Creative artificial intelligence has crossed a major threshold. What happens next?

Clip You’re listening to a frequency podcast network production in association with City News.

Jordan Heath-Rawlings When humans think of rapidly advancing artificial intelligence, their minds don’t often go “nice,” and that’s fair enough. We have been primed with decades of science fiction to fear a robot uprising and take over. And of course, to do that, the machines would have to become sentient, have to be capable of human things like learning, adapting human speech and language and even art. Maybe this is why what’s known as creative artificial intelligence is currently dominating coverage of the AI community. It makes sense, really, because one of the things that we like to think makes us most human is our capacity to create art that can stir powerful emotions. Wonder, fear, shock, disgust, inspiration, love all of these things. And to see AI programs like one called Dolly that is now open to the public, create those kind of works, is to wonder at the power of new technology and also to maybe feel a shiver creep up your spine. But is this new technology creating art, or is it still primarily created by the humans who are putting in prompts and selecting responses? Is that art? And if it is, do human artists then have a case, considering that it’s their work being used to train these machines? Or is this just the future? Messy, a little bit weird, a little bit wonderful, and with bigger questions looming every day? I’m Jordan Heath-Rawlings. This is The Big Story. Stephen Marche is a writer and a cultural critic. He is the author of a book called The Next Civil War and he has been covering creative AI for The New Yorker and The Atlantic. Hello, Stephen.

Stephen Marche Hey, Jordan. How are you?

Jordan I’m doing well. Why don’t you start for the purposes of this conversation? Just by defining creative artificial intelligence.

Stephen Marche I would say that it’s sort of a branch of computational art. So, like, there are algorithmic stories and there are stories derived from word clouds and so on, things like this. But I would say that the new brand of cultural artificial intelligence is transformer based text generation art primarily. And so that is both literary like. I’ve written some stories that are one of them is completely generated by text prediction software. But more importantly, I think, is the text to image generators like Daly Two and Stable Diffusion and Mid Journey, where you’re basically able to conjure images by means of prompts. And this is provoking a sort of entirely new form of creative practice, really, that is derived from these generative applications. Can you just give us a few examples of things people might use these programs to create? Just messing around with them casually? Well, I mean, the sort of text generation stuff, if you go to sudowrite or something like that, you can take any text that you want. Like I did for The New Yorker. I took a poem from Coleridge, the unfinished poem and asked it to finish it and it finished it in a way that was completely consistent with Coleridge. And if I told you it was Coleridge, you would absolutely believe the text image stuff. I mean, if you go in and put Cityscape, January 6, Washington DC. It will give you a completely generated picture that looks exactly like the January 6 riots in Washington. Or if you go in and say something incredibly abstract like sometimes people are coming out with prompts that are 5000 words long and it will generate really quite abstract forms. So I think one of the things to remember about this stuff is that it’s just like stable diffusion has been around for a month daily. Two was literally open to the public yesterday. So we don’t really know what it’s going to become. But it can generate images from text. And the consequences of that I think, are going to be pretty amazing.

Jordan How good are the images that it generates? Or to your point, the poems that it finishes? I realize art, if this is art, is subjective, but what’s the general consensus?

Stephen Marche There is no general consensus because you’re talking about very elite usages here. Like stable diffusion being used by a million people in a month. Now that’s incredible, but it’s still only a million people. I would say that the images are startlingly accurate. That would be eerily accurate, right? Like if you say I would like a tree that grows in Alberta coming out of crack of pavement, it will look exactly like you took a photograph of that. Right. And when you ask it to continue poems and stuff like that, sometimes it produces nonsense, but often it produces things that are completely fit and completely work and are completely meaningful and eerie. And I would say that as I’ve sort of interviewed engineers who’ve been working on this stuff, like including the people who invented Transformer, they’re quite often shocked with how eerie this tech is. Also, you have to remember the tech is getting much better basically every six months. So it will continue to get very much better.

Jordan So let’s talk about how it gets better then. How quickly? I know you mentioned every six months there is a new iteration of it. But how quickly does a program itself improve and how do they do that?

