Pushkin.
A few weeks ago, I went to Chicago to interview two people on stage about creative work they've done using artificial intelligence. One of the people was Stephen marsh He's a writer. He's done nonfiction books, novels, magazine articles, and earlier this year, he used AI to help him write a short novel called Death of an Author. That book, by the way, was published in audio form by Pushkin Industries, the same company that publishes this podcast. The other person
on stage with us was Lucas Contor. Lucas is a composer. Among other things, He's won a couple Emmys for his work scoring.
The Olympics for NBC.
He co produced a Lord song that was in one of the Hunger Games movies. And the reason he was there talking with us Lucas used AI to help him finish Schubert's unfinished symphony. It was a really interesting conversation and I thought it would make a great episode of What's Your Problem?
So here it is now, please join me and welcome our panel.
So let's do a thing you're never supposed to do in narrative. Let's answer the question right at the beginning. So, uh, the sort of headline question for this panel is will AI kill creativity? I want to ask both of you. I want you to answer in one word at the same time, on the count of three. It's gonna be it's not one two three go, You're gonna go on what it is? One two three go?
Yeah?
Okay, will AI kill creativity? One two three No? No, great done, let's go. Yeah, thank you, Thank you very much.
So I'm delighted to be here with both of you, in particular because you have made things with AI. Right there have been countless panels of people sort of waving their hands.
About the theory of AI or the future of AI.
But I love that we're here talking about things that you have made, creative work that you've made. And so what I want to do is, I want to start by talking a little bit about process. I love talking about how people make creative things. And we'll just do that in order, frankly, just because I want to get into first of the book and then the symphony, and then we can talk more generally about AI and creativity and humanity, and then we can wave our hands in
that classic can'd wave away. So Stephen, let's start with you. I want to read an excerpt from your book, in part because this book that was written with Ai has a very particular, I don't know quality to the pros. There's a really interesting feel to the pros, and I don't know if you'll quite get it from a paragraph, but I want to give you something to hold on to as we're talking about the book.
So I think I have this right.
This passage I'm going to read, it's in the first person, and it's actually in the book, spoken by a digital avatar, an AI avatar in the book, who is an avatar of a dead author whose death is the title of the book. So the passage in her voice goes like this, I learned the limits of machines when they wanted me to fly bombers. They were going to force me to push a button that would end the world.
I hope you can.
Understand that my stance as a pacifist wasn't cowardice or principle, but a confession. I could never bring myself to press that button. Human beings cannot stop making buttons, and once we've made them, we can't stop pushing them.
Pretty good for a machine, it really pretty good for a machine.
Yeah, I'm gonna read that last sentence again because I like it, and because it comes up a couple times in the book. Human beings cannot stop making buttons, and once we've made them, we can't stop pushing them. So maybe Stephen, we should actually start with that sentence.
Right.
It's a great sentence, I think, I or really interesting sentence. Sounds like a sentence a human being would write. It ends up being important in the book sort of thematically. How did the machine write that sentence?
Okay, let me see if I can get it exactly right. So that was the first person from the death of an author. So Jacob came to me in February and said, we need to release this thing.
This is Jacob Weisberg, actually the person who runs Yeah, pushed It. Let's actually start at the beginning of the sure and then we'll get to that sentence. So right, So Jacob Weisberg, who runs Pushkin, which is the company where I make a podcast, came to you in February.
And said, can you write a book that's AI that's generated by AI? In fact, he said, can you create an AI author and then have that author create a book. Now I'd been working on this for a while i'd been working. I'd wroteen my first you know, algorithmically generated story for Wired in twenty seventeen, which was before the Transformer, so the Dark Ages of AI really, and so I said, yes, I can definitely do that. It'll be about ninety five
percent computer generated. I don't want to if I want to change heat to the character's name or something like that, I want to be able to do that without forcing all these iterations and so on. And basically I used GBT four and I would use it to generate texts. I knew from having done AI AI text before that A is very poor at generating plots, okay, and it's very poor at certain other tasks.
It's incredibly good at style, okay.
