Sampling on the Data Exchange
I’m going to return to a distinction between opposing models of language I’ve made before (most recently here: http://gablog.cdh.ucla.edu/2020/06/toward-a-media-moral-synthesis/) and would like to drive to some new conclusions here. This is the distinction between, on the one hand, language as comprised of individual words, whose meanings can be defined, along with grammatical rules which determine the correct ways of articulating words and thereby creating an, in principle, unlimited number of possible sentences; and, on the other hand, language as “chunks” which are learned together in practice, and which become “commonplaces” and “constructions” as pieces from different chunks can be exchanged for pieces in other. The former model is shared by most of modern linguistics, most prominently in the generative grammar of Noam Chomsky; the latter is advanced by the social psychologist and primatologist Michael Tomasello and, within linguistics, advocates of “construction grammar” and perhaps “grammatization” theorists (who focus on grammatical changes over time) and no doubt others—anyone, really, who treats language as historical rather than structural.
We can now know, thanks to David Olson, that the words+grammatical rules mode of language is nothing more than a reification of the metalanguage of literacy developed to supplement the shift from reported speech in oral settings to recorded speech in literate settings. The metalanguage of literacy is useful only a posteriori, to examine, assess and correct already written prose—it can create nothing (but—an important qualification—very interesting literary experiments and hypotheses, a question I set aside for now). I think that the entire history of metaphysics is really nothing more than an attempt to derive the reality of conversation, thinking, inquiry and discourse from the metalanguage of literacy. This still goes on today, as the default criticism of another’s discourse is almost invariably its “illogic,” “irrationalism,” incorrect use of words, a failure to rely on “facts” and, of course, “critically assess” and “logically” articulate them. The fetish of “proof” is nothing but a mindless reiteration of the metalanguage of literacy, presuming, as it does, the universalization of legal language and the courtroom setting—since the law and the courtroom are largely dependent on precedent and traditions, even in the forms of “logic” and “proof” required, to demand logic and proof outside of that setting is to abstract from it only the metalanguage, which does certainly overlay and inform, and increasingly mar legal discourse. People end up arguing about the “amount” of “proof” they have (and the other side doesn’t) rather than asking the more normal question about some event: what, exactly, do you think happened? What do you make of the way this event played out, and the way the resulting scene is set and peopled?
The real action in language is in the “chunks,” a word which is interchangeable enough with my own preferred term for “sign” or “utterance”—“sample.” Thinking in these terms situates language in its history, including the hypothetical history of the derivation of the speech forms: ostensive, imperative, interrogative, and declarative. As Gans shows in The Origin of Language, what underlies or grounds this succession is the maintenance of “linguistic presence”—the exchange of assurances that we continue to interact on the same scene. So, the first imperative rushes to turn itself into an ostensive; the first interrogative presents itself as an elongated and softened imperative; and the first declarative “insists” that it has merely obeyed an imperative that included the one that has been resisted. In this way we can effect what the title of the post linked to above calls a “moral-media synthesis”—what is “good” is the maintenance of linguistic presence, and linguistic presence is maintained by making the various layers in the linguistic “stack” convertible and “interoperable” with each other. So, a good declarative, or, more broadly, a good discourse, is one that would propose a manner of “fielding,” “seeding” and “populating” the referenced territory with hypothetical observers, actors and inquirers who would redeem the ostensives implicit in all the nouns, adjectives and phrases of the discourse.
The model of good writing and good thinking proposed by the metalanguage of literacy is the “classical prose” explored by Mark Turner and Francis-Noel Thomas: writing (and the thinking which prepares one for, and is exemplified in, that writing) should provide the reader with a clear view to scene presented by the writer. The writer, and the writing itself, should be self-effacing. The writer should set up, even if necessarily from a particular position in relation to the scene, a “window” onto the scene that purportedly provides the view of the scene. One of the virtues of this approach to writing and thinking is its pedagogical promise: it provides a way of categorizing all mistakes and infelicities of prose, whether grammatical or stylistic, in terms of how they intrude the writer onto the scene, and draw attention to the writer and writing at the expense of the scene.
