Do Robots Write of Electric Sheep?

some way or participating within a story in some way. Now, let's talk a little bit about, um, what the stories that we wish existed but don't like. Can you imagine some of the greatest authors of all time, who who maybe died before you felt like they had really produced everything they could have produced? Like, do you have any favorite authors that you wish had lived longer to tell at least one more amazing story? Oh of course. Well,

actually what first popped into my mind was poets. I'd love to have more of Uncle Walton Anti m when I'm feeling when I'm feeling patriotic, you know, Walt Whitman, Emily Dickinson, I gotta feel the America, you know, coming up right bubbling up inside you. But they've been dead a long time, right, So, uh well, Lauren, are there any authors that you could think of off the top of your head where you're thinking, Man, I really wish I could have had one more book or story or

play or poem from this person. I'm a completely terrible person. The first thing that just popped into my head was George RR. Martin, who is not actually dead as of So you want you want an insurance policy to guarantee that a Song of Ice and Fire is complete at some point, and you want it in George RR. Martin's voice. That's really important. Obviously someone else could pick it up, but that wouldn't be the same. Yeah, no, no, no

one else would write. And I'm being actually very earnest here the kinds of descriptions of food and dresses into my political fantasy that he writes. Well, you guys have both made my choice of Shakespeare seemed pedestrian at this point. I'm, of course, I am a Shakespearean nut. I love Shakespeare's plays and I love his poetry as well. And uh we they're already some plays that may or may not

belong to Shakespeare. We'll talk more about some of those in a little bit, but uh, you know, there were there's some lost plays that perhaps we might actually have if he had lived longer to get to the point where he was committing these so that they could be published. You know, in Shakespeare's day, he wasn't It wasn't what

he was writing them down necessarily any anymore than too. Yeah, he would get the sides to the actors, but the whole purpose of the play was for performance, not for reading. So in Shakespeare's day, very few of his plays ever were published. It was only after his death when his contemporaries, Yeah, his contemporaries began to gather them all together. People who worked in the King's Men production company began to gather

as stuff together. And then even then there were points where scholars have argued whether other plays should have been included. Perhaps there were plays that Shakespeare wrote when he wasn't writing for the King's Men that weren't included. And if you have been attributed to to other authors. Sure now people think Shakespeare didn't write any of those plays, They're wrong. So I would love it if if Shakespeare could turn

out a few books. I would love it if Dickens could finish the Mystery of Edwin Drewd because there's nothing more irritating than a mystery that ends halfway through. He killed him? Or did he even die? Oh? Yeah, he may have just disappeared. There's a there's a science fiction book that C. S. Lewis never finished before his death that I would very much like to read the ending too.

I would have liked to have read The Simarillian as if Tolkien had actually written it in a form that was meant to be read by a human being, as opposed to a collection of things that his son pulled together from various works and notes and then put into some kind of bizarre encyclopedia. Yeah. I mean, don't get me wrong, I nuinely love The Simarillian, but I wonder what it would have been like had Tolkien actually himself

put it all out together. So, I mean there are lots of different examples, obviously, uh Hunter S. Thompson, Kurt. I mean, there's some authors out there I would love to have been able to read some more stuff. And Charles Dickens, Oh, I already mentioned him. My anglophile traits are starting to show, and there is always fan fiction. I mean, you know, some some fan authors can write very convincingly in an author's voice. Well, what if we could create a fan fiction style author that wasn't really

a person. You know. Before I get into that, we're gonna have to talk a little bit about some some terms here, like artificial intelligence. Artificial intelligence, What does that have to do with literature. Well, we're gonna get there. We'll take some steps first. So artificial intelligence or machine intelligence, this is a concept that's been around for more than

half a century at this point. Uh, And it was something that was kind of of played with early early early days in in the earliest days of computer science, in fact, before we even called it computer science. One of the fathers of the concept of artificial intelligence, you could argue, is Alan Turing. Now Touring was a computer scientist before there were was a word for such a thing,

who worked for the British government. He was working within the War Department to help decode messages that were sent by Germans using the Enigma machine you guys familiar with. He was one of he was on the team, and yes, he was highly responsible for cracking the code of the