Right, So I would, you know, have very clear ideas of where the narrative what's going. I'd give very specific grammatical and syntactical commands write a paragraph with high variability, like very very specific commands like wait, do the whole.
Give me an example in its entirety of a command.
It would be almost impossible to do because it's exactly like do it when you've seen them for visual stuff, where it's like they'll just to get really interesting AI generated pictures. You often have like one hundred different references.
Like it almost impossible, but just give me something.
Give me something.
Write a hard boiled detective story paragraph with a variability between short and long sentences and clear, elegant syntax, containing the following information, and then you write out information it would generate that. Then you would take that and I would put it into a program called pseudo.
Right, and wait just before we go to the next program, when you say containing the following information, like that.
One would be it would be like in this one.
The author says, well, that would be slightly different because with characters, I would use a whole different set of commands. So you know the author and here was basically a combination of Margaret Outwood and my dead father. Because I was writing this thing fast, so I needed to know something that I needed to have a character that I would automatically be interested in, and.
I should say, you're a Canadian, like basically the next closest thing after your.
If women were alive for you, yes, right, and so uh, and so that I would say write something like Sylvia Plath meets Philip Roth and meets a bunch of different other things and get hurt it.
So you're doing a very specific character. And then do you do all of the sort of exposition or plot points, like what what in terms of substance, what is an example of what you might put?
Well, I would that would probably actually be mostly the machine, but for plot details would be like she walks to a bridge.
And but this this paragraph about like the you know, the buttons. I wouldn't press the button, and it's like, how do you It would be something like you know it to be something like the character.
Reminisces about her times as a UH and and expounds philosophically on the difference between AI and being a fighter pilot.
And or the character expounds on being a pacifist in the military. Exactly okay, right, and so sometimes more, sometimes less. Tried to get as little as possible, but you know you want specificity here, like you're the more precise the command, the better information, the precise the command, the more it's just you writing it with the weird kind of intermediation is.
My creation, right, this is a tool which you will I will say the same thing. So just the same as if like this is the thing people don't understand, right, it's like, of course this is a creative act. It's just a different creative act, right, Like it's this is one hundred percent me. It's just I didn't write the words right like like like so that's like like that's weird, Like I am, yeah, it's very weird, like I am.
Don't you didn't write the words that ended up in the book word you weren't the words that were the instructions to the machine to write the words that.
Well, so good as any computer that's true, any computer program.
So so okay, so I want to get back to the specific sort of process narratives. So you put this very specific prompt into GPD four, which is basically chat GPT.
I would say, it's actually better fine, and chat EPT four is now it was better than what chatchat is now fine for creative stuff.
Uh. Then you get some output, you get the paragraph for it, and then what and then.
It usually it's very bad, right, And then you take that and you put it in a program called pseudo right. Okay,
and pseudo right is a stochastic writing instrument. So you could you then select the text and you say shorten lengthen you say and then it has another button, which is a customized feature, which is make it sound like X. So, make it sound like Ernest Hemingway, make it sound like f Scott Fitzgerald, and and and you know, the of course, the thing I figured out very quickly is that if you want something to sound like Margaret out With, the very last thing you should do is put in make
it sound like markered out.
That's not enough course to me.
Well, of course, because markered Outwood is in trying to sound like Margaret Out would she's trying to sound like Sylvia Plathmas Philip Roth meets, it meets a bunch of other things.
Right, then you ultimately always get back.
Yeah, And so that when you the way you get interesting things in this text is by essentially folding these layers of style onto each other.
Now I also use and then so pseudo right has some output. Yeah, and then is that output what we're reading in the book?
Correct?
Or you know, if I don't like it, I just try again, just refresh, refresh, refresh until I guess something that I like.
And so so this is very much a creative act.
And you're doing that basically a paragraph at a time.
Yeah, Well, with dialogue, it would go like die would be a lot longer, right, like, because you want flow and you want so I could do up to maybe five hundred words of dialogue at a time. Uh huh, So that would have been part of a much longer series of instructions.
