839 Speaker 1: risen to the surface, which is what will all this 95 00:06:31,040 --> 00:06:36,599 Speaker 1: mean for human art and human creativity? Personally, I'm working 96 00:06:36,640 --> 00:06:39,800 Speaker 1: on my next several books right now, and these are 97 00:06:39,839 --> 00:06:44,720 Speaker 1: all projects that have spanned years, and so I'm fascinated 98 00:06:44,839 --> 00:06:49,560 Speaker 1: and terrified about whether AI is going to replace me 99 00:06:49,640 --> 00:06:51,960 Speaker 1: as a writer. What does this kind of new AI 100 00:06:52,600 --> 00:06:57,640 Speaker 1: mean for writers, for visual artists, for musicians who studied 101 00:06:57,640 --> 00:07:00,680 Speaker 1: their whole lives to be able to compose beautiful piece 102 00:07:00,720 --> 00:07:06,200 Speaker 1: of music? Is human creativity destined for the dust bin 103 00:07:06,440 --> 00:07:11,160 Speaker 1: of history? So let's start with the downside of these models. 104 00:07:11,680 --> 00:07:14,640 Speaker 1: So in my book Live Wired, I talked about how 105 00:07:14,720 --> 00:07:20,320 Speaker 1: AI algorithms don't care about relevance they memorize whatever we 106 00:07:20,560 --> 00:07:23,080 Speaker 1: ask them to. So, now this is a very useful 107 00:07:23,120 --> 00:07:26,600 Speaker 1: feature of AI, but it's also the reason AI is 108 00:07:26,720 --> 00:07:31,640 Speaker 1: not particularly human like, because AI models don't have any 109 00:07:31,680 --> 00:07:35,400 Speaker 1: sort of internal model of the world. They have no 110 00:07:35,520 --> 00:07:38,880 Speaker 1: idea what it is to be a human and have 111 00:07:39,040 --> 00:07:44,080 Speaker 1: drives and concerns. They don't care which problems are interesting 112 00:07:44,600 --> 00:07:48,640 Speaker 1: or germane. Instead, they memorize whatever we feed them. So 113 00:07:48,720 --> 00:07:51,960 Speaker 1: whether that's distinguishing a horse from a zebra in a 114 00:07:52,000 --> 00:07:56,400 Speaker 1: billion photographs, or tracking flight data from every airport on 115 00:07:56,440 --> 00:08:01,160 Speaker 1: the planet, or composing music in the style of Brian Eno, 116 00:08:01,640 --> 00:08:06,360 Speaker 1: they have no sense of importance except in a statistical sense, 117 00:08:06,960 --> 00:08:09,679 Speaker 1: which is to say, which signals occur more often. 118 00:08:10,440 --> 00:08:13,160 Speaker 2: So contemporary AI could never. 119 00:08:13,080 --> 00:08:17,560 Speaker 1: By itself decide that it finds irresistible a particular kind 120 00:08:17,600 --> 00:08:21,600 Speaker 1: of ice cream, or that it abhors a particular kind 121 00:08:21,600 --> 00:08:26,120 Speaker 1: of music, or that it's heartbroken by King Lear's speech 122 00:08:26,320 --> 00:08:29,760 Speaker 1: over his dead daughter. So AI can dispatch, you know, 123 00:08:29,840 --> 00:08:34,000 Speaker 1: ten thousand hours of intense practice in ten thousand nanoseconds, 124 00:08:34,360 --> 00:08:38,200 Speaker 1: but it doesn't care about any zeros and ones over 125 00:08:38,240 --> 00:08:43,680 Speaker 1: any others. As a result, AI can accomplish incredibly impressive feats, 126 00:08:43,760 --> 00:08:48,760 Speaker 1: but not the feat of being quite like a human. 127 00:08:49,200 --> 00:08:52,320 Speaker 1: And so some critics of AI say, look, it's like 128 00:08:52,360 --> 00:08:55,760 Speaker 1: you want a sandwich, and what this transformer model does 129 00:08:56,200 --> 00:08:58,720 Speaker 1: is it looks at all the billions of sandwiches out 130 00:08:58,760 --> 00:09:02,079 Speaker 1: there in the world, and it gives you a slurry 131 00:09:02,480 --> 00:09:03,880 Speaker 1: and it pours it out in. 132 00:09:03,840 --> 00:09:05,240 Speaker 2: The shape of a sandwich. 133 00:09:05,559 --> 00:09:08,079 Speaker 1: A fellow writer gave me that analogy the other day, 134 00:09:08,120 --> 00:09:12,839 Speaker 1: and that doesn't sound particularly appealing, right, And yet these 135 00:09:13,000 --> 00:09:16,719 Speaker 1: ais have massively surprised us. 136 00:09:17,080 --> 00:09:20,280 Speaker 2: The text generation is so good, it's. 137 00:09:20,120 --> 00:09:24,319 Speaker 1: So complete, it's so human like that we find ourselves 138 00:09:24,360 --> 00:09:27,960 Speaker 1: not so much in the phase of invention like with 139 00:09:28,040 --> 00:09:31,239 Speaker 1: all the machines we've made before. Instead, the whole scientific 140 00:09:31,240 --> 00:09:36,760 Speaker 1: community is finding itself in a process of discovery. Everyone 141 00:09:36,880 --> 00:09:41,880 Speaker 1: is exploring to find out what these enormous models are 142 00:09:41,960 --> 00:09:46,240 Speaker 1: capable of, because nobody quite knows. They keep blowing our 143 00:09:46,280 --> 00:09:50,200 Speaker 1: minds with things they're able to do which weren't pre 144 00:09:50,240 --> 00:09:56,280 Speaker 1: programmed and not even foreseen. Have a friend who works 145 00:09:56,280 --> 00:09:59,800 Speaker 1: with a big city symphony, and she's trying to play 146 00:10:00,160 --> 00:10:03,520 Speaker 1: a program for the symphony several months out, which is 147 00:10:03,559 --> 00:10:07,760 Speaker 1: a typical timescale for symphony planning, but she's scheduling to 148 00:10:07,800 --> 00:10:11,520 Speaker 1: put on a program with music composed by AI, and 149 00:10:11,600 --> 00:10:14,240 Speaker 1: she's at a loss for how to plan this because 150 00:10:14,720 --> 00:10:18,360 Speaker 1: she's well aware that things are moving so fast that 151 00:10:18,400 --> 00:10:22,479 Speaker 1: the musical world and the skill level of AI composition 152 00:10:23,000 --> 00:10:25,320 Speaker 1: is going to be entirely different. In a few months, 153 00:10:25,320 --> 00:10:28,240 Speaker 1: it's can be more advanced. So she was telling me 154 00:10:28,320 --> 00:10:31,880 Speaker 1: that she doesn't quite know how to nail down plans 155 00:10:31,920 --> 00:10:36,600 Speaker 1: for this, because unlike every symphony planner who has come before, 156 00:10:36,760 --> 00:10:39,600 Speaker 1: she's now in a world where if she nails down 157 00:10:39,679 --> 00:10:43,080 Speaker 1: a choice of music and trains up the musicians, it 158 00:10:43,240 --> 00:10:47,360 Speaker 1: is guaranteed to be badly outdated some months from now. 159 00:10:47,760 --> 00:10:51,160 Speaker 1: And this is the world we're operating in now. So jennertive, 160 00:10:51,160 --> 00:10:54,920 Speaker 1: AI is moving so rapidly that we have entered this 161 00:10:55,120 --> 00:10:58,959 Speaker 1: massive revolution without most of us realizing that we were 162 00:10:58,960 --> 00:10:59,440 Speaker 1: going there. 163 00:11:00,280 --> 00:11:03,920 Speaker 2: Art and writing and music aren't. 164 00:11:03,679 --> 00:11:06,960 Speaker 1: Going away, but they're going to completely change from how 165 00:11:07,000 --> 00:11:08,079 Speaker 1: we know them today. 166 00:11:09,360 --> 00:11:09,640 Speaker 2: Now. 167 00:11:09,960 --> 00:11:13,000 Speaker 1: I told you earlier that AI doesn't have any idea 168 00:11:13,120 --> 00:11:14,800 Speaker 1: of what it is to be a. 169 00:11:14,880 --> 00:11:18,200 Speaker 2: Human, but I think it doesn't matter. 