Idiomatic Intelligence and the Black Box

A while ago, partly because my reading of Peirce got me interested in probability and knowledge, and partly because my work as a college writing instructor got me interested in the formulaic nature of language, I tried to think through a way of assessing and explaining any discourse, idiom, text or utterance in terms of degrees of predictability. Given the explosion of LLMs, this seems like a good time to revive this line of inquiry. I started with the assumption that the more predictable a “piece” of language was, the less meaningful and memorable it was, as it served merely to increase the predictability of all the other most predictable utterances. Such utterances have their place, of course, as “phatic” communication, but it’s not what I was interested in. Now, while the LLM presupposes its entire database against which to measure the predictability of the next utterance, we can be much more precise than that (so could LLMs as well, of course, if trained to be so)—what is most predictable in this situation, with these interlocutors, their history, separately and in interaction with each other, what kind of “genre” are they participating in, etc.—we’d have to account for unimaginable amounts of carefully curated data here. At this point, in fact, this conception is more of a “data sublime” sensibility than a real proposal—but the distance between these two is shrinking rapidly. Now, the most unpredictable utterance would be a randomly generated one, like those generated for us for passwords, and that also doesn’t teach us much about the workings of language. So, if not sheer unpredictability, what would be the “other” of the “predictable” utterance? The best way to think about this is to split the scene between the scene itself and the disciplinary space within the scene to be created by the utterance that is maximally unpredictable on the scene but maximally predictable for those joining the disciplinary space within, around, or over the scene.

What we have here is not so much a secret or coded language as a performative one, with its performativity measured by its scrambling effect on the scene—the way it alienates everyone from the position they are presupposing on the scene. It’s not secret or coded because it doesn’t exist prior to the utterance itself which, as an iteration of the originary sign, is cut and carved out of the semiotic material of the scene, therefore constituted by a great deal of improvisation. The effect is precisely to split the scene, and place everyone else on the scene on the boundary between the two scenes, upon which one acts differentially.  You could think of it as reworking the map and the terriotory simultaneously. This act is what I came to call “originary satire” in Anthropomorphics. This formulation has its roots in the implications of avant-garde writing and post-structuralist theory as much as in the emergence of algorithmic governance. This formulation of the doubled scene now provides for a position on the margin of LLMs, at its limits, and inside the black box within which extremely large numbers of parametrized tokens generate language in ways no one could map out without, it seems to me, another similar LLM which would then contain that same opacity, requiring yet another LLM to map it out, ad infinitum. The LLMs replicate the same articulation of transparency and opacity as language itself, where we can all point to the same thing and thereby render it transparent, but at the “price” of not being able to point to how exactly we are positioned to point at it in that particular space.

The end game of prompt engineering, then, is to enable us to continually position ourselves within the idiom that is most unpredictable on the scene while being most predictable for those taking up the idiom on the meta/para/infra scene. This is a position that can be programmed for, and the programmers can be programmed to program for, while being irreducible to any program. The purpose of all the programming is precisely to free us up to take this position, to create this unique transparency within opacity. And the opacity we generate will be the source of transparency for others, including us as we become other. It’s first of all a habit to get into, a habit of speaking and speaking about speaking, a “gesturification” of one’s every utterance, a sampling. A sample must itself be maximally typical but also maximally scenified, this sample place here and treated just so right now to extract from it maximum knowledge and representability. To ask someone why they said something so predictable, or whether they intended what they said to be new or interesting in any way is a hostile gesture, but to take up what someone has said or done as if it were unpredictable on the scene but predictable on the doubled scene is to create a new mode of reciprocity. Sometimes hostile gestures are necessary, but even then the model for the hostile gesture should be the disarming satirical one that invites the other to join a space of inquiry into the idiom we are in the process of creating. One does not thereby eliminate predictability (as if doing so were even imaginable): in repeating and reinscribing this split scene as a practice a broader predictability always remains a backdrop produced by the residue of any series of practices while predictability itself is always deferred. This is the main form of deferral available to us today, as bringing to bear the predictable upon the individual case unremittingly is the most destructive form of violence possible today and one which bears more than a faint family resemblance to scapegoating. The predictable will always bear down on the individual—that’s the nature of language and scenicity—but the predictable can only be rescued from its ultimate falsity if it is made to run through a gauntlet leaving it, at the end, predictable only to those designing the gauntlet, setting up the obstacles, etc., so as to distill a precedent-setting decision.

What one makes of the black box follows from what one makes of language—or, more precisely, whether one thinks language is elicited through exposure to a world somehow already nameable or one thinks language is learned by retracing, however idiosyncratically, the emergence of naming itself. If you think we’re hard-wired for grammar and merely need to have the words of a particular language plugged into the “mind,” the black box is an anomaly, even a scandal: machinic language generation should take place through logical elaborations of basic statements which could all be reproduced, assessed and made the basis for further assertions designed to withstand logical scrutiny. If you think language is learned by participating in human scenes where one follows the gestures, obeys the commands of and issues requests to, others, experiences the co-emergence of objects and their names, imitates, practices, mistakes, varies (through substitutions and combinations) “pieces” of language derived from one’s surroundings, then you realize that language cannot be separated from the events in which we interact with others and try to predict what they might say after we say something and discover what to say next only by improvisationally drawing upon tacit semiotic resources at the moment—in that case, we’re all black boxed.

