1 Technics and Text

Language and Digital Sense

There is little doubt that large language models (LLM) are good at predictive aspects of language processing.1 They scrape information from diverse digital registers and achieve a synthesis akin to reading and writing (Pavlick 2023). These models have emerged as part of our investment in neural network AI, enhanced by transformer algorithms that incrementally map the syntax of near-infinite digital resources. They train on enormous data sets, scouring the internet, sweeping through texts of dubious authority, and absorbing our every word. Generative algorithmic infrastructure allows them to simulate and mimic all that they have read. Perhaps this kind of AI is best suited to our distraction economy and post-truth conditions, as their power to fabulate diminishes our trust in the authenticity of texts. These algorithms embody a kind of neoliberal relativism in their performative attention to an internet of alternative truths. On the other hand, this might be the dawn of a new kind of digital or technical encyclopedism, and proof of an information ontology. Some advocates argue that large language models have achieved an “implicit embodied knowledge” of language, and that they are very effective at simulating a “grounded” reading, but there are all kinds of debates about how they do so and the extent to which their language use can be characterized by traditional linguistic and cognitive theories (Hayles 2023; Millière 2023). Rather than reasoning in a classical sense, they operationalize an algorithmic architecture called an attention mechanism that decomposes grammar and lexicon into a scheme of plausible “relevance” measures (Vaswani et al. 2017). Critics of large language models see them as poor examples of AI (Marcus 2022; Chollet 2019), because they are deemed stochastic parrots and inefficient learners, lacking symbolic and logical structure as well as any grounded semantics (Bender et al. 2021). And yet our understanding of how these technical models achieve their language skills is nascent, and it feels far too premature to dismiss their cognitive behavior as utterly baseless.

LLM are impressive at language performance. They are able to synthesize and compose texts as though they were reading. In one sense, the evolution of machinic or artificial readers lends support to the view that reading is a highly embodied and perceptual process, a matter of roaming over signals, in search of patterns, rather than a matter of innate grammatical structures assumed to be managing language production. The neural network topologically surfs the darkly folded terrain of plausibility (the hypothesis space) in a manner that remains opaque to humans who wonder at their perspicacious capacity. When we ask a large language model to read and synthesize Simondon’s seminal book, it does a decent job, just by brute empirical encountering of texts. Or does it? Which different perspectives have been compressed into a voice, and to what end? The challenge of reading and synthesizing a difficult text has always involved some manner of fabulation, even as one strives for an accurate rendering.

LLM use the units of language in different ways than humans. The extent to which they encode discrete constituent symbols and rules, organized within predicate-argument structures, is as yet unknown (Pavlick 2023). Methods for identifying internal structures and causal inferential patterns within the language practices of LLM suggest we might yet crack open the “black box” of their language use (Geiger et al. 2021; Meng et al. 2022). But critics argue that such AI models ultimately fail at systematic compositionality—whereby components can be reassembled so that meaning is preserved—because this requirement depends on the notion of discrete concepts, while LLMs struggle with individuating concepts and often entangle concepts (Mitchell et al. 2021). In other words, they are not good at separating and discretizing concepts (square, round, red) perhaps because they are simply not built to do symbolic reasoning and perhaps also because they read too widely; indeed, they must read obscene amounts of text in order to generate even the most minimal “understanding” of one text, and they struggle with compositional synthesis, despite their apparent fluency that is gained at huge computational cost. Perhaps difficulty with concepts makes current machine readers closer to humans, who tend to operate according to irrational associations and fuzzy concepts, and whose “attention mechanisms” are often spread too thin. And yet humans create concepts and generalizations that seem greater than the combinatoric aggregate, and are able to learn generalizations after only a few examples. Human learning is far more efficient, but also more speculative, as though operating with an entirely different approach.

Neural network models treat language as a stochastic being, something that is inherently statistical in distribution. The model searches for the weighted statistical significance of a text, tracing its citations and echoes, imagining its web of influence (Millière 2022). Notably, humans persistently diverge from the rules of statistical reasoning, as is evident from decades of decision theory research (de Freitas 2025). Such work suggests that embodied human reasoning about uncertain conditions does not follow the rules of classical probability. Chater (2023) claims that LLM language competency cannot be robust if it lacks a direct perceptual encounter with the “real,” by which he surely means an encounter with wet-organic-matter. Perhaps fleshy bodies enact a probability in the wild, and we might adopt a Wittgenstein perspective, whereby language-use determines meaning, to redefine “use” and claim that LLM are “as good a member of the causal chain as anyone.” (Pavlick 2023, 10). Like humans who learn through mimicry, these models track the residue that remains from human reading and use this residue to parrot and simulate human text. And yet their mimicry is not adequately concretized in the associated milieu of their sense making. They cannot compute the undigitized marginalia, nor the tempo and breath of the diverse readings of any one text. They reduce text to a passive, quiescent, accommodating object; the LLM finds use-value everywhere, and never seems to encounter a text outside of its purpose, never a question to which it will not respond.