170 00:11:18,880 --> 00:11:23,239 Speaker 1: AI doesn't need to feel anything to write great literature 171 00:11:23,320 --> 00:11:26,040 Speaker 1: or great art or great music, because while you can 172 00:11:26,120 --> 00:11:29,800 Speaker 1: think of it as a sandwich slurry. You can also 173 00:11:29,840 --> 00:11:34,160 Speaker 1: think of chat GPT as a remix of every human 174 00:11:34,200 --> 00:11:39,520 Speaker 1: writer that has come before. Its training set is humankind, 175 00:11:39,720 --> 00:11:43,640 Speaker 1: and so even if it's just statistical, it's generating the 176 00:11:43,720 --> 00:11:47,880 Speaker 1: expressions and the passions and the fears and the hopes 177 00:11:48,440 --> 00:11:51,760 Speaker 1: of millions of people. So it doesn't matter if it 178 00:11:51,880 --> 00:11:54,719 Speaker 1: feels or knows or has theory of mind, or if 179 00:11:54,760 --> 00:11:59,760 Speaker 1: it cries at king Lear's speech, because it can convincingly 180 00:12:00,559 --> 00:12:03,400 Speaker 1: tell you a story that breaks your heart. And it 181 00:12:03,440 --> 00:12:07,240 Speaker 1: does this by drawing on the best of human writing 182 00:12:07,440 --> 00:12:11,160 Speaker 1: over the centuries. So as a result, it's incredibly good 183 00:12:11,240 --> 00:12:14,600 Speaker 1: and it puts together things in a new way. And 184 00:12:14,679 --> 00:12:20,160 Speaker 1: I think part of understanding this requires acknowledging a really 185 00:12:20,200 --> 00:12:23,640 Speaker 1: important point, which is that the AI is really good, 186 00:12:23,840 --> 00:12:31,440 Speaker 1: but also that humans are so easily hackable. The phrase 187 00:12:31,880 --> 00:12:34,920 Speaker 1: humans are hackable is a phrase that I first started 188 00:12:34,920 --> 00:12:37,880 Speaker 1: hearing from my friend Lisa Joy Nolan, who with her 189 00:12:37,960 --> 00:12:42,440 Speaker 1: husband Joan Nolan, created the television show Westworld, and that 190 00:12:42,559 --> 00:12:44,840 Speaker 1: was a big theme in that show. The humans could 191 00:12:44,880 --> 00:12:48,400 Speaker 1: so easily get seduced by the robots, or convinced to 192 00:12:48,440 --> 00:12:51,679 Speaker 1: do bad actions or act violently and the robots were 193 00:12:51,720 --> 00:12:54,760 Speaker 1: just running AI. But if they say the right thing, 194 00:12:54,880 --> 00:12:57,559 Speaker 1: then they can get humans to do things, whether that's 195 00:12:57,760 --> 00:13:00,920 Speaker 1: fighting or fornicating or whatever. It's like turning the key 196 00:13:00,960 --> 00:13:03,000 Speaker 1: in the lock. Now, there's a point that I want 197 00:13:03,040 --> 00:13:06,600 Speaker 1: to dig into here. If you saw Westworld, you may 198 00:13:06,640 --> 00:13:09,520 Speaker 1: remember the scene from the first episode where a man 199 00:13:09,640 --> 00:13:13,560 Speaker 1: named William has just arrived to Westworld and he's greeted 200 00:13:13,679 --> 00:13:16,440 Speaker 1: in a room by a beautiful woman who guides him 201 00:13:16,520 --> 00:13:19,440 Speaker 1: to pick out his cowboy outfit and his gun in 202 00:13:19,440 --> 00:13:22,520 Speaker 1: his hat, and she makes it clear that she's available 203 00:13:22,559 --> 00:13:28,720 Speaker 1: for him sexually, and he uncomfortably asks her, are you real? 204 00:13:29,400 --> 00:13:34,200 Speaker 1: And she says, if you can't tell, does it matter? 205 00:13:35,480 --> 00:13:35,680 Speaker 2: Now? 206 00:13:35,760 --> 00:13:40,079 Speaker 1: This is a major theme throughout Westworld. Humans are hackable, 207 00:13:40,360 --> 00:13:43,520 Speaker 1: and if you can't tell the difference between something that 208 00:13:43,600 --> 00:13:47,199 Speaker 1: has evolutionary importance to you and a fake version of it, 209 00:13:47,600 --> 00:13:49,960 Speaker 1: then it makes no difference. And this is what we 210 00:13:50,080 --> 00:13:52,800 Speaker 1: see when we look at the text that is spit 211 00:13:52,880 --> 00:13:58,200 Speaker 1: out from chat GPT. It is statistically sound, meaning it 212 00:13:58,360 --> 00:14:01,559 Speaker 1: falls in the orders and rhythms of millions of people 213 00:14:01,559 --> 00:14:04,240 Speaker 1: who have written things like it before, and so we 214 00:14:04,320 --> 00:14:09,000 Speaker 1: can be just as compelled by the text, and therefore 215 00:14:09,080 --> 00:14:12,800 Speaker 1: the fact that AI can write a story that moves 216 00:14:12,880 --> 00:14:17,160 Speaker 1: us and impresses us is no surprise. It's easy to 217 00:14:17,280 --> 00:14:20,000 Speaker 1: move and impress us. In a sense, it's no more 218 00:14:20,080 --> 00:14:24,280 Speaker 1: surprising than drawing a pornographic cartoon that turns someone on. 219 00:14:24,720 --> 00:14:29,920 Speaker 1: You're just plugging into deeply carved programs. A human can't 220 00:14:29,920 --> 00:14:33,920 Speaker 1: mate with the cartoon. But nonetheless, it's easy enough to 221 00:14:34,120 --> 00:14:38,920 Speaker 1: activate the biological programs, so a story can make you 222 00:14:39,160 --> 00:14:43,240 Speaker 1: shed tears or laugh even if the transformer is just 223 00:14:43,320 --> 00:14:46,880 Speaker 1: pushing around zeros and ones. And therefore we shouldn't be 224 00:14:47,000 --> 00:14:51,600 Speaker 1: surprised that AI can write these really great pieces of prose. 225 00:14:51,680 --> 00:14:56,280 Speaker 1: It doesn't have to be real and it doesn't matter. 226 00:14:57,160 --> 00:15:00,440 Speaker 1: So now that we can write beautiful prose with AI, 227 00:15:00,720 --> 00:15:04,360 Speaker 1: what does this mean for the future of books. Well, 228 00:15:04,440 --> 00:15:07,200 Speaker 1: I think we can imagine a pretty cool future for 229 00:15:07,720 --> 00:15:14,240 Speaker 1: AI generated literature. We can imagine generating infinite, wonderful material. 230 00:15:15,040 --> 00:15:16,720 Speaker 2: And you know what, Back in the day. 231 00:15:17,080 --> 00:15:22,400 Speaker 1: Kings and emperors had poems written that were bespoke. The 232 00:15:22,560 --> 00:15:25,080 Speaker 1: poems were written just for them. And now it's going 233 00:15:25,120 --> 00:15:28,640 Speaker 1: to be trivial for us to all live as royalty, 234 00:15:29,160 --> 00:15:33,240 Speaker 1: having bespoke literature written just for us as much as 235 00:15:33,280 --> 00:15:36,880 Speaker 1: we want, as often as we want, in seconds, and 236 00:15:36,960 --> 00:15:41,120 Speaker 1: maybe we'll come to enjoy dynamic novels, by which I 237 00:15:41,160 --> 00:15:44,120 Speaker 1: mean a piece of literature that's not pre written, but 238 00:15:44,240 --> 00:15:48,480 Speaker 1: instead is written on the fly depending on the decisions 239 00:15:48,480 --> 00:15:51,720 Speaker 1: that you make, like a choose your own adventure. So 240 00:15:51,800 --> 00:15:54,200 Speaker 1: you say this is a good book so far. Now 241 00:15:54,240 --> 00:15:56,600 Speaker 1: I want to see what happens if I go in 242 00:15:56,640 --> 00:15:58,840 Speaker 1: the neighbor's door and get a view on his life, 243 00:15:58,960 --> 00:16:01,720 Speaker 1: or the mailman life who just passed by, or the 244 00:16:01,760 --> 00:16:05,360 Speaker 1: traffic cop and the book just keeps writing itself on 245 00:16:05,400 --> 00:16:09,320 Speaker 1: the fly, thousands of pages that end up being. 246 00:16:09,120 --> 00:16:14,000 Speaker 2: Unique for me, for you, for everyone as they go 247 00:16:14,080 --> 00:16:15,120 Speaker 2: on their own adventure. 248 00:16:15,600 --> 00:16:19,240 Speaker 1: Instead of having some poor author who has to write 249 00:16:19,480 --> 00:16:23,520 Speaker 1: every possible branching path, now there's no need to do that. 250 00:16:23,560 --> 00:16:25,120 Speaker 2: You just generated on the fly. 251 00:16:26,040 --> 00:16:30,200 Speaker 1: So now we'll all get to experience literary worlds that 252 00:16:30,240 --> 00:16:35,200 Speaker 1: are infinite in all directions. So in that light, it 253 00:16:35,280 --> 00:16:40,320 Speaker 1: certainly seems that AI is going to replace human creatives. 254 00:16:40,720 --> 00:16:43,600 Speaker 1: It can do things better and millions of times faster, 255 00:16:44,120 --> 00:16:46,680 Speaker 1: and it can be there to write the next pages 256 00:16:46,720 --> 00:16:51,400 Speaker 1: according to your wishes. So it looks like writers are 257 00:16:51,520 --> 00:16:53,680 Speaker 1: going the way of the mastodon? 258 00:16:54,720 --> 00:16:56,000 Speaker 2: Or are they? 259 00:16:56,600 --> 00:17:00,720 Speaker 1: I think the real story is not so simple. I'm 260 00:17:00,800 --> 00:17:05,320 Speaker 1: fairly sure that while AI will augment human told stories, 261 00:17:05,920 --> 00:17:09,080 Speaker 1: there's essentially zero danger that it's going to do a 262 00:17:09,119 --> 00:17:12,720 Speaker 1: wholesale replacement of human creatives. And I'm going to argue 263 00:17:12,720 --> 00:17:16,640 Speaker 1: this for four reasons. The first is that we care 264 00:17:16,840 --> 00:17:20,640 Speaker 1: about the overarching arc of a story, and at least 265 00:17:20,640 --> 00:17:24,400 Speaker 1: at the moment, AI can't even come close to constructing this. 266 00:17:24,800 --> 00:17:28,800 Speaker 1: And this is because of a fundamental limitation in its architecture. 