WEBVTT - Ep6 "What will AI mean for artists?" 0:00:05.559 --> 0:00:11.240 Will writers and artists and musicians become unemployed by AI? 0:00:11.920 --> 0:00:15.280 What are the new capabilities that we're seeing all around us, 0:00:15.360 --> 0:00:16.120 and what is this. 0:00:16.120 --> 0:00:18.840 Going to mean for human creativity? 0:00:19.200 --> 0:00:22.000 And what does this have to do with diamonds and 0:00:22.079 --> 0:00:27.120 Westworld and effort and Frankenstein in Beethoven and the Stark 0:00:27.200 --> 0:00:33.000 Family and Game of Thrones. Welcome to Inner Cosmos with 0:00:33.080 --> 0:00:37.639 me David Eagleman. I'm a neuroscientist and an author at 0:00:37.680 --> 0:00:41.480 Stanford University, and in this episode, I get to dive 0:00:41.520 --> 0:00:46.520 into something that's right at the intersection of science and creativity. 0:00:51.479 --> 0:00:55.720 Most of my podcasts are about evergreen topics about our 0:00:55.760 --> 0:00:59.920 brains and our psychology, but there's something so extraordinary happy 0:01:00.440 --> 0:01:01.080 right now. 0:01:01.560 --> 0:01:05.600 We're in the middle of a revolution with AI, and 0:01:05.640 --> 0:01:09.640 what's called generative AI in particular. So I'm going to 0:01:09.720 --> 0:01:12.960 do a two part episode on this. For today, I'm 0:01:12.959 --> 0:01:15.880 going to dig into what generative AI is and what 0:01:16.000 --> 0:01:20.080 it means for human creativity, and then in the next episode, 0:01:20.120 --> 0:01:24.360 I'm going to tackle the question of sentience. Are these 0:01:24.400 --> 0:01:28.720 ais conscious and if not, now, could they be soon? 0:01:29.400 --> 0:01:31.760 And how would we know when we get there? 0:01:35.000 --> 0:01:39.240 So let's start in twenty seventeen when almost no one 0:01:39.280 --> 0:01:42.440 in the world paid attention when a team at Google 0:01:42.520 --> 0:01:47.280 Brain introduced a new way of building an artificial neural network. 0:01:47.920 --> 0:01:51.040 So this was different than the architectures that came before it, 0:01:51.480 --> 0:01:55.560 which were called things like convolutional neural networks and recurrent 0:01:55.640 --> 0:01:58.800 neural networks. Instead, they presented a new model that was 0:01:58.840 --> 0:02:03.520 called a transformer. Now, transformer is not one of those 0:02:03.640 --> 0:02:06.760 robots that shapeshift into trucks and helicopters. 0:02:07.280 --> 0:02:10.040 Instead, a transformer model is. 0:02:10.000 --> 0:02:14.160 A way to tackle sequential data like the words that 0:02:14.200 --> 0:02:16.280 are in a sentence or the frames in a video. 0:02:16.800 --> 0:02:20.080 And a transformer model takes in everything at once, and 0:02:20.120 --> 0:02:23.239 it essentially pays attention to different parts of the data. 0:02:23.800 --> 0:02:29.400 And this allows training on enormous data sets, bigger than 0:02:29.440 --> 0:02:33.560 what was trained on before. Like now it's essentially everything 0:02:33.680 --> 0:02:36.960 that has been written by humans that is on the Internet, 0:02:37.160 --> 0:02:41.840 which is petabytes of data. So these models they digest 0:02:41.919 --> 0:02:44.840 all of that and what do they do. They essentially 0:02:44.919 --> 0:02:48.200 look at a sequence of inputs like the words and 0:02:48.240 --> 0:02:52.280 a sentence, and they ask what word is most likely 0:02:52.360 --> 0:02:56.440 to come next in that sequence. Now we'll come back 0:02:56.480 --> 0:02:58.000 to that in a second, but I just want to 0:02:58.040 --> 0:03:04.600 note that this transformer model is finding uses way beyond text. So, 0:03:04.680 --> 0:03:08.160 for example, a recent Nature paper used this kind of 0:03:08.200 --> 0:03:11.720 model to look at amino acids, which run in a 0:03:11.720 --> 0:03:14.800 sequence to make proteins, and they looked at these chains 0:03:14.800 --> 0:03:18.200 of amino acids like techt strings, and they set a 0:03:18.360 --> 0:03:21.840 major new water mark in determining how proteins