Enigma machine. The Enigma machine was a physical mechanical device that the Germans were using that you would set to a particular setting, had certain reels, and then what you do is I had like a typewriter and as you would type, the reels would channel a an electric signal through so that a particular lightbulb would light up on the other side. But it was designed in such a

way so that would encode eat letters. So if you typed a G, one thing you could be sure of is that the one bulb that's not gonna light up on that board is G. And in fact, that was one of the few things that Touring was able to figure out that that broke the code was knowing that whatever the letter was that you were looking at at an encoded message, it was definitely not exactly it'll and you would think eliminating one choice would not give you that big of an advantage, but that was kind of

the crack that really helped break it open. Okay, So what did Alan Turing have to do with the idea of artificial intelligence? So he was very much interested in this idea of what is what's the capability of the machine world and are there limits? Can it go as far as humans can go? Can machines think? Can machines think? He thought about that in such a way It's a little different than the way you might initially imagine a

machine thinking. Uh, Touring said, it wasn't so much. It wasn't really important if machines thought the same way that people do. In other words, what was necessary from Touring's point of view, was that the output of a machine would be indistinguishable to what a human would be capable of doing. But the process doesn't. Yeah. So in other words, if the machine was able to produce something that you seem the same as the way a human would produce it,

it didn't matter how the machine did it. It just meant that that machine possessed some form of intelligence. Uh, And I'll explain that in a second to In nineteen fifty he published a paper Computer Machinery and Intelligence and propose the idea of the touring test. Now, in general, today, the way we talk about the touring test is, uh. You. A simple version would be you have a computer in

front of your computer terminal. All you get is text, and your text is going to someone that's in another room, and that might be a person or it might be a machine. So you're you're like in a chat set up. Yeah, yeah, and you cannot see the other person at all all or the machine at all. All you see is whatever shows up on your screen and you type in a message and a message returns back to you that's generated

either by a person or a machine. The Touring test said that if you were unable to determine to a certain degree of certainty whether or not the the the thing on the other end was a person or a machine, that machine was said to pass the Touring test. So if it could imitate human speech through text, yes, So if the machine was able to carry on a conversation in such a way that it would convince you that it's another person, then it could pass the Touring test.

It would at least seem to possess intelligence. And here's how Touring kind of thought about this. It's kind of a philosophical way of thinking about intelligence. So I, Jonathan Strickland, I know that I possess at least some level of intelligence because I have that personal experience. I know what it means in an abstract way to myself. When I talked to you, Lauren, and you start to display several

of the same things I associate with an intelligence. I then assume you too, possess intelligence of some level, and that that is just something that I am granting to Lauren. From my perspective, because I cannot inhabit Lauren, I can't know what her experience really is. I can only base

that upon the observations I make of her behavior. Touring says, well, if we do that, if you assume that every person you encounter has some level of intelligence, why would you not offer a machine the same courtesy If it seemed to display the same sort of behaviors, You might as well say it has intelligence, because there's no way that you can inhabit that machine, just as there's no way

you can inhabit another human being. And in fact, saying that we know the machine is different because we programmed it might be an example of the genetic fallacy, right that the idea that because you know where a phenomenon came from, that proves it's not genuine. Sure. Yeah, and so when we get to the actual touring test. We've had plenty of people design various kinds of of software.

Usually they're called chatbots. These are just programs that are designed to interact with people and respond in a conversational way through some context. And uh, and some of them have been pretty successful. Surely they could never really trick somebody, right, I mean, I've talked a smarter child on instant Messenger. I don't think that's a person. It all depends on again,

the context and how well designed the program was. So for example, one of the earliest that was designed was Eliza, which was created by Joseph Wisenbaum, and Eliza was effective within a certain context, and that it was it was presenting a point of view of almost kind of a naive person, someone who has very little real world experience, and so as long as the real human being had

that same kind of perspective, it was fairly effective. There was another one called Perry p a r R Y that was made by Kenneth Colby, and that one simulated

a paranoid schizophrenic responding to questions. And what they did was they ran a whole bunch of conversations between an interviewer and actual patients who suffered from paranoid schizophrenia and conversations with Perry, and then printed out all those conversations handed that to a panel of judges who were made up of psychologists, and they had a success rate of forty eight percent of saying which one was real in

which one. Now, in that case, you're talking about a specific context that has its own limiting factors, right sure, yeah, yeah, you're you're you're talking about two different, um, you know, hypotheticals of a human person that aren't the kind of conversation that you would normally have straight And in the case of Perry, you're talking about a conversation that was

not conducted by the actual panel of judges. Right, So their experience is very different from someone who say, it's down at a terminal and it's actually having the conversational moment. They're reviewing something that's already happened. That's different too, I mean that experience is different. But there are other examples too. There's the artificial linguistic Internet computer entity also known as Alice, and there's a one that was made by Rollo Carpenter

called Jabberwocky. Was brilliant, the slightly topes uh that one was meant to simulate human chat in a very kind of humorous way. I actually tried this one out before we came to the podcast. I wish I had been able to print out our conversation because I mean, you could tell that it was or I could tell it was a chat bot. But for one thing, I already knew, So that that's problematic, right, I mean I already know

going into it, So yeah, very scientific, right. I would need to have like a double blind test if I really wanted to do this properly. But also there were other giveaways, like you would it would ask you a question and I would answer because it was being snarky. I was being snarky. I answered in a snarky way, and then it didn't know how to deal with that, like it didn't have enough keywords to work off of, so it would go with a uh, like a stock