267 00:17:29,240 --> 00:17:31,760 Speaker 1: And this isn't just a question of pouring more money 268 00:17:31,800 --> 00:17:34,560 Speaker 1: in and getting more massive computers on the job. It 269 00:17:34,640 --> 00:17:40,199 Speaker 1: has to do with the exponentially increasing computational cost of 270 00:17:40,320 --> 00:17:45,760 Speaker 1: representing longer pieces of work. So currently with chat GPT four, 271 00:17:46,320 --> 00:17:50,120 Speaker 1: it looks at the past four ninety six tokens, which 272 00:17:50,160 --> 00:17:53,240 Speaker 1: is about three thousand words, and it decides what the 273 00:17:53,280 --> 00:17:57,280 Speaker 1: most likely next word is. But without getting into the 274 00:17:57,320 --> 00:17:59,719 Speaker 1: details of the math, I want to point out that 275 00:17:59,760 --> 00:18:03,920 Speaker 1: this requires a matrix. Think about it like a big 276 00:18:03,960 --> 00:18:07,000 Speaker 1: spreadsheet that has four thousand ninety six rows in four 277 00:18:07,040 --> 00:18:07,919 Speaker 1: thousand ninety. 278 00:18:07,640 --> 00:18:10,160 Speaker 2: Six columns and an entry in every cell. 279 00:18:10,200 --> 00:18:13,760 Speaker 1: That represents something about the probability of those words going 280 00:18:13,800 --> 00:18:14,360 Speaker 1: with each other. 281 00:18:14,840 --> 00:18:17,680 Speaker 2: Now, this matrix will grow larger. 282 00:18:17,280 --> 00:18:20,560 Speaker 1: With time, but the size of the output is inherently 283 00:18:20,640 --> 00:18:25,160 Speaker 1: constrained by this structure, and as a result, chat GPT 284 00:18:25,359 --> 00:18:28,760 Speaker 1: is perfect for poems or blonde posts or small articles, 285 00:18:29,240 --> 00:18:33,280 Speaker 1: but not something the size of a novel. Why because 286 00:18:33,320 --> 00:18:38,680 Speaker 1: a novel has arcs and plot twists and cleverly planted 287 00:18:38,880 --> 00:18:42,639 Speaker 1: clues and cliffhangers, and all of these operate at a 288 00:18:42,720 --> 00:18:48,080 Speaker 1: longer timescale. So a human author mentally zooms in and 289 00:18:48,119 --> 00:18:52,720 Speaker 1: out such that their stories have this sweeping arc to them. So, 290 00:18:52,840 --> 00:18:55,159 Speaker 1: for example, in a mystery novel, we get to the 291 00:18:55,400 --> 00:18:58,560 Speaker 1: end and we realize that all the clues and the 292 00:18:58,600 --> 00:19:02,560 Speaker 1: red herrings we saw or subservient to the solution to 293 00:19:02,600 --> 00:19:05,440 Speaker 1: the mystery, which of course the author knew from the beginning, 294 00:19:05,680 --> 00:19:08,439 Speaker 1: and the author was just spooling out clues to you 295 00:19:08,480 --> 00:19:11,080 Speaker 1: one at a time. In writing, you often have to 296 00:19:11,200 --> 00:19:14,639 Speaker 1: know the end to structure the beginning in the middle. 297 00:19:14,920 --> 00:19:18,320 Speaker 1: And this is, by the way, why chat GPT can't 298 00:19:18,359 --> 00:19:20,720 Speaker 1: make up a new joke, even though it can repeat 299 00:19:20,800 --> 00:19:22,160 Speaker 1: jokes that are already made. 300 00:19:22,320 --> 00:19:25,760 Speaker 2: But it's because to construct a joke, just like a 301 00:19:25,800 --> 00:19:29,159 Speaker 2: mystery novel, you have to know the punchline first, and 302 00:19:29,200 --> 00:19:33,119 Speaker 2: then you construct the joke backwards. But these large language 303 00:19:33,119 --> 00:19:37,639 Speaker 2: models are simply constructing everything in the forward direction. It 304 00:19:37,680 --> 00:19:41,879 Speaker 2: does statistical calculations on what the most probable word to 305 00:19:41,920 --> 00:19:45,359 Speaker 2: come next is given all the words before it. So, 306 00:19:45,480 --> 00:19:49,240 Speaker 2: coming back to the long arc, if you watched all 307 00:19:49,280 --> 00:19:52,119 Speaker 2: eight seasons of Game of Thrones, for example, or you 308 00:19:52,160 --> 00:19:55,480 Speaker 2: read those books, you come to care about these characters 309 00:19:55,560 --> 00:19:58,879 Speaker 2: because you've been with them through so many trials and 310 00:19:58,920 --> 00:20:01,600 Speaker 2: you feel like you know the and understand them, and 311 00:20:01,640 --> 00:20:05,200 Speaker 2: you can predict things about their behavior, and you're invested 312 00:20:05,320 --> 00:20:09,399 Speaker 2: in their long term trajectories. So all the children of 313 00:20:09,440 --> 00:20:13,359 Speaker 2: the Stark family end up scattered in different directions in 314 00:20:13,400 --> 00:20:17,320 Speaker 2: the world, and then in the final season, they end 315 00:20:17,400 --> 00:20:21,639 Speaker 2: up reconvening. After what seems like a lifetime of adventure. 316 00:20:21,680 --> 00:20:26,119 Speaker 2: They're all back together for the final big showdown with 317 00:20:26,280 --> 00:20:29,280 Speaker 2: the Knight King. And when we watch the series and 318 00:20:29,320 --> 00:20:32,399 Speaker 2: we get to season eight, we think, wow, I didn't 319 00:20:32,480 --> 00:20:35,359 Speaker 2: see that coming, that they're all back together now, and 320 00:20:35,400 --> 00:20:38,640 Speaker 2: now this story has a beautiful shape to it. 321 00:20:39,119 --> 00:20:42,800 Speaker 1: I'm really in the hands of a professional here. At 322 00:20:42,840 --> 00:20:46,960 Speaker 1: least with our current AI architectures today, it's impossible to 323 00:20:47,080 --> 00:20:50,679 Speaker 1: achieve that, except possibly in a few thousand word version, 324 00:20:51,119 --> 00:20:54,800 Speaker 1: because chat ept is playing its statistical game, and of 325 00:20:54,800 --> 00:20:57,320 Speaker 1: course it's playing it extremely well and successfully. 326 00:20:57,560 --> 00:20:59,960 Speaker 2: But the trick to recognize here is. 327 00:21:00,000 --> 00:21:02,800 Speaker 1: That it is amazing at the level of paragraphs and 328 00:21:02,840 --> 00:21:07,000 Speaker 1: possibly a few pages, but not at the level of 329 00:21:07,080 --> 00:21:10,119 Speaker 1: thinking about the details of a five hundred page novel, 330 00:21:10,480 --> 00:21:15,160 Speaker 1: or a two hour movie screenplay or an eight season epic. 331 00:21:15,920 --> 00:21:18,800 Speaker 1: It's great at this small stuff because it can do 332 00:21:18,840 --> 00:21:22,560 Speaker 1: that with statistics, but it's fundamentally limited for the longer 333 00:21:22,600 --> 00:21:26,320 Speaker 1: stuff because it has no way to zoom out and 334 00:21:26,480 --> 00:21:30,159 Speaker 1: think about the crops that it wants to plant for 335 00:21:30,240 --> 00:21:34,120 Speaker 1: the long game. Okay, you might say, fine, maybe we'll 336 00:21:34,160 --> 00:21:36,760 Speaker 1: get there at some point, but even for now, couldn't 337 00:21:36,800 --> 00:21:40,800 Speaker 1: you build a big story out of smaller chunks. So 338 00:21:41,240 --> 00:21:44,720 Speaker 1: one idea is to make this form of storytelling in 339 00:21:44,760 --> 00:21:46,760 Speaker 1: which the world is infinitely big. 340 00:21:47,280 --> 00:21:48,639 Speaker 2: Let's come back to this picture. 341 00:21:48,680 --> 00:21:51,520 Speaker 1: I painted a moment ago of a choose your own 342 00:21:51,560 --> 00:21:56,199 Speaker 1: adventure in which the AI generates plot points on the 343 00:21:56,240 --> 00:21:59,800 Speaker 1: fly for you. So I say, okay, open that door 344 00:21:59,840 --> 00:22:03,480 Speaker 1: to my left and the story continues as though it 345 00:22:03,560 --> 00:22:08,320 Speaker 1: were all prescripted, as though I have an author, let's say, 346 00:22:08,320 --> 00:22:12,720 Speaker 1: in the style of Henningway or Nibokov for Morrison, who 347 00:22:12,760 --> 00:22:16,560 Speaker 1: has pre written every possibility. In certain ways, this would 348 00:22:16,560 --> 00:22:20,320 Speaker 1: be amazingly cool, But I think the problem here is 349 00:22:20,320 --> 00:22:25,680 Speaker 1: that a story like that would just equal randomness, and 350 00:22:25,720 --> 00:22:29,040 Speaker 1: that's not actually what we want in a story. Instead, 351 00:22:29,080 --> 00:22:32,159 Speaker 1: we want to feel like we're putting our trust into an. 352 00:22:32,040 --> 00:22:33,840 Speaker 2: Author who sees the big picture. 