fold, which 0:03:21.880 --> 0:03:25.400 is a very difficult problem. And people are using transformers 0:03:25.400 --> 0:03:30.800 for everything from making music to reading giant reams of 0:03:30.840 --> 0:03:35.040 medical records and so on. These transformer models are built 0:03:35.080 --> 0:03:37.440 into search already, and soon they're going to be in 0:03:37.520 --> 0:03:40.120 your phone and in your car, and in your bank 0:03:40.160 --> 0:03:48.080 and in your doctor's office. So what everyone in Silicon 0:03:48.160 --> 0:03:50.800 Valley is talking about is how this new kind of 0:03:50.880 --> 0:03:55.120 AI is going to disrupt the workforce. And a lot 0:03:55.160 --> 0:03:58.320 of people are thinking about white collar jobs that have 0:03:58.440 --> 0:04:04.200 traditionally required memorization of long textbooks, and these jobs, whether 0:04:04.280 --> 0:04:08.440 they're legal or medical, suddenly seem to be kind of outmoded. 0:04:09.000 --> 0:04:11.880 And so we're all thinking about what this means for 0:04:11.960 --> 0:04:15.120 the economy because so many jobs are going to be 0:04:15.160 --> 0:04:19.920 displaced by this new technology. Now, there's nothing totally new 0:04:19.960 --> 0:04:24.039 about this kind of worry, because every generation sees new 0:04:24.080 --> 0:04:28.240 technologies take over old jobs. That's natural, and we don't 0:04:28.320 --> 0:04:32.840 lament the fact that we don't have elevator operators anymore, 0:04:33.120 --> 0:04:38.000 or switchboard operators at telephone companies, or factories that make 0:04:38.240 --> 0:04:43.800 VCRs or eight track tape players, because new technologies continuously 0:04:43.880 --> 0:04:48.560 replace the old, and industries change and people adapt. But 0:04:48.680 --> 0:04:52.120 the concern that we're seeing with the AI revolution is 0:04:52.200 --> 0:04:55.559 the speed of it. It's probably the case that we've 0:04:56.040 --> 0:05:00.479 never before had a move forward in technology that's so 0:05:00.960 --> 0:05:07.080 unbelievably rapid. So this is why everyone's talking about this 0:05:07.240 --> 0:05:09.919 with a different point of view than we did with 0:05:10.080 --> 0:05:13.080 previous innovations. But I want to zoom in on something 0:05:13.120 --> 0:05:15.360 a little different for this episode. I want to know 0:05:15.400 --> 0:05:19.719 what this all means for human creativity, because the thing 0:05:19.800 --> 0:05:23.080 to note is these models have been trained up not 0:05:23.240 --> 0:05:26.960 just on the handful of novels and conversations and schoolwork 0:05:27.040 --> 0:05:31.599 that you have experienced on your thin trajectory through space 0:05:31.640 --> 0:05:34.880 and time, but they have been trained with everything that's 0:05:34.960 --> 0:05:40.479 ever been written by humans. Every textbook, every article, every poem, 0:05:40.520 --> 0:05:46.800 every blog post, every novel. We're talking seventy one billion 0:05:46.880 --> 0:05:52.719 web pages and hundreds of trillions of words, It's something 0:05:52.760 --> 0:05:57.680 that's so far beyond any human's capacity to consume even 0:05:57.720 --> 0:06:01.360 a fraction of it, or to really imagine a corpus 0:06:01.480 --> 0:06:04.839 of text that large. Oh and by the way, it 0:06:04.880 --> 0:06:08.400 has a perfect memory for every word that it's read. 0:06:08.520 --> 0:06:12.080 So now you're talking about a system that's not the 0:06:12.120 --> 0:06:17.560 same as a brain, but is incredibly powerful at generating 0:06:17.720 --> 0:06:22.159 text or visual art or music and soon video. And 0:06:22.200 --> 0:06:25.000 so while we'll talk about sentience next week, this week, 0:06:25.080 --> 0:06:28.080 I want to address a social point that has quickly 0:06:28.160 --> 0:06:30.839 risen to the surface, which is what will all this 0:06:31.040 --> 0:06:36.599 mean for human art and human creativity? Personally, I'm working 0:06:36.640 --> 0:06:39.800 on my next several books right now, and these are 0:06:39.839 --> 0:06:44.720 all projects that have spanned years, and so I'm fascinated 0:06:44.839 --> 0:06:49.560 and terrified about whether AI is going to replace me 0:06:49.640 --> 0:06:51.960 as a writer. What does this kind of new AI 0:06:52.600 --> 0:06:57.640 mean for writers, for visual artists, for musicians who studied 0:06:57.640 --> 0:07:00.680 their whole lives to be able to compose beautiful piece 0:07:00.720 --> 0:07:06.200 of music? Is human creativity destined for the dust bin 0:07:06.440 --> 0:07:11.160 of history? So let's start with the downside of these models. 