Genealogies of Neural Networks

Simondon’s attention to technicity and the “ontogenesis” of technical objects was composed in the 1950s and was offered as a correction to both cybernetic materialisms and formal “molding” idealisms (De Boever et al. 2012). He was critical of Marxist treatments of technology for their lack of historical ontogenetic analysis, and their failure to adequately theorize agentic matter (Bardin 2021). For Simondon, tendencies to fear or fetishize information technology fail to comprehend the nature of technical existence (Bardin 2015). We fail to reckon with our earthly environment if we only react with fear of alienation when faced with a technical automaticity that mimics human outputs (Barthélémy 2015). Voss (2019) critiques Simondon for not attending adequately to the economic forces that shape developments in technology, and she is concerned that people use Simondon to “celebrate the apparently self-sufficient, intrinsic aspect of technological functioning” (2019, 5). I think this critique is unfair if it means we abandon attempts to make sense of technicity at the scale of the technical element or the algorithm. The challenge is to think with Simondon at multiple scales, exploring the elemental technicity (of the neural network in this case) while also exploring the socio-technical reality that emerges, a challenge that he himself meets by integrating biological analyses of auto-kinesis with cross-cultural studies of larger social formations. To answer Voss’s (2019) concern is to extend this study, tracing the impact across individual and transindividual scales, revealing how algorithmic technicity loops into larger scales of mattering.

Simondon’s ideas influenced Gilles Deleuze tremendously—the key concepts of virtuality, singularity, and minor/major, among others, are firmly linked to the work of Simondon (Bowden 2012). His genealogical method begins with a search for technical elements, recognizing that there is a lack of necessity, coherence, and measure at the level of the technical element. Nonetheless this elemental scale forms a powerful structuring schema or figuration that patterns the information, long before anything is assembled into an industrial purpose with utility. Innovation and invention occur at this elemental scale, in the associated milieu out of which new technical objects and ensembles emerge, and which circle back to inform the associated milieu. Relatively abstract technical objects are precarious because they are “part” of a disarticulated collection, not yet concretized into conjugated functionalisms. As they further concretize, they become more like natural beings, and can be studied inductively over time and in empirical environments. “The necessity of adaptation, not to a milieu defined as exclusive, but to the function of relating two milieus that are both evolving, limits adaptation and gives it more precision in the direction of autonomy and concretization” (Simondon 2017, 56). Adaptation is not properly speaking just a compromise: “The different aspects of the technical being’s individuation constitute the center of an evolution, which proceeds via successive stages, but which is not dialectical in the proper sense of the term, because, in regard to it, negativity does not play the engine of progress” (2017, 71).

Notably, the publication of On the Mode of Existence of Technical Objects coincided with the 1958 launch of the artificial neural net, by Cornell psychologist Frank Rosenblatt who created the “perceptron” machine, a five-ton computer that taught itself how to distinguish punch cards that were marked on the right or the left, through sampling and self-correcting. A New York Times headline stated: “New Navy device learns by doing: Psychologist shows embryo of computer designed to read and grow wiser” (1958). Our current machine learning models are derived in part from this earlier attempt to perform a kind of automated reading. This connectionist neural network approach to AI was displaced for decades by efforts in symbolic reasoning, but since the 2000s, layered neural networks have returned to dominate AI models. These models can only operate with large scale datasets, and they are hugely inefficient learners compared to humans. These models mobilize a calculus of gradient descent, which allows them to minimize error as they pursue an emergent plausible text production. These “creative” machine readers are capable of artful expression and simulation and deep fakes, just as humans are also capable of parroting, mirroring, copying and forgery (de Freitas 2022). In the 1950s, Rosenblatt was trying to build a machine that behaved like other neural networks in humans and animals, as part of a cybernetic paradigm shift. Initially clunky and limited in the way chance was concretized in its shifting statistical weights, the connectionist paradigm has evolved into neural network infrastructures that are now extremely impressive in their ability to simulate human language-use. The neural network technicity of the Perceptron persists in these newer generative models, because access to excessive amounts of computational power and data has changed the technical milieu (Buckner 2019).