Stephen Marche Essentially take this program invented by Canadians called the Transformer, which is a way of processing huge amounts of data and then they scale it into basically all of the texts they can find. So in the case of GPT-3, which is probably the most famous text prediction software, is the one that’s accessible because it’s with OpenAI, they use this massive billion dollar supercomputer that Microsoft lent them basically to create this incredible 175,000,000,000 parameters program. That’s GPT 375,000,000,000 parameters is a lot, but like Google’s Palm, which can do truly crazy things like chain of command reasoning, in which you can basically teach it how to think rather than compute that has 540,000,000,000 parameters. And then, of course, God only knows what actually Google has and what Facebook have. This is not open to the public like Google’s Palm is. I’ve only seen it because the engineers took me on a tour of it. No other researchers outside of Google have access to it. So transformational things are happening behind closed doors.

Jordan We’ve been talking about artificial intelligence and what it would mean maybe since the beginning of science fiction, I don’t know. But we’ve been working on this stuff for quite some time. You mentioned the creation of Transformer. Was there a tipping point? Why is it different now? Why has this ramped up so quickly?

Stephen Marche Well, I would say that the transformer paper, which is called Attention is the Only Thing You Need, which was created by Toronto and Montreal researchers at Google DeepMind. That to me, is like pre and post. There’s before and after that. Transformer is the T in GPT-3. It’s the key behind OpenAI. OpenAI is Transformer based technology, and it is very different than other forms of machine learning, and certainly very different from symbolic machine learning. Can you explain how it’s different without going like 6ft over my head? It’s honestly about 12ft over my head. Like, I’ve talked to the guys who invented it and have them explain it to me. I can sort of explain it to you, but if you take it with a massive grain of salt and understand that I’m dealing with things I’ve had the people who invented talk me through it. And I don’t consider myself an idiot, but it’s extremely I would be remiss if I said that I understood it right. Essentially what it does is it scours language and creates and finds in it patterns. That in itself is a gross oversimplification. But then it inputs these patterns into ways that make things make sense. And all it does, you have to remember all of this, is only derived from text prediction. All it is about is creating the next word in a way that makes sense. But that capacity, the capacity to make the next word make sense, leads to all of this crazy stuff, including figuring out how proteins work on a level that we could never imagine. And all the text image generation stuff is also the same thing. It’s just that tied to words, tied to images, and then you create images based on the words that you’ve tied to text. So it actually reveals something very profound in the nature of language. But to grossly oversimplify in a way that’s basically unacceptable, but is the limit of maybe what we could talk about conceivably this machine takes vast quantities of language, like scouring everything in the Internet, and it breaks it into tokens, which are not words, they’re actually just bundles of letters. And then it finds the patterns between these tokens, which are the parameters, and then the more and more parameters you have, the more sensible things become when you ask them to predict to the point where you can get it to create, like, meaningful paragraphs and meaningful stories that are just derived strictly from prompt. Is that helpful?

Jordan Yeah. No, that is very helpful, and I think I have a grasp on how it’s working.

Stephen Marche Okay. Honestly, that’s as close as I’ve been able to get. That’s fine. And the technical details are not necessarily something that we need to obsess over right now, because what you’re writing about is the implications of this technology and the stuff that it produces, which I want to ask you, is this stuff art? And how do you even answer that question? Of course it’s art. Why of course? We’re in the 21st century now. Like, over 100 years ago, du Shawn signed a Urinal and put it in a museum and said it was art. Barnett Newman painted, like, a strip of black, a strip of red, and another strip of black, and it’s the voice of Virus, an incredibly powerful painting. We have art left behind handicraft a long time ago. And I mean, to me, the best model is the camera here, where a camera is just a machine and all you do is push a button on the machine. But the difference between me pushing a button on a camera and Annie Leibowitz pushing a button on a camera is vast. It’s absolutely enormous.

Jordan Right.

Stephen Marche And the idea that the machines are there’s always been this fear that the machines are going to replace creativity and technology, and they never, ever do. They, in fact, just lead to more explosive forms of creativity.

Jordan What about the actual human artists here who have in the past, or maybe even now, are currently producing the works of art that these programs need in order to learn and get better and get closer to replicating what human creativity can do?

