Is literature superior to philosophy?

Hi from Queens, where examples of Fisherian weirdness and eeriness lie in wait at every turn. A good place to write a newsletter called Elf Theory. Fair warning: this post is more theory than elf, at least after the introduction.

Big thanks to everybody who reached out after the last post, which introduced The Lenox Institute for Advanced Study. At the moment we are planning our inaugural event, which will be in NYC in June. Details and registration to come. Exciting stuff!

To get on the Lenox mailing list, email info@lenoxinstitute.org. Elf Theory is not the mailing list: I will share updates here only occasionally. And while we’re on the topic, if you’re a recent subscriber, please know that most posts don’t include nearly this much news. Some will, out of necessity, but I won’t make it a rule. Thanks for understanding.

News

• I am running a major discount. If you follow that link by April 30, you get access to paid subscriptions at the cost of $1/month forever (real forever pricing!). After that, it goes back to the lowest tier that Substack allows, $5/month. I use paywalls sparingly, mostly for a little income, but also to protect my privacy (some posts are very personal) and boost visibility (the algorithm likes paid subscribers, hence the current sale.)

• The latest LEPHT HAND episode features my friend Sujit Thomas. Sujit is a medical anthropologist who studies how psychedelic therapists address the spiritual concerns surrounding death and mystical experiences in medical settings. We talk about the clinical “death transcendence” scale, Aldous Huxley’s heroic final LSD trip, and whether a medicalized “good death” destroys the excess and irrationality that makes life worth affirming. Deleuze, Foucault, and Nietzsche mentioned (on brand :) ).

• Acid Horizon Research Commons’ summer offerings are now open for enrollment. Starting in August, I am teaching a 5-week course on anti-civilization thought / techno-pessimism. Here’s the full description and week-by-week outline, and I’m always happy to answer questions.

• Last but not least: if you’re in New York, please consider joining me for a free event this coming Thursday (4/30) at Columbia University. Leif Weatherby will be discussing his new book Language Machines: Cultural AI and the End of Remainder Humanism, and I’ll be part of a panel of interlocutors. Columbia affiliation is not required, but you do need to RSVP, which you can do here.

This post collects a few thoughts inspired by Language Machines. Onto it now.

I’ll begin with what Weatherby describes as the “internalist” and “externalist” positions on language. He introduces these concepts in an early section that reviews the grounding debate in linguistics, cognitive science, and AI. The debate is about what invests language with meaning: assuming LLMs aren’t conscious, and that their interactions with the world are mediated in very different ways than ours, what features and mechanisms render their outputs meaningful?

A more precise way of stating the question: what are the general features of the “ground” that saves LLM-generated text from decomposing into a semantic void? It’s worth noting here that the problem pertains to meaning as a concrete, real-world phenomenon, not a transcendental ideal, which is why we can’t be more fine-grained with our terminology. Accounting for meaning in every possible sense forces us to stick with generic terms like “meaning” and “semantic,” since getting narrower would restrict our scope. “Meaningful” has valences that don’t readily map onto better-defined concepts such as “useful” and “truthful.” It’s important to be clear about this, since meaning’s expansiveness is the central concern of this post.

According to the internalist position, meaningful language begins with communicative intent. Communicative intent can only come from inner life, or a remit of experience, consciousness, and awareness that we typically associate with human minds. If you’re an internalist who believes that AI has no inner life, you’d have to deny it the ability to produce meaning in any form (language, images, etc). Externalism gives meaning fully to the “outside” or “world.” For externalists, the signifying function of language only begins upon expression.

Citing linguist Ellie Pavlick, Weatherby argues that these positions share the same fatal flaw: both presuppose mechanisms of testimony that are too pure, too scientific, to encompass meaning across all of its different registers. The passage below begins with Weatherby addressing the externalist argument that LLM-generated language is meaningful by dint of its outside source material. Since LLMs don’t produce their training data endogenously, the argument goes, they’re grounded by the world. As with the internalist position, this argument identifies the meaning of language with a particular site of origin (mind or world.) But neither is based on a theory that satisfactorily explains the rules that govern how language maps onto or directly instantiates that site.