353 00:22:33,880 --> 00:22:38,040 Speaker 1: We want the Stark children to reconvene such as we 354 00:22:38,119 --> 00:22:40,720 Speaker 1: feel the overarching pattern of the story and we have 355 00:22:40,760 --> 00:22:44,960 Speaker 1: a sense of completeness. If you just wanted randomness, you'd 356 00:22:45,240 --> 00:22:47,760 Speaker 1: go out into the world and find it there. You 357 00:22:47,760 --> 00:22:51,480 Speaker 1: wouldn't sit on your couch and read about meaningless characters 358 00:22:51,520 --> 00:22:57,320 Speaker 1: who are just in Brownian motion. And I think this 359 00:22:57,400 --> 00:23:00,800 Speaker 1: is the same issue with AI music, at least as 360 00:23:00,840 --> 00:23:01,520 Speaker 1: it stands now. 361 00:23:02,240 --> 00:23:04,080 Speaker 2: Recent examples show. 362 00:23:03,840 --> 00:23:07,560 Speaker 1: That it can compose incredible sounding music moment to moment. 363 00:23:07,720 --> 00:23:10,119 Speaker 2: But the reason it doesn't beat out. 364 00:23:09,920 --> 00:23:13,560 Speaker 1: A real human composer, at least today, is because it 365 00:23:13,560 --> 00:23:16,880 Speaker 1: doesn't have any long term vision, and so the whole 366 00:23:16,920 --> 00:23:21,240 Speaker 1: piece of music just hangs together. Statistically, moment to moment, 367 00:23:21,640 --> 00:23:25,840 Speaker 1: and that's perfectly good for composing things like elevator music, 368 00:23:25,880 --> 00:23:28,760 Speaker 1: which is for a short ride, or commercial music which 369 00:23:28,800 --> 00:23:31,960 Speaker 1: only needs to be twenty seconds. But it won't for 370 00:23:32,040 --> 00:23:36,359 Speaker 1: now replace a human composer who writes with the long 371 00:23:36,640 --> 00:23:39,439 Speaker 1: arc in mind. For example, I was just talking with 372 00:23:39,480 --> 00:23:42,800 Speaker 1: my friend Tony Brandt, who's a composer, and he was 373 00:23:42,840 --> 00:23:46,040 Speaker 1: explaining to me that when Ludwig and vad Beethoven died, 374 00:23:46,520 --> 00:23:50,679 Speaker 1: he left behind sketches for a tenth symphony. So a 375 00:23:50,720 --> 00:23:55,280 Speaker 1: few years ago some computer scientists used AI to complete 376 00:23:55,359 --> 00:23:59,479 Speaker 1: the symphony, to finish what was unfinished. Now did they 377 00:23:59,520 --> 00:24:02,320 Speaker 1: do a good job. In one sense, it was an 378 00:24:02,440 --> 00:24:08,320 Speaker 1: incredible feat. They extracted the statistics of Beethoven's choices and 379 00:24:08,480 --> 00:24:12,040 Speaker 1: preferences from everything he'd written, and they used that to 380 00:24:12,640 --> 00:24:16,320 Speaker 1: statistically guess what moves he would have made next had 381 00:24:16,359 --> 00:24:20,920 Speaker 1: he lived, What notes, what chords, what instruments. But even 382 00:24:20,960 --> 00:24:23,920 Speaker 1: with this feat, it was clear that the AI didn't 383 00:24:23,960 --> 00:24:27,960 Speaker 1: know how to think long term. For example, Beethoven's Ninth 384 00:24:28,040 --> 00:24:31,840 Speaker 1: Symphony ends with a chorus, which was such a surprise 385 00:24:31,960 --> 00:24:34,000 Speaker 1: to end a symphony this way. It had not ever 386 00:24:34,040 --> 00:24:37,960 Speaker 1: been done before, so the team training the AI decided 387 00:24:38,000 --> 00:24:41,560 Speaker 1: Beethoven would have found a similar novelty to end his 388 00:24:41,720 --> 00:24:45,960 Speaker 1: tenth Symphony, so they instructed the AI to include an organ, 389 00:24:46,480 --> 00:24:49,040 Speaker 1: a church instrument that had also never been used in 390 00:24:49,040 --> 00:24:52,440 Speaker 1: a symphony before. So at the start of the last movement, 391 00:24:52,720 --> 00:24:54,679 Speaker 1: the AI generates an organ. 392 00:24:55,119 --> 00:24:58,639 Speaker 2: But when we zoom in, we see the difference. 393 00:25:01,119 --> 00:25:05,000 Speaker 1: The real Beethoven laid all sorts of clues in the 394 00:25:05,119 --> 00:25:09,040 Speaker 1: Ninth Symphony to set the groundwork for the chorus. Like 395 00:25:09,160 --> 00:25:12,640 Speaker 1: the orchestra plays a type of music called a recitative 396 00:25:13,200 --> 00:25:18,680 Speaker 1: before the choir enters. Why because recitatives are found in opera, 397 00:25:18,720 --> 00:25:22,640 Speaker 1: and opera has voices, So he was laying clues down. 398 00:25:22,800 --> 00:25:26,200 Speaker 1: But in the AI tenth Symphony, there was no build 399 00:25:26,240 --> 00:25:29,320 Speaker 1: up to the organ. There was no suspense, no hidden 400 00:25:29,400 --> 00:25:30,399 Speaker 1: clues about. 401 00:25:30,119 --> 00:25:30,919 Speaker 2: What was coming. 402 00:25:31,480 --> 00:25:35,840 Speaker 1: The AI didn't know how to prepare the organ's arrival, 403 00:25:36,240 --> 00:25:39,760 Speaker 1: how to give it the significance that's there for experts 404 00:25:39,760 --> 00:25:46,080 Speaker 1: who listen for arcs that build through time. So, at 405 00:25:46,160 --> 00:25:50,120 Speaker 1: least for now, AI is useful at writing brief articles 406 00:25:50,200 --> 00:25:53,680 Speaker 1: and composing short ditties, but it doesn't have the architecture 407 00:25:53,720 --> 00:25:59,960 Speaker 1: to write long pieces that humans love to create, and consume. 408 00:26:13,480 --> 00:26:14,159 Speaker 2: So as I'm. 409 00:26:14,000 --> 00:26:17,840 Speaker 1: Writing my next books, these large language models don't feel 410 00:26:17,840 --> 00:26:21,640 Speaker 1: to me like a real threat, at least not yet. 411 00:26:21,920 --> 00:26:26,080 Speaker 1: But let's imagine that we cut to ten years from 412 00:26:26,119 --> 00:26:29,840 Speaker 1: now and some hardworking programmers have figured out how to 413 00:26:29,880 --> 00:26:33,639 Speaker 1: build an AI with the right sort of architecture that 414 00:26:33,880 --> 00:26:36,200 Speaker 1: zooms in and out on the scope of a story, 415 00:26:36,480 --> 00:26:41,080 Speaker 1: and it can successfully generate a novel with cliffhangers and 416 00:26:41,200 --> 00:26:44,520 Speaker 1: overarching themes and so on. It's certainly not impossible that 417 00:26:44,560 --> 00:26:47,439 Speaker 1: we're going to get there, and it'll probably happen sooner 418 00:26:47,480 --> 00:26:50,440 Speaker 1: than we expect. So let's imagine we get there in 419 00:26:50,480 --> 00:26:53,160 Speaker 1: a year or five or ten. An AI can generate 420 00:26:53,200 --> 00:26:58,359 Speaker 1: a million good novels in an hour. Then what Well, 421 00:26:58,400 --> 00:27:01,560 Speaker 1: there are several directions in which things can go, And 422 00:27:01,600 --> 00:27:04,520 Speaker 1: the possibility that I mentioned earlier is that novels might 423 00:27:04,560 --> 00:27:10,000 Speaker 1: become bespoke, totally personalized to you. So you prompt your 424 00:27:10,080 --> 00:27:13,440 Speaker 1: AI to make an adventure story of exactly the type 425 00:27:13,480 --> 00:27:15,960 Speaker 1: that you might like. So you say, tell me a 426 00:27:16,200 --> 00:27:20,240 Speaker 1: murder mystery about a basketball player who's killed by someone 427 00:27:20,280 --> 00:27:23,480 Speaker 1: who appears to be his girlfriend. But then it turns 428 00:27:23,480 --> 00:27:26,960 Speaker 1: out it's actually a CIA plot. That opens the door 429 00:27:27,000 --> 00:27:30,840 Speaker 1: to a cover up involving a pharmaceutical company. Let's assume 430 00:27:30,880 --> 00:27:33,280 Speaker 1: that the AI then spits out a book to your 431 00:27:33,320 --> 00:27:36,800 Speaker 1: exact specification, and it does an amazing job, and it 432 00:27:36,800 --> 00:27:39,440 Speaker 1: gives you a colorful story just how you wanted it, 433 00:27:39,720 --> 00:27:42,000 Speaker 1: and you can enjoy that on the beach seconds later. 434 00:27:42,280 --> 00:27:45,720 Speaker 1: Well that's cool, But I assert that this is never 435 00:27:45,920 --> 00:27:49,600 Speaker 1: going to replace literature. And this is my second point 436 00:27:49,640 --> 00:27:53,120 Speaker 1: why artists don't need to worry, because when you define 437 00:27:53,160 --> 00:27:56,920 Speaker 1: your own plot, the surprise is diluted. 438 00:27:57,440 --> 00:27:59,520 Speaker 2: The joy of literature is diluted. 439 00:28:00,080 --> 00:28:03,800 Speaker 1: After all, even if you are a creative prompter, you 440 00:28:03,880 --> 00:28:07,399 Speaker 1: are limited to versions of what you have experienced or 441 00:28:07,520 --> 00:28:10,600 Speaker 1: read before. And much of what we love in literature 442 00:28:10,720 --> 00:28:14,200 Speaker 1: is this surprise that comes from a particular point of 443 00:28:14,280 --> 00:28:18,800 Speaker 1: view that you have never considered, like characters or plot 444 00:28:18,880 --> 00:28:22,720 Speaker 1: points that would never be generated by your own limited 445 00:28:22,800 --> 00:28:26,000 Speaker 1: point of view. In the end, I think we don't 446 00:28:26,119 --> 00:28:30,199 Speaker 1: want to be limited by the parochial fence lines of 447 00:28:30,240 --> 00:28:31,440 Speaker 1: our own imagination. 