So this sentence human beings cannot stop making buttons, and once we've made them, we can't stop pushing them. A nice sentence, you know, big idea. I certainly didn't think of that.
You didn't. It just came out of some refreshment, yeah, fresh, and.
It was in some I mean, obviously I made it, and I authorized it too. You know, I've compared it in the Atlantic to doing hip hop in the sense that you're you're folding things on top of each other, right, You're folding styles and metrics and effects on top of each other until you get something new and weird.
Right.
And I would say about twenty times during the course of writing it, I felt like I was, you know, putting my hand up against something new and weird.
That's fun, right, like something.
But you know this is for most of the process, it's just a writing tool, right, Like, it writes it for you. You decide if it works, right, and you tell it's you tell it what to write in.
A very granular way.
The more granular, just like writing normally, the more you know about the bigger planning. The more planning you have for any essay, the better the essay is going to be right. And in this case, so you have a plan and then you have the editing process, and in between there's this machine. But how much of that, how much does that matter? Is actually I don't know if it's like twenty times it did matter where it was like, oh that's not something I would have written, but.
It's very beautiful.
Yeah, and it's very strange, and it's you know, there's a there's a Danish journalist who deals with go players who play ai go against each other, and they say it's like listening to an alien make music right, because it's like it's not how they would play go, it's not how a human could play go, but it's obviously makes sense on some level. Similarly, that's how I felt like most of the time, it's just a writing machine that does what I tell it and then I correct it.
But then maybe twenty times you feel this new presence. That's what's exciting.
We'll be back in a minute to hear how Lucas Contour used AI to help him finish Schubert's unfinished symphony. Okay, back to the conversation in Chicago with Stephen Marsh and Lucas Contour. Lucas's story of using AI to finish Schubert's unfinished symphony goes back to twenty nineteen. He was approached by a Chinese tech company called Huawei. They said, we want our phone, which runs AI, to finish Schubert's on
Finnish symphony. And they didn't know what that meant. They had a tech team in place that was running the AI and I knew those people.
That's why they, I think brought me in. I was told that. So my friend, the technologist who brought me in on this project, told me that he thought that I would be a good fit because I have a corporate friendly bio where they could say, oh, he can do it. And he said, I know they you don't have to say that part. You don't have to say that part. He said, uh, But he said I. He said that you, I know you can command an orchestra, but I don't think you'll be precious about the project,
meaning that I won't be. He didn't think I would say like, oh, well, this is heresy. We shouldn't take Schubert's perfect work that was so perfect that he didn't even finish it and do something with it. And uh yeah. So I think they thought they would just that I would press a button on the phone and a symphony would come out and somehow a bunch of musicians would play.
So they need you for it. They just pushed the button.
So this is the conversation we had, and eventually I had to I was on a call with them and I said, look, this is this is not I mean, what you're asking for in principle doesn't exist, like you can't And I mean, what do you even want the machine to do? Do you want it to generate audio for you? Do you want it to generate a score? Do you want it to perform the score? So, I mean, right off the bat, this was a fascinating project because I had to think about the very nature of music
to even really get started. I don't know if that answers the question about I think it does.
I mean, I just wanted you to set yourself up, and I think you've done it.
You want to I think I'm set up, so I'm gonna try something new for you today. So on the on the prep call for this event, we discussed I said something that I don't often say out loud, but I realized as a hallmark of my presence on stage, is that I like to do things that might spectacularly fail in the hopes that they will be entertaining to an audience. So I'm going to do one of them for you.
Now.
I'm going to I wrote a little thing about the Unfinished Symphony. I'm going to explain it while I'm playing some music in the background and basically scoring it as i'm talking. So you know, wish me luck and hopefully it'll be interesting. This is how the Unfinished Symphony starts. A symphony has four movements, but Schubert only wrote two and sketched a third of his eighth Symphony, the Unfinished Symphony.
No one knows why he abandoned the Unfinished Symphony, but he did, and now it's probably his most famous work, along with his greatest hit, Ave Maria. Some scholars believe that Schubert couldn't find a way to fit the Eighth Symphony into the orthodoxy of the time. Which forbade three movements in a row in triple meter meters like three, four and sixty eight.