Here’s why that’s a problem: if language takes its meaning from a source that’s separate from language in its capacity as “immediate data,” i.e. language construed as identical with its formal features (like phonetics and the structure of alphabetic symbols), we would want to know what principles of selection it observes as it includes or excludes certain features of the source terrain. To use an analogy, maps make sense to us because their relationship with territory has robust theoretical grounding. This foundation allows us to understand the nature of the connection between maps’ represented content and the real-world entities they represent, such as topographical features or political borders. Along these lines, the internalist and externalist positions should come with general accounts of what language’s meaning-giving features have in common with each other. Otherwise, internalism and externalism will lead to nothing other than endless games of example-giving disguised as resolvable arguments.

With that, here’s Weatherby’s account of the rot at their core:

LLMs learn from “social” data and thus participate in long-­range causal chains of external grounding. But the same problem applies to this externalism as to the internalism: no amount of language, generated or otherwise, testifies to any such “natural history”… We use words all the time that neither bottom out in some absolute scientific knowledge of the world nor to which we commit some crystalline belief, 100 percent confidence. Language must be more than ground, and grounding must be some activity within that language (even if it involves other systems as well.) (12-13)

In this context, “testifying” means providing evidence that accounts for every step along the path from meaning’s origin (or “ground”) to the content of specific meaningful claims, as in a logical proof. But words don’t follow this rule: they don’t necessarily exist at the end of a chain of reference made up of links that build logically on prior links.

In a lyrical register, for example, meaning is inextricable from form. It’s all surface, or at least largely surface. Phrases we might find aesthetically pleasing, like cellar door or liquid acrobat as regards the air, don’t have the exact same resonance in translation. There’s always going to be some variation, at least at the level of phonetics. The fact that both come from popular works of art tells us that their aesthetic effect isn’t arbitrary, but rather consists through some apparatus of perception that multiple people share. This meaning-registering faculty is “internal,” while the formal characteristics of cellar door and liquid acrobat as regards the air are “external.” What we consider meaningful obtains between the two poles.

A question: to what degree, if any, should philosophy deliberately incorporate non-referential or “unchained” forms of meaning? Deliberateness is key, because all bodies of language take at least some meaning from non-referential properties. (I would argue that this is true even in the case of language generated exclusively by and for computers, even when human beings never interact with such statements, but explaining why would be a long tangent.) Put differently, how far can we wander into poetic territory while still claiming philosophy as our rightful home?

The question seems to turn on systemizability: while nobody aspires to systematic philosophy these days (Peter Wolfendale being the exception that proves the rule), we still want at least some commensurability between our philosophical claims and others. If philosophy is going to have points of real contact, it needs a source of meaning that doesn’t boil down to poetry. I could suggest that semantic commensurability based on referential meaning is what distinguishes philosophy from literature, but when our understanding of meaning is thrown into existential uncertainty (an unavoidable effect of LLMs), we have to go back to the drawing board.

Provisional answer: literature, as the “unchained meaning” area of research and practice par excellence, always yields philosophy. By that, I mean that literature expresses meaning that’s theorizable even as it exceeds human-legible representation. Here, we might think of Kant’s definition of beauty as an excess of representation over knowledge. Literature, unlike philosophy, is art. It produces beauty above and beyond knowledge. But if the line between beauty and knowledge is given by beauty’s resistance to theorization (since theorization yields knowledge), generative AI challenges the idea that literature and philosophy (which I’m analogizing with beauty and knowledge) are formally separable.

One of Language Machines’ central arguments is that we have to develop a “general poetics” fit for the LLM age. This is because LLMs threaten the idea that language’s poetic function is generally subordinate to its referential function. This argument hinges on a mathematical phenomenon known as vector space, which is central to how LLMs (and all other neural network-based forms of AI) work. The key thing about vector space is that it simulates continuousness, or infiniteness, in the nominally discrete space of the digital. Vector space reveals registers of meaning so multiple that we need a poetics that at least partially expresses the insights they yield in terms that human beings can understand.