Another model of writing and thinking follows from language as sampling. Classic prose presupposes the existence of a scene, and a vantage point on it; language as sampling refuses to take linguistic presence for granted, and therefore must set the scene, and provide a way of checking everyone on it. Something more like Leo Strauss’s esoteric/exoteric writing, which implants a pedagogical relation within the prose, is more appropriate here. Classical prose imagines “One Big Scene” that all readers can share; language as sampling imagines something more like a single, localized scene, with participants on that scene turning away following its closure to offer testimony regarding what they saw and heard on that scene, with the participants on those scenes in turn creating new ones to testify on, and so on. This set of “staggered” scenes is built into the writing and thinking itself. To put it simply, if we take the claims of classical prose seriously, one should never have to reread a text (other, perhaps, than to remedy a memory lapse)—one should read it once, take in the entire scene to which you have been given a full view, and be able to reproduce it for oneself on demand. You wouldn’t even have to talk about it, because any competent reader would see and take in the same scene as yourself. This is, in a more prosaic way, the “metaphysics of presence.” Language as sampling wants to produce texts that will be discussed, argued over, taught, redirected and recontextualized—although it’s more accurate to say that it recognizes that this will always be the case with any significant text, and that this possibility should be maximized rather than reduced to a minimum as nothing more than the need to clean up regrettable misreadings.
Now, it’s very interesting to point out that algorithmic, computational approaches to textual production adopt, seemingly counter-intuitively, the “chunks” approach rather than the words+rules (maybe words x rules?) one. Google translate began to work when it stopped trying to replace the words in the source language with synonymous words from the target language and then program the translation to arrange the words in the grammar of the target language. Rather, what works is the following: perform a search for all the existing translations of specific phrases and sentences from the source to the target language, and replace the phrases and sentences in the original text with existing translations from the target language. Of course, if phrases and sentences have been translated in different ways, or not every phrase or sentence in the original has been translated (obviously the case with a genuinely original phrase or sentence) the program needs to be further modified, and judgments need to be made—an important thing to point out, but something that only in very rare circumstances of extremely innovative texts would create an unworkability. Similarly, the new language generator GPT-3 seems to work by being fed an opening passage and then proceeding to draw upon the corpus to generate the most like next passage and, then, given those two passages, the most likely third, and so on. It’s a question of generating sample texts that would best (according to some parameter) represent the entire population. To realize that humans don’t think all that differently—following an assignment to fuse text fragments together from various sources, adapting the fragments to make them interoperable with each other (according to an emergent parameter)—is also to take on the possibility of thinking differently.
The same is true of algorithmic “governance” or, we could say, “haunting,” in general. Amazon doesn’t abstract a set of qualities or characteristics of me as a reader and then apply those characteristics or qualities to books that “embody” them—it identifies who else has bought the books I’ve bought and then identifies which other books they’ve bought, and then, after determining what counts as “similar” books and how much “similarity” is necessary, and how many other samples need to be surveyed, is able to get back to me and let me know that I might also like… The more I buy, the more accurate its choices will get. Similarly, if one wants to compose an algorithm for an app (which has already probably been done) to inform its users of the relative “safeness” of the various neighborhoods in some city, it wouldn’t begin with definitions of “safeness” and work from there. It would start with models of the kind of events one wants to avoid when looking into “safety”—muggings, rapes, murders, gang activity, etc., and composes an algorithm that will let you know the likelihood of encountering one of these situations in given times and places, based on police records and other evidence.
To compose a model, you need to start with an event, real or imagined (or real events imagined in a particular way). To start with an event is to start on a scene—it is to enter the scene. There’s no neutral position on a scene, but scenes produce their own means of adjudicating conflicts that might arise on them, and those provide for roles one might occupy—one is “interested” rather than neutral insofar as the overriding imperative is to preserve the scene itself. As a reader, writer, thinker, learner, you don’t learn to “think logically” or “critically,” or to cut your thinking to the contours of specific canons of evidence and proof. You work with models—first of all, some text, which you know had a thinker behind it, which interferes with your existing “store” of linguistic samples sufficiently to compel you to work with and within it to further rework your “store” or “stock” so as to achieve a new consistency. You go through a certain number of texts and authors, and a synthesis, or competing syntheses, emerge—some thinkers get elevated into the judge of other thinkers, some thinkers corner some restricted domain of your sampling—perhaps no one else has yet ventured there—one thinker fills in gaps that made it difficult to integrate yet another text, and so on. You become, yourself, a model of reading, writing and thinking, and you get to the point where you can model your own adjudications between competing syntheses. You become a “self” by populating yourself. Of course, along the way, you can stop and ask questions about “logic” and “proof,” but these will always be questions immanent to the body of texts you have familiarized yourself with and which you also use to defamiliarize yourself—they will really be you either responding to objections you’ve already heard or, increasingly, those you can conjure up for yourself. Your touchstone, though, is never logic or proof in themselves; it is linguistic presence to yourself as a self ensuring your own succession in time as a sample that might spread and be replicated in certain ways.