448 00:28:31,960 --> 00:28:32,960 Speaker 2: I suspect that. 449 00:28:33,040 --> 00:28:35,919 Speaker 1: No matter how far in the future we look, we 450 00:28:35,960 --> 00:28:39,720 Speaker 1: are still going to want stories that surprise us, plot 451 00:28:39,800 --> 00:28:44,120 Speaker 1: twists that we don't see coming. Okay, fine, you might say, 452 00:28:44,160 --> 00:28:46,719 Speaker 1: so you agree that it's more exciting if we go 453 00:28:46,800 --> 00:28:49,920 Speaker 1: on rides that we didn't predefine. But you might point 454 00:28:49,960 --> 00:28:53,040 Speaker 1: out there's another thing that AI can do. So let's 455 00:28:53,120 --> 00:28:56,760 Speaker 1: address the next issue, the idea that AI could someday 456 00:28:56,880 --> 00:29:02,000 Speaker 1: generate millions of highly creative versions of a single story, 457 00:29:02,240 --> 00:29:04,560 Speaker 1: so there'd be no need to stick with just one 458 00:29:04,680 --> 00:29:08,440 Speaker 1: version of stories anymore. Instead of George R. R. Martin 459 00:29:08,840 --> 00:29:13,200 Speaker 1: writing Game of Thrones over decades, future AI could generate 460 00:29:13,360 --> 00:29:17,160 Speaker 1: thousands of fascinating versions in a second, and we wouldn't 461 00:29:17,200 --> 00:29:21,520 Speaker 1: depend on him for the next slow novel. But I 462 00:29:21,600 --> 00:29:24,120 Speaker 1: suggest that's not going to catch on either. 463 00:29:24,760 --> 00:29:25,120 Speaker 2: Why. 464 00:29:25,480 --> 00:29:30,440 Speaker 1: It's because we care about shared adventure. Would Game of 465 00:29:30,520 --> 00:29:33,840 Speaker 1: Thrones have been so popular if we each saw our 466 00:29:33,920 --> 00:29:38,240 Speaker 1: own version of it? In my version, John snow dies early, 467 00:29:38,440 --> 00:29:42,680 Speaker 1: and in your version, danaris Mary's Tyrian lanister, and in 468 00:29:42,720 --> 00:29:45,920 Speaker 1: your neighbor's version, Ariya marries into a royal family in 469 00:29:46,000 --> 00:29:49,200 Speaker 1: some subplot island that never even appears in my version. 470 00:29:49,720 --> 00:29:53,560 Speaker 1: If this sounds less appealing to you, to have mutually 471 00:29:53,720 --> 00:29:57,360 Speaker 1: exclusive worlds, it illustrates the point that I want to make, 472 00:29:57,400 --> 00:30:01,600 Speaker 1: which is a big part of story is this social aspect, 473 00:30:01,720 --> 00:30:06,040 Speaker 1: the shared experience. We certainly could use AI to generate 474 00:30:06,080 --> 00:30:09,520 Speaker 1: a million different versions of west Ros, and in the 475 00:30:09,560 --> 00:30:12,640 Speaker 1: future we can generate instant video around these plots with 476 00:30:12,800 --> 00:30:17,360 Speaker 1: terrific special effect. But as a society, I think we 477 00:30:17,440 --> 00:30:22,160 Speaker 1: wouldn't want to each consume our own version. You want 478 00:30:22,200 --> 00:30:25,680 Speaker 1: your John Snow to do the same thing as my 479 00:30:25,880 --> 00:30:28,680 Speaker 1: John Snow. And this is because a huge part of 480 00:30:28,800 --> 00:30:34,840 Speaker 1: story is this shared experience. We enjoy sharing fantasy worlds 481 00:30:34,880 --> 00:30:37,959 Speaker 1: because we talk about them. This is why we do 482 00:30:38,000 --> 00:30:41,200 Speaker 1: book clubs, so we can sit around and discuss something 483 00:30:41,240 --> 00:30:44,400 Speaker 1: we all shared together. All the time, I hear people say, hey, 484 00:30:44,440 --> 00:30:47,720 Speaker 1: did you see the latest episode of The Peripheral or 485 00:30:47,800 --> 00:30:51,320 Speaker 1: Jack Ryan or Severance or Star Trek or whatever. And 486 00:30:51,400 --> 00:30:55,840 Speaker 1: our love of communal stories stems partially from our need 487 00:30:56,360 --> 00:31:00,080 Speaker 1: for shared references. For example, I'm always making reference and 488 00:31:00,360 --> 00:31:04,040 Speaker 1: is to how Neo in the Matrix saw in slow motion, 489 00:31:04,200 --> 00:31:07,040 Speaker 1: and that's decades after that movie came out, but it 490 00:31:07,120 --> 00:31:10,440 Speaker 1: serves as a quick, culturally shared way that we can 491 00:31:10,480 --> 00:31:14,800 Speaker 1: talk about concepts. We all have quick cultural references for 492 00:31:15,040 --> 00:31:18,680 Speaker 1: time travel, where people say met me up Scotti when 493 00:31:18,680 --> 00:31:22,920 Speaker 1: they're talking about teleportation, or we reference Obi wan Kenobi 494 00:31:22,960 --> 00:31:25,240 Speaker 1: when we say may the force be with you, or 495 00:31:25,360 --> 00:31:29,520 Speaker 1: we reference ex Machina or Westworld as a shorthand for 496 00:31:29,640 --> 00:31:30,680 Speaker 1: AI going bad. 497 00:31:31,120 --> 00:31:33,040 Speaker 2: And take this as an example. 498 00:31:32,920 --> 00:31:37,400 Speaker 1: Imagine that you could generate a fantasy football game with 499 00:31:37,440 --> 00:31:41,080 Speaker 1: your favorite players from any decade on one team versus 500 00:31:41,160 --> 00:31:43,880 Speaker 1: players on another team, and you can now watch a 501 00:31:43,960 --> 00:31:47,680 Speaker 1: full football game from stem to stern. But would you 502 00:31:48,360 --> 00:31:52,000 Speaker 1: if no one else ever saw that game? In other words, 503 00:31:52,320 --> 00:31:55,840 Speaker 1: would you follow teams all the way through the World 504 00:31:55,920 --> 00:32:00,000 Speaker 1: Series if it was purely AI generated plays and games. 505 00:32:00,760 --> 00:32:02,800 Speaker 1: I know that people might have different opinions on this, 506 00:32:02,880 --> 00:32:06,120 Speaker 1: but to me, that sounds not the least bit appealing. 507 00:32:06,560 --> 00:32:09,880 Speaker 1: Why it's because a giant part about sports is the 508 00:32:09,920 --> 00:32:12,960 Speaker 1: culture of talking about the game. Hey did you see 509 00:32:12,960 --> 00:32:15,560 Speaker 1: that play last night? Can you believe that shot he took? 510 00:32:15,800 --> 00:32:19,400 Speaker 1: Can you believe the call that refmade? And stories are 511 00:32:19,480 --> 00:32:22,200 Speaker 1: analogous to sports in this way. We come to our 512 00:32:22,240 --> 00:32:25,160 Speaker 1: book clubs to take the world that we read in 513 00:32:25,320 --> 00:32:28,840 Speaker 1: solitude and find a community with other people who were 514 00:32:28,880 --> 00:32:32,160 Speaker 1: there with us from their own living rooms. So I 515 00:32:32,200 --> 00:32:34,800 Speaker 1: suggest that as a culture, we are always going to 516 00:32:35,440 --> 00:32:40,360 Speaker 1: desire and need a shared vocabulary, and the only way 517 00:32:40,360 --> 00:32:43,440 Speaker 1: to grow that is to watch the same movies and 518 00:32:43,520 --> 00:32:45,280 Speaker 1: read the same stories. 519 00:32:45,760 --> 00:32:47,360 Speaker 2: And that's why I predict that. 520 00:32:47,400 --> 00:32:53,240 Speaker 1: While individualized stories might find niche audiences, it won't replace 521 00:32:53,440 --> 00:32:58,000 Speaker 1: our need for shared stories. This is an interesting dimension 522 00:32:58,040 --> 00:33:05,280 Speaker 1: of literature that's not typically canered. Story gives us social glue. Okay, fine, 523 00:33:05,320 --> 00:33:08,160 Speaker 1: so let's assume that at some point AI could write 524 00:33:08,160 --> 00:33:12,520 Speaker 1: a story that's so evocative and beautiful that it becomes 525 00:33:12,560 --> 00:33:17,400 Speaker 1: a shared story, an adventure which everyone taps into and enjoys. 