But I don't believe this.
Schubert showed little reverence for orthodoxy during his short life, and the AI that I used to finish Ubert's on Finnish Symphony didn't believe it either. At first, we trained the AI on recordings of Schubert's entire catalog, then prompted it with the first two movements of the unfinished symphony. Seems like a reasonable strategy, right, This was the result sounds like Kat's walking on a piano, But this was
actually pretty logical. Recorded music has almost no mathematically discernible patterns to it, so from the AI's perspective, the input was nonsense, so more nonsense was a logical output. Music as an abstraction is math, but music in practice is convention. Music is understood by groups of humans, and like any language, music doesn't have objective meaning. Music is emotionally inert left myself. A water break is symphony. A symphony is like a skyscraper.
It's enormous, but every inch of it is designed in meticulous detail. It's beautiful on the outside, but the inside is filled with utilitarian solutions to simple problems. A skyscraper has electrical columns to distribute power throughout the building, It has plumbing, it has elevators, but you don't see any of this essential detail when you admire the building from outside. A symphony is like a skyscraper, but a recording of
a symphony is like a skyscraper's facade. There is no way to tell from photos of even a million facades that skyscrapers should have electricity, bathrooms and a way for humans to move from one floor to another. Similarly, there is no way to tell from the morass of frequencies that is a piece of recorded music which frequencies are the most important.
There we go.
So analyzing recorded music got us nowhere, and I thought that the best way to proceed was to simplify the task and just train the AI on the blueprints of music rather than a finished building. So train the AI on a blueprint rather than a finished building. So what you just heard, what you're hearing now is the main theme from the unfinished symphony. Here it is again, just really listen and try to listen for the melody. And here is that same theme reduced to its blueprint. This structure,
this blueprint in music, is just a simple melody. So my team and I went to work extracting just the melodies from as much of Schubert's music as we could get our hands on. These are some examples of the melodies we extracted. These sound robotic because they are. They sound emotionally inert. But these are Schubert's melodies reduced to their simplest forms, the forms that human composition students would
use when beginning a study of Schubert. Your ear knows how to pick a melody out of a dense arrangement, but an untrained AI cannot do this. The reason that, since the results we wanted were simple, we needed to train the AI on simple data. We trained on hours of these simple melodies and then prompted again. We prompted it with the unfinished symphony reduced to its blueprint, and these were some of the results. So this is what it suggested might be something that Subert would have written.
These are simple, but much more musical than the cats walking on a piano that came from the audio only training data. This one, for some reason, caught my attention. Let's hear it again. I liked it, so I selected it for embellishment. I decided to use this. I decided to use this blueprint. This melody is a bit more modern sounding than any of Schubert's work. If Schubert lived to old age, these sonorities would have been available to him.
The orthodoxy around triple meters and other constraints of form would have given way to the exploration of the Romantic period. Providing simple singable melodies is perhaps not how most people would imagine that an AI would be useful in writing a symphony. But what is a symphony? Typically people think about a symphony as something that you hear, while the score is just a byproduct of the notated sounds. But to me, the sound is a byproduct, and the symphony
is something that you see. It's something that you read. It's a collection of abstract ideas in abstract notation. It's markings on a page that serve as instructions for how to create sounds. A symphony itself is a blueprint, and those instructions that blueprint will be executed differently at every performance. Let me just check out this music. It's pretty cool.
The sounds are a byproduct of the abstractions that are expressed in the notation, and that byproduct is what the audience experiences as a symphony.
The byproduct is what you hear.
I didn't know that I thought about music in this way until I had to explain how I think about music to a machine. This project taught me to question the assumptions I make when thinking about my own craft. I think this is the job of the AI assisted composer today to think about what we know and to guide our audience to rethink what happens inside their own minds. I think it's our job to question orthodoxy. I think it's our job to use new tools to make new art.
