We don’t commonly call painters “brush artists” or or voice actors “microphone artists” or Substack writers “keyboard artists” (although keyboard *warriors* can certainly fit the bill, at times).
Still, I appreciate the depth of your analysis here. Nice to see someone taking the issue seriously 👍
Thanks for reading, I appreciate it! Re: terminology…there are some artists who explicitly describe themselves as "machine learning artists" (and maybe "new media artists" is also similar?) but I agree that "AI artists" is an awkward term…especially because AI could be applied to a lot of different mediums.
Some artists do identify themselves as "textile artists" though! This is probably the most minor point in all the debates about AI art, but I'm so curious how our terminology around it will evolve
i feel like this is the most careful consideration of this conflict I've read so far. I think the tech can rly help ppl better hone their observing skills and can help ppl get the core structures of like paintings and pictures done more reliably so they can focus on more of the details .
and with writing, being able to simulate the voice of a certain writer to learn more about how they might write something can help people better predict analogous translations in other situations when they dont have the model available. like i made a custom LLM based on Joyce's Finnegan's wake to just simulate more kinds of word puzzles that i can use to figure out interesting ways to take a story that i might have taken a lot longer to think of on my own or would never thought of because of writer's block
Thank you for your incredibly kind comment! And I agree—AI can offer a different way of looking at information, in a strange/fascinating/surreal aggregate form that gives insight into the STRUCTURE of, say, a particular artist's work, and not just the specificity
2 or 3 years ago, a friend of mine did this brilliant exercise where he asked DALL-E to generate "runway model with flower on head" in the style of 4 different fashion designers: John Paul Gaultier, Rei Kawakubo/Comme des Garçons, and 2 others I can't remember. The images were fascinating and very, very different—and for anyone who's familiar with these designers, there was a real pleasure in recognizing how the images reflected the distinctive tendencies of Gaultier's (more colorful) punk extravagance, and Kawakubo's (mostly black) avant-garde silhouettes
Your Finnegan's Wake LLM sounds fascinating—and I definitely think projects like that are extremely interesting and rich for writers and artists!
If I start commenting on the intricacies of your post then I'll be here a while.. you've given so much to think about/digest, but I'd rather just give kudos and say thanks. I can only imagine the editing, organization and second-guessing that went into writing this piece. Just subscribed, seems like we think/write about similar things - looking forward to more good stuff!
ah, I'm so happy you found your way to this! it was really incredibly fun (though daunting) to write…I did not plan to try to answer "what's the social function of art" in order to discuss Fluxus x AI, lol
and likewise, looking forward to reading your thoughts on technology/art/culture as well!!
The distinction you make between art as a commodity and art as an experience is absolutely crucial and would bring a lot of clarity to most of the discourse around AI art that I've seen. It seems like people are constantly telling users of AI art tools to just pick up a paintbrush/Cintiq and start creating their own pieces, without realizing that there are plenty of situations where someone might want to be able to "use" art without becoming an expert in it, just like I need to be able to use my car to drive without learning the skills a mechanic. It's never easy for someone who has invested the time in learning how to make and appreciate art to hear that, though, just as most mechanics would probably cringe to hear me try to describe something that's happening with my car.
I'm not an expert in transfer learning, but I did use it briefly in a graduate level class, so: my understanding is that it's primarily used in tasks like image recognition, where the layers of pattern recognition that let a model break an image down into its most basic features can be repurposed, because the basic features that can be analyzed to recognize e.g. species of birds can also be analyzed to recognize e.g. breeds of dog. In generative AI, the inability to naturally make connections across domains of knowledge has been one of the most curious limitations of current models. Given that ChatGPT has memorized so many facts from so many fields of science, making it more than competitive with the breadth of knowledge of human experts, why are we still relying on humans to do new research and discover new molecules? It seems like large language models are limited in their ability to create "new knowledge" instead of just matching their training data, whereas a human mind would naturally make connections during the process of learning so many different facts. This is something that big-name AI podcaster Dwarkesh Patel has pointed to as a potential fundamental limitation of language models, although we'll see if that ever changes and researchers manage to fuse together different fields of knowledge within the mind of ChatGPT.
Thank you so much for reading, and also for your very helpful explanation!