0:07:11.680 --> 0:07:14.640 So in my book Live Wired, I talked about how 0:07:14.720 --> 0:07:20.320 AI algorithms don't care about relevance they memorize whatever we 0:07:20.560 --> 0:07:23.080 ask them to. So, now this is a very useful 0:07:23.120 --> 0:07:26.600 feature of AI, but it's also the reason AI is 0:07:26.720 --> 0:07:31.640 not particularly human like, because AI models don't have any 0:07:31.680 --> 0:07:35.400 sort of internal model of the world. They have no 0:07:35.520 --> 0:07:38.880 idea what it is to be a human and have 0:07:39.040 --> 0:07:44.080 drives and concerns. They don't care which problems are interesting 0:07:44.600 --> 0:07:48.640 or germane. Instead, they memorize whatever we feed them. So 0:07:48.720 --> 0:07:51.960 whether that's distinguishing a horse from a zebra in a 0:07:52.000 --> 0:07:56.400 billion photographs, or tracking flight data from every airport on 0:07:56.440 --> 0:08:01.160 the planet, or composing music in the style of Brian Eno, 0:08:01.640 --> 0:08:06.360 they have no sense of importance except in a statistical sense, 0:08:06.960 --> 0:08:09.679 which is to say, which signals occur more often. 0:08:10.440 --> 0:08:13.160 So contemporary AI could never. 0:08:13.080 --> 0:08:17.560 By itself decide that it finds irresistible a particular kind 0:08:17.600 --> 0:08:21.600 of ice cream, or that it abhors a particular kind 0:08:21.600 --> 0:08:26.120 of music, or that it's heartbroken by King Lear's speech 0:08:26.320 --> 0:08:29.760 over his dead daughter. So AI can dispatch, you know, 0:08:29.840 --> 0:08:34.000 ten thousand hours of intense practice in ten thousand nanoseconds, 0:08:34.360 --> 0:08:38.200 but it doesn't care about any zeros and ones over 0:08:38.240 --> 0:08:43.680 any others. As a result, AI can accomplish incredibly impressive feats, 0:08:43.760 --> 0:08:48.760 but not the feat of being quite like a human. 0:08:49.200 --> 0:08:52.320 And so some critics of AI say, look, it's like 0:08:52.360 --> 0:08:55.760 you want a sandwich, and what this transformer model does 0:08:56.200 --> 0:08:58.720 is it looks at all the billions of sandwiches out 0:08:58.760 --> 0:09:02.079 there in the world, and it gives you a slurry 0:09:02.480 --> 0:09:03.880 and it pours it out in. 0:09:03.840 --> 0:09:05.240 The shape of a sandwich. 0:09:05.559 --> 0:09:08.079 A fellow writer gave me that analogy the other day, 0:09:08.120 --> 0:09:12.839 and that doesn't sound particularly appealing, right, And yet these 0:09:13.000 --> 0:09:16.719 ais have massively surprised us. 0:09:17.080 --> 0:09:20.280 The text generation is so good, it's. 0:09:20.120 --> 0:09:24.319 So complete, it's so human like that we find ourselves 0:09:24.360 --> 0:09:27.960 not so much in the phase of invention like with 0:09:28.040 --> 0:09:31.239 all the machines we've made before. Instead, the whole scientific 0:09:31.240 --> 0:09:36.760 community is finding itself in a process of discovery. Everyone 0:09:36.880 --> 0:09:41.880 is exploring to find out what these enormous models are 0:09:41.960 --> 0:09:46.240 capable of, because nobody quite knows. They keep blowing our 0:09:46.280 --> 0:09:50.200 minds with things they're able to do which weren't pre 0:09:50.240 --> 0:09:56.280 programmed and not even foreseen. Have a friend who works 0:09:56.280 --> 0:09:59.800 with a big city symphony, and she's trying to play 0:10:00.160 --> 0:10:03.520 a program for the symphony several