I have essentially no choice but to begin with what is probably the most quoted line from William Gaddis's The Recognitions:
That romantic disease, originality, all around us we see originality of incompetent idiots, they could draw nothing, paint nothing, just so the mess they make is original… Even two hundred years ago who wanted to be original, to be original was to admit that you could not do a thing the right way, so you could only do it your own way. When you paint, you do not try to be original, only you think about your work, how to make it better, so you copy masters, only masters, for with each copy of a copy the form degenerates… you do not invent shapes, you know them… by heart.
There are almost too many things going on in this line for me to analyze in a single essay. What is a work of art? What is it meant to accomplish? Does it communicate something real about the world that we cannot understand, cannot articulate, through propositional ways of thinking and speaking?
Wyatt Gwyon, an artist who could execute masterful reproductions of the great painters, but produce nothing “original” of his own, received this wisdom from his teacher, Herr Koppel. If we take Koppel literally, I think we might find some of the same issues with his conception of copying the masters as we might with Plato’s account of knowledge of the Forms: we have innate knowledge of the Forms because we knew them in a past life, and we need to merely “relearn” what we already know. But how did we come upon that knowledge in the past life? From where did that knowledge originate? In the same way, how did the masters become masters? If you do not invent shapes but know them by heart, who first knew those shapes?
But it’s worth noting that Koppel’s words are less meant as a prescription—note the “when you paint”—as a description of the state of a culture’s romantic originality. Does not the talented violinist who could not reach Heifetz’s level call his tainted Tchaikovsky “interpretation”? Do I not preface my writing with “this is not original” in self-deprecation, an attempt to shed the burden of taking my own work and writing seriously, an attempt to pretend it’s “something different” because I’m saying something and I’m self-aware? Do I not fear that someone else should tell me about the shortcomings of my writing? How dare I take pride in my work?
A generative AI system, trained on the world’s production of text or images or sound, can produce outputs that result in soundscapes, perhaps in vibes, that none of us have experienced before. In an important way, these outputs are recombinations. Are they “original”? There’s something “new” to putting together a string of words that no one has ever put together before, but perhaps that string of words is entirely meaningless, or fails to leave any “meaningful” impact on the reader.
Ted Underwood wrote about neural language models as models of culture: their outputs are pastiche-collages of every one of us. We can understand a lot of the powers and dangers of vision-language models by looking at their latent spaces as compressing the “visual culture” of their training data, and I think we can also understand a bit about “originality” as well.
Underwood’s more recent “The Empirical Triumph of Theory” explores how language models have begun to cast empirical light on debates about topics that concern humanists. Underwood notes that the term “artificial intelligence” was coined to organize inquiry around a concept, intelligence, that was assumed to be an attribute of individual agents. But large language models, rather than being developed as models of any specific, intelligent person, inhaled everything, learning to imitate large corpora of writing:
The success of this “misdirected” effort has tended to support theories of meaning that explain it instead as a collective phenomenon—like Lévi-Strauss’s “universe made up of meanings” or Foucault’s Archaeology of Knowledge (1969).
But, to really speak to us, these models could not remain kaleidoscopes with no organization. In turning GPT-3 into ChatGPT, instruction-tuning trained models to distinguish their own language from interlocutors’—a “rhetorical reconfiguration.”
Even language models, it turns out, need to make inferences about speakers and intentions. If GPT-3 dispensed with authors entirely, ChatGPT has been compelled to reconstruct a provisional “author-function”—rather as Foucault eventually did.
AI models that produce art, too, seem to have author functions of a sort, or at least “concepts” that correspond to an author’s style: you can ask for paintings in the style of Botticelli or the Simpsons and expect something that hews close to your desired artist’s “way.”
But if “all they do” is recombine, find and reproduce patterns, can they do anything original? Is everything we do detection and recombination of patterns? Nao Tokui, a Japan-based artist and researcher, has an interesting technical take on this—the idea of a “genre” is a fuzzy one, but could you produce music that doesn’t fit neatly into any given genre?
