Das Andere der Zählbarkeit (The Other of Countability). Villa Vigoni, Italy, Nov. 28, 2023.
Slides here.
I think I’m here because I’ve written now three books where I’ve counted things in the titles, mostly monolingualism and multilingualism. I’d like to thank my old friend Till [Dembeck] for always having great ideas for ways to bring people together. Till has been a patient and thoughtful comrade in matters of multilingualism for me for 13 years or so, though it feels like longer. I’m going to start by telling you what I’m about to go on about for 30 minutes and then turn to a bit of a touchstone literary moment that’s been helpful to me this past year in framing these questions.
First, “language-agnostic” is a term that has flooded the many, many technical journals of the Institutes for Electrical and Electronic Engineers since around 2016 and the phrase is something that I want to dialogue with today. It piqued my interest first when I was heading up a German Studies Association seminar on “linguistic indifference in German Studies”, which ultimately led to this forum in the German Quarterly, in the introduction to which my friend Chantelle Warner and I [next slide] wrote:
“German Studies researchers working in various disciplines face a range of potential decisions about how and whether to acknowledge the role of language(s) in their inquiries. Plausibly, a novel could be analyzed with or without regard for the (socio)linguistic contexts of its production and reception. History could be presented with or without attention to the linguistically specific discourses and communicative media that facilitate or deter social and political change. Inquiries in political science and economics could be forged without even a mention of the linguistic repertoires and information infrastructures that make or break ‘the economy’ or ‘the state.’ ”
So the forum was generally about how and whether we choose to acknowledge language(s) as a constitutive feature of multidisciplinary research. But as we were finishing up this forum, I was noticing this term language-agnostic rising up quite potently but vaguely in the computational engineering literature, and I wanted to know what the rhetorical relationship was between linguistic indifference and language-agnostic modeling, but more importantly what the political economy was that made language-agnosticism such an emphatic new focus of attention on the supply side in the past 7-10 years. So, it’s important to me that we think of this as a supply-side discourse, made by developers and sold to clients, but more or less undesired and incomprehensible among lay users.
Anyway, let’s turn to countability for a moment and to valuation, which are the other key terms in the title I promised to talk about on the program for today. As a bit of a preface, for anyone who’s spent time around critical applied linguistics in the last decade, countability and nameability of languages are bad things in those contexts. Making a language countable and nameable reflects badly on the counter and the namer; it reveals him to be on the side of neocolonial prescriptivism / positivism, rather than on the side of translanguaging variationism,
And our friend Li Wei describes translanguaging in the following terms
“using one’s idiolect, that is one’s linguistic repertoire, without regard for socially and politically defined language names and labels. This is not to say [adds Li Wei] that the speakers are not aware of the existence of the idealized boundaries between languages and between language varieties. As part of the language socialization process, we become very much aware of the association between race, nation, and community on the one hand, and a named language on the other, and of the discrepancies between the boundaries in linguistic structural terms versus those in sociocultural and ideological terms.”
This stance was perhaps best acknowledged by a title of a book by Robin Sabino [next slide] called Languaging without Languages—which isn’t just a great title, the book is also a fantastic empirical study of how languages come to be countable, named entities and about what counterevidence has to be ignored in order for that consecration to succeed historically.
But now, for our brief literary interlude about countability, nameability, and the alleged modesty of agnosticism. I’d like to turn quickly to Alice Zeniter’s [next slide] lovely novel The Art of Losing, about decolonial memory and French Algeria, and I stumbled onto this passage which
gets at the crux of these matters I’ve been addressing in this introductory section, and then adds a little something extra for us to chew on, which is the theological dimension of countability. So here goes Alice Zeniter:
“Une ancienne tradition kabyle veut que l’on ne compte jamais la générosité de Dieu. On ne compte pas les hommes présents à une assemblée. On ne compte pas les oeufs de la couvée. On ne compte pas les grains que l’on abrite dans la grande jarre de terre. Dans certains replies de la montagne, on interdit tout à fait de prononcer des nombres. Le jour où les Français sont venus recenser les habitants du village, ils se sont heurtés au silence des vielles bouches: Combien d’enfants as tu eu ? Combiens sont restés vivre avec toi ? Combien de personnes dorment dans cette pièce ? Combien, combien, combien… Les roumis ne comprennent pas que compter, c’est limiter le futur, c’est cracher au visage de Dieu.”