months out, which is 0:10:03.559 --> 0:10:07.760 a typical timescale for symphony planning, but she's scheduling to 0:10:07.800 --> 0:10:11.520 put on a program with music composed by AI, and 0:10:11.600 --> 0:10:14.240 she's at a loss for how to plan this because 0:10:14.720 --> 0:10:18.360 she's well aware that things are moving so fast that 0:10:18.400 --> 0:10:22.479 the musical world and the skill level of AI composition 0:10:23.000 --> 0:10:25.320 is going to be entirely different. In a few months, 0:10:25.320 --> 0:10:28.240 it's can be more advanced. So she was telling me 0:10:28.320 --> 0:10:31.880 that she doesn't quite know how to nail down plans 0:10:31.920 --> 0:10:36.600 for this, because unlike every symphony planner who has come before, 0:10:36.760 --> 0:10:39.600 she's now in a world where if she nails down 0:10:39.679 --> 0:10:43.080 a choice of music and trains up the musicians, it 0:10:43.240 --> 0:10:47.360 is guaranteed to be badly outdated some months from now. 0:10:47.760 --> 0:10:51.160 And this is the world we're operating in now. So jennertive, 0:10:51.160 --> 0:10:54.920 AI is moving so rapidly that we have entered this 0:10:55.120 --> 0:10:58.959 massive revolution without most of us realizing that we were 0:10:58.960 --> 0:10:59.440 going there. 0:11:00.280 --> 0:11:03.920 Art and writing and music aren't. 0:11:03.679 --> 0:11:06.960 Going away, but they're going to completely change from how 0:11:07.000 --> 0:11:08.079 we know them today. 0:11:09.360 --> 0:11:09.640 Now. 0:11:09.960 --> 0:11:13.000 I told you earlier that AI doesn't have any idea 0:11:13.120 --> 0:11:14.800 of what it is to be a. 0:11:14.880 --> 0:11:18.200 Human, but I think it doesn't matter. 0:11:18.880 --> 0:11:23.239 AI doesn't need to feel anything to write great literature 0:11:23.320 --> 0:11:26.040 or great art or great music, because while you can 0:11:26.120 --> 0:11:29.800 think of it as a sandwich slurry. You can also 0:11:29.840 --> 0:11:34.160 think of chat GPT as a remix of every human 0:11:34.200 --> 0:11:39.520 writer that has come before. Its training set is humankind, 0:11:39.720 --> 0:11:43.640 and so even if it's just statistical, it's generating the 0:11:43.720 --> 0:11:47.880 expressions and the passions and the fears and the hopes 0:11:48.440 --> 0:11:51.760 of millions of people. So it doesn't matter if it 0:11:51.880 --> 0:11:54.719 feels or knows or has theory of mind, or if 0:11:54.760 --> 0:11:59.760 it cries at king Lear's speech, because it can convincingly 0:12:00.559 --> 0:12:03.400 tell you a story that breaks your heart. And it 0:12:03.440 --> 0:12:07.240 does this by drawing on the best of human writing 0:12:07.440 --> 0:12:11.160 over the centuries. So as a result, it's incredibly good 0:12:11.240 --> 0:12:14.600 and it puts together things in a new way. And 0:12:14.679 --> 0:12:20.160 I think part of understanding this requires acknowledging a really 0:12:20.200 --> 0:12:23.640 important point, which is that the AI is really good, 0:12:23.840 --> 0:12:31.440 but also that humans are so easily hackable. The phrase 0:12:31.880 --> 0:12:34.920 humans are hackable is a phrase that I first started 0:12:34.920 --> 0:12:37.880 hearing from my friend Lisa Joy Nolan, who with her 0:12:37.960 --> 0:12:42.440 husband Joan Nolan, created the television show Westworld, and that 0:12:42.559 --> 0:12:44.840 was a big theme in that show. The humans could 0:12:44.880 --> 0:12:48.400 so easily get seduced by the robots, or convinced to 0:12:48.440 --> 0:12:51.679 do bad actions or act violently and the robots were 0:12:51.720 --> 0:12:54.760 just running AI. But if they say the right thing, 0:12:54.880 --> 0:12:57.559 then they can get humans to do things, whether that's 0:12:57.760 --> 0:13:00.920 fighting or fornicating or whatever. It's like turning the key 0:13:00.960 --> 0:13:03.000 in the lock. Now, there's a point that I want 0:13:03.040 --> 0:13:06.600 to dig into here. If you saw Westworld, you may 0:13:06.640 --> 0:13:09.520 remember the scene from the first episode where a man 0:13:09.640 --> 0:13:13.560 named William has just arrived to Westworld and he's greeted 0:13:13.679 --> 0:13:16.440 in a room by a beautiful woman who guides him 0:13:16.520 --> 0:13:19.440 to pick out his cowboy outfit and his gun in 0:13:19.440 --> 0:13:22.520 his hat, and she makes it clear that she's available 0:13:22.559 --> 0:13:28.720 for him sexually, and he uncomfortably asks her, are you real? 0:13:29.400 --> 0:13:34.200 And she says, if you can't tell, does it matter? 