526 00:33:17,800 --> 00:33:21,640 Speaker 1: And now we arrive at my fourth point about why 527 00:33:21,680 --> 00:33:25,880 Speaker 1: AI won't totally displace creatives, and that is the question 528 00:33:25,960 --> 00:33:28,840 Speaker 1: of whether we get something more out of a piece 529 00:33:28,880 --> 00:33:32,360 Speaker 1: of literature or art if we feel there's. 530 00:33:32,120 --> 00:33:34,120 Speaker 2: A heartbeat behind it. 531 00:33:34,760 --> 00:33:37,640 Speaker 1: I read a beautiful quotation in The Atlantic about a 532 00:33:37,680 --> 00:33:42,080 Speaker 1: decade ago quote one of the only requirements for literature 533 00:33:42,640 --> 00:33:46,240 Speaker 1: is that the reader can feel a heart pulsing back 534 00:33:46,280 --> 00:33:49,600 Speaker 1: from them on the other side of the page. The 535 00:33:49,840 --> 00:33:54,120 Speaker 1: heartbeat matters because when we read, we consider the intention 536 00:33:54,240 --> 00:33:57,719 Speaker 1: of the author. We think, oh, this is Mary Shelley, 537 00:33:57,760 --> 00:34:00,400 Speaker 1: whose mother died a couple of weeks after she was born, 538 00:34:00,480 --> 00:34:03,560 Speaker 1: and she had a troubled childhood, and her father homeschooled her. 539 00:34:03,560 --> 00:34:07,600 Speaker 1: And she married the romantic poet Percy bish Shelley, and 540 00:34:08,000 --> 00:34:10,880 Speaker 1: he was already married and his wife committed suicide, and 541 00:34:10,880 --> 00:34:13,160 Speaker 1: they moved to France, and she came back pregnant, and 542 00:34:13,200 --> 00:34:16,040 Speaker 1: they were destitute, and their daughter died. And then they 543 00:34:16,040 --> 00:34:18,839 Speaker 1: went to spend a summer in Geneva with friends, and 544 00:34:18,880 --> 00:34:21,000 Speaker 1: they each set out to write a ghost story, and 545 00:34:21,080 --> 00:34:23,120 Speaker 1: she ended up writing Frankenstein. 546 00:34:23,640 --> 00:34:24,800 Speaker 2: So we read her. 547 00:34:24,719 --> 00:34:28,200 Speaker 1: Novel and we think, this is her voice, and this 548 00:34:28,360 --> 00:34:31,080 Speaker 1: is her viewpoint on the world, and these were the 549 00:34:31,120 --> 00:34:33,839 Speaker 1: things that she knew and the things she didn't know, 550 00:34:33,880 --> 00:34:35,279 Speaker 1: and the things she couldn't know. 551 00:34:35,800 --> 00:34:38,640 Speaker 2: It isn't just the piece of art itself. 552 00:34:38,719 --> 00:34:43,760 Speaker 1: It is the artist behind the art that colors our experience. 553 00:34:44,239 --> 00:34:48,480 Speaker 1: So imagine we get Chad Gpt to adopt Mary Shelley's 554 00:34:48,600 --> 00:34:52,720 Speaker 1: style and write a story involving cell phones and electric cars. 555 00:34:52,960 --> 00:34:56,200 Speaker 1: It might be interesting and amazing, but I suggest we 556 00:34:56,239 --> 00:35:00,239 Speaker 1: wouldn't enjoy it as much because we would recognize there's 557 00:35:00,360 --> 00:35:05,239 Speaker 1: no unique human, no unique beating heart who had the 558 00:35:05,360 --> 00:35:09,760 Speaker 1: experiences and slaved over the words. Now, you could argue 559 00:35:09,760 --> 00:35:14,120 Speaker 1: that almost all of the authors we enjoy. We live 560 00:35:14,160 --> 00:35:17,080 Speaker 1: apart from them in space or time, and we'll never 561 00:35:17,160 --> 00:35:19,520 Speaker 1: meet them, and we just have the vaguest sense of 562 00:35:19,560 --> 00:35:20,400 Speaker 1: their existence. 563 00:35:20,719 --> 00:35:23,320 Speaker 2: And that might be true, but it's still worth. 564 00:35:23,120 --> 00:35:27,600 Speaker 1: Noting that we know fundamentally that they are human and 565 00:35:27,640 --> 00:35:30,480 Speaker 1: they are like us in some way. They may be 566 00:35:30,840 --> 00:35:34,400 Speaker 1: more successful, or more impoverished, or maybe from a different country, 567 00:35:34,800 --> 00:35:38,879 Speaker 1: but we know that fundamentally they are fellow travelers with 568 00:35:38,960 --> 00:35:55,600 Speaker 1: us on the human journey. Now, obviously we love a 569 00:35:55,600 --> 00:35:59,000 Speaker 1: lot of things that aren't real, like Spider Man or Batman, 570 00:35:59,560 --> 00:36:02,399 Speaker 1: but we all I also love the actors behind them. 571 00:36:02,440 --> 00:36:04,759 Speaker 1: If you had a chance to have dinner with or 572 00:36:04,800 --> 00:36:07,719 Speaker 1: even to shake the hand of the actor behind some 573 00:36:07,920 --> 00:36:11,080 Speaker 1: fantasy character that you love, you'd be thrilled about this. 574 00:36:11,640 --> 00:36:13,040 Speaker 2: Now, I think that leads. 575 00:36:12,760 --> 00:36:17,000 Speaker 1: To an interesting open question about some of these new 576 00:36:17,400 --> 00:36:20,600 Speaker 1: avatars that are hitting the scene with hundreds of thousands 577 00:36:20,600 --> 00:36:24,600 Speaker 1: of followers on Twitter. Even though they're fake. They're just avatars, 578 00:36:24,600 --> 00:36:27,360 Speaker 1: they're not real people. The part that strikes me is 579 00:36:27,400 --> 00:36:30,480 Speaker 1: really interesting is that the ones who get all the 580 00:36:30,520 --> 00:36:34,080 Speaker 1: attention are the creators behind the avatar. In other words, 581 00:36:34,360 --> 00:36:37,360 Speaker 1: if I told you there was an avatar on Twitter, 582 00:36:37,360 --> 00:36:39,359 Speaker 1: with a one hundred thousand followers, and you could get 583 00:36:39,360 --> 00:36:42,040 Speaker 1: the chance to meet the young woman behind all this, 584 00:36:42,520 --> 00:36:45,040 Speaker 1: you'd be thrilled. What this tells me is that we 585 00:36:45,080 --> 00:36:49,600 Speaker 1: are compelled by the heartbeat that is just behind the 586 00:36:49,640 --> 00:36:54,120 Speaker 1: actor or the avatar. In many ways, that's more interesting 587 00:36:54,280 --> 00:36:58,080 Speaker 1: to us than the actor or the avatar themselves. Now, 588 00:36:58,640 --> 00:37:00,719 Speaker 1: I don't think this goes on in so let me 589 00:37:00,760 --> 00:37:03,680 Speaker 1: just address the counterpoint. You might say, well, does that 590 00:37:03,719 --> 00:37:07,200 Speaker 1: mean that if AI generated a thousand novels in a second, 591 00:37:07,239 --> 00:37:09,920 Speaker 1: that I'd be really interested in meeting the team of 592 00:37:10,000 --> 00:37:13,840 Speaker 1: young programmers behind that. I don't think so, because meeting 593 00:37:13,880 --> 00:37:18,319 Speaker 1: the programmers doesn't expand your understanding of the story. But 594 00:37:18,440 --> 00:37:21,880 Speaker 1: meeting an author who poured her heart into the story 595 00:37:21,960 --> 00:37:27,200 Speaker 1: for years that does shape and color and expand your understanding. 596 00:37:27,560 --> 00:37:29,680 Speaker 2: And by the way, beyond writing, I think. 597 00:37:29,480 --> 00:37:33,239 Speaker 1: This applies to musical composers and visual artists in the 598 00:37:33,280 --> 00:37:38,120 Speaker 1: same way, and in fact, to all human endeavors. I 599 00:37:38,239 --> 00:37:40,600 Speaker 1: was just talking with a neighbor of mine. He and 600 00:37:40,640 --> 00:37:43,759 Speaker 1: I spend a lot of time on airplanes flying to 601 00:37:43,840 --> 00:37:47,040 Speaker 1: some city in the world to give a talk. He 602 00:37:47,160 --> 00:37:51,000 Speaker 1: just got a three D scan and a high resolution 603 00:37:51,280 --> 00:37:54,680 Speaker 1: avatar of himself made and he can combine that with 604 00:37:54,800 --> 00:37:58,920 Speaker 1: Chad GPT to make his avatar give little speeches. And 605 00:37:58,960 --> 00:38:01,240 Speaker 1: so he and I were really chewing on this because 606 00:38:01,280 --> 00:38:04,400 Speaker 1: the question is, the next time he gets invited to 607 00:38:04,480 --> 00:38:08,400 Speaker 1: speak on some stage and some random city around the world, 608 00:38:08,800 --> 00:38:12,160 Speaker 1: can he just have the avatar give the speech online instead? 609 00:38:12,560 --> 00:38:15,880 Speaker 1: Will conferences still want him to fly across. 610 00:38:15,520 --> 00:38:16,960 Speaker 2: The globe to give a talk. 611 00:38:17,040 --> 00:38:19,719 Speaker 1: Or will the avatar be good enough and save a 612 00:38:19,760 --> 00:38:24,280 Speaker 1: lot of expense and plane fuel? Possibly, But the flip 613 00:38:24,320 --> 00:38:29,040 Speaker 1: side is do people value going to the talk because 614 00:38:29,080 --> 00:38:30,640 Speaker 1: of the beating heart. 615 00:38:30,520 --> 00:38:31,440 Speaker 2: On the stage? 