Indeed, it seems you can. He built a rhythm generation model based on the generative adversarial network (GAN), but introduced a new wrinkle. GANs, which try to produce realistic images or, in this case, sound spectrograms, typically have two components: a generator that tries to produce outputs that could be said to be “like” its training data, and a discriminator that tries to distinguish between items that are and aren’t from the model’s training data. Nao adds a second discriminator that classifies generated critique into a genre, then asks the generator to generate patterns that are realistic (“fooling” the first discriminator into thinking the sound sample belongs to the training data), but don’t belong to any genre in the training data.
Originality by negation is a tricky endeavor, but it intuitively feels like it’s barking up the right tree. How about a text generator with an author function which writes something that, according to some discriminator, was not written by any living writer? The collages we’ve achieved by instruction tuning large language models could be read as a version of this—blurring and blending—but they sound so bereft of anything resembling a “voice” that it’s hard to attribute any positively-valenced originality to them.
I don’t know how to settle the question of originality, and I’m not sure it’s possible. But I do think it’s worth considering the role of originality in our conception of art: is “originality,” whatever it might mean, a worthwhile pursuit? How does it fit into what we think art is, what we think art should be?
Tolstoy attempted to answer this question in his 1897 book What Is Art?, first published in English due to difficulties with Russian censors. He rejects beauty as a criterion for art, and instead argues that “real” art communicates emotion—it bears a sort of infectiousness. Note that neither beauty nor emotive infectiousness seems to demand originality as a necessary component. Before considering the relevance of Tolstoy’s conception of art, I think it’s worth observing how he sets up the stakes of his inquiry.
In the beginning of What Is Art?, Tolstoy makes an effort to convince us why the question he seeks to answer is an important one. Indeed, a great deal of human effort is expended on many endeavors that we label “art.” And this effort is not just on the part of the recognized “artist,” but on the part of the many people whose labor supports that artist and their work.
There is an inherent originality that shows up in Tolstoy’s conception of art—a novel work of art articulates a new feeling that had not been articulated in art before. Does this merely boil down to pattern recognition? I’m a little skeptical. Again: if an artist could communicate the feeling associated with their work simply in words, they would do so simply in words. If the construction and artistic communication of new, “original” feelings were a matter of applying a decision procedure that could be reverse-engineered and written as an algorithm, there’d be no need (for Tolstoy) for that new art in the first place.
And this might be considered a first reason why AI art cannot be art. But this way of getting there seems shaky: describing the procedure used to create a thing is not the same as describing that thing itself. One could, in theory, describe a (complicated) algorithm for creating a Michaelangelo painting, but that’s not the same as being able to describe the feeling associated with the end result itself. And describing rules to combine components of pieces of art or to combine feelings, or to combine styles, doesn’t describe the subjective experience of looking at the end result, either. What’s happy + annoyed in unequal proportion?
But there’s another reason worth discussing: it’s probably very hard, or impossible, to communicate the depth of a feeling (especially through art) without being able to feel in the first place.
For Stevan Harnad, “meaning” is something like: T3 + sentience. T3 is a robotic version of the classical Turing Test: a system that “passes T3” possesses all the sensorimotor capacities a human has, and could convince you through its physical behavior (not just through a chat) that it is a human. It walks, talks, and quacks just like you. And, through its ability to interact with the world, to “do the right things with the right kinds of things,” it develops grounded representations of objects in the world. Sentience, to Harnad, is the capacity to feel—that unobservable, untouchable, unmeasurable capacity in the background of it all.
I’m sympathetic to Harnad’s view here, and it colors my idea of whether AI systems, as they exist today, are even the sort of things that could be in the business of doing art. Debates about whether LLMs refer to things in the world often commit what I think is a similar category error. There is sometimes a focus on the content of LLMs’ utterances and whether those statements are directionally correct, whether the patterns of LLMs’ linguistic representations correlate with patterns in the world, whether LLMs’ representations “stand in the right kinds of relations” to things in the world.