I'm extremely curious whether AI has the capacity to actually create new knowledge (not just seemingly-fluent aggregations and syntheses of existing knowledge). Right now, it feels like AI can do so WITH a human who's very patiently and actively engaged in prompting and shaping the model. But prompting is a science (or art?) unto itself…so it makes sense that people still prefer their existing processes for creating knowledge, instead of designing totally novel and somewhat unreliable processes that incorporate AI.
I really love this idea of 'escaping your own taste' through aleatory art -- where the process of making the art is weird and unorthodox and delightful, and that means digressing far away from procedure and moving toward play and experimentation. I have to remind myself that this sort of generative and improvisational aspect occurs whenever my writing gets stale. Eno was also great -- I am reminded of one clip where Brian mentions that music has to fulfill and live a life of its own, beyond its creator.
Kudos to another incredibly detail oriented and well researched piece, Celine!
Thank you so much!!! I obviously respect your thoughts on AI, art, creativity, and process so much—so really happy to hear that you enjoyed this
To me, one of the fundamental qualities of ambitious, non-commodity art (that aspires to change the creator, and change the culture exists in) is that it is always trying to be novel, always trying to occupy the cutting edge in some way…the artist making it wants to do something beyond their present capacities, to grow and flex their practice; and they also want it to be surprising, fresh, and original in some way. (And originality doesn't mean this work is totally disconnected from the past, in the same way that original/groundbreaking research is always built off of past discoveries.) And I really think it is SO hard to not get trapped in preexisting ways of making/thinking. Having collaborators helps a lot here, so you're exposed to new ideas—and AI can serve as a collaborator in some ways! It's not the same as working with actual people, but actual people are not immediately available at all hours of the night when you're working on a project…
this was awesome — have you seen Jon Rafman's "Counterfeit Poast" ? from 2022 and using GANs, the only AI video stuff I've really loved, partially because it is SO WEIRD / the sort of thing impossible to produce without the tools and, of course, bears all of the flaws of the medium: too many fingers, unexplainable artifacts, etc
thank you for sharing this! I haven't seen it before and really enjoyed it—totally agree that the flaws of the medium (the strange digital quality of the first woman speaking; how she's awkwardly superimposed over the backing video) are so present here, in a really delightful way…the still images of couples kissing with weirdly distorted faces are some of my favorite parts tbh
I saw another Rafman film last year and was not a fan (doesn't really make use of AI, mostly 3D modeling https://www.180studios.com/exhibitions-and-performances/minor-daemon), partly because the flaws of the medium were a bit TOO evident and the overarching narrative felt pretty weak—but "Counterfeit Poast" is much more compelling; just enough smoothness, just enough roughness to showcase the delightfully specific qualities of GAN art
I also love that the first woman narrating is describing a very specific meta-narrative of AI: she's describing a film that NO one else has seen except her, to the point where we, the viewers, wonder if the film exists at all or if it's this deepfake fantasy that she's imagined up, and is now spreading "fake news" by talking about this film at all these dinner parties
I really enjoyed this take on AI and art—especially the idea of embracing AI’s randomness and flaws. It's not about replacing artists but exploring new ways to create and think about art. Nonetheless, I think that in practice it does eliminate many creative jobs where the point is not necessarily artistic expression or craftmanship, but rather some kind of 'creative productivity' - the commodity you mention. Great article overall! 👏
Thank you, Maria! And yes—I think at its best it can be a way to shift perspectives and perceive the world (especially the symbolic/abstract world we've created with data) in novel ways, which is a very artistic function…but it can also serve as a way to automate away jobs that are (for now) necessary for many people to have a dignified living.
The automation is a political/economic/policy problem, and I'm really curious (and nervous as well, tbh) about how it will play out
I think this is a very down-to-earth explanation. Even further, I appreciate the attempt to properly align ai art with conceptions we already have. I think that's great and needed.
I think randomness isn't exactly a thing but a perception and definitionally one you can't define much less seek. Same goes with unexpectedness and chance. These don't actually succeed as being categories of art in themselves and that sorta follows your argument that Cage can't be copied (along with your 5th proposition).