0:13:35.480 --> 0:13:35.680 Now? 0:13:35.760 --> 0:13:40.079 This is a major theme throughout Westworld. Humans are hackable, 0:13:40.360 --> 0:13:43.520 and if you can't tell the difference between something that 0:13:43.600 --> 0:13:47.199 has evolutionary importance to you and a fake version of it, 0:13:47.600 --> 0:13:49.960 then it makes no difference. And this is what we 0:13:50.080 --> 0:13:52.800 see when we look at the text that is spit 0:13:52.880 --> 0:13:58.200 out from chat GPT. It is statistically sound, meaning it 0:13:58.360 --> 0:14:01.559 falls in the orders and rhythms of millions of people 0:14:01.559 --> 0:14:04.240 who have written things like it before, and so we 0:14:04.320 --> 0:14:09.000 can be just as compelled by the text, and therefore 0:14:09.080 --> 0:14:12.800 the fact that AI can write a story that moves 0:14:12.880 --> 0:14:17.160 us and impresses us is no surprise. It's easy to 0:14:17.280 --> 0:14:20.000 move and impress us. In a sense, it's no more 0:14:20.080 --> 0:14:24.280 surprising than drawing a pornographic cartoon that turns someone on. 0:14:24.720 --> 0:14:29.920 You're just plugging into deeply carved programs. A human can't 0:14:29.920 --> 0:14:33.920 mate with the cartoon. But nonetheless, it's easy enough to 0:14:34.120 --> 0:14:38.920 activate the biological programs, so a story can make you 0:14:39.160 --> 0:14:43.240 shed tears or laugh even if the transformer is just 0:14:43.320 --> 0:14:46.880 pushing around zeros and ones. And therefore we shouldn't be 0:14:47.000 --> 0:14:51.600 surprised that AI can write these really great pieces of prose. 0:14:51.680 --> 0:14:56.280 It doesn't have to be real and it doesn't matter. 0:14:57.160 --> 0:15:00.440 So now that we can write beautiful prose with AI, 0:15:00.720 --> 0:15:04.360 what does this mean for the future of books. Well, 0:15:04.440 --> 0:15:07.200 I think we can imagine a pretty cool future for 0:15:07.720 --> 0:15:14.240 AI generated literature. We can imagine generating infinite, wonderful material. 0:15:15.040 --> 0:15:16.720 And you know what, Back in the day. 0:15:17.080 --> 0:15:22.400 Kings and emperors had poems written that were bespoke. The 0:15:22.560 --> 0:15:25.080 poems were written just for them. And now it's going 0:15:25.120 --> 0:15:28.640 to be trivial for us to all live as royalty, 0:15:29.160 --> 0:15:33.240 having bespoke literature written just for us as much as 0:15:33.280 --> 0:15:36.880 we want, as often as we want, in seconds, and 0:15:36.960 --> 0:15:41.120 maybe we'll come to enjoy dynamic novels, by which I 0:15:41.160 --> 0:15:44.120 mean a piece of literature that's not pre written, but 0:15:44.240 --> 0:15:48.480 instead is written on the fly depending on the decisions 0:15:48.480 --> 0:15:51.720 that you make, like a choose your own adventure. So 0:15:51.800 --> 0:15:54.200 you say this is a good book so far. Now 0:15:54.240 --> 0:15:56.600 I want to see what happens if I go in 0:15:56.640 --> 0:15:58.840 the neighbor's door and get a view on his life, 0:15:58.960 --> 0:16:01.720 or the mailman life who just passed by, or the 0:16:01.760 --> 0:16:05.360 traffic cop and the book just keeps writing itself on 0:16:05.400 --> 0:16:09.320 the fly, thousands of pages that end up being. 0:16:09.120 --> 0:16:14.000 Unique for me, for you, for everyone as they go 0:16:14.080 --> 0:16:15.120 on their own adventure. 