response that was just a generic response. And in fact, that's the way a lot of these chatbots work is that they search for keywords and the things that the person the interviewer is typing in, and then it generates a response based upon those keywords. And if none of the keywords that it normally quote unquote knows as in it keywords that are in its database, then it will generate some other form of response, either a generic response, or it will repeat something that it had already said

previously in the conversation. So what are some of the ways that we think a chat bought could get really convincing? Because I'd imagine that we're not there yet, right, but we're getting better. We are getting better, um um. And obviously there's some kind of hurdle we'd have to get over there. There's some like strategy for supplying these things with conversation asian rules that we haven't quite achieved yet. I mean, would uh, would machine learning have anything to

do with it? Sure, like the idea of maybe mining thousands of other conversations to establish rules about how real people interact, or even having every time the machine has a conversation that it ends up reviewing that afterward and learning from its own experiences. I mean, these are right because you know, it's human. Human vocabulary and human understanding. Don't stop when you learn a language. We're continually learning

our own languages. We're learning new rules for it, we're learning new vocabulary for it, we're creating things like metaphor. And similarly, we're creating these ideas, these abstract ideas that makes sense to people who are native speakers of that language, who have had exposure to this sort of thing, But to something that is just creating, you know, working out of say, a very strict dictionary of words, it would be meaningless or confusing. This reminds me of that time

that that IBM t um. Remind me to remind me the name of the computer. Um. They yes, they taught Watson Urban dictionary. Oh no, yeah, oh no. Yeah. They then then after after Watson started talking back at them a little bit too much, they were just like, let's just nuke that portion of Watson from more. Yeah, it's the only way to be sure. Yeah. It turned out that Watson Watson developed a bit of a potty mouth, like, Um, are you familiar with the show Breaking Bad? Do you

know how Jesse ends most of his sentences. That's pretty much how Watson was ending his sentences. Um. So yeah, they found out it was I think Urban Dictionary, and it was one other thing too that I think it might have been. It may have been Wikipedia, but I'm not sure. But they fed that into how to talk. Yeah, well then it just becomes unintelligible, right, hacks or yeah, okay, so uh I think I've seen I could be wrong.

About this. What I think I remember seeing is that most recently, Ray Kurzwild said he thought that chat bots would reliable, you be able to beat the Turing test. And to be fair, when we say beat the Turing tests, there's no hard and fast rule. I think what that means, reliably fool the judges, right, yeah, because there's it all depends upon whom you ask, Right. Some people say you need to fool people at least thirty percent of the

time and then you beat the Turing test. Other people say, well, Touring had suggested that you show it to a panel, and if the majority of the panel thinks that it's a real person, when in fact it's a computer, then the computer beats the Turing test. It all depends on how you frame it. There is not like a magical This is the way the test should be administered, and this is the only way you can tell if you

pass or fail. I just want to make that clear because when we say passing the Turing test, that's so fuzzy. We really are talking about if you the listener, you you who are listening right now, if you were to have a conversation with one of these machines, you would not be completely certain whether or not that was a machine or a person. We see this actually used a lot in corporate settings for things like UM customer service.

So if you ever have one of those customer service things pop up when you are on a website and it says do you need some help? Often this ends up being a chatbot that actually just has a very deep series of keywords that relate back to the products and services of that company, so that when you start typing things in, it can start sending you two links that possibly can solve your problem, but more likely will cause you to go into a red hot rage, burning

brighter than a thousand exploding suns. Of course, they're the goal is to get you to the information or or service that you need. It's not really to trick you into thinking like, wow, I really just formed an emotional connection with this person, or the goal is to cut down on the cost of manning actual human beings and customer support roles. But that your your point stands either way. UM, so I would like to imagine a future where we're

proposing the next Turing test. Okay, so we've gotten to the point where we can have a versation and not be sure if it's a person or a machine. Yes, so, either because the machines have become really smart or people aren't just dumb as bricks. Imagine we get there. Let's say it's nine kurs wild was right. Um now we've got chat bots that people cannot tell the chat bought from a human that every single time, you know, it's statistically and significant. Um uh. Once we're there, what's the

next step. What's the next hardest thing in that same sort of realm of imitating human intelligence? Well, Joe, since this whole podcast is about storytelling, I'm going to take a wild stab in the dark and say it's storytelling. That's my guess. I think that'd be really interesting. I mean, so we can sort of imagine, like the chat bots we've interacted with today, they can't really do conversation, but

we can imagine it. We should see it storytelling. Now, on one hand, I can imagine how an AI story bought would work. But on the other hand, it seems so alien into me. Surely they couldn't really create a piece of AI that could tell stories that seemed as good as human stories, right, I think it would be a huge challenge, But I don't see anything that's fundamentally