A side point about the psychoanalyzing bit, it's within the training set at least right? It doesn't extend to society and even if it did, most people don't swear but most people do react strongly to swearing. Without a lot of band-aid coding you'd probably see swearing pop up a bit more given how evocative it is. There's not a transference of semantics with ai which was your main point in the article. Also psychoanalysis requires a semantical transference as it's specifically based on interpreting later behaviors by past ones, generally ones that rely on first social relations.
I think art is meant to be timeless to some degree. Joyce represents turn-of-century Dublin. Not exactly timeless but it is participating in modernism as a theory for all human societies which again isn't exactly but it's more timeless. It's on some fundamental values. Even more, his technique, or ability to fulfill those values, are expository. There can be a transformative factor but that's entirely derivative of the values and the techniques to fulfill them. AI is operationally seeking a mean with maybe some forced re-learning at the ends but it's very much incapable of seeking a timeless thing. Obviously it doesn't know what that means but it can at best only work within its training set and towards that average. If there is AI that can appeal and deal with things outside an average (and still be functional) then we still have to discover it but we're decades into using averages as an epistemic framework for ai.
Shawn, thank you so much for reading and engaging so deeply—it's very much appreciated
Your point about randomness is so interesting—I cut out a paragraph about how what we perceive as random is often different than true randomness. The classic example is how, when people are asked to write out a hypothetical sequence of 30 coinflips, they tend to avoid sequences of 3 heads or 3 tails in a row—even though it's actually quite likely for it to occur. I think I wanted to make some point about how the aesthetic experience of novelty/surprise/randomness actually requires some more human intentionality…we are constantly confronted by situations that feel improbable to our minds, but are actually (statistically) totally probable! They just FEEL off…
Your description of "band-aid coding" is also extremely evocative—have to think about this further, but it feels like such an interesting description of how anti-bias work feels (a tendency in the data becomes explicit in the LLM response, people notice and panic, and try to find a way to repress it). I'll admit btw I don't know a ton about psychoanalysis so your point about semantical transference is escaping me rn, but I appreciate you raising it!
Lastly, very intrigued by this point: "we're decades into using averages as an epistemic framework for ai"…I do think art that escapes the mean tends to feel more surprising/distinctive/novel, but right now it doesn't feel that you can easily get there with models?
Oh yeah, no doubt. Between deep learning and bayesian inference and regular statistics you're just fundamentally appealing to an average which is just something our brains can't handle by itself. We have hopes and dreams tied into how we perceive things and we obviously can't be conscious of averages of anything we do deal with. I think Noam Chomsky said it really well, ai can speak in human impossible languages. That and if there weren't any humans they couldn't really ever speak in human languages no matter how much computing power they had. I'd posit that animals would be predisposed to using similar human language if they had the ability.
This was really interesting. I am overall more bearish on the usefulness of LLMs, but I loved the discussion here about the artistic possibilities of treating them as our collective textual unconscious. This is the first essay that has made me feel that indeed perhaps we can query them from that standpoint to get new information and see things in a new way, ie, make art.
Thank you as always for reading—you are always such a thoughtful and perceptive writer, so it means a lot!
I love the Helen DeWitt piece you shared, and she's actually someone who feels ideally suited experimenting with the literary possibilities of LLMs. This passage is also quite similar to Brian Eno's ideas around generative music and the possibilities of creating truly unique aesthetic experiences for each listener:
"The disempowerment of the author seems to arise, in part, from the fact that the publishing world is oriented toward derivatives — in the first instance, copies of an object that is already at one remove from the original, in the second transformations of that object into translations, films, and so on."
I also love her discussion of a Tuftean approach to fiction, and this observation: "There’s no place for the practices of the art world, in which gallerists, curators, dedicated collectors visit the artist’s studio, in which it’s a mark of engagement to extend the artist’s scope by supporting technically challenging projects." Maybe literary residencies are a bit like this—but the support is usually in terms of providing community, space, isolation, and a small bit of funding…not the more involved level of funding and technical/production/execution support that DeWitt is visualizing.
Your comment is also reminding me that I forgot to mention a very interesting literary essay, by Vauhini Vara and published in 2021, that was composed with AI: https://www.thebeliever.net/ghosts/ (it really illustrates why the presence of the writer is essential to really moving works with AI, imo)
This was SO thoughtful and SO well-timed; “what counts as original in your art when you’re inspired by so many things and also allow unplanned things in as one of them” has been on my mind a bunch lately. Thank you for the buffet of nuance, ma’am! Feeling very “seen” and all!