0:16:15.600 --> 0:16:19.240 Instead of having some poor author who has to write 0:16:19.480 --> 0:16:23.520 every possible branching path, now there's no need to do that. 0:16:23.560 --> 0:16:25.120 You just generated on the fly. 0:16:26.040 --> 0:16:30.200 So now we'll all get to experience literary worlds that 0:16:30.240 --> 0:16:35.200 are infinite in all directions. So in that light, it 0:16:35.280 --> 0:16:40.320 certainly seems that AI is going to replace human creatives. 0:16:40.720 --> 0:16:43.600 It can do things better and millions of times faster, 0:16:44.120 --> 0:16:46.680 and it can be there to write the next pages 0:16:46.720 --> 0:16:51.400 according to your wishes. So it looks like writers are 0:16:51.520 --> 0:16:53.680 going the way of the mastodon? 0:16:54.720 --> 0:16:56.000 Or are they? 0:16:56.600 --> 0:17:00.720 I think the real story is not so simple. I'm 0:17:00.800 --> 0:17:05.320 fairly sure that while AI will augment human told stories, 0:17:05.920 --> 0:17:09.080 there's essentially zero danger that it's going to do a 0:17:09.119 --> 0:17:12.720 wholesale replacement of human creatives. And I'm going to argue 0:17:12.720 --> 0:17:16.640 this for four reasons. The first is that we care 0:17:16.840 --> 0:17:20.640 about the overarching arc of a story, and at least 0:17:20.640 --> 0:17:24.400 at the moment, AI can't even come close to constructing this. 0:17:24.800 --> 0:17:28.800 And this is because of a fundamental limitation in its architecture. 0:17:29.240 --> 0:17:31.760 And this isn't just a question of pouring more money 0:17:31.800 --> 0:17:34.560 in and getting more massive computers on the job. It 0:17:34.640 --> 0:17:40.199 has to do with the exponentially increasing computational cost of 0:17:40.320 --> 0:17:45.760 representing longer pieces of work. So currently with chat GPT four, 0:17:46.320 --> 0:17:50.120 it looks at the past four ninety six tokens, which 0:17:50.160 --> 0:17:53.240 is about three thousand words, and it decides what the 0:17:53.280 --> 0:17:57.280 most likely next word is. But without getting into the 0:17:57.320 --> 0:17:59.719 details of the math, I want to point out that 0:17:59.760 --> 0:18:03.920 this requires a matrix. Think about it like a big 0:18:03.960 --> 0:18:07.000 spreadsheet that has four thousand ninety six rows in four 0:18:07.040 --> 0:18:07.919 thousand ninety. 0:18:07.640 --> 0:18:10.160 Six columns and an entry in every cell. 0:18:10.200 --> 0:18:13.760 That represents something about the probability of those words going 0:18:13.800 --> 0:18:14.360 with each other. 0:18:14.840 --> 0:18:17.680 Now, this matrix will grow larger. 0:18:17.280 --> 0:18:20.560 With time, but the size of the output is inherently 0:18:20.640 --> 0:18:25.160 constrained by this structure, and as a result, chat GPT 0:18:25.359 --> 0:18:28.760 is perfect for poems or blonde posts or small articles, 0:18:29.240 --> 0:18:33.280 but not something the size of a novel. Why because 0:18:33.320 --> 0:18:38.680 a novel has arcs and plot twists and cleverly planted 0:18:38.880 --> 0:18:42.639 clues and cliffhangers, and all of these operate at a 0:18:42.720 --> 0:18:48.080 longer timescale. So a human author mentally zooms in and 0:18:48.119 --> 0:18:52.720 out such that their stories have this sweeping arc to them. So, 0:18:52.840 --> 0:18:55.159 for example, in a mystery novel, we get to the 0:18:55.400 --> 0:18:58.560 end and we realize that all the clues and the 0:18:58.600 --> 0:19:02.560 red herrings we saw or subservient to the solution to 0:19:02.600 --> 0:19:05.440 the mystery, which of course the author knew from the beginning, 0:19:05.680 --> 0:19:08.439 and the author was just spooling out clues to you 0:19:08.480 --> 0:19:11.080 one at a time. In writing, you often have to 0:19:11.200 --> 0:19:14.639 know the end to structure the beginning in the middle. 