A number of analogies between the human capacity to refer and the capacities of LLMs seem to miss the point. Debates about humans consider systems who can refer to things under good conditions and explore the limits of their capacities to refer, whether humans can reference in non-ideal scenarios (am I referring if I say I invented the peanut?). But this is independent from the question of whether AI can refer at all. AI systems aren’t in perpetually non-ideal conditions, either. But one way to think about it, perhaps, is that they can’t even fail to refer—they don’t operate on the “intentional plane” (h/t Kyle Thompson for an email that is basically this whole paragraph).
In the same way, AI systems like Stable Diffusion are not in the business of feeling. Yet, there is a tricky juncture to navigate at this stage: though Stable Diffusion and its ilk cannot feel, they might still be able to produce artifacts that cause the people looking at them to feel things. I am reminded of Ken Liu’s story about Semafor, the movie-making company that developed movies like an optimization problem, tuning films based on the real-time reactions of a live audience.
While a generative AI system may not itself feel things, it is conceivable that its generations could produce feelings in people. But these feelings could be read as “synthetic” or “un-grounded.” The artist who experienced a feeling of love or misery in their real-life experiences communicates, through their art, a feeling that did exist for them in the real world. The “feeling” produced by an AI system is a sort of reaction in the observer, but is not grounded in any real, feeling being’s experience of the world.
Does this, however, make the “feeling” experienced by someone looking at a generation produced by an AI system any less real than a feeling attained by looking at a real person’s artwork? Maybe not. If you believe that mental states are reducible to a set of chemical-biological interactions—if you are a physicalist about mental states—you might think the observer’s feeling is just as real because there’s stuff going on, but that (depending on your views about AI systems) this feeling isn’t something being communicated.
This said, what I find important in Tolstoy (and elsewhere) is the ineffability of certain mental states, of certain feelings. Indeed, you probably can analyze certain mental states by looking at neural correlates, but is this a useful level of abstraction to really understand what’s going on? This person had such-and-such a feeling, which coincides with such-and-such happenings in the brain—is this informative? Does it explain anything that can help us learn more about other mental states, about our own interactions with the world and how those experiences manifest in feeling? I’m doubtful.
So, then, even if we restrict ourselves to a “higher” level of abstraction in considering the reality of AI-produced “synthetic” feelings, do we arrive at a different answer to the question? I think the picture is complicated, but, interestingly, we may arrive at the same answer. The person looking at an AI-generated artwork is actually experiencing something as an artifact of their interaction with the world—that artwork is something in the world, after all, and the person’s observation of that artwork is a real experience. But that experience is not a receipt of another’s experience; it occurs for the first time in the person observing the generation.
And, perhaps, the feeling is a novel one. An AI system could, conceivably, synthesize a feeling that, by some definition, is more original than a feeling that is being relayed through a human’s artwork. That feeling does not “connect” someone observing the system’s artwork to its author (so long as that author, the AI system, cannot feel), but, as always, different people observing the same artwork can experience a shared feeling.
On the question of AI art, I took Tolstoy’s writing as a possible argument that the concept of AI art, or at least the “declarative” AI art whereby a user only passively generates an artwork by specifying a desired result, is a category error. But I think the picture gets more complicated when we start to consider how feelings manifest in observers. I don’t have a final, principled answer to this, but I do think it is more complicated than it seems at first glance.
And what of originality? Is originality a mere excuse for not doing something “the right way”? I don’t really have an answer here. But, just as it was surprising for Harnad that ungrounded AI systems can do the things they do, it’s fascinating that we can create systems so unlike us that can, at least in principle, cause us to experience feelings. Perhaps this says more about us than it does about them.
—Huge thanks to the incredible Sheon Han for giving feedback on a draft of this !