Stephen Marche Presumably, I just don’t understand their argument. If you are a student of, like, I have read every book by Charles Dickens. I did a PhD in Shakespeare. Am I then ripping off these people? This is the creative process. The creative process is taking what, you know, saw. Bellow said, a writer is a reader struck to emulation who wants to emulate readers who wants to emulate what he or she reads. Right? And this is so obvious to me. We exist in traditions. We exist in archives. This is just another way of processing the archive in a very discreet way. I mean, do you think a person who makes a building hasn’t taken from every other building that’s been built before? Or that somebody who paints or takes a photograph isn’t taking from what everyone who’s ever taken a paragraph before? I mean, the better you are as an artist, the more references you have. Right. So the idea that this is the creative process, this is the idea that is the essence of creativity that we’re living in archives of references and making things out of it. The idea that it’s some kind of ripoff to me is just ludicrous. So who’s the artist then? The computer program or the person that puts in the text prompt? The person who manages to conjure something out of the machines that’s meaningful to people. Just the same way with a photograph, people ask these questions when they haven’t used it, but when you use it, it’s just so obviously a tool. It’s not going to replace anywhere interesting. It’s going to change the way certain things are done. But only an idiot is going to say, you know what, I’m just going to go to Stability of Fusion and whip up an ad campaign. I don’t need a designer. I mean, what they’re going to say is, we need a designer who knows how to use stable diffusion, right. And the mastery of the technique will of course, be a skill. Already you can buy and sell prompts on prompt exchanges for Dal-e2 and other things. So it’s just a new form of creativity.

Jordan Does the person who puts in the prompt own what’s being created from it?

Stephen Marche Well, yeah, if you put in something in Stability of Fusion and it comes up, you just have it. Yeah. So if I do that with Dal-e, too, which, as you mentioned, just open to the public, if you were using a photography, like editing software for photographs, and you change a photograph using the tech and change the tone, you own that photograph. Right. That’s the way you’ve changed it. It’s the same thing with this. If you use this stuff and change it, you own it. Right. It’s just a new tech to make it. Now, I think there are OpenAI because I think at first they had some kind of restriction on like they owned a part of it too. But I think that’s gone. Actually, I would have to check on that. I’m not sure.

Jordan So one thing that really sticks out to me and I have played with this stuff just a little bit, obviously not in as much depth as you have is the difference between a traditional prompt that like to your point, you say you want I want a tree growing out of a crack in the pavement. And it will give you something that looks extremely realistic to when you put in something that manages to return an image that can surprise or delight or shock and terrify you, which to me is like the essence of a piece of art, right? Like an unrestrained human emotion. So I want you to tell me about how that works and maybe you want to tell me about LoAb, because this is a crazy story.

Stephen Marche Yeah, well, LoAb is this Swedish artist named Super Composite who was fooling around with negative weight prompts. So this is the kind of thing, when people work on cameras, they’re like, oh, well, you just put the light there and take a picture. But actually manipulating this text is quite something. Right. It is not nothing. So she put in a negative weight prompt, which was Brando negative one. Right. So it gave the opposite of Marlon Brando. And this was actually a terrifying like a woman straight out of a horror movie. Right. And then whenever she put that in and various other scenarios, this terrifying woman would kind of come to these scenes to the point where she stopped doing it because it created horrible, snuff images and things that were really unacceptable. And so this is like a monster that’s lurking in the archives of all these images. Right. That’s sort of a different thing than anything we’ve ever seen before. It’s a new figure out of this tech, a monster of the data set, if you will. And so that’s something new. That’s a new form of creativity. And I think that the kinds of creativity that are going to emerge from this are going to be quite like this.

Jordan What is that? I don’t know what that is. Is this something the person created or is it something they discovered? It’s kind of neither here nor there. What do we call this kind of art as a medium? You mentioned earlier when you were talking about word clouds, that it’s computational art. I can describe, like, that’s an oil on canvas, or this is a musical composition for the piano. Or even just like this is a finger on iPad drawing. What medium is this?

Stephen Marche Well, this is creative AI. Right. I mean, that’s what it is. I think in the case of text to image generation, each form of this, like, Dally Two and stable diffusion and midjourny, they’re all quite different. They all have quite different aesthetics. And I think eventually what will happen is people will build their own data sets and use AI out of them. But the data is the art form, if you see what I mean. Right. The manipulation of the data is kind of almost lesser than the data that you’re manipulating. I would just call it creative AI. That’s what everyone’s calling it. I don’t really think there’s a better definition. I mean, you could call it algorithmic art, but I think algorithmic art is actually something quite different because that’s not necessarily self generating. Right. It’s not necessarily as intellectually automated, where you’re having something here that is not consciously you’re not in control of it. Right. Which is what’s really fascinating here, is that you’re essentially creating art that you’re not in control of.

