This is a transcript of a presentation I gave on campus today. Thanks for the Critical AI Studies humanities research group for inviting me to speak. Some UB folks mentioned they couldn’t make it.

My work has investigated how nonhuman actors, particularly media technologies, participate in our rhetorical practices. Frontier AI represents a significant new actor in those arrangements. But that description immediately raises a question: what exactly are we identifying as the actor when we speak of frontier AI?

In the context of contemporary products offered by frontier AI companies, artificial intelligence has had notoriously ambiguous definitions ranging from semi-divine superintelligence to a measure of labor output. The resulting cultural imaginary of AI has focused on the large-language models that OpenAI, Anthropic, and others build. That is, in our imaginary, the models are the AI. That said, frontier AI systems operate with a software envelope that harnesses the LLM. Harnesses regulate AI function through specialized software tools, database curation, output instructions, market-responsive guardrails, and so on. As such, they reduce and shape the world rendered to the LLM, as well as the output it delivers.

The problem of locating AI becomes more acute as these systems participate in their own development. Frontier AI corporations point to increasing AI participation in their own development as milestones toward “recursive self-improvement” (RSI). Theoretically this participation might extend to the point when AIs would fully direct their own improvement. RSI serves as a milestone for AGI. In addition to self-improvement through programming, AI systems also operate broadly in their own infrastructural improvements. AI participates in the design, production, and operation of GPUs, data centers, and related sectors.

From a different perspective, this emerging media situation underlines Friedrich Kittler’s observation that, in the era of computation, “nobody writes anymore” (220). As he notes in his 1992 essay “There is no software,” by the early days of the PC, chip design could no longer be accomplished without computer-aided design software. And increasingly, especially with frontier AI, it is these chips, designed by machines, that do the reading and writing in our world. RSI imagines a media environment in which humans play no necessary role. Hence, nobody writes anymore. Kittler’s assertion that “there is no software” follows on this media situation.  While software can be composed and read as static code, it would be anthropocentric to conceive of human-readable code as what software “really is.” Terms like LLM, harness, and model may be a useful shorthand for some. And this shorthand extends to the tendency to assign agency to AI.  But materially these terms represent machine states being transformed by calculations. Kittler writes that the “decisive capacity of computers… depends only on the degree to which a given item of hardware can house something like a writing system” (226). That is, as Turing’s Universal Machine defined, the capacity of a computer is its computational mutability, its power to write any computational state. As such, when frontier AI labs assert their models are self-improving, we might equally say that AI systems are calculating new machine states.

Within the frontier AI infrastructure, data centers are “something like a writing system.” AI-optimized data centers constitute a new material basis of planetary computation and thus a new planetary writing system. But rather than situating the LLM at the center and the user in or on the loop, we can trace a series of operational inscriptions and circumlocutions. At the prompt or interface, software in the harness processes our inputs in the context of system instructions, conversational history and related data before submitting tokens to the LLM. Depending on the model, additional software operations may precede the inference, and inference might produce structured instructions that triggers the harness to call software tools (e.g., a web search, database query, code execution). The outcomes of those operations are parsed by the harness again before being supplied to the model for further inference. When the user-facing response is generated, it also moves from the LLM through software layers to be rendered as the message.

While the technical distinctions between model and harness are significant, they operate together upon an infrastructure of servers, data centers, networks, electrical grids, markets, laws, users, and so on. To function, AI operates upon infrastructural strata that record, secure, translate, and transform discrete states across their technical layers. And these computations are not free. Computers “pay a price for their decomposability [Zerlegbarkeit] or discretization” (Kittler 227). Programmability depends upon producing and maintaining discrete distinctions as they are translated across technical-material layers.

Frontier AI meters these operations in tokens, but I prefer the related term: inference. While there is no set exchange between tokens and inference, the sociotechnical value of AI derives from our uptake of AI’s inferential output. Whether inferences are produced in tracking a factory production cycle, compiling reports for a presentation, or directing the movement of a child’s virtual companion, their value shapes and is shaped by their sociotechnical encounters. These encounters feed back into harness software, infrastructural hardware, corporate and state practices, policies and regulations, and a network of sensors. In my work, I describe these processes as aspects of operational pragmatism.

Operational pragmatism analyzes the discrete computational operations of AI systems as they are attributed values of intelligence. At the technical level, this mediation is specifically computational: expressions are tokenized, encoded, and processed through discrete operations. Operational pragmatism describes these computational operations as their output comes to be recognized as intelligent, as trustworthy and actionable.

In technical terms, this commitment to recognizability encounters computation’s material operation. As Kittler notes, computers only recognize terms rendered in discrete, decomposable states. As such, artificial intelligence cannot simply model intelligence; its computational inscriptions must be differentiated, calculated, and translated across discrete computational operations. However, nothing in these operations requires intelligence to persist in a recognizable form as those transformations occur. That is, there is no sense that the intelligence of the AI exists in the electricity that powers its GPUs.

There is here, a simple question. Where is the intelligence? It is a question with many answers. My interest is in the notion of recognizability itself. When we locate intelligence, how do we recognize it?  Descartes’ famous dictum provides an example of this situation. As Gilles Deleuze puts it, “when the philosopher says, ‘I think therefore I am’, he can assume that the universality of his premises – namely, what it means to be and to think… – will be implicitly understood, and that no one can deny that to doubt is to think, and to think is to be” (Difference 130). That is, Descartes can reference the experience of thinking because we all already recognize thought for ourselves. As this critical observation moves beyond Deleuze’s critique, there is extensive scholarship on the historical development of recognition as an empirical science through discrimination, classification, and repetition as they are smoothed into common sense practices. That said, to be clear, Kittler’s analysis of computational discretization and Deleuze’s critique of recognition cannot be equated. As I have said, computational operations do not require intelligence to be recognized at each step. Given that, my interest is in how these distinct processes become intertwined: how these computational operations become recognized as intelligible, and sometimes as intelligent.