And I love that passage so much that I’ll read it now in the English, by Frank Wynne:
“According to ancient Kabilye tradition, you should never quantify God’s generosity. You should never count the men at a meeting, the eggs in a clutch, the grains in an earthenware jar. In some parts of the mountains, it is even forbidden to say numbers aloud. On the day the French came to take a census of those living in the village, their questions were met with dumb mouths. How many children do you have? How many still live with you? How many people sleep in this room? How many, how many, how many? The roumies do not understand that to count is to circumscribe the future. To spit in the face of God.”
And this exhortation to modesty in the midst of colonial power relations brings me finally to the question, or the value, of language-agnosticism in our moment. Now these [next slide] are just a smattering of the research articles that have been published in the last 24 hours or so, where the concept of language-agnostic modeling features prominently in the first paragraph of the research writeup. And let me reiterate that this wouldn’t have been the case 10 years ago, because agnosticism was not a material market value yet, at least not in matters of language.
Back then we were still trying very hard to name and optimize [next slide] certain entities in these processes of machine translation, like JFIGS (Japanese, French, Italian, German, Spanish) or CKJ (Chinese, Korean, Japanese). This era is over, and has been replaced by modeling where the emphasis is allegedly on generously servicing low-resourced languages for vague social justice purposes, but not specific languages.
And most of the hot LLM models are now speculatively “language-agnostic”, on the premise that such agnosticism will most benefit low-resource languages—from Tigrynha to Shona to Tibetan. So, it seems to me like a lot is at stake in this equation, especially for the future of how exactly less commonly translated languages are going to be iterated and resourced or extracted in the coming ten years.
And it also seems that just as Applied Linguists got around to articulating a critique of what they call the “entity-hypothesis”, i.e. the pernicious model under which languages are nameable, countable entities—I think in the Luxembourg School of Multilingualism [next slide] this is called Sprachigkeit (as opposed to Sprachlichkeit)—, at exactly this time (so around 2018 or so) computational engineers quite suddenly lost their fondness for the entity hypothesis, becoming instead “agnostic” both about when and whether the large-batch data they are working from comes from or is benefiting an “entity” called Catalan or Hongkongese or Ewe. But also about whether their efforts are at all designed to service any pre-specified languages or language markets.
And this shift certainly followed a broad inclination that has been long addressed in the Varieties of Capitalism [next slide] literature, about how new political economies since the 1980s favor switchable rather than non-switchable assets, i.e. whether they base themselves in specific bilateral, or rather generically multilateral, arrangements for envisioning and planning supply-chains, value-chains, and manufacture logistics.
So it is interesting to me that critical applied linguistics, literary studies, computational linguistics, and supply-side manufacture logistics are each at very different moments regarding their respective investment in the entity/countability hypothesis—and that investment in each of these disciplines or sectors is not merely one of belief, ideology, and conceptual adequacy, it is a pragmatic investment in what can actually done, in what “value-creation” can be achieved (to use the business language) by virtue of such an allegedly colonial concept as a nameable / countable “linguistic entity”—or rather through an allegedly more supple and switchable gambit as linguistic agnosticism.
So, what does the pretext of linguistic agnosticism allow actors to purvey multilaterally, even when those underlying actions purveyed may continue to rest profoundly on the same old extractivist, monolingual, and sometimes arguably epistemicidal procedures? What I mean is: if you’ve already based your system of analysis on an entity/countability framework for three hundred years, you can’t abruptly back out of it into agnosticism and pretend you don’t know about those named-entity constructions and their practical effect on speakers, language, institutions. This makes me think a bit of Simon Kasper’s upcoming argument later today that ChatGPT doesn’t just work on blank text, it works on the discourses and ideologies that humans actually enunciate.
Anyway, let’s back up and ask some silly basic questions [next slide] of this ill-defined term “language-agnostic”, because it’s been vigorously operationalized in a deluge of computational literature but rarely spelled out. I want to know exactly what is being achieved with this phrase. Not just what is meant, but what is being done. Are these engineers embracing a spirit of modesty, i.e., like Alice Zeniter suggested the French should do instead of trying to name and count things all the time? Or is something more furtive taking place? Anyway, let’s figure it out: “Agnostic” about what?
· agnostic about which language the data comes from, or whether it comes from any language entity at all?
· agnostic, perhaps, about which linguistic communities, polities, and markets are intended to benefit from, or ultimately have to deal with, the downstream effects of the innovation, or which may eventually benefit regardless of intention?
· agnostic about what language is, vs. what is paralinguistic or nonlinguistic?
· agnostic about what language is for: information-conveyance, meaning-making, obfuscation, poesis, documentation, conviviality, intellectual activity, political subjectivity, prayer, or the facilitation of socio-commercial goods transfer
· agnostic about whether, in linguistic matters, authority and order exists or should exist, i.e., an agnosticism about the standpoint and relevance of prescriptivism or descriptivism?