impossible about it. Now, well, I mean if you take the fact that there are programs these days that can create music or or create works of visual art, then you know, and those are based on a few thousand rules about what makes a good piece of whatever media. Right. Okay, Well, well let's let's set it up. Now, let's describe the literary touring test. Okay that what what kind of test

would a story bought have to pass? Well, you'd have to be able, I would say, to read the story and feel like it was saying satisfied, for lack of a better word, organic that that it did not it was not just a series of choppy sentences where a person goes through some mundane task. It would need to be uh, it would need to be engaging. I know you love that word, Joe, I keep using it because

of that. But it would need to It would need to captivate an audience in some way, either because of the actual plot or the characterization or some combination thereof. And it would need to It would need to have a narrative flow. It would need to have internal consistency. So, in other words, things that happen earlier in the story could not be contradicted later in the story. Let's say that you have a story about a father and uh, the father experiences the heartbreak of a child dying, and

that's a terrible thing. This is something that we would encounter in a story on novel whatever. But then three chapters later, the child's there and nothing has gone wrong, and there's it's not like it's mystery or I think it's just that the child had never died. And and either you know, either adhering to those rules, or if it breaks those rules, to be breaking them in a way that seems conscious and purposeful. Exactly, it couldn't just

be breaking rules indiscriminately. Here's my blanket rule for the literary touring test I've imagined. Um, it doesn't have to create stories that are as good as your favorite author. It just has to do in the same way that the original Turing test didn't have to be as great a conversationalist as your most interesting friend. It just has to be good enough that you think this is probably written by a human being. Sure, I've read some awful stories,

so the bar is set fairly low. Um, we're going back to that fan fiction thing now, But you've read some awful stories. But even people who are bad writers write better stories than than we could generate now with a I some really bad stories. I've read some really

bad things. Well, let's let's just play this way. The way the way a computer would generate a story right now, the only way I think you could have one where it would have any sort of of actual sense to it, unless you were to have an incredibly powerful computer and hundreds and hundreds and hundreds of hours of programming time built into it, would be if you went the mad Libs approach, where you had a story outline that already existed and the computer was just filling in the blanks.

And even then it's not necessarily going to make a whole lot of sense. Of course, that that wouldn't be generating, No, that would. In fact, I've seen generators that do this

kind of thing where they they follow very simple rules. So, for example, my father created a program in Apple Basic back in nine five, I want to say, so, yes, we had computers back then, uh, And he created this program that was a science fiction novel title generator, and it had just very simple rules where it would take some ridiculous adjective and ridiculous nouns and pair it up together so that you would get a sentence or or title that would just sound like yeah, I could totally

see that being on the store shelf in the science fiction fantasy section, the burning lamp, Shade of Venus and exactly that would be. That would be a great example. It reminds me of I saw online a while back at James Bond movie title Generator, but that one spit out horrible things. It was like the Last Day to Gun or like, oh yeah, because it was my favorite was ball Eyes. It's because it was taking it was taking existing James Bond titles and breaking them up and

then putting them back together. So instead of instead of creating a really deep, uh database of words to work from, it was just like, let's put in every title of j so Thunderball and Golden I maybe or something like that. Yeah. So so. So the components that we're trying to smush together here to to create this literary Turing test passing machine are are that that vocabulary. You want to have enough vocab so that you're not just ball Eyes to the next century. The most basic it would just have

to know what words mean in a robust way, right right. Uh, you know, it needs to be able to work with the tropes and metaphors of of the culture that it's writing for. I mean, because you know, a lot of books will contain a lot of stories, contain references to countless other stories, and are using um, the bits and pieces the devices and conventions of storytelling that a culture is used to write. There either either deliberately playing on

those tropes or they're deliberately defying them. But in either case, it's something where the awareness is important both for the storyteller and the audience. It's got to have an advanced level of machine learning, which you guys mentioned earlier, which is which is the science of getting a computer to do something without explicitly having told it to do it, right, because there are just too many rules you could not

possibly code them, right. Sure, Sure, it's it's you know, machine learning is what's behind something like telling a Google car to go down a street without it ever having seen that street before. Um, you know, it's it's also what goes into into web search and lots of other you know, the Human g Project, everything like that is based on machine learning. UM. And and then a basic I mean not basic because this is actually really huge.

And ontology a semantic or abstract model of data UM that that builds upon databases which are logical or logical or physical UM. So you know, it's the term comes from philosophy, where it's used to describe studies of nature or existence, and it was picked up by early AI researches. Researchers to um to define the the objects, concepts, and other entities that are presumed to exist in in some area of interest, and the relationships that are held among them. Right, so, uh,

let's make this concrete. Let's imagine a story. Okay, say there's a story where there's like a murder or something. Uh, you know, Colonel Mustard kills Mr. Body and and this is a murder mystery about the story, Well it the computer that's generating the story would have to have such a robust understanding not just to what a word like knife means, like not just that it's a noun or whatever, but that it has relationships. So it can be found in the kitchen, it can be used for murder. Maybe

it's found in Mr. Body. Yeah it can. But it can also be a clue. Sure, but there's also metaphors about being on a knife's edge or so to understand. Yeah, it wouldn't just have to have knife in like its library, but it would have to have a connected web of relationships to other terms. Furthermore, you would have other relationships you'd have to determine, like the fact that you would have uh, you know, uh, Colonel Mustard as murderer. Mr Body is victim, but you might have Mr Green as suspect.