Given that, an investigation of how output is recognized as intelligent does not fully explain the operations that produced the output. To investigate that side of the question, I turn to Wolfgang Ernst’s radical media archaeology.  His concept of time-critical media describes a set of technologies that must be operated if their stored information is to be retrieved.  For example, an LP must be played. A software program must be executed. From Ernst’s perspective, artificial intelligence does not alter his primary investigation. He writes “For the precise identification of these scenes, and for locating their technological momentum in both its spatial (topological) and temporal (time-critical) unfolding, media analysis stays close to the actual technological event, in quest for the arché of logos in matter” (180). That is, Ernst analyses computation’s operation as it organizes reason in matter. When a frontier AI model generates a response, what has materially occurred in the AI system to produce this result? Radical media archaeology locates “the specificity of media enactment which occurs disruptively between technical hardware, its symbolic diagram, and abstract computation” (181).

Ernst’s radical media archaeology sits across a divide from other media studies methods. As he describes it “media phenomenology concentrates upon mostly screen-based media effects on humans… media archaeology intends to increase media awareness by making its technology transparent, opening the proverbial ‘black box’” (27).  Without entering this debate, there are recognizable methodological differences. In his articulation of “ambient engineering,” Thomas Rickert’s approach is phenomenological, following on Heidegger and Merleau-Ponty. His development of Heidegger’s remoteness addresses the phenomenological encounter with increasingly sophisticated devices. As he writes, “through the use of powerful computational and predictive technologies, we see the emergence of the capacity to adjust at scale and in micro-moments the foreground and background constituting the warp and woof of existence” (427). His analysis of ambient engineering in terms of habits and practices focuses on the capacity of these technologies to influence our behaviors. Despite acknowledging their increasing cultural and rhetorical power, Rickert writes “the machinic is strong in calculative potency but, if I may phrase it this way, poor in world. The sense of world it achieves through datafication and machine learning is derivative of the human quotient of world” (427). If the achievement of AI is “poor in world,” can that poverty be remedied by further computational achievements? Whatever value may be ascribed to the “derivative of the human quotient of the world,” Rickert remains skeptical. Instead, he attends to developing rhetorical habits and practices of remoteness that respond to AI’s poor understanding.

Building upon radical media archeology, operational pragmatism attends to computational media as it participates in calculative power and understanding. In one sense, an LLM exists as one or more files of weighted values. However, as we know, the operation of frontier AI LLMs requires a significant and growing infrastructure of data centers, power generation, and related supports. This is in addition to the manufacturing infrastructure from mining to silicon chip fabrication. Seen from another angle, starting with a human typing into a chatbot on a smartphone, there are many layers of discretization, programming, signaling, transforming, and translating. But there is no moment of abstraction when intelligence appears.

Across these events of prompting and outputting, media enactment is both continuous and transitional. That is, when AI operates, it is operating among transitions and synchronizations in media. Ernst’s analysis of the organization of reason into matter finds media enactments recurring across the continuous and transitional media of the AI system. They occur in each programmed step. However, these materializations of computational reason do not, in themselves, locate intelligence.

In terms of operational pragmatism, what we can trace across these synchronized, multidimensional transformations is the production of information. Artificial information. In a technical sense, AI media systems are informed by the calculative processes of LLMs. Historically, informing has two meanings. In one, a sculptor informs the marble. In the other, the sculptor informs her students. HR prefers that the sculptor maintains the boundary between these two informational methods, as does the sculptor. In the sculptor’s case, we have little trouble associating either kind of informing with intelligence because we already recognize the sculptor as intelligent.

In the sculptor’s case, we take intelligence for granted. But what if we begin with artifice rather than with an intelligence we already recognize? We do not deny other animals have artifice, though I think it is an interesting question to consider the ways we equate artifice with intelligence. Animals migrate; they build dams, hives, and colonies; they socialize. An ant colony, as an artifice, informs itself. Is it an AI? Could an ant colony be trained to function as an AI that produces information? If we could network ant colonies to produce AGI would we say that ants are smarter than humans? Adrian Tchaikovsky’s novels offer a speculative answer.

That said, for my purposes, it is sufficient, as a starting point, to say that the answers to “where and when is intelligence?” are historical and material matters of concern. Operational pragmatism investigates our tendencies to imagine the virtual-computational world as another universe where the superintelligent AI speaks from the outside. But in fact, there’s just the one world.

Works Cited

Deleuze, Gilles. Difference and Repetition. Translated by Paul Patton, Columbia UP, 1994.

Ernst, Wolfgang. Technológos in Being: Radical Media Archaeology and the Computational Machine. Bloomsbury Academic, 2021.

Kittler, Friedrich A. “There Is No Software.” The Truth of the Technological World: Essays on the Genealogy of Presence, edited by Hans Ulrich Gumbrecht, translated by Erik Butler, Stanford UP, 2014, pp. 219–29.

Rickert, Thomas J. “Ambient Engineering: Hyper-Nudging, Hyper-Relevance, and Rhetorics of Nearness and Farness in a Post-AI Algorithmic World.” Rhetoric Society Quarterly, vol. 54, no. 5, 2024, pp. 413–30. https://doi.org/10.1080/02773945.2024.2407263.

Tchaikovsky, Adrian. Children of Time. Tor, 2015