Today's artists are not on the verge of being replaced. On the contrary, we are possessed of powers so great that we will expose more truth about the human mind and the human soul than any generation before us. We stand on the shoulders of giants. They have given us the language, they have given us the blueprints, they have given us the technology. What we build with these tools will be more powerful, and more beautiful, and more profound
than anything we can now imagine. Artificial intelligence is nothing like us than a prosthetic for the human mind. It will enhance art the way writing enhanced memory, the way printing enhanced literature, the way the steam engine enhanced travel. Artificial intelligence is an automobile. We're only beginning to emerge from the age of horse and buggy. Artificial intelligence helped me write the music that you're hearing right now. So will AI kill creativity?
No, that's really rather Good's that more or less worked? I think that's really rather good.
Thanks.
We'll be back in a minute to wave our hands a little bit about the future of AI and creativity.
That's the end of the ads.
Now we're going back to the show.
The reason I knew AI was going to take off was when I was writing a piece for The New Yorker about GPT three and I got it to finish off Coleridge's Kubla Khan is great unfinished poem, and it did it perfectly well. Like I mean, if somebody told me, yeah, this is how it ended, I would have been like, great, right and so, And it did it like that like one second.
I mean, it was just so incredible to me.
Just to sort of close this part of the conversation, I'm curious. I mean, both of these projects. We were very AI forward, right, They were like high concept, you know, sort of let's explicitly wrap this thing in AI.
Fine.
Interesting, But presumably the real action comes in the things that are just what you guys are working on that just happens to have AI as a tool, the same way say a Google search, which by the way, is a kind of AI, is also a tool, right, And so I'm curious in your work now on other projects that are not like, hey, look this was made with AI kind of projects. Are you guys using AI? And if so, how what do you want to go first?
Yeah? Yeah, first, so yeah, obviously of course, like it's in everybody's pockets, you use it all the time. And AI has done nothing so far other than help my career. And I don't mean just by doing this, which was fantastic. But when I write a piece of music and put it on Spotify, the reason you hear it is because an AI recommended it to you. You know, that's the
only reason you're going to find it. And so and these types of algorithms that are generating that are keeping people out on apps longer and keeping people on Netflix and on Spotify longer, are putting money not enough money, and that's another panel discussion, but putting money in our pockets directly?
Let me let me ask a more precise version of the question in response to that clever answer.
Do you use generative AI?
Yes?
And also this is a terminology problem.
But you know what do you use music?
Do you use AI to generate musical ideas for you?
Yes?
But also like what is a musical idea? I use a parametric eque that I mean they were using a they were using this was there was probably good. I'm trying, well the answer the answer is yes.
I know what you're saying. But I feel like you know what I'm saying.
Well, yes, I'm The reason I'm trying to drill down here is because this there tell me how to ask the question I want to asking doesn't have the answer that you want?
Right, So fair, what's the what? What's the smarter version of the question? I'm not well enough equipped to ask.
I don't know if I can.
I don't know if I can help you with that.
I don't let.
Me ask the question to you.
Thank you for your Stephen.
Do you you use generative AI when you're writing with other things?
Okay, here's the thing, and I think this is sort of where we're going. Like I would when I write something for a magazine or newspaper or novel that I'm working on, I would never use chatchipt.
Even to get an idea because here or whatever they because I'm so much smarter than chat GPT.
Right, And I'm like when you and what you have to also have to understand is chatchypt. The reason it's so successful is exactly that it has been banalified, like when you use other generative ais that we have access to, because you realize that like these are the ones that the public uses are very poor creatively, like they're actually.
But you have access to the good ones, to the good stuff.
Here's the thing you can't get on when when you use the good stuff.
What the good stuff is going to be used for stuff that doesn't exist yet. What we're seeing here is the birth of a new medium, right and what and so when it comes to write an essay, what people want when they write, when they read an essay, is a human being communicating their thoughts and feelings, right, they don't want like they don't That's why they go to it. And a generative AI cannot do that generative Like it's sort of like asking, like do you use film to
make theater? Like at first, you know, when you when film was invented, all they did was cannibalized theater and they were putting on weird shows or they were recreating news events and things like this. That's where we're at right now. This is going to be used for new art forms that don't exist, and that's that's the exciting stuff. And it's also why it's almost impossible to do.