616 00:38:32,040 --> 00:38:35,960 Speaker 1: And my long bet is that conferences will continue to 617 00:38:36,080 --> 00:38:40,879 Speaker 1: invite flesh and blood humans because audiences are humans who 618 00:38:41,200 --> 00:38:46,239 Speaker 1: care about other humans. So when it comes to legal documents, 619 00:38:46,280 --> 00:38:48,879 Speaker 1: if AI can do it better, awesome, when it comes 620 00:38:48,920 --> 00:38:52,000 Speaker 1: to medical diagnoses, if AI can do it better awesome, 621 00:38:52,600 --> 00:38:56,800 Speaker 1: when it comes to hearing a speaker on the stage 622 00:38:57,239 --> 00:39:01,840 Speaker 1: with his or her imperfections and limited knowledge and fundamentally 623 00:39:02,280 --> 00:39:05,600 Speaker 1: human nature, I'm going to take the bet that that 624 00:39:06,040 --> 00:39:10,600 Speaker 1: is going to last and beyond just appreciating the reality 625 00:39:10,719 --> 00:39:13,560 Speaker 1: of another human. This maybe for another reason as well, 626 00:39:14,120 --> 00:39:17,920 Speaker 1: an interesting psychological effect that I think is going to 627 00:39:17,920 --> 00:39:20,279 Speaker 1: be at play here. This is what I'm going to 628 00:39:20,280 --> 00:39:23,880 Speaker 1: call the effort phenomenon. I'll give you an example of this. 629 00:39:24,320 --> 00:39:27,320 Speaker 1: A well known colleague of mine here in Silicon Valley 630 00:39:27,360 --> 00:39:31,240 Speaker 1: recently announced that he had published a book half written 631 00:39:31,280 --> 00:39:34,680 Speaker 1: by him and half written by AI. And when I 632 00:39:34,719 --> 00:39:37,960 Speaker 1: first heard about this, I thought, I wish I wanted 633 00:39:38,000 --> 00:39:41,839 Speaker 1: to read this, but I don't now. I did take 634 00:39:41,880 --> 00:39:44,640 Speaker 1: a look at the book, and there are clever insights, 635 00:39:44,680 --> 00:39:48,680 Speaker 1: and it's well written. But I'm simply not that inspired 636 00:39:48,760 --> 00:39:53,040 Speaker 1: to read something that's even half written by AI, because 637 00:39:53,400 --> 00:39:56,600 Speaker 1: it makes me feel, perhaps unfairly, that. 638 00:39:56,640 --> 00:39:58,800 Speaker 2: He didn't put in the normal amount of effort. 639 00:39:59,400 --> 00:40:02,719 Speaker 1: My analogy you would be if Picasso said, hey, will 640 00:40:02,760 --> 00:40:05,719 Speaker 1: you buy this painting? My students painted most of it, 641 00:40:05,760 --> 00:40:07,600 Speaker 1: but then I finished it off and put my signature 642 00:40:07,640 --> 00:40:10,200 Speaker 1: on it. It feels like it would be slightly less valuable. 643 00:40:10,800 --> 00:40:14,000 Speaker 1: So let's return to that scene in Westworld where William 644 00:40:14,120 --> 00:40:18,000 Speaker 1: asks the host are you real? And she says if 645 00:40:18,040 --> 00:40:21,600 Speaker 1: you can't tell, doesn't matter, Because this is the question 646 00:40:21,640 --> 00:40:22,480 Speaker 1: that comes up. 647 00:40:22,600 --> 00:40:24,040 Speaker 2: About a novel. 648 00:40:24,320 --> 00:40:27,680 Speaker 1: If I spend seven years writing a novel, and if 649 00:40:28,000 --> 00:40:31,439 Speaker 1: Chad Gpt or google bart spits out a novel that's 650 00:40:31,520 --> 00:40:33,000 Speaker 1: word for word equivalent. 651 00:40:33,719 --> 00:40:34,560 Speaker 2: Does it matter? 652 00:40:35,120 --> 00:40:39,680 Speaker 1: And I think, perhaps surprisingly, the answer is yes, it matters. 653 00:40:40,200 --> 00:40:43,279 Speaker 1: We care about the effort that went into it. If 654 00:40:43,320 --> 00:40:45,440 Speaker 1: I were to show you two pieces of artwork that 655 00:40:45,560 --> 00:40:48,280 Speaker 1: someone had done, and one of them just involves painting 656 00:40:48,320 --> 00:40:51,440 Speaker 1: a single dot on the middle of a big white canvas, 657 00:40:51,480 --> 00:40:55,920 Speaker 1: and the other one is the person carefully gluing marbles 658 00:40:55,960 --> 00:40:58,399 Speaker 1: one on top of each other until they balance eight 659 00:40:58,440 --> 00:41:01,200 Speaker 1: feet high. You may have a p for looking at 660 00:41:01,239 --> 00:41:03,279 Speaker 1: one or the other, but just think about how much 661 00:41:03,320 --> 00:41:06,000 Speaker 1: money you would, in theory, be willing to pay for 662 00:41:06,080 --> 00:41:08,919 Speaker 1: each of these. If you're like most people, you think 663 00:41:08,960 --> 00:41:12,439 Speaker 1: the thing that took a lot of effort is worth more. 664 00:41:13,080 --> 00:41:16,640 Speaker 1: There have been psychology studies on this since the nineteen fifties. 665 00:41:17,000 --> 00:41:20,040 Speaker 1: It's difficult for people to separate out the effort that 666 00:41:20,120 --> 00:41:24,080 Speaker 1: went into something from its value. In other words, the 667 00:41:24,160 --> 00:41:29,640 Speaker 1: effort is used as a shortcut for understanding quality. For example, 668 00:41:29,640 --> 00:41:33,200 Speaker 1: in one paper done by Krueger at All, they had 669 00:41:33,400 --> 00:41:37,200 Speaker 1: people rate a poem, or rate a painting, or rate 670 00:41:37,239 --> 00:41:40,239 Speaker 1: a suit of armor, and the people generally thought it 671 00:41:40,320 --> 00:41:43,480 Speaker 1: was better quality and worth more money, and they liked 672 00:41:43,520 --> 00:41:46,839 Speaker 1: it better if they thought it took more time and 673 00:41:46,920 --> 00:41:50,160 Speaker 1: effort to produce a friend of mine. Uses the example 674 00:41:50,280 --> 00:41:54,040 Speaker 1: of diamonds. People will pay much more money for a 675 00:41:54,200 --> 00:41:58,640 Speaker 1: real diamond with flaws than they will for a synthetically 676 00:41:58,800 --> 00:42:02,799 Speaker 1: grown diamond from laboratory that has no flaws at all. Now, 677 00:42:02,800 --> 00:42:06,759 Speaker 1: why would you pay extra money for flaws? Part of 678 00:42:06,800 --> 00:42:09,239 Speaker 1: this has to do with the notion of effort. The 679 00:42:09,280 --> 00:42:13,480 Speaker 1: real diamond was produced by mother nature over millions of 680 00:42:13,680 --> 00:42:17,719 Speaker 1: years of compression, so it's a very special thing that 681 00:42:17,800 --> 00:42:20,680 Speaker 1: took quote unquote effort on the part of mother nature. 682 00:42:21,160 --> 00:42:23,920 Speaker 2: But the lab grown diamond that can be done in 683 00:42:23,960 --> 00:42:25,040 Speaker 2: a day and a half. 684 00:42:25,440 --> 00:42:28,640 Speaker 1: And so even though it's more perfect, it is less 685 00:42:28,719 --> 00:42:31,480 Speaker 1: valuable because it just took less time to make it. 686 00:42:32,000 --> 00:42:33,840 Speaker 2: We actually pay for flaws. 687 00:42:34,600 --> 00:42:37,000 Speaker 1: Now, I'm not arguing that we can't be fooled at 688 00:42:37,000 --> 00:42:41,760 Speaker 1: some point into loving AI generated literature. It seems quite 689 00:42:41,800 --> 00:42:44,000 Speaker 1: possible to me that in the future there will be 690 00:42:44,120 --> 00:42:48,080 Speaker 1: novels written by AI, and we might not always know it, 691 00:42:48,360 --> 00:42:53,080 Speaker 1: because the AI will also generate a false story about 692 00:42:53,200 --> 00:42:57,480 Speaker 1: the author, complete with a biography and a generated photograph. 693 00:42:57,800 --> 00:43:00,920 Speaker 1: My assertion is simply that FA it is going to 694 00:43:00,920 --> 00:43:03,680 Speaker 1: be an important part of what the AI will need 695 00:43:03,719 --> 00:43:08,080 Speaker 1: to do, because it's more difficult to become invested in 696 00:43:08,160 --> 00:43:12,360 Speaker 1: something that we think is simply doing massive statistical calculations 697 00:43:12,920 --> 00:43:18,320 Speaker 1: rather than having a private, limited internal life. We care 698 00:43:18,600 --> 00:43:23,240 Speaker 1: about other humans, So what's the big picture. My friend 699 00:43:23,520 --> 00:43:27,400 Speaker 1: Kevin Kelly suggested to me the other day that generative 700 00:43:27,480 --> 00:43:31,400 Speaker 1: AI may play a role that's analogous to the invention 701 00:43:31,560 --> 00:43:35,120 Speaker 1: of the camera. What happened at that moment in history 702 00:43:35,239 --> 00:43:39,200 Speaker 1: was that painters lamented that this was the end of 703 00:43:39,320 --> 00:43:43,799 Speaker 1: painting because you could now capture anything instantly with the 704 00:43:43,840 --> 00:43:46,080 Speaker 1: click of a button, and you could capture it with 705 00:43:46,160 --> 00:43:48,960 Speaker 1: zero mistakes. So why would you sit there with a 706 00:43:49,000 --> 00:43:53,920 Speaker 1: paint brush and painstakingly try to capture every detail by hand. 707 00:43:54,480 --> 00:43:58,839 Speaker 1: At that moment in history, it seemed clear that painters 708 00:43:59,280 --> 00:44:03,319 Speaker 1: were done for But as it turns out, photographs ended 709 00:44:03,400 --> 00:44:05,200 Speaker 1: up filling a different niche. 