This passage begins with a technical breakdown of vector space and vector semantics, followed by a more detailed explanation of their implications for linguistic meaning:

Vector semantics is a form of word embedding in which dependencies in meaning are laid out in a “continuous” vector space. The generalization “capital city” is easy for me to impose and search for, if necessary, conceptually. But what length of string would allow a net to find the relationship China:Beijing::Portugal:x? It turned out that not “context” but rather computational similarity was the issue here. As Mikolov et al. explain, “somewhat surprisingly, it was found that similarity of word representations goes beyond simple syntactic regularities. Using a word offset technique wherein simple algebraic opera­tions are performed on the word vectors, it was shown for example that vector(“King”) − vector(“Man”) + vector(“Woman”) results in a vector that is closest to the vector representation of the word Queen. This example has become famous, as it demonstrates the ability of vector addition to find not formal regularity between variables (grammatical functions) but meaningful relations between words. These vector-words can have “multiple degrees of similarity,” including (in the case of inflected languages) the intricacies of conjugation, cases, and so forth, so that a single lexical entry can have dozens of forms. As the team observed, there are many examples of phrases that are hard to learn if we impose some representational assumptions on language. (144)

As an aside, vector semantics might help us understand Sam Kriss's claim that analytic philosophy, which is premised on representational assumptions about language, is closer to fiction and poetry than continental philosophy, despite all appearances to the contrary. (This is not to say that the analytic / continental binary makes sense.)

If all this is true, everyday speech and writing acts count as “philosophy,” or expressions that can only be properly grasped through rubrics that recognize them as intrinsically open to interpretation. But the reverse doesn’t hold: not all philosophers yield so many different forms of meaning as an essential part of their output. Formal logicians actively work against this effect. Some would say that all philosophers should.

Literary writers fall squarely on the side of “everyday meaning,” which is profoundly multiple and impure, but perhaps more theorizable than we could’ve appreciated before the advent of LLMs. And to get a little loose for a sec, perhaps literature could in turn be considered a higher calling (more meaning-generating) than philosophy. David Foster Wallace’s Wikipedia page quotes him as saying that The Broom of the System used “97 percent of him,” whereas writing philosophy used only 50 percent. Shots fired.

97 percent of me wants to pin a lot on that remark. I’m instinctively drawn to the idea that literature is more meaningful than philosophy. I can’t do that, though, because this relative valuation participates in what Weatherby calls “remainder humanism.” Among other effects, remainder humanism blinds us to the limited purchase of intuition against the sheer scale of the dimensions of meaning rendered computationally tractable by AI. Resisting remainder humanism means admitting that David Foster Wallace’s remark tells us about nothing other than David Foster Wallace. Maybe we can stretch it to apply to other people, but it has nothing to say about philosophy or literature as such.

So the answer to the title question is that the question doesn’t make sense, since all distinctions between the two disciplines break down under scrutiny. A better question might be: what’s the most refined version of the lines that demarcate philosophy from literature, at least in terms of what human beings can understand?

To say that LLMs can assist us in answering this question isn’t to make a political claim about their externalities. But if we care about externalities, the move isn’t to ignore the unique possibilities for intellectual work that they potentiate. We should be directly engaging these possibilities in order to marshal them in politically salient accounts of AI as a world-making force (as Weatherby does, in a chapter on how LLMs produce ideology.)

This is a good note to wind down on: a reminder that when it comes to the AI culture wars, if we can entertain the idea that they’ll get us anywhere, I’m on the same side as ever.

It’s easy to misread Language Machines as a celebration of AI rather than a corrective to the thin criticism that dominates today. It isn’t techno-optimism or determinism, but a sober acknowledgment that culture and technology are mutually constitutive. To refuse to engage with one is to be willfully blind to the other, and any wishfulness or sentimental attachment to the idea of a culture unsullied by it does more harm than good to human life. This is as far from capitulation to Big Tech as it gets.