0:19:14.920 --> 0:19:18.320 And this is, by the way, why chat GPT can't 0:19:18.359 --> 0:19:20.720 make up a new joke, even though it can repeat 0:19:20.800 --> 0:19:22.160 jokes that are already made. 0:19:22.320 --> 0:19:25.760 But it's because to construct a joke, just like a 0:19:25.800 --> 0:19:29.159 mystery novel, you have to know the punchline first, and 0:19:29.200 --> 0:19:33.119 then you construct the joke backwards. But these large language 0:19:33.119 --> 0:19:37.639 models are simply constructing everything in the forward direction. It 0:19:37.680 --> 0:19:41.879 does statistical calculations on what the most probable word to 0:19:41.920 --> 0:19:45.359 come next is given all the words before it. So, 0:19:45.480 --> 0:19:49.240 coming back to the long arc, if you watched all 0:19:49.280 --> 0:19:52.119 eight seasons of Game of Thrones, for example, or you 0:19:52.160 --> 0:19:55.480 read those books, you come to care about these characters 0:19:55.560 --> 0:19:58.879 because you've been with them through so many trials and 0:19:58.920 --> 0:20:01.600 you feel like you know the and understand them, and 0:20:01.640 --> 0:20:05.200 you can predict things about their behavior, and you're invested 0:20:05.320 --> 0:20:09.399 in their long term trajectories. So all the children of 0:20:09.440 --> 0:20:13.359 the Stark family end up scattered in different directions in 0:20:13.400 --> 0:20:17.320 the world, and then in the final season, they end 0:20:17.400 --> 0:20:21.639 up reconvening. After what seems like a lifetime of adventure. 0:20:21.680 --> 0:20:26.119 They're all back together for the final big showdown with 0:20:26.280 --> 0:20:29.280 the Knight King. And when we watch the series and 0:20:29.320 --> 0:20:32.399 we get to season eight, we think, wow, I didn't 0:20:32.480 --> 0:20:35.359 see that coming, that they're all back together now, and 0:20:35.400 --> 0:20:38.640 now this story has a beautiful shape to it. 0:20:39.119 --> 0:20:42.800 I'm really in the hands of a professional here. At 0:20:42.840 --> 0:20:46.960 least with our current AI architectures today, it's impossible to 0:20:47.080 --> 0:20:50.679 achieve that, except possibly in a few thousand word version, 0:20:51.119 --> 0:20:54.800 because chat ept is playing its statistical game, and of 0:20:54.800 --> 0:20:57.320 course it's playing it extremely well and successfully. 0:20:57.560 --> 0:20:59.960 But the trick to recognize here is. 0:21:00.000 --> 0:21:02.800 That it is amazing at the level of paragraphs and 0:21:02.840 --> 0:21:07.000 possibly a few pages, but not at the level of 0:21:07.080 --> 0:21:10.119 thinking about the details of a five hundred page novel, 0:21:10.480 --> 0:21:15.160 or a two hour movie screenplay or an eight season epic. 0:21:15.920 --> 0:21:18.800 It's great at this small stuff because it can do 0:21:18.840 --> 0:21:22.560 that with statistics, but it's fundamentally limited for the longer 0:21:22.600 --> 0:21:26.320 stuff because it has no way to zoom out and 0:21:26.480 --> 0:21:30.159 think about the crops that it wants to plant for 0:21:30.240 --> 0:21:34.120 the long game. Okay, you might say, fine, maybe we'll 0:21:34.160 --> 0:21:36.760 get there at some point, but even for now, couldn't 0:21:36.800 --> 0:21:40.800 you build a big story out of smaller chunks. So 0:21:41.240 --> 0:21:44.720 one idea is to make this form of storytelling in 0:21:44.760 --> 0:21:46.760 which the world is infinitely big. 0:21:47.280 --> 0:21:48.639 Let's come back to this picture. 0:21:48.680 --> 0:21:51.520 I painted a moment ago of a choose your own 0:21:51.560 --> 0:21:56.199 adventure in which the AI generates plot points on the 0:21:56.240 --> 0:21:59.800 fly for you. So I say, okay, open that door 0:21:59.840 --> 0:22:03.480 to my left and the story continues as though it 0:22:03.560 --> 0:22:08.320 were all prescripted, as though I have an author, let's say, 0:22:08.320 --> 0:22:12.720 in the style of Henningway or Nibokov for Morrison, who 0:22:12.760 --> 0:22:16.560 has pre written every possibility. In certain ways, this would 0:22:16.560 --> 0:22:20.320 be amazingly cool, But I think the problem here is 0:22:20.320 --> 0:22:25.680 that a story like that would just equal randomness, and 0:22:25.720 --> 0:22:29.040 that's not actually what we want in a story. Instead, 0:22:29.080 --> 0:22:32.159 we want to feel like we're putting our trust into an. 0:22:32.040 --> 0:22:33.840 Author who sees the big picture. 