by TJ Price
I haven’t been writing lately. I’m fully aware that sometimes there are cycles to these things; that there’s no sense in feeling guilt or shame about it—and yet. An author said once in an interview that if they went too long without writing, they “turned into a bit of a jackass.” I resonate with this response; it’s almost as if, sometimes, the pressure of not-writing backs up into my brain, clogging me up with unformed material. It is only the act of writing—or sitting down and letting the words come—when relief can arrive, for both me and for anyone who has had to suffer through my grouch.
Not-writing, though, is also in vogue on social media, and everywhere you look: there are those who say that not-writing is also writing. This, for me, has a unique ability to both irritate and inspire in equal measures. It implies that the “work” of writing happens constantly, always generating itself in one’s head, even though a record of the language combinatorics is physically absent, which I agree with—but is it actually writing? It mystifies things a bit to look at it this way, to me, which isn’t entirely undesirable, but it also—by nature of its inversion—seems to de-emphasize the actual act of writing, seems almost to destabilize the definition of the word itself.
Even worse, in recent days, I see views which hold that prompt generating is also writing, viz., pulling up one’s chair to the computer and invoking an LLM to write for them. It’s rather like pulling up one’s chair to the vanity and adjusting the lights; the machine readily reflects back to you exactly what you want to hear—though it is frightening how many people seem to think this is naturally also what others want to hear. Even worse, there’s even those who believe that they have performed labor equal to the work of “writing,” that they have produced something worthy of respect—even worthy of (paid!) publication.
It’s not just publications for fiction, or for poetry, that are afflicted by this scourge of not-writing. I see it everywhere, now—users of the machine believe, in some deluded fashion, that the output is suitable for a mind-boggling range of occasions—from public apologies or announcements to keynote speeches, letters to loved ones—even eulogies. (Heaven help anyone who uses the LLM for this; I imagine the decedent returning from beyond to enact a swift and terrible vengeance, appalled at the insincerity on display.) It makes me blanch to see the comments on the cardboard prose used in these posts, spilling over with adulation from readers: “beautifully said,” or “wow, so well-written.”
Donald Barthelme, in his essay “Not-Knowing,” talks a great deal about the act of writing, but even more than this, about language, speaking, and expression. Toward the end of the piece, Barthelme mentions that
“The combinatorial agility of words, the exponential generation of meaning once they’re allowed to go to bed together, allows the writer to surprise himself, makes art possible, reveals how much of Being we haven’t yet encountered. It could be argued that computers can do this sort of thing for us, with critic-computers monitoring their output. When computers learn how to make jokes, artists will be in serious trouble.”
This, written with a startling degree of prescience—or perhaps just clear-headedness—in 1987, is widely known to those who have attended to the canon of postmodern literature—or at least to as many of its rhizomes as can be withstood. Barthelme argues passionately (and then elliptically) and then passionately again for the aspect of writing that remains cloaked in mystery. In fact, if you were to search for the word mystery on the vast plains of the internet, you’d discover that this sentiment has been echoed for decades by many, though phrased differently by each, and each has come to a similar conclusion: it is very rarely when a mystery is solved that we feel any type of magic. In fact, it’s quite the opposite—with the solving of a mystery, the magic tends to dissipate entirely, extirpated until a new query rises, buoyed by the excitement of, you guessed it…
Not-knowing. Again, Barthelme:
“The not-knowing is crucial to art, is what permits art to be made. Without the scanning process engendered by not-knowing, without the possibility of having the mind move in unanticipated directions, there would be no invention.”
For me, the only mystery engendered by the machines, however, is why everyone continues to use them, why there is so much stock put in the bland unity of their output—indeed, even funneling millions of dollars into funding their idiot hegemony, imposing a dreary kind of worship. And lord, the machines have so many damned acolytes, jostling at the silicon teat, lips pre-pursed to suck. Teats for all, if greige-flavored, pre-masticated pablum is to your liking—prose filled with the irritating whine of obvious rhetorical device deployed via formula rather than a sense of individual authorial style.
“[The writer] discovers that in being simple, honest, and straightforward, nothing much happens: he speaks the speakable, whereas what we are looking for is the as-yet unspeakable, the as-yet unspoken.”
Barthelme’s forty-year old essay taps the vagus nerve of modern discourse: an LLM operates via algorithm. It does not generate; it predicts—based on patterns of what has already happened—what is most likely to happen, given an enormous volume of extant sample. The issue is that with this output (even though its end result may seem mysterious) is entirely explicable by probability, which measures—along with past data—a spectrum of future likelihood. However, probability can only ever tell you what might happen within the bounds of the recorded data, which is of course useful in some disciplines.