Long read, but worth it. Your breakdown on the two main perspectives on "what art is" reminds me of an interview with Rebecca Giblin at ACMI a couple years ago. She was mainly talking about generative AI's effect on the creative labor market, but shared a story where someone told her that "artists are like buggy-whip manufacturers; we don't need them anymore."
For people who only view artists as the producers of artistic commodities, that makes perfect sense! For people who can point to a single museum exhibit and describe how it changed their life, how could that ever be true?
Thank you for your patience in reading this post!! I absolutely could have and should have edited it down a bit further—still, I really wanted to incorporate quite a bit about the role of art and my obsession with Fluxus event scores, so here we are…
The story you shared is so interesting…and yes, a great example of the importance of distinguishing between different kinds of art (and also different relationships people have to art)! For many, art will never become obsolete, and trying to automate it away is like trying to automate away conversations with a loved one—why would you try to free yourself from meaningful, joyful labor?
About 10 years ago, someone gave me a quote in this vein and it never left my heart. Mostly because it intrigued me in my disdain of the notion at that time. I had not started dealing with my own traumatic past.
"..many artists agree that you need some knowledge of a field’s history in order to do radical work.."
Through going back to school and writing here on Substack, I'm captivated by the revelations that have reconciled this idea. I had to experience the importance of historical foundations in a very intimate way.
The stuff about AI gives me a nervous energy. BUT it lines up with so much that I've been looking into recently on learning behavior and creativity.
Ooh, a lot to digest here!!
Thank you for this very precious exposition, Celine ♥️.
Licinda, thank you so much for reading!! Re: "learn the rules in order to break them"—this has really been THE principle motivating a lot of how I read and write and learn…it's so helpful to understand the history of a field in order to do something interesting in it!
Really appreciate hearing from you and knowing what resonated with your own creative/intellectual experiences!
Anyway this was such a good read, I kept feeling like it connected deeply to my own approaches to AI and art, though I have no hesitations about being critical of how it’s been developed. :)
omg—I wish I'd come across your work when I was writing this!! I loved reading your first link, on "reverse diffusion" and Brecht's attempt to construct an image entirely free of human bias. Thanks so much for sharing, I'm really looking forward to reading more of your work in this area. And I appreciate you putting so much of your AI images course material online!
The passage below, from your first link, is an (imo) nearly perfect summary of how chance in AI works today—there's sometimes novelty at the image level, especially when the AI makes mistakes; but there's not enough novelty in aggregate, because models bias towards the most frequent outcomes.
"When we look at an AI generated image…we might see all kinds of variety, all kinds of unexpected decisions, as a result of this random starting point. Noise is essential to novelty and variety. Glitches, errors, mistakes and misinterpretations (as above) are where the novelty comes from…
But if we scale outward, and look at what is created from a distance, what we see remains in the known visual category of images: these images are still constrained…[and] depict similar things in similar ways. AI generated images, free of human intervention, rarely offer new perceptions for us. It simply references what else exists: fashion photography, video games, advertisements. If we create enough images, we will find patterns in the images we make: repetitions that point us to the central tendencies of the datasets they trained on. We can work within these boundaries, and many post-photography AI users create beautiful, thoughtful work. But we are unlikely to create a new visual experience or mode by relying on existing references."
“ai artists” is a contradiction in terms imo.
We don’t commonly call painters “brush artists” or or voice actors “microphone artists” or Substack writers “keyboard artists” (although keyboard *warriors* can certainly fit the bill, at times).
Still, I appreciate the depth of your analysis here. Nice to see someone taking the issue seriously 👍
Thanks for reading, I appreciate it! Re: terminology…there are some artists who explicitly describe themselves as "machine learning artists" (and maybe "new media artists" is also similar?) but I agree that "AI artists" is an awkward term…especially because AI could be applied to a lot of different mediums.