· agnostic in the sense of who does the knowing, the gnosis, and who doesn’t need to know anything at all for the model to be successful
Since the 1940s [next slide], it has been a working premise of this kind of linguistic engineering that its engineers likely should not learn the living languages they are working to manage and translate, because such ignorance or indifference to languages and their meanings are themselves the arm’s-length test of the effectiveness of their code-breaking solutions. The wartime mathematician Warren Weaver recalled the decoding work of a particular “distinguished mathematician whom we will call P, an ex-German who had spent some time at the University of Istanbul and had led an experiment where the most important point is that the decoding was done by someone who did not know Turkish, and did not know that the message was in Turkish” (1947, 1–2).
So, with the rise of language-agnosticism, we witness a renaissance of this common methodological badge of honour among linguistic engineers since World War II: to outwit the languages they intend to decipher and manage, rather than to merely learn the languages—in a somatocentric, analog fashion. This is also a kind of chauvinistic agnosticism.
And then there is the general agnosticism about downstream consequences typical of “Don’t ask, don’t tell” technologies development, since before the beginning of the War on Terror, and indeed the age of nuclear armament too. The agnosticism that says We engineers are not responsible for the downstream consequences of these tools and their iterative presence in actual communities of practice, we are just fulfilling demand. We’re just here to make the models and test them out in a clean-room somewhere. The rest we’ll leave to the policy makers and rational-actor end-users in the Culture.
And then also: “language”: what does that mean here in this phrase? And, without being coy, I will say simply that it is very difficult for an educated lay person like me to read a plain text version of most of these computational engineering research writeups from the IEEE (Intl Electric and Electronic Engineers) and come out of that reading experience knowing for sure, whether the article is talking about so-called “natural languages” or about programming languages. I can understand all of the text on a propositional level, but often it is left ambiguous whether this or that model is about Luxembourgish, Scots, and Inner Mongolian, or whether it is about Python and other programming languages. There seems to be a mix of both.
Most often, the text reads logically with either sense of “language” in mind, and this relative interchangeability in the normative treatment of the matter suggests we have yet another layer of agnosticism on our hands, a will to not differentiate between programming languages and human languages, which reminds me of some of the narrative threads in Neal Stephenson’s 1999 novel Cryptonomicon.
I am also intrigued by the regularity that the alternate formulation “language-configurable”, which you also see frequently in these write-ups, has not surpassed “language-agnostic” in general use. What is it about the word agnostic that surges past “configurable” in a kind of totemic way? Does agnosticism just come bearing the proper sublime affect and reverence for dealing with 7000+ planetary languages, while configurable sounds too much like the assembly instructions for a child’s board game?
If you haven’t noticed quite yet, I am a non-engineer who cares a lot about how engineers talk about language—not because I intend to call them out and criticize them, but because I think these usages are really important, as both performances and indicators for what Dan Slobin calls “thinking for speaking”, how certain concepts get us speaking and acting in certain ways. There is a linguistic relativism at work in this literature, similar I think to what Carol Cohn [next slide] talked about in the 1980s regarding how defense intellectuals talked about nuclear weapons, where the terms, logics, and shorthands used to approach a question end up having quite a profound normative effect on the outcomes of that approach.
That is, if you talk about language in a certain way, you are indeed going to engineer over time certain kinds of new linguistic systems for us, new material realities. And as Dourish and colleagues [next slide] point out, we are in an age of profoundly intensive iteration and socialization around engineering models that used to be kept in the lab until they were complete. Now, a beta model for cross-language management can be sent out into the world of users to be validated long before it’s actually ready, and we users—particularly marginalized communities and low-resourced language users—are expected to deal with all of the effects of having “being iterated” in a rather crass and unforesightful way. All the mistakes, mischaracterizations, optimization errors, etc., we are expected to correct for the model in this social process for the modelers, and we are not of course remunerated for doing so. Because it’s a monopsony.
But anyway, agnosticism seems to lend to this age of computational engineering around language a new “strategically deployable shifter” as Bonnie Urciuoli calls such terms that can do diametrically opposing work at once: here, with “language-agnostic”, projecting a spirit of modesty while enabling more extensive capillary power for the proposed model to materially impact linguistic commerce and reality. There is also the potential that this discourse deftly endeavours to duck the kind of expertise culture that philologies traditionally cultivated for themselves based generally on the premise, advanced by economists like William Davies at Goldsmiths, that the very nature of power and expertise is changing under advanced neoliberalism, where decision-makers and infrastructural actors, so-called “cyborg intermediaries”, much prefer to be invisible and assiduously un-elite in their style of intervention, than to be associated with classic elites who “know” things authoritatively. And perhaps the best thing a powerful capillary intermediary can do to duck such attention and scrutiny is to allege to be “agnostic”.