But what about the hapax legomenon? This ominous-looking expression sounds like an Unforgivable Curse from the Potterverse, but is actually used to categorize a word (or phrase) which occurs singularly in a given context. Transliterated from Greek—ἅπαξ λεγόμενον—it means “said once.” This extremophile on the scale of probability adheres to something called “Zipf’s law,” which is an empirical observation stating that when a set of measured values is sorted in decreasing order, the value of the nth entry is often approximately inversely proportional to n.
Disregarding the fact that something called “an empirical law” contains the words ‘often’ and ‘approximately,’ the instance of Zipf’s law occurs most commonly within the corpus of natural (read: human) language—though it has also been observed in a host of other logarithmic plotting, such as the rate at which we forget, or the firing patterns of neural networks. Valences and inflections aside, a list of the most-commonly used words in the English language (with some minor possible variation due to evolving observation) would read as follows:
The of and to a in is I that it for you was with on as have but be they.
On their own, without any additional insertions or extrapolations, this is gibberish—these are articles, pronouns and particles, rubbery ligaments with no bones. Much like a boat which has capsized under enemy fire, whose flotsam and jetsam now bobs haplessly on the waves, we might be able to intuit the shape of the ship which once sailed—but that would require the insertion of our own knowledge to build the mental image. Like the Mad Libs of old, we would need to yawn blanks between each of the words to furnish ourselves with some kind of explicable story.
In a recent email conversation with a colleague, discussing the scourge of LLM and its blighted procession across the landscape of fiction writing, my colleague reminded me of the approach of the L=A=N=G=U=A=G=E School to poetry. As a response to the sometimes-daunting strictures of traditional verse—with its spondees, dactyls, and trochees, sounding more like a list of saurian anatomy than metrical syntax—the L=A=N=G=U=A=G=E School of poetry sprang from the roots of such experimentalists as Gertrude Stein (notably, of “If we knew everything ahead of time, all would be dictation not creation.”), and fruited from the branches of Frank O’Hara and John Ashbery, prizing first and foremost the role of the reader, as opposed to the author. There’s something quite beautiful about this approach, I think, which gets at the very heart of what I love so much about writing, literature, and even reading, and I think this very human instinct is being corrupted by LLMs—these machineries of our own creation. As my colleague said, “I love poetry for its striking and bizarre imagery and its unexpected comparison. But the latter hits hardest when the comparison is both unexpected and revelatory for the reader.”
This last part, “revelatory,” sank into my brain like a depth charge. With LLM output, it’s the revelation that’s in question. Revelation is probably best known in its Scriptural incarnation as the final book of the Bible, wherein a seeming apocalypse is shown to John of Patmos, in the form of a lurid vision. So descriptive are its contents that I found myself at a tender age (being marched through the process of Lutheran confirmation) asking my pastor about it. I remember the pastor’s answer, too—she told me that Revelation had been written during a time when the writers were strictly persecuted and could not worship freely, and that the book itself was coded so as not to arouse the suspicion of anyone who might happen upon it. Needless to say, this was somewhat disappointing, but it also led me to investigate further the title. I discovered that its etymology is metaphorical even on its own: ἀποκάλυψις, from Koine Greek, meaning “unveiling.”
And yet: the “unexpected comparison that is also revelatory,” as my colleague put it, is exactly what creates the beautiful mystery of poetry, creating room to interpret and find resonance in text that isn’t necessarily intentional—but which might reveal itself with each new reader, and with each new reader. LLMs, I have found, use that instinct, that desire, to hide the emptiness they manufacture—such elegantly devised con artists, who need do so little in order to swindle their marks!
For the thirsty swindler, what better model could there be than capitalism, which Marx once called “dead labor, which, vampire-like, lives only by sucking living labor, and lives the more, the more labor it sucks.” What LLMs (and their owners) feed upon, however, is the human desire for knowability. We do not wish to Not-Know—to Not-Know is to be Caught Unawares. Popularly, LLMs are thought of as a kind of hyper-“auto-correct,” finishing our sentences for us, while also convincing us that our input is integral to that process.
But the only thing the vampire needs from its victim is entrance; it does not need further permissions to bleed them dry.