Some artists do identify themselves as "textile artists" though! This is probably the most minor point in all the debates about AI art, but I'm so curious how our terminology around it will evolve
Maybe "AI ripoff artists" is more accurate.
i feel like this is the most careful consideration of this conflict I've read so far. I think the tech can rly help ppl better hone their observing skills and can help ppl get the core structures of like paintings and pictures done more reliably so they can focus on more of the details .
and with writing, being able to simulate the voice of a certain writer to learn more about how they might write something can help people better predict analogous translations in other situations when they dont have the model available. like i made a custom LLM based on Joyce's Finnegan's wake to just simulate more kinds of word puzzles that i can use to figure out interesting ways to take a story that i might have taken a lot longer to think of on my own or would never thought of because of writer's block
Thank you for your incredibly kind comment! And I agree—AI can offer a different way of looking at information, in a strange/fascinating/surreal aggregate form that gives insight into the STRUCTURE of, say, a particular artist's work, and not just the specificity
2 or 3 years ago, a friend of mine did this brilliant exercise where he asked DALL-E to generate "runway model with flower on head" in the style of 4 different fashion designers: John Paul Gaultier, Rei Kawakubo/Comme des Garçons, and 2 others I can't remember. The images were fascinating and very, very different—and for anyone who's familiar with these designers, there was a real pleasure in recognizing how the images reflected the distinctive tendencies of Gaultier's (more colorful) punk extravagance, and Kawakubo's (mostly black) avant-garde silhouettes
Your Finnegan's Wake LLM sounds fascinating—and I definitely think projects like that are extremely interesting and rich for writers and artists!
My mind is still whirring after this one. Thank you!
thank you for reading and leaving this comment!!
If I start commenting on the intricacies of your post then I'll be here a while.. you've given so much to think about/digest, but I'd rather just give kudos and say thanks. I can only imagine the editing, organization and second-guessing that went into writing this piece. Just subscribed, seems like we think/write about similar things - looking forward to more good stuff!
ah, I'm so happy you found your way to this! it was really incredibly fun (though daunting) to write…I did not plan to try to answer "what's the social function of art" in order to discuss Fluxus x AI, lol
and likewise, looking forward to reading your thoughts on technology/art/culture as well!!
The distinction you make between art as a commodity and art as an experience is absolutely crucial and would bring a lot of clarity to most of the discourse around AI art that I've seen. It seems like people are constantly telling users of AI art tools to just pick up a paintbrush/Cintiq and start creating their own pieces, without realizing that there are plenty of situations where someone might want to be able to "use" art without becoming an expert in it, just like I need to be able to use my car to drive without learning the skills a mechanic. It's never easy for someone who has invested the time in learning how to make and appreciate art to hear that, though, just as most mechanics would probably cringe to hear me try to describe something that's happening with my car.
I'm not an expert in transfer learning, but I did use it briefly in a graduate level class, so: my understanding is that it's primarily used in tasks like image recognition, where the layers of pattern recognition that let a model break an image down into its most basic features can be repurposed, because the basic features that can be analyzed to recognize e.g. species of birds can also be analyzed to recognize e.g. breeds of dog. In generative AI, the inability to naturally make connections across domains of knowledge has been one of the most curious limitations of current models. Given that ChatGPT has memorized so many facts from so many fields of science, making it more than competitive with the breadth of knowledge of human experts, why are we still relying on humans to do new research and discover new molecules? It seems like large language models are limited in their ability to create "new knowledge" instead of just matching their training data, whereas a human mind would naturally make connections during the process of learning so many different facts. This is something that big-name AI podcaster Dwarkesh Patel has pointed to as a potential fundamental limitation of language models, although we'll see if that ever changes and researchers manage to fuse together different fields of knowledge within the mind of ChatGPT.
Thank you so much for reading, and also for your very helpful explanation!
I'm extremely curious whether AI has the capacity to actually create new knowledge (not just seemingly-fluent aggregations and syntheses of existing knowledge). Right now, it feels like AI can do so WITH a human who's very patiently and actively engaged in prompting and shaping the model. But prompting is a science (or art?) unto itself…so it makes sense that people still prefer their existing processes for creating knowledge, instead of designing totally novel and somewhat unreliable processes that incorporate AI.
I really love this idea of 'escaping your own taste' through aleatory art -- where the process of making the art is weird and unorthodox and delightful, and that means digressing far away from procedure and moving toward play and experimentation. I have to remind myself that this sort of generative and improvisational aspect occurs whenever my writing gets stale. Eno was also great -- I am reminded of one clip where Brian mentions that music has to fulfill and live a life of its own, beyond its creator.