So, in the remaining time, I’ll share just one quick instance of how computational engineers talk about language agnosticism at present. For one thing, there is the phenomenon that the putative nameability of languages is in some ways being replaced by the desirous naming of language models [next slide]. Much like the nuclear weapons of the early 1980s, they have cute names like BERT, which stands for Bidirectional Encoder Representations from Transformers, out of which spring even cuter next-generation names like XLM-RoBERTa. So, the models themselves are being cathected with the kind of totemic intensity with which a previous generation of engineers were cathecting natural language names like “Japanese”.
I love this passage because it is a persuasive preamble about persuasion techniques, and I’ve coded it lightly to reflect words that signal revenue (green), social justice (purple), and control / domination (blue):
“Arabic, with its rich morphology and diverse dialects, presents unique challenges for text analysis. Our proposed model has the ability to capture the underlying structure and semantics of persuasion technique in text, regardless of language. The results obtained in our analysis demonstrate that XLM-RoBERTa can adapt effectively and perform well on such intricate tasks, even in languages that are structurally different from the ones they were originally trained on. This not only underscores the versatility of XLM-RoBERTa but also sets a promising direction for further research in detecting persuasion techniques across various languages. In future work, we plan to accommodate more languages in the dataset, and fine-tune other multilingual models for this task.”
In some ways, this just feels like such a fun effort to be a part of, a kind of gold rush culture full of ebullience and epic wins. What makes this such a goldrush though? It’s not just the AI-capacity building, it’s also fulfilling some major economic goals that have been identified as desirable for almost a hundred years. Since Hayek [next slide] in the 1940s, microeconomics has suspected that information-sharing is the “chief efficiency of competitive markets” (Farrell and Rabin 1996), such that ordinary, informal talk plays an ill-recognized but profoundly material role for the future structuration of global economic activity. Since the 18th century even, workforce strategists and workplace designers have aimed to refine forms of “cheap talk” across local languages that would yield optimal economic results across markets—or what economists sometimes call Pareto-optimal Nash equilibria (Farrell and Rabin 1996).
Reading any research writeup from an IEEE-affiliated research group gives a vivid sense for the optimization culture that fuels this tenacious and confident project to produce “cheap talk” across languages. The resulting product, cheap meaning [next slide ], is I think likely the eighth “thing” to add—alongside nature, money, work, care, food, energy, and lives—to Raj Patel and Jason Moore’s History of the World in Seven Cheap Things (2017).
And, in this way, we can see how language has ultimately crossed what economists like Marianna Mazzucato call the production-boundary into being secuiritized value creation processes. (Previous to this era, non-commoditised language use had been generally blurred out of the global economic picture, no matter how evidently essential it had been to economic successes in any era, region, and sector we can think of.)
As I wrap up though, I think one of the underlying meanings of Agnostic here is a discourse that flirts with the general notion that “I don’t have to learn languages and neither do you.” Which takes me back to the title of the conference, ambivalenza della calcolabilità, Das Andere der Zählbarkeit. It appears that this turn to language-agnosticism in computational engineering is in fact exactly an answer to that question, what is the other of countability? Agnosticism is, not knowing is, not counting, not naming. But I’m not sure if that is a good thing, as what agnosticism means essentially here is having access to omnilingualism (Kellman 2016), to all the languages, without actually learning any of them. Maybe if we are not actually learning languages, we should at least be counting and naming them, if only out of respect. So, my modest parting thought is: maybe not all Others of Countability are better than countability.
Maybe countability is a diverse wide-ranging domain, and what matters is how you count, how fast you count, why you count, and for whom you count. And maybe abstaining from doing so isn’t the transcendent revelatory stance it’s cut out to be.
I’ll also let you know, in the spirit of this conference about calcolabilità, that I counted my steps up from the lake below and there were 1801. Of course, there were also a lot of beautiful things along the way, a lot of erhabene views, moments of distraction, social encounters, and of course false steps. But I did count them, and I wanted to, not as a form of replicable measurement but as a way of establishing a reverent relationship with a new place, making a memory of it, and of noting exactly what were the barriers, the necessary errors, and the superordinate phenomena or the real and the symbolic that made such counting increasingly senseless the farther I got up the hill. But, also, how such counting was a way of accounting in some modest way for my own fleeting subjectivity in this place.
Thanks very much!
