It is language: layered, hidden, promising, specified for certain ears, routed through models and back out again, draped over the city like a second surface that has no edges and no seams to look under. Imagine kids billowing a parachute on the playground. The pitches, the tweets, the bios, the decks, the public personas, the calls tuned for different rooms, all of it produced and reproduced toward no terminus, only the next iteration of growth. The fog is the condition where language has become the final medium, where fluency substitutes for competence, where the performed coherence converges not on a person but on story.

In 2004, Matthew Kirschenbaum published "Extreme Inscription: Towards a Grammatology of the Hard Drive,"¹ an essay that made an argument against what he diagnosed in new media theory as "screen essentialism:" the tendency to treat digital text as a visual surface floating free of any physical substrate. Kirschenbaum instead insisted on following the bits all the way down to the metal. The hard drive, he argued, was the inscription technology with the single greatest impact on computing in the latter half of the twentieth century, and virtually no one in the critical humanities had looked at it. They focused on the screens that N. Katherine Hayles labelled as flickering signifiers.² However, beneath the screen, sealed in a chamber that could hardly be opened without destroying its contents, a platter spun at ten thousand revolutions per minute while a read/write head floated nanometers above its surface, flipping magnetic domains smaller than a red blood cell. That was where the writing actually was.

"Digital inscription is a form of displacement," Kirschenbaum wrote. "Its fundamental characteristic is to remove electronic objects from the channels of direct human observation."³ The inscription was hidden from the writer, but it did not vanish. It was a durable mark made by physical means on a physical surface. He noted the prepositions: we write to a disk, not on it, because we cannot see the surface, because the disk has been semantically refigured as a volumetric receptacle, a black box with a closed lid.⁴ The preposition to still implied a destination where mark arrived even though its physicality was abstracted away from the author.

Kirschenbaum was apt in labeling the abstraction of writing devices and its impact but could not see where it was leading. The technology with the single greatest impact on writing in the first quarter of the twenty-first century is not a storage device. It is the generative language model, and unlike the hard drive, it does not inscribe. It is more like those flickering signifiers that pop into and out of meaning.

The apparatus that generates text retains no trace of having generated it. The prompts disappear and the weights are read-only during inference. No inscription occurs in the mechanism of production. You could forensically examine every parameter in the model and find no evidence that any specific sentence had ever been produced. The act of generation passes through the apparatus and leaves it unchanged.⁵

What prompting more closely resembles is unlocking — finding the language that opens a pathway through the model's latent space, not writing on a surface but turning a key in a lock that has no fixed shape until the key arrives. The output arrives as an edited version of the context that created it, and the key returns from the lock reshaped, carrying something that was latent in the mechanism, not something that was inscribed by the hand that turned it. The result is emergent from unrepeatable conditions and dense in its circumstance. It lays on a heavy fog, blurring the original source. A specific cloud is not a record of the atmospheric conditions that produced it. A specific generated sentence is not a record of the prompt that elicited it.

Reading "Extreme Inscription" now, twenty years later, what stands out is how much of the argument depended on that persistence and resistance of the inscription. The whole essay is organized around the claim that digital writing is material despite appearing immaterial, that the traces are real despite being nearly inaccessible to the human eye, that the hard drive is interesting precisely because it forces us to confront the gap between the perceived weightlessness of digital text and the physical reality of how that text is stored. There was still a layer of language and translation constantly becoming but it was material.

The generative model dissolves both anchors of authorship and inscription. All changes are now in the language layer. The writer is no longer an inscriber but a manager of context, their task: gathering, arranging, applying pressure. It is pressure on the model, not inscription in it. The relationship between what was said and what changed is irretrievable.⁶

The unit of cultural production in the time of Kirschenbaum was the meme: the fragment, the clipped image, the decontextualized quote, existing in digital storage. Now the unit has shifted from fragment to surface, from meme to myth: the wholistic ability to produce not a piece but a complete narrative, a performed coherence that's tailored for every room. A single author now has many specific voices for all their audiences. "Here's a generated draft for you and your agents." That is the new competence, and it's hard to see through.

Generative language models gave the fog its own mode of production, not productive in the sense of producing things, but in the sense of producing itself, reproducing its own conditions, thickening. The fog proliferates in a culture where self-mythologizing is one of the only routes to meaning making. Now it is an industry. The commodity is the fog.

This is a weather-space with a new kind of labor. The work is gathering context and voices, applying them to agents, ensuring their legibility to other agents.

I want to call this generative pressure.

The prompt that generates a draft that is edited into a new prompt that generates a revision that is pasted into a deck that is used to secure funding that generates a mandate that is summarized into a memo that is drafted by a model that is prompted by someone who read the deck. Generative pressure is the name for this labor and its structure: the recursive circuit of prompt and output in which no stage constitutes a terminus and every surface is produced for the purpose of producing the next surface, because that is the demonstration of value. The decks cannot build the company. But the fog thickens regardless, because the labor that produces it does not require the company to be built. It requires only the next surface, the broader the better. Apply enough generative pressure and you can win anything.⁷

Henri Lefebvre, in The Production of Space (1991), argued that space is not a neutral container in which social life occurs but a product actively manufactured by the social relations and modes of production that inhabit it. "(Social) space is a (social) product," he wrote, and every mode of production produces its own.⁸ Feudalism produced feudal space: the manor, the commons, the cathedral. Industrial capitalism produced abstract space: homogeneous, exchangeable, organized for extraction. The production of space structures the practices of the people within it, and the practices in turn reproduce the space.