0:22:33.880 --> 0:22:38.040 We want the Stark children to reconvene such as we 0:22:38.119 --> 0:22:40.720 feel the overarching pattern of the story and we have 0:22:40.760 --> 0:22:44.960 a sense of completeness. If you just wanted randomness, you'd 0:22:45.240 --> 0:22:47.760 go out into the world and find it there. You 0:22:47.760 --> 0:22:51.480 wouldn't sit on your couch and read about meaningless characters 0:22:51.520 --> 0:22:57.320 who are just in Brownian motion. And I think this 0:22:57.400 --> 0:23:00.800 is the same issue with AI music, at least as 0:23:00.840 --> 0:23:01.520 it stands now. 0:23:02.240 --> 0:23:04.080 Recent examples show. 0:23:03.840 --> 0:23:07.560 That it can compose incredible sounding music moment to moment. 0:23:07.720 --> 0:23:10.119 But the reason it doesn't beat out. 0:23:09.920 --> 0:23:13.560 A real human composer, at least today, is because it 0:23:13.560 --> 0:23:16.880 doesn't have any long term vision, and so the whole 0:23:16.920 --> 0:23:21.240 piece of music just hangs together. Statistically, moment to moment, 0:23:21.640 --> 0:23:25.840 and that's perfectly good for composing things like elevator music, 0:23:25.880 --> 0:23:28.760 which is for a short ride, or commercial music which 0:23:28.800 --> 0:23:31.960 only needs to be twenty seconds. But it won't for 0:23:32.040 --> 0:23:36.359 now replace a human composer who writes with the long 0:23:36.640 --> 0:23:39.439 arc in mind. For example, I was just talking with 0:23:39.480 --> 0:23:42.800 my friend Tony Brandt, who's a composer, and he was 0:23:42.840 --> 0:23:46.040 explaining to me that when Ludwig and vad Beethoven died, 0:23:46.520 --> 0:23:50.679 he left behind sketches for a tenth symphony. So a 0:23:50.720 --> 0:23:55.280 few years ago some computer scientists used AI to complete 0:23:55.359 --> 0:23:59.479 the symphony, to finish what was unfinished. Now did they 0:23:59.520 --> 0:24:02.320 do a good job. In one sense, it was an 0:24:02.440 --> 0:24:08.320 incredible feat. They extracted the statistics of Beethoven's choices and 0:24:08.480 --> 0:24:12.040 preferences from everything he'd written, and they used that to 0:24:12.640 --> 0:24:16.320 statistically guess what moves he would have made next had 0:24:16.359 --> 0:24:20.920 he lived, What notes, what chords, what instruments. But even 0:24:20.960 --> 0:24:23.920 with this feat, it was clear that the AI didn't 0:24:23.960 --> 0:24:27.960 know how to think long term. For example, Beethoven's Ninth 0:24:28.040 --> 0:24:31.840 Symphony ends with a chorus, which was such a surprise 0:24:31.960 --> 0:24:34.000 to end a symphony this way. It had not ever 0:24:34.040 --> 0:24:37.960 been done before, so the team training the AI decided 0:24:38.000 --> 0:24:41.560 Beethoven would have found a similar novelty to end his 0:24:41.720 --> 0:24:45.960 tenth Symphony, so they instructed the AI to include an organ, 0:24:46.480 --> 0:24:49.040 a church instrument that had also never been used in 0:24:49.040 --> 0:24:52.440 a symphony before. So at the start of the last movement, 0:24:52.720 --> 0:24:54.679 the AI generates an organ. 0:24:55.119 --> 0:24:58.639 But when we zoom in, we see the difference. 0:25:01.119 --> 0:25:05.000 The real Beethoven laid all sorts of clues in the 0:25:05.119 --> 0:25:09.040 Ninth Symphony to set the groundwork for the chorus. Like 0:25:09.160 --> 0:25:12.640 the orchestra plays a type of music called a recitative 0:25:13.200 --> 0:25:18.680 before the choir enters. Why because recitatives are found in opera, 0:25:18.720 --> 0:25:22.640 and opera has voices, So he was laying clues down. 0:25:22.800 --> 0:25:26.200 But in the AI tenth Symphony, there was no build 0:25:26.240 --> 0:25:29.320 up to the organ. There was no suspense, no hidden 0:25:29.400 --> 0:25:30.399 clues about. 0:25:30.119 --> 0:25:30.919