Jordan Do you think and I realize this may come off as kind of a stupid question, but that we will ever have creative AI. Artists who become mainstream have galleries and shows. Like, would you go to a Creative AI?

Stephen Marche They already have galleries. They already do. Oh, yeah. There’s one in the National Pharaoh Island Museum. Somebody sent me a link on, like it’s coming very quickly. I do not have any doubts that there will be prominent AI artists. I mean, there are already artists who are working in this space, were prominent in themselves. I don’t think anyone’s ever has broken through yet, but it’s tough to break through. I’m not sure there are a lot of painters who have broken through. So the last thing I want to ask then is so far, we’ve kept this discussion to art, and that’s because that’s what these programs are made for. Well, they’re not made for that. They’re actually made for a huge number of other functions. Art is like the afterthought of the afterthought.

Jordan Why are we paying so much attention to it, then?

Stephen Marche Well, because it is totally fascinating and sort of interferes in something that we think should be specifically human. Ie creativity, right? But this is absolutely coming for weaponry. It is coming for every form of human conversation. Art is the echo of a change that’s already happened, which people kind of don’t talk about because they’re not exposed to it, or they don’t, unlike Search and unlike social media, where everyone had an opinion, because everyone used this stuff, almost nobody has used this. And when they do it’s in a very specific way, it’s not as integral to life as these other technological changes, though I think it will be every bit as urgent.

Jordan Well, that’s why I was asking that question, to get at exactly that, which is this technology has profound implications way beyond art. But as you look around and as you look at sort of, I guess, the more prominent press coverage the technology gets, and you’re part of that, obviously, in a massive monthly magazine, is the art stuff kind of the PR face of a really profound technological change that’s going on way underneath that, that nobody’s really examining?

Stephen Marche Well, I think art is one thing that’s very profoundly human, and in fact, is the way that we connect with our humanity on a basic level. Right. I think that’s one of its functions. Maybe it’s most important function to me when you see that what you’re actually talking about is when you call up Canadian Tire and complain about your tires five years from now. And you have a lovely conversation with a woman on the phone. And you hang up. And then you realize that it’s a robot. Or you don’t realize it’s a robot until much later. That’s a very profound change in the nature of humanity and our relationship to language. And that’s coming right like that’s coming on like a bunch of levels. It might already have happened to you. Right. And what we’re talking about, art, is it’s really a syndicate for humanity. And so I think that’s a perfectly natural way to talk about it, although I actually don’t think it’s talked about anywhere near enough. Right. When you write something about Facebook, you can be assured many millions of readers, when AI is a much smaller audience interested in this stuff, for sure. I mean, I find it, I’m totally obsessed with it, but it does not have anywhere near the kind of public attention that it deserves. I mean, there’s a lot going on, right? Like the threat of nuclear war in Ukraine and the collapse of the American political system and so on. But it is really a profound change into our foundational concepts of how language works. We’re about to enter a world where just because something talks in a meaningful way doesn’t mean it’s a human being. Maybe this thing can save us from those other problems you discussed. We’re not having any luck solving these things. Technology isn’t going to solve any of our problems. Our problems are ours.

Jordan Stephen, thank you for this. It’s really fascinating and I would urge anyone listening, if they haven’t tried it out, to go and play around with Dali or one of these other ones, because you will surprise yourself.

Stephen Marche Yeah, no problem. That was really fun.

Jordan Stephen Marche writes about all sorts of things, but recently about creative AI in The New Yorker and The Atlantic. If you haven’t read his book, The Next Civil War, it is somewhat prescient given current events in America. That was the big story. If you want more big stories, including previous interviews with Steven, you can head to The bigstory podcast CA. You can talk to us on Twitter at the Big Story. FPN. You can email us anytime. Hello at the Big Story podcast. Dot CA and you can call us. We won’t pick up the phone, but you can leave a voicemail 416-935-5935. The Big Story is available in your favourite podcast app or on the web or via smart speaker. Just ask it to play the big story podcast. Thanks for listening. I’m Jordan Heath Rawlings. We’ll talk tomorrow.