The use of generative language models is a mode of production that is generating its own spatial logic and the space it produces is fast, full of hyperparameters, convincing, and everything is possible here. It stretches is toward infinite open-endedness as an aim, as the designed condition of a system that must never stop producing in order to justify its own existence and grow the myth. They are tools that expand a haze toward unbearable spaciousness.

The textual layer over San Francisco (the pitches, bios, investor decks, product copy, tweets, the performed coherence of an entire industry) is a whole space itself, retreating from disclosure, developing something that never breaks. A founder has a voice for the pitch and a voice for the memo and a voice for the tweet and none of them are lying exactly, and between them there is no person reconciling the differences, just a fog of doubt thick enough where they can do nothing but benefit.⁹

Lefebvre was cautionary of the "overestimation of texts, written matter, and writing systems, along with the readable and the visible, to the point of assigning to these a monopoly on intelligibility." His point was that space exceeds its representations; there is always a material remainder that textual analysis cannot reach. But growing generative pressure produces diminishing material remainder. It is (con)text. Lefebvre warned against treating space as text, and the dominant mode of spatial production has now made space (and product) into text.

The representations of space — the decks, the memos, the performed coherences — are the space itself, colonizing the lived experience of the place the way Lefebvre's abstract space colonizes representational space: by tending toward homogeneity, toward the elimination of existing differences or peculiarities, by replacing the particular with the fluent, the contingent with the coherent, the inhabited with the generated.¹⁰ The homogenization hides itself. It presents as clarity and familiarity, leaving no place for critique.

Space becomes place again through experience.¹¹ The pause in movement turns a locality into a center of felt value for its inhabitants. Familiarity is how the undifferentiated becomes inhabited: slowly, through fumbling encounters, over time. The generation through a model produces a false familiarity disguised as fluency, a resemblance that is its opposite. Knowing without inhabitation is from outside. What was differentiated through experience is being smoothed back into space. Real clarity and familiarity (truth) come from roughness, brokenness, an unconcealment. ^{somewhere in Heidegger}

Kei Kreutler notices that "machines have a talent for producing cultural patterns around their use, and practices also have an ability to produce machines," applying that observation to time and devices of temporality.¹² The observation applies, as well, to devices of communication, and with particular force to generation, because it is a self-referential loop in a way that no prior machine-practice cycle has been. The mechanical clock produces the practice of punctuality, but punctuality does not produce clocks; it produces demand for clocks, which is different. The hard drive produces the practice of archiving, but archiving does not produce hard drives.¹³ In the generative loop, the output is of the same kind as the input. The practice of prompting produces models optimized for prompting, and the models produce practices, workflows, templates, agent architectures that feed back into the next generation of models. It's markdowns all the way down. The cycle accelerates because there is no convserion between stages, no moment where text must become something other than text in order to continue.

Kreutler's essay preserves a remark from Federici that serves as a diagnostic for the generative condition. Federici notes that astrology can return to cultural prominence with "the certainty that even the most devoted consumer of astral charts will automatically consult their watch before going to work."¹⁴ Magical practices can proliferate within serialized, industrial time because they do not threaten its enforcement mechanisms. An analogous condition holds for generative pressure. The loop can proliferate from text to text from text, because it operates inside the enforcement mechanisms of product cycles, quarterly reports, and shipping deadlines, not against them. Everyone still ships on time. The fog runs inside the schedule. And the longing for craft, for material, for the handmade will make a comeback, but it returns, as Federici might predict, with the certainty that the romantic will consult their model before going to work.

Applying generative pressure is the new deliverable. Guiding my agents with context and prompts to make documents for your agents to read is the new workflow. We are now reading and writing for efficient agent understanding.

David Abram, in The Spell of the Sensuous: Perception and Language in a More-Than-Human World, argues that the act of reading and writing restructures the reader's perceptual relationship to the world. The written word is not a neutral instrument for description but an active force reshaping the sensorium of its user, redirecting attention from the sensuous world toward the self-referential world of text.¹⁵

The generative loop intensifies this to the point of dissolution in the vapors of the fog. The textual environment is increasingly generated, and the people who read this language are also the people who prompt it, and the prompts they write are shaped by the generated language they have read. The observer-observed boundary folds. Read enough generated text and the writing starts to come back tinted. Write like generated text and the models trained on that writing generate more of the same. The feedback is the operating condition.

Kirschenbaum notes, almost in passing, that the engineers at IBM's San Jose lab borrowed the same iron oxide paint that gives the Golden Gate Bridge its distinctive hue to coat their first disk platters.¹⁶ The detail is small but the coincidence is structural: the same substance that made the city's most visible piece of infrastructure visible also made the hard drive's inscription invisible. International Orange on the bridge, iron oxide on the platter; one seen from every hill, the other sealed in a chamber no one would ever open. The coating was the medium, it held the traces. These sentences were passed through it on their way to the page.¹⁷