You might have Scarlett as suspect. You know this. These would be these would be different definitions. And then not only do you define everything, but then you have to relate them all to one another. Right. That's the big thing that would make it even more complex, because in order to have good fiction you have to have strong

characterization and relationships between the characters. So you need ontologies that defined characters by like is in love with or has a grudge against, is afraid of that kind of thing? Or uh, you know, for example, in our our clue or clue, oh example, we could go back to uh, not only is Mr Green a suspect and Colonel Mustard is the murderer, but Colonel Mustard views Mr Green as

a patsy. Now, the other characters would not view Mr Green as a patsy, and within the narrative of a story, that would be clear to the audience but not clear to the characters within the context of that story. It's a very simple thing for us to think about as

human beings. This is something that comes very naturally to us. Obviously, when you're watching a movie or reading a book, you are aware of certain things that other characters are not aware of, and if those characters magically become aware of them, that pulls you out moment. I could see that being a very challenging thing to to quote unquote teach a computer to know that you have traumatic Irony exists and can be used in these ways, but not in these ways.

And because and just because one character is aware of something does not necessarily mean that another character is aware of that same thing. I mean I I had an experience like that this week where I was watching something and I thought, wait a minute, how does this character even know this? And I totally pulled me all the story. Yeah, um,

so it's clear just how difficult this is. How how I mean it, I'm gonna say frankly, it seems impossible to me now, but I wouldn't say it's actually impolsible. I would have said that a computer beating Jeopardy champions at Jeopardy would have been impossible, you know, a couple of years ago, and then IBM with Watson totally proved me wrong. And they did it in a way where Watson was not even connected to the Internet. It was

all self contained. So the fact that there could be a self contained database of enough information to be able to anticipate practically anything that Jeopardy can throw at you. And Jeopardy is a game that includes things like wordplay, where it's not just a standard here is your answer, what's the what's the correct question? You may have to interpret that answer beyond just what jokes and and et cetera.

Now there might be puns, there can be references. So the fact that we were that we the fact that IBM was able to create a computer that was capable of doing this, I I hesitate to say impossible. I think it's an incredibly difficult task. I would hesitate to go so far as to say it's it's an uncrackable task. Yeah, so so, so what's it like right now? This is what we are imagining in the future. I hear that

you have some examples of computer generated poetry. Yeah, well, okay, so obviously I think it's pretty clear poetry is easier I think for machines to generate than fiction at this point, because because the rules of poetry are a little bit looser, Um, we expect poetry to be a lot of people expect poetry, you know, you can expect it to be associative rather

than narrative. And it's actually I mean, it's got to be really difficult to get a computer to generate a coherent narrative and less it's just following like a pre supplied skeletal structure. Yeah, yeah, to tell a computer, Hey, tell me a fantasy story. Here are four characters you have, and here is one, uh, one important piece of information that you need to incorporate. Go. That would be like like you said, Joe, right now, that's really hard to image.

It's impossible, you know, it wouldn't be able to I mean, the computer just doesn't understand events and the progression of events enough to tell a coherent narrative from one you know that that feels meaningful in any way. I mean why didn't the eagles just give them a ride to more to or at the very beginning if they could? All? Right? Right, but but but speaking speaking as one who basically majored in poetry in college and very nearly went on to do an m f a in poetry, Um, writing good

poetry is really hard. It's hey, I am, I am right there with you. I also studied poetry and I love poetry. Um. The difference is that you can get away with more associative stuff. Now, it might not be poetry that anybody really cares to read more than once, but it can pass for a poem. Um. So this is a I found this book. It was published in two thousand eleven by Pure Press, edited by a computational

linguist named are Eli Hirtelo. And this is a book of poems called Discourse dot c PP that was created by a computer that was designed to generate ontologies. So what we were talking about before, you know, the idea of association, establishing meaning for terms that it finds on the internet and relationships with other terms. Uh. And so the way that this book was created by it was

by machine learning. It was by looking at tons and tons of text on Wikipedia and trying to establish relationships between terms and then using that to auto generate poems based on the semantic relationships that it had determined. I bet it's poem about citation. Question mark is amazing. Okay. The umbrella. Okay, this is when we have not heard yet. We heard one earlier that actually Lauren and I both

agree like we can imagine a terrible poet. We'll get to it, Jonathan, No, no, no, no, I don't want you to read that one. I want you to read just ones I have not heard yet. The umbrella. You want an umbrella and all you have is a flannel handkerchief and a sponge. That's that's it. That's that's probably bicycles. No, no, no, I'm still I'm still taking an umbrella, flannel handkerchief and a sponge that's written by a computer. Go another one.

