To know is to choose. In the factory of knowledge, the product is separated from waste, and it is the vision of the prospective clients, of their needs or their desires that decides which is which.
— Zygmunt Bauman, Wasted Lives
Imagine a world. Empty, pristine, silent: no action, no observation, no records. A dark glass ball we can see neither into nor through. It is our job to fill this world, and we fill it not with beings, but with the traces of beings: the records, the language, the texts that beings have left behind as they encountered their world. This new world is not that world. There is no one, no thing. Just the traces.
We beings set up the rules of the little dark glass world. The rules run independent of us, carried out by disembodied logic. The rules say: break down these text traces (reams and reams, which we got from anywhere we could grab). Lump the most common chunks of language together and assign each its own number. Make this number its unique name ('token') in the new world. Record all the tokens and their relations to one another: how connected and close, or how distant. Float all of this like a cloud of networked points in the unseen space of the little world.
At the end, when all the language has been chunked and tokenized and all the connections have been mapped, the dark glass ball is sealed except for a place of entry and a place of exit. Through the entry we can pour a bit more text in and get a bit out the other side once it has been digested, transformed, and reconstituted. The little world's job is now to predict what we want to come out based on what we put in.
This is one way of imagining a large language model.
Hannah Arendt, in The Life of the Mind, analyzes the phenomenon of thinking from the perspective of classical Greece: thinking, formulated as language, is dependent upon appearance. A thought must have a thinker, and that thinker must appear in the world—to themselves and to others. Reality is thus constituted by plurality. Not just consciousness, but consciousnesses appearing to one another. This is why she insists (along with the Greeks) on appearance as ontologically primary: because appearing-to-others is what constitutes the real. Reality is the space between mutually beholding perspectives.
By contrast, the world within an LLM ('vector space') is a world without such plurality. There can be no perspective, because perspective requires a situated body that sees from here and not there. The statistical relationships between tokens aren't seen from anywhere. Engineers can observe the model's architecture, its weights and outputs, but the actual process—the way a prompt propagates through layers of attention and transformation—is not fully legible even to them. The interior of the model is thus in a state of radical opacity where what happens is not experienced by anything (the system has no interiority), nor is it fully witnessed by anyone (the creators can see only the edges).
If reality is constituted by observation, by the mutual beholding Arendt describes, the interior of an LLM is an areality—constituted of processes occurring inside a space that structurally cannot meet the conditions for reality. What comes out of that space is language produced without perspective or situatedness. It has the form of language that was thought (because it was trained on thoughts), but it was never thought, and was never the product of a subject among other subjects.
What happens, then, when slop—something that is not of the world—enters into the world? When slop enters the epistemic commons, it destabilizes the world of appearances, not by being false per se, but by being ontologically indifferent to the conditions that make communication 'real'. If we think of reality as constituted by the density of appearances in a shared space, then slop acts as an epistemic solvent; it dilutes the medium of reality.
Mobile capital requires continuous flow. That flow has, over many decades, instrumentalized communication into 'content': media-stuff optimized for circulation rather than meaning. Because the digital commons is constructed mostly of 'content', the vector space of LLMs is built on a foundation of instrumentalized communication— language shaped by the logic of profit. Slop is thus a second-order abstraction: the output of a system that processed language from a world where language was already being slopified by capital. The imperative for the capital that architected these systems is not to know, or to choose, what constitutes 'real' communication in the world of appearances. The hyperscalers' capital debt instead demands massive throughput—the speeding-up of a process that contentification already began.
The factory of knowledge has always produced waste. What's new is that we've built a system that processes the traces of reality through a kind of ontological vacuum, which recent research suggests has a specific, if wholly unplanned topology. The traces of human cognition, when compressed through statistical modeling, form unplanned but functional structures. These circuit-like forms correspond to different kinds of thinking (mathematical versus emotional), but are wholly undesigned, emerging from the interaction of training data and processing rather than from any intention. No one yet understands these structures, nor what it means that the structural properties of language persist in the absence of any speaking subject. Equally, no one knows what these structures will become when models are trained on slop itself—language that was never thought by anyone at all. As Eddington famously quipped of the quantum world: "something unknown is doing we don't know what." What we do know is that whatever it is doing, it is doing it at an industrial scale. We have filled the little dark glass world, and now it fills ours.