Kudos to another incredibly detail oriented and well researched piece, Celine!
Thank you so much!!! I obviously respect your thoughts on AI, art, creativity, and process so much—so really happy to hear that you enjoyed this
To me, one of the fundamental qualities of ambitious, non-commodity art (that aspires to change the creator, and change the culture exists in) is that it is always trying to be novel, always trying to occupy the cutting edge in some way…the artist making it wants to do something beyond their present capacities, to grow and flex their practice; and they also want it to be surprising, fresh, and original in some way. (And originality doesn't mean this work is totally disconnected from the past, in the same way that original/groundbreaking research is always built off of past discoveries.) And I really think it is SO hard to not get trapped in preexisting ways of making/thinking. Having collaborators helps a lot here, so you're exposed to new ideas—and AI can serve as a collaborator in some ways! It's not the same as working with actual people, but actual people are not immediately available at all hours of the night when you're working on a project…
this was awesome — have you seen Jon Rafman's "Counterfeit Poast" ? from 2022 and using GANs, the only AI video stuff I've really loved, partially because it is SO WEIRD / the sort of thing impossible to produce without the tools and, of course, bears all of the flaws of the medium: too many fingers, unexplainable artifacts, etc
https://dis.art/counterfeit-poast
thank you for sharing this! I haven't seen it before and really enjoyed it—totally agree that the flaws of the medium (the strange digital quality of the first woman speaking; how she's awkwardly superimposed over the backing video) are so present here, in a really delightful way…the still images of couples kissing with weirdly distorted faces are some of my favorite parts tbh
I saw another Rafman film last year and was not a fan (doesn't really make use of AI, mostly 3D modeling https://www.180studios.com/exhibitions-and-performances/minor-daemon), partly because the flaws of the medium were a bit TOO evident and the overarching narrative felt pretty weak—but "Counterfeit Poast" is much more compelling; just enough smoothness, just enough roughness to showcase the delightfully specific qualities of GAN art
I also love that the first woman narrating is describing a very specific meta-narrative of AI: she's describing a film that NO one else has seen except her, to the point where we, the viewers, wonder if the film exists at all or if it's this deepfake fantasy that she's imagined up, and is now spreading "fake news" by talking about this film at all these dinner parties
Read this all in one big bite and felt my outlook being broadened in real-time; what a thoughtful, balanced, optimistic piece of writing! I love it!
this is truly such a kind comment, I'm so thrilled that you enjoyed the essay and found it expansive!! thank you so much for reading
I really enjoyed this take on AI and art—especially the idea of embracing AI’s randomness and flaws. It's not about replacing artists but exploring new ways to create and think about art. Nonetheless, I think that in practice it does eliminate many creative jobs where the point is not necessarily artistic expression or craftmanship, but rather some kind of 'creative productivity' - the commodity you mention. Great article overall! 👏
Thank you, Maria! And yes—I think at its best it can be a way to shift perspectives and perceive the world (especially the symbolic/abstract world we've created with data) in novel ways, which is a very artistic function…but it can also serve as a way to automate away jobs that are (for now) necessary for many people to have a dignified living.
The automation is a political/economic/policy problem, and I'm really curious (and nervous as well, tbh) about how it will play out
But it IS about replacing artists whose work is stolen. Mine was.
I think this is a very down-to-earth explanation. Even further, I appreciate the attempt to properly align ai art with conceptions we already have. I think that's great and needed.
I think randomness isn't exactly a thing but a perception and definitionally one you can't define much less seek. Same goes with unexpectedness and chance. These don't actually succeed as being categories of art in themselves and that sorta follows your argument that Cage can't be copied (along with your 5th proposition).
A side point about the psychoanalyzing bit, it's within the training set at least right? It doesn't extend to society and even if it did, most people don't swear but most people do react strongly to swearing. Without a lot of band-aid coding you'd probably see swearing pop up a bit more given how evocative it is. There's not a transference of semantics with ai which was your main point in the article. Also psychoanalysis requires a semantical transference as it's specifically based on interpreting later behaviors by past ones, generally ones that rely on first social relations.