710 00:44:05,960 --> 00:44:09,879 Speaker 2: Absolute realism wasn't the only end goal of art. 711 00:44:10,360 --> 00:44:15,360 Speaker 1: People didn't only want a maximumly realistic print of a scene. 712 00:44:15,440 --> 00:44:19,480 Speaker 1: They also wanted swirls, an amazing color, and more importantly, 713 00:44:19,600 --> 00:44:23,360 Speaker 1: things that didn't exist in the outside world. So canvas 714 00:44:23,400 --> 00:44:28,560 Speaker 1: painting remained an active field, even while photography grew and 715 00:44:28,719 --> 00:44:33,720 Speaker 1: ended up flowering on a neighboring field. So one possibility 716 00:44:34,280 --> 00:44:38,560 Speaker 1: is that AI generated literature will not foment it takeover, 717 00:44:39,040 --> 00:44:42,719 Speaker 1: but instead it's going to fill a new niche, one 718 00:44:42,760 --> 00:44:45,319 Speaker 1: that we don't quite see yet, but it isn't the 719 00:44:45,360 --> 00:44:48,839 Speaker 1: same plot of land. And I think there's one more 720 00:44:48,880 --> 00:44:51,879 Speaker 1: possibility for where this could go for writers, not now, 721 00:44:51,960 --> 00:44:54,920 Speaker 1: but in the coming years. And for that, I want 722 00:44:54,960 --> 00:44:57,919 Speaker 1: to tell you what happened with the world champion Go 723 00:44:58,080 --> 00:45:02,000 Speaker 1: player Could Jig. He was the world's number one player 724 00:45:02,160 --> 00:45:04,799 Speaker 1: at Go, which is the game in which you use 725 00:45:04,880 --> 00:45:08,680 Speaker 1: those small black or white rocks to define your territory 726 00:45:08,680 --> 00:45:11,520 Speaker 1: and try to surround your opponent. So in May of 727 00:45:11,600 --> 00:45:17,080 Speaker 1: twenty seventeen, he faced off against an AI program called 728 00:45:17,280 --> 00:45:21,160 Speaker 1: Alpha Go, which was designed by Deep Mind, and Alpha 729 00:45:21,200 --> 00:45:24,239 Speaker 1: Go had been trained on millions and millions of games 730 00:45:24,280 --> 00:45:28,400 Speaker 1: of Go, so it had deeply absorbed the statistics of 731 00:45:28,600 --> 00:45:33,960 Speaker 1: possible plays. So they played the first game and Jiu lost. 732 00:45:34,520 --> 00:45:38,960 Speaker 1: Alpha Go had pulled moves that none of his human 733 00:45:39,000 --> 00:45:42,799 Speaker 1: opponents had ever thought of, and then Jua lost the 734 00:45:42,880 --> 00:45:46,319 Speaker 1: second game. The AI had won over a human in 735 00:45:46,360 --> 00:45:50,279 Speaker 1: a game that's way more complex than chess, and subsequent 736 00:45:50,440 --> 00:45:53,480 Speaker 1: versions of the AI are no doubt going to continue 737 00:45:53,520 --> 00:45:56,759 Speaker 1: to win evermore. But that's not the interesting part of 738 00:45:56,800 --> 00:46:01,640 Speaker 1: the story. The interesting part is what happened next. So 739 00:46:01,960 --> 00:46:06,799 Speaker 1: Jig got over his embarrassment and he became mesmerized by 740 00:46:06,960 --> 00:46:11,920 Speaker 1: what had just transpired, and he studied the games. 741 00:46:11,560 --> 00:46:12,320 Speaker 2: That he lost. 742 00:46:13,400 --> 00:46:17,520 Speaker 1: Before he played Alpha Go, Jia had won a majority 743 00:46:17,680 --> 00:46:21,920 Speaker 1: of the games against his human opponents, but afterwards he 744 00:46:22,000 --> 00:46:25,240 Speaker 1: found he was able to beat his human opponents even 745 00:46:25,360 --> 00:46:31,160 Speaker 1: more easily. After his species shaming defeats in twenty seventeen, 746 00:46:31,520 --> 00:46:35,160 Speaker 1: he went on to play twelve straight matches against humans and. 747 00:46:35,160 --> 00:46:38,560 Speaker 2: He won them all in a row. So what had happened. 748 00:46:39,360 --> 00:46:43,400 Speaker 3: He had been exposed to new kinds of moves and 749 00:46:43,600 --> 00:46:47,320 Speaker 3: strategies that had been pulled by Alpha Go, and these 750 00:46:47,600 --> 00:46:51,279 Speaker 3: all lay outside of traditional ways of doing it. 751 00:46:51,600 --> 00:46:54,080 Speaker 2: All these moves that Alpha Go had done. 752 00:46:53,920 --> 00:46:57,719 Speaker 1: Were legal and possible, but they were just different from 753 00:46:57,719 --> 00:47:01,040 Speaker 1: what had been played over the last twenty five hundred years. 754 00:47:01,400 --> 00:47:02,799 Speaker 2: If you're a Go officionado. 755 00:47:02,840 --> 00:47:07,000 Speaker 1: This included things like playing a stone directly diagonal to 756 00:47:07,520 --> 00:47:12,319 Speaker 1: your opponent's loan stone, or playing six space extensions, while 757 00:47:12,400 --> 00:47:13,680 Speaker 1: humans tend to prefer. 758 00:47:13,600 --> 00:47:15,080 Speaker 2: Five space anyway. 759 00:47:15,440 --> 00:47:21,320 Speaker 1: Joe reported that playing against the AI was like opening. 760 00:47:20,920 --> 00:47:22,560 Speaker 2: A door to another world. 761 00:47:22,840 --> 00:47:27,080 Speaker 1: Once he was exposed to these alien game plays, he 762 00:47:27,200 --> 00:47:33,120 Speaker 1: incorporated them, and this story I suspect typifies the future 763 00:47:33,719 --> 00:47:37,480 Speaker 1: as humans and machines interface. Some people are worried that 764 00:47:37,560 --> 00:47:41,160 Speaker 1: AI is going to take over, but we will continue 765 00:47:41,200 --> 00:47:45,359 Speaker 1: to adapt as well. We will become better writers as 766 00:47:45,400 --> 00:47:48,880 Speaker 1: we see examples that are allowed by the language but 767 00:47:49,120 --> 00:47:52,960 Speaker 1: no one had ever tried it, or visual art techniques 768 00:47:52,960 --> 00:47:56,759 Speaker 1: that involve moves that are allowable, but culturally we just 769 00:47:56,880 --> 00:47:59,879 Speaker 1: never thought to do it, Or musical moves that are 770 00:48:00,160 --> 00:48:00,839 Speaker 1: possible to. 771 00:48:00,840 --> 00:48:03,799 Speaker 2: Do with notes, but no one does. 772 00:48:03,560 --> 00:48:06,800 Speaker 1: Them because traditionally we just wouldn't think of going there. 773 00:48:06,920 --> 00:48:10,160 Speaker 1: Because fundamentally, as a writer, I think I'm doing all 774 00:48:10,280 --> 00:48:13,640 Speaker 1: kinds of original things, but there's a very real sense 775 00:48:13,680 --> 00:48:18,799 Speaker 1: in which I'm simply remixing what I've absorbed before. I 776 00:48:18,880 --> 00:48:22,960 Speaker 1: interpolate between examples that I've seen. So even if AI 777 00:48:23,160 --> 00:48:28,280 Speaker 1: is just interpolating, it's read billions of times more texts 778 00:48:28,320 --> 00:48:33,600 Speaker 1: than I have, and so it can do very clever interpolations, 779 00:48:33,640 --> 00:48:37,080 Speaker 1: and I can learn from that a lot of people 780 00:48:37,080 --> 00:48:40,840 Speaker 1: are worried that AI is going to leave humans far behind, 781 00:48:40,960 --> 00:48:44,720 Speaker 1: and in many respects that's true. But as computers improve, 782 00:48:45,560 --> 00:48:49,840 Speaker 1: so will we. In the battle of man and machine. 783 00:48:50,600 --> 00:48:53,600 Speaker 1: Both are going to get better, and as we continue 784 00:48:53,640 --> 00:48:58,399 Speaker 1: to adapt in parallel, the future definition of AI may 785 00:48:58,440 --> 00:49:04,880 Speaker 1: well shift from our official intelligence to augmented intelligence. In 786 00:49:04,920 --> 00:49:07,680 Speaker 1: the best case scenario, this isn't going to be a war, 787 00:49:08,160 --> 00:49:12,440 Speaker 1: but a collaboration. It's going to be an ongoing, guided 788 00:49:12,600 --> 00:49:19,560 Speaker 1: tour into areas that were previously just beyond our view. 789 00:49:22,920 --> 00:49:24,080 Speaker 2: That's all for this week. 790 00:49:24,360 --> 00:49:26,839 Speaker 1: To find out more and to share your thoughts, head 791 00:49:26,840 --> 00:49:30,680 Speaker 1: over to eagleman dot com, Slash Podcasts, and you can 792 00:49:30,719 --> 00:49:34,280 Speaker 1: also watch full episodes of Inner Cosmos on YouTube. 793 00:49:34,640 --> 00:49:36,480 Speaker 2: Subscribe to my channel so you can. 794 00:49:36,320 --> 00:49:40,000 Speaker 1: Follow along each week for new updates until next time. 795 00:49:40,360 --> 00:49:43,719 Speaker 2: I'm David Eagleman, and this is Inner Cosmos.