Bicycles by the computer. The cycles dominate the street, an infantry of two wheeled implements, tricycles and rickshaws, most of them far left crutch carriages. The pilots are no duffers. Visitors rent higher priced engine powered recumbent three seaters, flashlight and tulip included. All commute as one would skim. Pretty sure, E Cummings wrote something like that. Um, but hey, no, I want to read the good one, this one, this one actually, this was kind of amazing, the only one

that we had heard actually like this one. If if a human wrote this, I would think it was kind of interesting. Okay, it's called love to love rather like prefer and wish to want first, bother then approach, chase, catch, eat and kill, thank you, to love of to remind and remember, to know and to forget. Oh, that last

line is kind of a chiller. I just like to know the bother part because, as I said earlier in our pre meeting, I said, this sounds like the middle school approach to romance, like bother the person first and then get the kind of or the Anakin Skywalker school

of courtship. Um, so that's where we're where. Yeah, And I think that what we can kind of take from this is that it's easier to write poetry about um, about something abstract like love, than it is about something concrete like a bicycle, right, like a narrative right right to see. I would really be curious to see what the first computer generated stab at something like bail Wolf would be like an epic saga. Well, I would say probably,

So you have poetry on two sides. It's probably easier or a computer to write abstract or lyrical poetry than it is to write prose fiction. But it's probably even harder for a computer to write narrative poetry than it

would be for the computer to write prose fiction. Right, Yeah, I would imagine so, because then you've got you've got to deal with all those difficulties of coherent progressive narrative, and you've also got to deal with whatever, you know, poetic tropes you're using, like not to makes you have a length of a line, right, exactly, like how much alliteration, especially for something like Old English where it was all

alliterative and not rhyming poetry. So one of the things that you brought up in the video, and I think it's a really interesting idea, is let's say that we get to the point where computers are able to simulate to a certain extent the ability to tell a story, and it will really depend heavily upon what rules the programmer creates. Right, So, in other words, if I were

to program and artificially intelligent story Heller. The stories that would come out of that computer would reflect the rules I had created, which might be different from the rules you create, Joe. You might think that, you know, especially something as simple as whether or not you value one aspect of storytelling over another could make wildly different stories. But we could get really granular, like to the point of how frequently do I use the word the versus

how frequently you use it? And that actually matters, That actually contributes to somebody's authorial voice. All right, you can you can actually, um take a take a genetic imprint of an author's voice based on I think, Joe, I think you've got a bunch of notes on this one, based on like the number of of times that they use particular words, and the scope of their vocabulary and like yeah, go ahead, I'm sorry, and the grammatic structure

of their sentences. Okay, Well, let's start with an idea. Um. In N six, I think it was, somebody published a novel called Imary Colors. The story, well, primary Colors was a political novel. It was a Romano left you know, which was a thinly veiled story about the Bill Clinton campaign with the names changed and stuff like that. But it was published anonymously. The author didn't reveal his or her name, gotcha um, And so there was a big question like, well, you know, we don't know who wrote this,

but people were really curious. They wanted, you know, all they were positing all these names who were insiders on the Clinton campaign and stuff like that. Um. There was a Vassar College shakespeare professor named Don Foster, who I've got this interesting scene in article from two thousand about him, uh, and about how he used literary textual analysis to determine who he thought the author was, and he thought it was the columnist Joe Klein. Uh. And it turned out

years later Klein admitted that Foster had been correct. So how did Foster identify him? While it was just straight up textual comparison. He was looking at the text of the novel and looking at the idiosyncrasies of the kinds of vocabulary and sentenced construction, just how the text read, comparing that with all kinds of other famous writers. And when he finally stumbled across Joe Klein's column, he said,

ah ha, here we go. It's this guy, because he he uses the same kinds of words in the same order. The voice you can just easily identify and he was correct. Now, actually you don't need a really skilled Shakespeare professor to do something like this these days. So it helps. It helps because because as we were talking about at the very top of the show, Shakespeare produced a lot of work, and some of the work that has attributed to Shakespeare

is probably a collaboration with other authors. In some cases they're known collaborations. In other cases, Um, there are plays where uh, it's suspect said that Shakespeare did not start

the play, but he finished it. And so having someone who is uh specialized in examining the way a particular writer wrote and being able to compare other works against that body of known work or attributed work is very very useful when you're trying to determine like, is this actual play that we discovered that has no name attached to it? Is that actually a lost Shakespearean play or whatever? Yeah. Um,