In that case, the GPT is the other/meta/infra scene of our scene of language learning, with the machine learning and human learning proceeding in parallel and reciprocally informing each other. We learn language by seeking to predict the actions of others on the scene and trying to be predictable ourselves, but also by learning that predictions often fail from both directions, which requires us to seek and transmit more information facilitating the reading of a particular sample—this in turn entails stretching the range of predictability so that one could, for example, surround one’s sample with markers of extreme unpredictability that are deferred within the sample itself (this, among other things, can be called “irony”). This approach to language learning as approximation to an elusive norm that must be elicited and framed against the non-normal suits the “scribal theory of history” I’ve been exploring. The ancient scribal community was a pedagogical order, focused on introducing a small minority into literacy—and, at the highest levels, an extremely advanced form of literacy. The mode of teaching writing was, no doubt, by training learners in the formulas and commonplaces of the culture, themselves drawn from and feeding back into both the wisdom literature and what I will now call “narratives of succession” comprising the team. Narratives revolve around exemplary figures whose stories are also a source of proverbs, by-words, sayings, and so on. We can assume that scribal learning begins by having students repeat and memorize the most widely shared versions of these idioms, which comprise, more than an analysis of grammar or lexicon, the language of the order. This training prepares the student to handle more complex and ambiguous versions of the idioms, and eventually—perhaps only the most advanced and trusted—to participate in revising, filtering, collating and archiving the collected texts of the scribal community. Treating the ancient Hebrew scribal communities as exemplary (yes, I am choosing them), I feel increasingly confident in asserting that this textual legacy amounts to the preservation, refinement and legitimation of a deed of inheritance, ultimately anchored in land and traced back to a trust, or bailment. Literary culture is a kind of legal and accounting culture, through which debts are ascertained, sorted out, repaid and deferred.

As language learners, then, we always begin with what is most predictable within the space or upon the scene within which we are situated—we “begin with” in the sense of repeating and trying to get it right, which is to say, to say the same thing everyone else is saying. If we could continue to do that indefinitely we probably would because that is what it means to be on a scene, which is to say, saying the same thing as everyone else so that everyone doesn’t do the same thing at the same time. But we can’t, because the lines between saying and doing get blurred and one finds that one does not know whether what one is saying is the same as everyone else, in which case you fall back on what seems most like what at least some others, maybe one other, maybe one particularly important other, is saying, and repeating that, while stretching what will count as repeating with variations that elicit transitions into new predictabilities. Ultimately, this entails being the same thing you are saying. So, what was very much like saying the same thing as everyone else starts to become perilously close to everyone doing the same thing at the same time precisely because someone, maybe you, is not saying the same thing. That just means that everyone must be saying the same thing as everyone else on some other scene, meta or infra, and you approximate what everyone else is saying until everyone can see and say that it is the same. That other scene might be a prolonged one, comprised of widely distributed doings, so one takes out derivatives on its completion. The thinner the line between maximally predictable upon the scene and maximally unpredictable upon that scene but maximally predictable upon the new scene being erected the better—this is where the oscillation of attention between sign and thing at the center (this thing) accelerates to the point where motion is undetectable. On this scene, heightened engagement with other intelligences confirms or authenticates the human as the only intelligence that is always saying this sample is the same.

LLMs, like all technology, are revelatory because they replace things humans might have done with things machines can now do, which raises the twinned questions of what, exactly, were we doing that got automated, and what are we doing now, confronted with the automation. If we say that AI can’t “communicate” (a word that is itself surely an instance of new technologies being retroactively applied to human activity) or “understand” (ditto, probably, even if we’d have to go back further) then we are doing two things: identifying something presumably distinctly human and doing so against, but also because of, the technology that we want to insist has not yet crossed that threshold; but, also, we are identifying some new capacity that we can try and program computers to simulate effectively enough that we couldn’t tell the difference—and in this case we are acknowledging that distinctly human capacity to have been “always already” technological. The originary hypothesis interferes similarly with its object of inquiry, though, by “reducing” any human interaction to the “deferral of violence” upon a “scene” in some “event,” which confirms but also deflates such treasured words as “faith,” “transcendence” and so on. It would represent a failure of theoretical ambition to stop there and try to emphasize the “confirmation” and downplay the deflation or demystification—GA then just becomes another modern conservativism “proving” that something like “religion” is good for holding society together, etc. The approach I’m arguing for here (and everywhere else) is to knowingly treat everything we say and do as deferrals carried out on scenes in events, with those deferrals getting more prolonged and discretized as the center is amplified—somehow remembering that there will always be something unknowing about that knowing that others will point out to us. Originary hypothesizing is, then, a technology that we program to propose and implement new modes of deferral, and to gather data (detect, record, and analyze) through the techno-scenes we construct that feed back into that programming. Rather than distinguishing the human from the rest, the originary event transforms the world into the imminent architecture of deferral. We are never done anthropomorphizing ourselves and each other because we always do so on the model of the center, which means the reciprocal imperatives (commands and petitions) crisscrossing center and periphery, and which are always revealing new layers of the human and new elements of the human to be deposited upon the increasingly articulated acene. When we look at each other and the world we’re making, we see ever new features of the human that are simultaneously signs of our becoming technological and therefore the prospect of future iterations of the human. It has become routine to think in terms of technological solutions to human problems, like, a blockchain to eliminate bureaucratic and autocratic or just flawed human mediation and judgment, but this just shows us what we thought mediation and judgment look like and how they “works” while creating a new field for new modes of mediation and judgment that will have been “human” while never having quite existed before the technical intervention. And then “blockchain” will start to work as a verb in sentences, something we do as a matter of course and that in turn suggests new modes of technicized deferral.