Dramatics aside, this metaphor gives agency, even ulterior motive, to a machine, which (at least for now, however) seems impossible. I think this says more about us as humans at this moment in history; that we are desperate to see revelation where the veil has yet to be lifted. We’ve certainly seen evidence enough in certain political corners that many are willing—and some, to a frighteningly fanatical degree—to look past the nonsensical to satisfy a hollow (and hollowing) feeling. George Zipf, himself, whose law ranks by frequency, even attempted to explain the perplexing predictable patterns he observed by detailing a Principle of Least Effort—even wrote a whole book on it. Speakers, he said, naturally prefer the route of least resistance—fewer words—to articulate their thinking, whereas listeners actually prefer the opposite: more words, to help them understand what is being communicated. Predictability of vocabulary, Zipf theorized, would allow the listener to gloss over smaller, more frequent words, thus conserving energy for focusing on those more salient.
I’m not a mathematician—far from it, in fact—though even the most prominent of those in that field have been unable as yet to understand why Zipf’s law is such a constant. It’s been theorized that “preferential attachment processes” are a primary driver for the phenomenon, which are also based entirely on history and frequency of use. The more something is said, the more likely it is that someone else will say it. The less something is said—evidencing the dissenter, the outlier—the less likely it is to be repeated. This mimesis, or representation of our surrounding environment (and the mystery of its dis/continued replication) is another example of the path of least resistance, lined with primroses.
For a long time, I held out against calling LLMs (Large Language Models) “AI” just because it seemed—empirically—like the wrong word, especially once I started understanding a little bit more about it. I lasted about two or three years, but lately I’ve noticed I just call it what everyone else does: “AI.” I have moments of doom-like stupor, where I think it’s actually rather fitting to be so un-technical about it: to glaze over it with such a vague, misleading term. A sloppy, universal title for something which we have convinced ourselves fulfills the requirements to be named such. If I follow the morbid, masochistic spiral of doom-like stupor further down, it could be argued that this misalignment of title and criteria is a more expansive phenomenon than previously noted—in the words of a movie trailer for any great Spy Thriller—it could go all the way to the top. Could, in fact, have metaphorical origins there. Another vocabularian holdout, and one I won’t be stinting from, is to only ever refer to the person in the highest office in the country by their official title: the current President of the United States. This is a person for whom the mimetic power of their name—as blatting and trombone-like as it sounds—is wildly powerful, a name which has become a kind of sigil, and whose bearer greedily laps up the energy it drains.
Soon enough, I fear, there will be a generation of children for whom the term “LLM” will mean nothing, for whom pablum is a greige ambrosia dispensed cheerlessly in an arid, barren wasteland.
And yet, then there is this—Barthelme’s argument against the consuming blight:
“But artists will respond in such a way as to make art impossible for the computer. They will redefine art to take into account (that is, exclude) technology—photography’s impact upon painting and painting’s brilliant response being a clear and comparatively recent example.”
This, maybe, is the thing that is not known—a hapax legomenon of hope. We don’t know what it is, and we can’t predict it, because it hasn’t happened yet. We can’t ask the LLMs what it is, either, because all they can do is tell us about the past. But the past is where we are most comfortable, is it not? It has already happened, so—ironically, it is “predictable.” We know it can happen, because it already has. What we need is what hasn’t happened yet—therefore, it is impossible to know that it can happen. Perhaps it is radical faith, then, where we must put our hopes—faith, with its own paradoxical (but also strangely beautiful) definition—the “evidence of things not seen.”
So then—is not-seen the same as seen, the same as not-knowing is knowing and not-writing is writing? No matter how I look at it, to say that I’m writing while I’m not writing, inserting a hyphen between “not” and “writing,” does not bridge the fact that I’m describing a negative quality. It is a similar feeling I get from reading a piece of work someone has attempted to pass off as their own, but which has been clearly generated by a predictability machine. How do I know this, you might ask? Well, I don’t. Not for certain. I don’t have some supernatural talent for discerning the gradient of machine and human, like those tests which display narrowingly similar shades of color. Eventually, I imagine, it might become so difficult that I won’t be able to tell the difference between greige and greige.
As I said at the outset, I’m fully aware these things come and go in cycles. Writing/not-writing. Fertile/fallow. Input/output. Our fascinations, too, wax and wane. Hopefully this current paradigm, informed as it is by machines, duplicity and our own our worst narcissistic tendencies, will pass as soon. Until then, let’s not accept half-formed things, which only mimic the structure of dreaming. Even if we do not dream, I think it might be best we don’t attempt to convince ourselves we are not-awake, only for the chance.
Enjoying the content on 3QD? Help keep us going by donating now.