words combined
May it be an evening star
Shines down upon you
May it be when darkness falls
Your heart will be true
You walk a lonely road
Oh, how far you are from home
—yay for soothing LOTR music
consumed
he is playing Tchaik with TWO FINGERS what is going on
found a recent-ish recording of Hilary playing Bach Partita No. 3 and she did an experimental (no orch) solo performance of Sibelius!! it’s rly fun to listen to this way (if god doesn’t exist how does she sound like that)
!!!!!!!!! this entire concert is threatening to yeet me into earthly transcendence but also what happened at 16:15 :( anyway going to go dissolve in a puddle of my own tears listening to this now
(if you want to listen to my favorite part of the Tchaik go to ~ 21:25 (i’m basic ok); Tchaik’s Romeo & Juliet fantasy overture is an absolute headbanger you’ll look like you’re listening to metal or something if you put this on in public and have no sense of shame; Strauss Death & Transfiguration is just ᝰᏪຊ)
Boaz Barak’s reflections on The Making of the Atomic Bomb I now want to read this book
especially interesting: “If humanists and social scientists can find cognitive processes analogous to experiment — processes where a well-documented simulation of learning is the same thing as learning — we will be in the enviable position Robin originally thought he occupied: students who can simulate the process of doing an assignment will effectively have completed the assignment.”
I don’t know how long it will be until AI systems are able to simulate metacognition (if they do get there) well enough that you could convince a teacher that you’d done all the work of ideating an essay, drafting, revising, re-revising, nuking the damn thing and starting from scratch.
But, if Prof Underwood’s thesis is correct, then the academy as a site for knowledge production and not regurgitation and recombination may have to become a reality earlier and earlier (of course, maybe not all teachers will want to stand up to the challenge and you end up in a world where many are, in fact, learning nothing).
“Why Mathematical Proof is a Social Compact” in Quanta — the YouTube video is also enjoyable and quite different from the article!
first off I love to see an interview w/ a number theorist; I never got to take a full Number Theory class in college but it was always one of my favorite branches of mathematics—I did this talk for a Math Forum class on a randomized algorithm for some NT problem relating to primes or something and was having so much fun with it but in my peer feedback I realized my classmates kind of didn’t think it was as cool as I thought it was and that made me sad
on blush, this is an interesting idea! Mathematical proof as not a revelation of eternal proof but instead a social construct—though I worry about this going too far?
this sounds quite a bit like Popper on scientific theories in Conjectures and Refutations—not exactly, but you get the sense of “accepting” a mathematical proof and it’s always in this state of conditional acceptance until/if it’s proven wrong (you can never prove a thing right you can only disprove it): “The best verification system we have in mathematics is that lots of people look at a proof from different perspectives, and it fits well in a context that they know and believe. In some sense, we’re not saying we know it’s true. We’re saying we hope it’s correct, because lots of people have tried it from different perspectives. Proofs are accepted by these community standards.”
The Digital Tolkien Project !!!!!!!!!!!!! i am just happy this exists
“The New Old Age” in The Atlantic — the ideas in here didn’t feel that novel but the idea of allocating some part of your career, esp after your “productive” years, to something that’s (almost) wholly focused on giving seems like a nice idea.
“Elon Musk’s Shadow Rule” in New Yorker lol i mean
“We Don’t Need a New Twitter” in New Yorker kind of what you’d expect from Cal Newport — though I am interested in his contrasting of Twitter / similar social media as for everything vs very niche forums for dedicated fans of X thing and how single social media platforms could achieve both of these things, stand in between, etc.
books / short stories
finally got to Lorrie Moore’s “People Like That Are the Only People Around Here” and cried at various points. this was tough to get through to be honest
Popper’s Conjectures & Refutations my hs debate coach would talk about Popper / falsifiability a lot and I’d never actually read Popper until now
Tolstoy’s What Is Art? i guess this one was obvious
Natsume Sōseki’s Sanshirō — nice bildungsroman I think the vibe of a dude who is too afraid to actually do anything about his desires or impose his will on the world in any way at all (people will say stuff to him that makes no sense and he’ll just be like “ok that’s how it is ig) is so frustrating and deeply relatable at the same time
in yet another canon event I finally watched Peaky Blinders (like almost all of it whoops)
a paper from RT McCoy + Tal Linzen (and I think something Chalmers wrote a while back?) has what I think amounts to an interesting refutation of Fodor & Psylyshyn on the limits of neural/distributed vs symbolic architectures in that, empirically, u can get an NN to do tasks (like reversing a string) that would require their (distributed!) representations to be functionally symbolic (in this case, we have a tensor-product representation in an RNN) — theorizing about representations / showing they must have some particular form feels very very very hard to me though idk