The fact that this very model lies at the basis of our algorithmic surrounds and hauntings is what enables us to—not “resist,” not desperately assert our “humanity,” which means nothing more than asserting the subject formed by print against the user/interface being formed by planetary computation—enter into exchanges with what is the center that speaks to us today. I tossed out an example in my previous newsletter to bring the question into focus: the algorithm has developed to the point where, as you are writing an essay, your computer will inform you that since you have written that sentence, it is urgent for you to read this text (or body of research). And, keep in mind, your computer will already know whether you have done so. If the computer is, in fact, right, and you couldn’t justify having written that sentence without demonstrating familiarity with a particular region of texts, could the next step—the computer revising your sentence—and then the next step—your computer writing it, and all the other sentences, itself—be next? Couldn’t the program know as well as any inquirer what, given the existing status of discourse in its totality, would most advance the discussion here and now?
Once that became the case, we would of course know it to be the case, and we would therefore no longer “write essays”—if we were interested, we could just ask the program: what’s next? Why would we be interested, though, if we weren’t actually participants in the inquiry? Why, for that matter, would the program be running if no one were interested? And wouldn’t the computers themselves then “lose interest”? It seems that some kind of exchange between program and user is irreducible—the question then becomes, what kind? If the program relies upon human ‘input,” it can’t calculate for “linguistic presence” past a certain point. What only humans can contribute to the center, then, is forms of linguistic presence. I have formulated this many times, in part because it always needs new formulations: originary satire; hypothesizing the present; hypothesizing the whole from a single sample, maximal addressability within the field of sample utterances; testing through enacting the meaning of words; the primacy of the present tense; and so on. We can speak, in more familiar terms, of maximizing self-referentiality. An example: Johanna Drucker has written a book called Diagrammatic Writing, which is a “book that is about itself as much as is possible.” Sentence by sentence it refers not only to its words and sentences but to everything that makes a book a book, including much that we never think of—title page, margins, spacing, print, etc. It’s an interesting book to read and would be very interesting to study, but it is to be read iconically, as a model of enacting literacy. To examine what you are doing, following the “assignment” to identify everything that is “automatic,” sheer repetition, or pre-programmed about it is to create a practice that exceeds its programming and contributes to future programming. The implication is that acting morally is presenting ourselves as a sample within a continually reconfigured field of populations so as to model ways for, eventually, everyone to be doing so all the time. You draw upon all the fields of inquiry within your range and even those beyond it to present yourself as a sample and in doing so give yourself as a sample over to fields of inquiry yet to come.
I return to the definition of “technics” I proposed in my previous newsletter: the perfection and generativity of the imperative. The technical question is, how to form an imperative that will be performed as specified. And, then, can be repeated, with identical results. And, then, can be supportive of and supported by other perfected imperatives. And, then, made to be “imitated” by matter, in such a way as to eliminate some links and add others in new imperative chains. As technological users and interfaces, our primary duty is to participate in this process of perfecting imperatives so they can be disappeared so as to present us with new imperatives. Without becoming proficient in this duty, anyone will be completely ineffectual, socially and politically. “Users” and “interfaces” should be able to itemize all the ways they are fulfilling programmable commands. Indeed, that’s an imperative all can issue to themselves, and work on perfecting. As designers of declarative sentences, or “prosaicists,” we can project and enter the scenes upon which the ostensives confirming the completion of imperative sequences are played out. We can present ourselves as fulfilling programmable, but not yet programmed commands—the programmable can be made to run parallel to and engage in reciprocal translation with the programmed. I could imagine, more or less vaguely, the algorithm that could have searched and sampled through the archive so as to have written this essay; if I were to enter that algorithmic imaginary with an eye to perfecting it, I would write a very different essay, or a new one.