I think art is meant to be timeless to some degree. Joyce represents turn-of-century Dublin. Not exactly timeless but it is participating in modernism as a theory for all human societies which again isn't exactly but it's more timeless. It's on some fundamental values. Even more, his technique, or ability to fulfill those values, are expository. There can be a transformative factor but that's entirely derivative of the values and the techniques to fulfill them. AI is operationally seeking a mean with maybe some forced re-learning at the ends but it's very much incapable of seeking a timeless thing. Obviously it doesn't know what that means but it can at best only work within its training set and towards that average. If there is AI that can appeal and deal with things outside an average (and still be functional) then we still have to discover it but we're decades into using averages as an epistemic framework for ai.
Shawn, thank you so much for reading and engaging so deeply—it's very much appreciated
Your point about randomness is so interesting—I cut out a paragraph about how what we perceive as random is often different than true randomness. The classic example is how, when people are asked to write out a hypothetical sequence of 30 coinflips, they tend to avoid sequences of 3 heads or 3 tails in a row—even though it's actually quite likely for it to occur. I think I wanted to make some point about how the aesthetic experience of novelty/surprise/randomness actually requires some more human intentionality…we are constantly confronted by situations that feel improbable to our minds, but are actually (statistically) totally probable! They just FEEL off…
Your description of "band-aid coding" is also extremely evocative—have to think about this further, but it feels like such an interesting description of how anti-bias work feels (a tendency in the data becomes explicit in the LLM response, people notice and panic, and try to find a way to repress it). I'll admit btw I don't know a ton about psychoanalysis so your point about semantical transference is escaping me rn, but I appreciate you raising it!
Lastly, very intrigued by this point: "we're decades into using averages as an epistemic framework for ai"…I do think art that escapes the mean tends to feel more surprising/distinctive/novel, but right now it doesn't feel that you can easily get there with models?
Oh yeah, no doubt. Between deep learning and bayesian inference and regular statistics you're just fundamentally appealing to an average which is just something our brains can't handle by itself. We have hopes and dreams tied into how we perceive things and we obviously can't be conscious of averages of anything we do deal with. I think Noam Chomsky said it really well, ai can speak in human impossible languages. That and if there weren't any humans they couldn't really ever speak in human languages no matter how much computing power they had. I'd posit that animals would be predisposed to using similar human language if they had the ability.
This was really interesting. I am overall more bearish on the usefulness of LLMs, but I loved the discussion here about the artistic possibilities of treating them as our collective textual unconscious. This is the first essay that has made me feel that indeed perhaps we can query them from that standpoint to get new information and see things in a new way, ie, make art.
All this, combined with this piece from Helen DeWitt in the LARB, makes me wonder what DeWitt would do with them. We should give her loads of funding. https://lareviewofbooks.org/article/the-wrong-stuff/
Thank you as always for reading—you are always such a thoughtful and perceptive writer, so it means a lot!
I love the Helen DeWitt piece you shared, and she's actually someone who feels ideally suited experimenting with the literary possibilities of LLMs. This passage is also quite similar to Brian Eno's ideas around generative music and the possibilities of creating truly unique aesthetic experiences for each listener:
"The disempowerment of the author seems to arise, in part, from the fact that the publishing world is oriented toward derivatives — in the first instance, copies of an object that is already at one remove from the original, in the second transformations of that object into translations, films, and so on."
I also love her discussion of a Tuftean approach to fiction, and this observation: "There’s no place for the practices of the art world, in which gallerists, curators, dedicated collectors visit the artist’s studio, in which it’s a mark of engagement to extend the artist’s scope by supporting technically challenging projects." Maybe literary residencies are a bit like this—but the support is usually in terms of providing community, space, isolation, and a small bit of funding…not the more involved level of funding and technical/production/execution support that DeWitt is visualizing.
Your comment is also reminding me that I forgot to mention a very interesting literary essay, by Vauhini Vara and published in 2021, that was composed with AI: https://www.thebeliever.net/ghosts/ (it really illustrates why the presence of the writer is essential to really moving works with AI, imo)
This was SO thoughtful and SO well-timed; “what counts as original in your art when you’re inspired by so many things and also allow unplanned things in as one of them” has been on my mind a bunch lately. Thank you for the buffet of nuance, ma’am! Feeling very “seen” and all!