so they're all kinds of anonymous works throughout history. Think about the Federalist papers, right that they were all all these uh what were the articles published in newspapers? I guess they were bad um written by John Jay Uh, Alexander Hamilton's and James Madison, Right, they're looking at me. I'm the English history guy. Yeah. Well okay, but so

they were published anonymously. UM, but I wonder if you could get right, a computer, your program that can look at these three works and group them into different author sortings, and then compare those different groups two works that we know by each of those three authors and figure out

which ones wrote which. Well, bam, we've got it. So this is essentially like fingerprint analysis, where you're looking for points of comparison that are identical against multiple criteria, whether of using the both the the anonymous work and then the known body of work. Right. Yeah. So the one example of this type of software is a free piece of software called Signature. It's the Stick Signature stylo Metrics System UM, and it's a computer program used for textual analysis.

And one of the things that's used for is to uh uh, is to assess authorship in a book where the author is questionable, or to identify an author in the case of been anonymous work like we were talking about with the Federalist papers. And so how does it work. Well, it does the same thing that Don Foster did with primary colors, except it has an organized system of analyzing

text in a in a machine readable way. So it can look at idiosyncrasies in the way you structure sentences, or in the frequency with which you use certain words, or it can look at the frequency of common function words like we were talking about before, just using too and from and but and if you if you frequently misuse a word, that's a dead giveaway. And it does happen,

you know. UM, And all of the things that go together to form an author's voice, you know, a distinctive authorial writing style are things that can be used to identify the author. And and we're getting better and better at this. There's another textual analysis tool that was created by grad students at Drexley University, UM, and it called Jay Style. Oh and let's say you've got an anonymous

work here, UM. And what it does is it let's say the work is about five hundred words, and you have a pool of maybe fifty potential authorial candidates, and you have maybe sixty five hundreds of words of text from each of the candidates. They say, that's enough for it to identify the five hundred words sample. The author of the five five hundred words sample with a very high degree of certainty. Um and all funny enough, this is kind of a side note, but they also created

something that works the other way. Uh So, if we're getting better and better at using software to identify the author of anonymously written text, that seems like it could put people like whistleblowers at risk. So they came up with a program that does the opposite. It's called anana Mouth, and it takes text that you write and helps you retranslate it out of your voice. Oh. I see, I was just about to say, like I could see this being a thing that that that hurts anonymity online, something

that we already know is becoming a precious commodity. Sure, sure, but instead of just running it through babel fish and into Russian and then Chinese and then back to English. Then what you're thinking, like, I think I know what you're saying. So this is really interesting. How does this all relate to our Ai storyteller? Well, yeah, this idea of if if we can work backwards, then can we work backwards to say, create the Simari in the way that Tolkien would have written it exactly? So you can

look at the rules that generate an author's voice. And if you can derive rules enough to identify an anonymous work, you can potentially derive rules well enough to generate new work. Right, Um, but you know can can we can we generate a story the way that an author would have created a story, Because that's that's emergent behavior right there. I mean, that's that's something more than the some of its parts. Again, it's it's one of those things where I'm I'm pulled

in both directions. On one hand, I can see how this would work. On the other hand, it seems so impossible. Also, I mean, I I edit novels freelance, and like I said, I was a writing major in college, and the idea of a computer writing a story is personally offensive to me. Like coming coming into this podcast, I was angry about the very thought. Like I was, I was totally ready

to kill that computer before it took my job. But and but but but but doing research for it, I found something that Ray Kurtzwil wrote that that kind of struck me. And uh. He was talking about Harold Cohen, who was a computer programmer and artist who tat a program a thousand rules for drawing and and had it draw some stuff and kurts while asked, you know, like, Okay, so who's the artist? Uh. Cohen claims that he is, and that his computer has not been programmed to complain.

And I just thought that that was a very neat way of looking at it, you know, looking at creating a program that can itself be regarded as a work of art, and that what it creates being something new and not necessarily you know. Yeah, I think, uh, Well, first of all, I think the idea of recreating work in the voice of another author and existing author is very interesting. I would be eager to see the first,

what we would deem truly successful example of that. Uh. I think if it's possible at all, it's a long way off. Again, I think it's probably possible, But I do agree this a long way off. I think actually, I think that creating something I think I think that the I think actually creating something in a particular person's voice is in some ways easier because you are limiting the uh, the choices that your computer program can make.