You mean, like the book that is never done, the book where it can or like what like.
I'm written that I have written a short story that is infinite art forms?
Like what do you have in your mind when you say it?
Well, like, for example, I'm working with cohere to recreate the Oracle at Delphi. Right there's a large amount of information that you can glean from that, and there's also pretty interesting historical record.
And so you'll ask it a question and it will answer, yes.
We're try and recreate the experience of going to the Oracle at Delphia as closely as we can use effects.
Yeah, it's a perfect use of AI and so oracles. This is one of the things that has come up in my research is that we use oracles because we're bad at doing things randomly. So if we're out in the wilderness, we'll just go hunt in the same place over and over and over again, right, And eventually animals figure it out and they say, just don't hang out there.
And you won't get eaten by the humans.
And so when we like consult an oracle, or roll some dice, or like ask the sacred chickens if we should go to war, they're basically giving us a random answer.
That's right. There are randomization engines, see, and it's.
Things of this nature that I think will be that I'm excited about to use it. We're cannibalizing forms. That's what I do writing short stories to It's very interesting. But the truth is that what this can be used for we don't know yet, and what it's going to be used for is some weird and the problem is there's absolutely no institutions to do it with, right, Like.
Nobody will buy your oracle of Veli's.
Supposed to take oracle of Hi.
My name Stephen, I'd like to recreate the oracle at DELFI using generative AI. I'm sorry, sir, this is a key mark, you know what I mean, like like it like it's not that's that's not like there's no one to go to. So that's that's where we're at. To me, Like, I think the the the thing that I think is very obvious is that when you use generative AI, what it is very good at is the most stock answer, right.
And that's why it's so such a threat to like the undergraduate essay, right, because that there you're basically looking for the fulfillment of a stylistic you know, set pattern that it can do.
Right.
But people respond to human like there's this weird idea that art is something external to our experience of it. It isn't. It's just we we have we create tools. As the moment we find tools, all we're thinking of is can we do something weird with it? And I think, I mean, one thing that I've really learned doing this
is that creativity is instructible. Like it it doesn't matter what comes down technologically, what comes down politically, what Like, we are creative animals and we have to understand that that's just our nature and nothing is gonna kill it, nothing, not certainly not chat gept.
Great I can I can sum up the history of music from the year sixty thousand before present to now with one sentence, and maybe you'll agree that this sums up the history of art already. It's the search for new sounds. Yeah, that's it. That's all there is to it. If something exists, nobody cares and chat geept I will chat. Chapet doesn't do music. But there are many music generative ais, and they generate music that, like charitably would call insipid.
Yeah it's fine, like it's music. You would recognize it as music, but nobody. You wouldn't listen to it. It'll get bad music. It won't, so it'll sound better, it'll sound better. So this is the but, but nobody cares
about that. So as soon as like, as soon as you can have so Jacob for your podcast, as soon as you can have beautiful sounding orchestral music like this for free, you're gonna want something else because this is available and it's everywhere, and so what you're gonna what you're gonna want is like the thing where like Lucas
plays a guitar with a really nice sounding reverb. That's gonna be the style and you can trace and we have a we have a composer in the audience who could, hopefully will agree with me on this, and a professor of this kind of thing. But you can trace musical styles in media, and it's like whatever is ubiquitous just falls out of fashion and then that whatever the opposite of it is becomes becomes fashionable. So yeah, that's my that's my two cents the search for new sounds.
Thanks you guys. This is closure. Yeah.
My conversation with Lucas Contour and Stephen Marsh was organized.
By Chicago Humanities.
Today's show was edited by Karen Chakerji, produced by Edith Russolo, and engineered by Amanda k Wong.
You can email us.
At Problem at pushkin dot fm. We are always, always, always trying to find interesting new guests for the show, So if there's somebody who think we should book, please let us know. I'm Jacob Goldstein and we'll be back next week with another episode of What's Your Problem.