This is such a kind comment, thank you!! So happy to hear this resonated and relates to some of your artistic concerns too
Long read, but worth it. Your breakdown on the two main perspectives on "what art is" reminds me of an interview with Rebecca Giblin at ACMI a couple years ago. She was mainly talking about generative AI's effect on the creative labor market, but shared a story where someone told her that "artists are like buggy-whip manufacturers; we don't need them anymore."
For people who only view artists as the producers of artistic commodities, that makes perfect sense! For people who can point to a single museum exhibit and describe how it changed their life, how could that ever be true?
Thank you for your patience in reading this post!! I absolutely could have and should have edited it down a bit further—still, I really wanted to incorporate quite a bit about the role of art and my obsession with Fluxus event scores, so here we are…
The story you shared is so interesting…and yes, a great example of the importance of distinguishing between different kinds of art (and also different relationships people have to art)! For many, art will never become obsolete, and trying to automate it away is like trying to automate away conversations with a loved one—why would you try to free yourself from meaningful, joyful labor?
"learn the rules in order to break them"
About 10 years ago, someone gave me a quote in this vein and it never left my heart. Mostly because it intrigued me in my disdain of the notion at that time. I had not started dealing with my own traumatic past.
"..many artists agree that you need some knowledge of a field’s history in order to do radical work.."
Through going back to school and writing here on Substack, I'm captivated by the revelations that have reconciled this idea. I had to experience the importance of historical foundations in a very intimate way.
The stuff about AI gives me a nervous energy. BUT it lines up with so much that I've been looking into recently on learning behavior and creativity.
Ooh, a lot to digest here!!
Thank you for this very precious exposition, Celine ♥️.
Licinda, thank you so much for reading!! Re: "learn the rules in order to break them"—this has really been THE principle motivating a lot of how I read and write and learn…it's so helpful to understand the history of a field in order to do something interesting in it!
Really appreciate hearing from you and knowing what resonated with your own creative/intellectual experiences!
You have been sooo generous in your disposition, Celine. I also apologise (implicitly) -I'll make a greater effort to appreciate your work♥️.
God I wish I wrote this. Bowing down. 🙌🏻
thank you so much for reading, Jasmine!! 💌
Hi! You might be interested in some of my work exploring ties between Fluxus and generative AI. I’ve been looking at George Brecht’s Chance Operations (See: https://www.cyberneticforests.com/news/reverse-diffusion-chance-vs-prediction-2024) and I trained an LLM on Fluxus Performance Workbook scores (https://www.cyberneticforests.com/news/the-prompt-is-an-event). I also taught a course for AI images that focused significantly on the process of art as opposed to the product, which I agree is a crucial distinction — AI overemphasizes the idea of art as product manufacture, rather than as an artifact of a thought process. (https://www.cyberneticforests.com/ai-images).
Anyway this was such a good read, I kept feeling like it connected deeply to my own approaches to AI and art, though I have no hesitations about being critical of how it’s been developed. :)
omg—I wish I'd come across your work when I was writing this!! I loved reading your first link, on "reverse diffusion" and Brecht's attempt to construct an image entirely free of human bias. Thanks so much for sharing, I'm really looking forward to reading more of your work in this area. And I appreciate you putting so much of your AI images course material online!
The passage below, from your first link, is an (imo) nearly perfect summary of how chance in AI works today—there's sometimes novelty at the image level, especially when the AI makes mistakes; but there's not enough novelty in aggregate, because models bias towards the most frequent outcomes.
"When we look at an AI generated image…we might see all kinds of variety, all kinds of unexpected decisions, as a result of this random starting point. Noise is essential to novelty and variety. Glitches, errors, mistakes and misinterpretations (as above) are where the novelty comes from…
But if we scale outward, and look at what is created from a distance, what we see remains in the known visual category of images: these images are still constrained…[and] depict similar things in similar ways. AI generated images, free of human intervention, rarely offer new perceptions for us. It simply references what else exists: fashion photography, video games, advertisements. If we create enough images, we will find patterns in the images we make: repetitions that point us to the central tendencies of the datasets they trained on. We can work within these boundaries, and many post-photography AI users create beautiful, thoughtful work. But we are unlikely to create a new visual experience or mode by relying on existing references."