They cannot choose something that would go outside of that. Uh. Now, with some authors, with some people who have created stuff that's particularly tricky for example, again I go back to Shakespeare. And the reason I do is because not that Shakespeare was creating brand new plots. We all know Shakespeare took almost every single one of his plots either from history and then he revised it heavily, or he took it from pre existing work but then put it into his

own voice, with his own motivations and characters. Um, so that's not you know, that wouldn't surprise me to be able to create another Shakespearean style just based upon the previous existing plot. The problem is that Shakespeare was also known for creating words. He created words that we use in our vocabulary. He created phrases and metaphors that are now part of our common vocab in in English speaking nations, and UH to be able to recreate that would be

particularly difficult. So that it would feel natural that you would create new metaphors and new UH phrases that would become something that people you could see people quoting. I mean people are going to quote it. It's a play. You gotta quote it if you're doing it as a play. So that part is it's hard for me to imagine. But I still would be reluctant to say it's impossible. Right, So your question is basically, just can a work that is created based on a set of rules have meaning?

Can it have meaning? And can it be true? And you know, and I think you know. Meaning is one of those things that is found within the reader. You get to readers to read the same piece of work, and it doesn't matter if that piece of work is regarded as the best literature ever written or if it's just you know, pulp fiction or anything in between. Some one reader might get a lot of it and another reader might get nothing out of it, or they might

get too totally different things, both of value out of it. Right. Um, maybe I'm just a sucker, but I felt a little bit of meaning in the computer's poem about love. I sucker that it wasn't poetry, it didn't even rhyme um, But no, I I you know. The The interesting thing here again goes back to the touring test. Right, maybe the stuff that's produced has meaning, Maybe the machine never

realizes that there's any meaning. And touring would say, what's the problem because we just assume that the person who is creating something is seeing meaning there and we cannot

be sure that that's the case. Great examples there when you have literature classes that say, like, the teacher is going on and on about how a certain passage in a book actually represents this one particular thing, and then you talk to the author and like, uh no, that part is about a guy having breakfast, which is pretty much what I wrote. Actually, this is not a problem for me to imagine at all. We are so eager to find meaning, right and personifying machines to begin with.

We find meanings in things that weren't designed by anybody at all. I mean, we find meaning in like if I walk outside and I'm feeling a certain way about I don't know about a relationship I'm in, or my job or something, and I see a hawk swoop down and grab a mouse. I mean the hawk isn't trying to create meaning or entertainment, right, but I will interpret. I mean that will become meaningful to me. I'll see

meaning in this utterly random event. Well, sure, I mean people also will create patterns, whether or see things within patterns, whether there was no actual thing. They're so looking up at the clouds and saying, oh, look, it's very like a whale. So it's a Shakespeare reference for you. I don't think a person had to necessarily choose a word to go in a certain place and a sentence for that word to have a meaningful effect on me. Sure, yeah, I mean it'll be interesting to see if this, if

this ever becomes a reality, I assume it will. If it is possible it will happen that that we can go ahead and just say, because I think it's one of those things you can say about the future, is that if something is possible, someone will do it. So it's just a question curiosity. So it's just a question of when, assuming that it is possible, I'm curious to

read it. Uh, it may I may feel that it is a much better novel than and let's see, I'm just gonna pick one of the novels I hated reading in school, and it's a toss up between Tests of the Revels and Ethan Frome. So take whatever you like. But anyway, I might read it and think it's a novel ten times better than either of those. Well, I'm going to guess that I'm I might get some convincingly dickinson Ian No, no no, no, not Charles Dickenson. Dickinson. I

was saying, Emily Dickinson Dickinson. That's why I'm sooner to get some convincingly Dickinsonian poetry than you are to get some convincing prose narrative. I I think you're I think you're on target there. I would imagine that we will see this develop over time in various fields, and the ones that will be conquered first will be the ones that require the least narration. The ones that will be

conquered last will be really, really good poetry. Because obviously, one of the rules here is that the longer the form is for a human anyway, the longer the form is in general, the easier it is to write. Not that it's going to be easy to write a good one, but writing a good novel is easier than writing a good short story because you have to make every word count. Writing a good short story is way easier than writing a really good poem because you have even fewer words

to work with. Here's the thing, too, though, if we're considering creating stuff by generating rules in a dead author's voice, how much work they have available is going to be a big deal. Here, somebody who's got, you know, seventy novels that we can feed in to generate rules from is going to create a more robust AI voice than somebody who published one novel and that's it. Sure, yeah,

that's very true. So, uh you know, maybe we'll eventually get to a point where we have these computers just generating things in their own voices, in which case it will all be zeros and ones be a fascinating read binary the novel. All right, So, uh no, this is a fun discussion. I am curious to hear what our listeners have to say about this. You should go to our website FW thinking dot com. That's where we have all the blog post, podcast videos, articles all relating to

these topics are right there. Um find the one for this podcast, the entry we have for the for the episode, and let us know what you think about the idea of computer generated fiction, and tell us if you think that's ever gonna be a reality or what do you think the first computer generated book will be about? Now? Will it be The Mouse who Loved Me? Maybe it's I have no mouse and I'm a screen who knows apologies.

All right, so let us know and we will talk to you again really soon for more on this topic and the future of technology. Visit forward thinking dot Com, brought to you by Toyota Let's Go Places