Trans-in-dividual Technical Reality

Combes (2013) emphasizes that Simondon’s concept of transindividuality is not confined to the human species, but is a field of relations among psychical, biological and technical processes. Simondon’s genetic ontology attends not just to human processes but also to organic and physical processes by which complex technical ensembles come and go. The challenge is to understand exactly how the human in the loop is situated and agentic in this technical reality (Amoore 2020). This kind of inquiry cannot be the work only of technicians (at one scale) and sociologists (at another scale), says Simondon, for only philosophy does the hard work of conjoining vastly different scales of mattering. A philosophy of technology must grapple with the unscripted ontogenesis and inventive nature of human technical activity, and discern its relevance to material and cultural practices at different scales. Such work cannot be axiomatized: “It could be that ontogenesis is not able to be axiomatized, which would explain the existence of philosophical thought as perpetually marginal with respect to all the other studies, since philosophical thought is what is driven by the implicit or explicit research of ontogenesis in all orders of reality” (Simondon 2020, 256). Although his work sometimes fuels proposals for a “transcendental technicity” (Hörl 2015, 9) that may seem to abandon the human, Simondon is hopeful that philosophy will find a way to comprehend the complexity of our relationship with technology. In the midst of strong socioeconomic forces that often shape invention, Simondon seeks a “reciprocity of exchange” and even a “social relation of sorts” between the human and technology. The task of philosophy is then to trace technicity across element and ensemble, to study the socio-technical ecology as it evolves, and to cultivate a new technological culture (Simondon 2011, 2012).

If we are to make sense of LLM reading habits, we need to analyze the technical ensemble (platforms, software, public hype, policy) in relation to the underlying technicity of connectionist computing. We neglect the scale of the technical element (the algorithm, for instance) at our peril, distracted by the fumbling software assemblage and its inadequate sense making, riven by fear that machines will replace us. In other words, we are distracted by and in thrall to the automation of reading, the simulation of writing, and the learning machine, as that which resembles and replaces the human. And yet, Simondon and more recent commentators focused more explicitly on LLM note that machines are not operating in human-like fashion, as far as we know, and that our anxiety about their replacing our labor is not about an essential resemblance between human and machine, but about their output or performance (Chater 2023). As usual, the humans are distracted by appearances, and need to attend to the process of engendering. Getting under the hood helps reveal the technicity at work, but so does an archaeology of media and a genealogy of technological descent, revealing the processes of technical individuation and entanglement (de Beistegui 2012).

Simondon suggests that human beings are essentially transducers in the sense of being a dynamic mediator of technics and the world. Following a somewhat chastened Bergson, Simondon characterizes the living thing as that which modulates a temporal process of contraction (life actualized) toward the formal concretizations and yet more dissipated contexts of the technical milieu (Simondon 2017, 156). The relation between organic life and machine is transduction. But a living transducer goes beyond the image found in machines, where mechanical transducers manage indeterminacy and procedural randomness. The living thing is inventive in ways that are not merely combinatoric; life “has the capacity to give itself information, even in the absence of all perception, because it possesses the capacity to modify the forms of the problems to be resolved; for the machine, there are no problems, only data that modulate the transducers . . . There is no true virtuality in a machine; the machine cannot reform its forms in order to solve a problem” (2017, 56). Ultimately, for Simondon, the distinction between human and technical thought rests on the processual dynamics of virtuality and the associated problematics of the milieu.

Maladaptive Technology

Reading brings me pleasure. It seems as though my reading reanimates a text, and links me to other readers across a transindividual collective. But there are other kinds of nonhuman readers, like the large language model. By lifting the hood and searching for the essential technicity of LLM, we can offer philosophical insight into the associated technical milieu. The core technical element of this widespread AI technology is the artificial neural network. Building larger and larger neural networks that consume more and more computing power, scraping more and more websites, and requiring more and more lithium will surely lead to what Simondon calls “maladaption.” Technical maladaption results in socio-technical collapse, when forms of technical life diminish the environmental milieu. It is not difficult to imagine a population, like we have today, fixated on screens and keyboards, without any other material skills; we live an all-encompassing digital life. Technical maladaption leads to nonstop tweaking of a machine that is no longer open or capable of more adaptive purposes. To avoid such hypertelec technical fetishism, one must understand the ontogenesis of the undergirding elemental technicity, and then find a way to get at the “open machine” that is buried in the bloated technology. Only with open machines is the environment itself brought forth as responsibly agentic, rather than scraped for its resources (Simondon 2017, 58). This would be a learning machine that does not entail deadly mining of the milieu, extracting lithium and rare earth minerals to serve our digital desires; this would entail birthing a technicity that is in reciprocity with humans and ecology.

Notes

This essay contains some writing previously published in E. de Freitas, “Fragile Books and Machine Readers: Trans/in/dividual Reading Tactics in a Complex Technical Milieu,” International Journal of Qualitative Studies in Education 37, no. 6 (2024): 1655–65.

1. 1. LLM include, for instance, ChatGPT, where GPT stands for “Generative Pretrained Transformer.” I will use “large language models” throughout this chapter to cover these kinds of technologies.

References