Latent Writing
Language Models, Historical Memory, and the Mediation of Self-Consciousness
A text that answers back
Text has been a primary means of communication among thinkers since at least the late Hellenistic period, increasingly dominating and expanding discourse over time. A text provides a way to preserve an argument, something of an author’s mind, and a way to reach across the pages to question a thinker directly. Of course, the page doesn’t actually converse or present an argument in response to your question, instead acting only as a calcified form of the original intent. More often than not, that conversation would be simulated internally or between other readers of the text.
As writing increasingly displaced oral traditions as a primary mode of intellectual communication, it expanded in scope but not necessarily in kind. The printing press and the internet were technologies that reinforced the medium and extended its reach. This spread could be imagined as a process in time and an intellectual dynamic across readers. The text itself is dead, but the interpretation and communication between readers of a text create understanding and further text. Now language models present an interesting question: are they like the printing press or the internet, a means to an end? Or are they a change in kind, not just a mechanism for spreading text, but an extension of interpretation and communication itself? The model responds, makes connections, draws relations, and offers revisions. It searches between texts, sometimes hallucinating, just as another person misreading a text could misread a passage, moving in wrong directions.
An author of a text cannot respond directly to a question from a reader; they can only anticipate it. A language model gives the appearance of dialogue. The model is not a resurrected author, and an answer is not always a faithful reconstruction. Yet this new kind of encounter points to something important: language models may be understood as a new stage in a much older history of intellectual mediation. Rather than reinventing or encapsulating the externalization of thought, they extend it. We go from a technology of library science, with information merely preserving written expressions, to the generation of new expressions conditioned on patterns learned from prior writing across nearly all of recorded history.
Is this space something we could call latent writing? Unseen intellectual history, that which was only suggested, context between the works, requiring further discovery by future writings. Is this now surfaced in new ways? Before, we had new spaces of ideas or meaning explored and expanded between humans using writing as an artifactual basis. Now we have real interaction between humans and machines, and between machines themselves. In latent writing, books become references for a model trained on historical language, within a generative process and medium where unrealized expressions are inherited and can be composed in response to novel questions. Can historically conditioned generative computing systems make previously unrealized text available through interaction?
The claim in one sentence
Writing stores a form of what was said. Libraries, and then the internet, help us find what was said. A language model can generate what might be said, conditioned by traces of what has been said. This new operation has predecessors, but it changes what the medium itself contributes.
Our minds have not developed in isolation. Tools, for humans, are extensions of ourselves. Sentences turn ephemeral thoughts into durable artifacts, now open to inspection, debate, criticism, and further development. Writers themselves combine novel intuitions with thoughts conditioned by prior experience; the ideas they have encountered, struggled with, disagreed with, and transformed. Reflection is always mediated by the external.
History is a process across generations, growing and compounding this effect. Texts survive their authors; future readers then encounter new distinctions when comparing their experience to a world that no longer exists. In this way, writing is more than communication. It is the foundation and material support for the persistence and transformation of social thought itself.
Of course, writing has its own direct antecedent. Oral traditions were the technology that allowed societies and civilizations to organize. Our most powerful traditions were rooted in this world, only later being solidified into something that appears more concrete by scribes, sometimes thousands of years later. Some of this transition in societal technology is itself recorded, and Plato recognized both the power and the danger of this development. In the Phaedrus, Socrates complains that written words cannot defend themselves when questioned: they can only endlessly repeat the same thing. His concern is not only that writing is a poor way to store and communicate ideas, but that it can produce an illusion of understanding, a hallucination that can compound over time. Direct oral dialogue might be a better machine for wisdom.
How do language models fit into this? If the old contrast was a simple one between a fixed, written statement and a living, breathing speaker who answers and revises in real time, then what do we have now? The question seems unsettled, and is often unsettling for users and skeptics alike. The machine responds, revises, and appears to have understanding, but that does not make it an accountable author or guarantee knowledge. It offers responsiveness without the original speaker. We call some hallucinations when they are obviously in error, but in effect it is all hallucination. And to what extent is our own creativity a hallucinatory process? Is this accidental wisdom, or does it require mediation itself? Do novel intuition and the opportunity for understanding really exist in latent spaces?
Our relationship to communication has been altered repeatedly over time by technology. Manuscripts and commentary added layers of interpretation; indexes and commonplace books reorganized excerpts; printing dramatically multiplied copies; encyclopedias and libraries gave inherited writing new structures and meta-organization; digital search and hypertext made associations rapidly navigable. These all coexisted and framed the ways new readers and writers wrestled with textual history. But the sequence clarifies what is changing: not just a progression from memory to speech, to writing, and then to sophisticated forms of retrieval, but a change in the actual process of recombining and articulating our inheritance.
External memory shapes internal thought
This understanding of external memory as a foundational component of thought has a substantial philosophical history. Bernard Stiegler, drawing on Husserl, provides a particularly useful vocabulary.. Drawing on phenomenology he identifies three kinds of retention. Primary retention is the moment-to-moment experience we could call primary experience, and that we would express through oral speech. Secondary retention is memory, the recall of organized thought. He emphasizes tertiary retention. These are memory traces held outside individual consciousness, embedded in technical forms ranging from writing to all kinds of modern media. These traces are not just containers for the thoughts we already have. They provide the basis for what we see, make note of, can remember, and use for new thinking.
Using this approach, maybe the novelty of an LLM can be framed more easily: the external medium itself now produces new textual intermediates while a person is thinking. It can mediate and simulate oral tradition itself, offering a counterexample, inventing analogies, or pulling distinct authors into conversation. The medium is still dependent on earlier human inscription, but part of the reconstructive work once performed by readers, teachers, and editors now takes place inside a technical process. This is still a medium, though, and latent writing has the same limitations of a text, and in some ways perhaps more.
From an archive of sentences to a field of possible sentences
A library contains identifiable documents organized by parameterized systems for easier retrieval. Many texts, such as encyclopedias, act to further structure ideas for simple lookup. Likewise, a language model contains parameters learned through training on data, but these are well outside the scope of normal human bibliographic systems. It is not simply a searchable copy of the original data, though models sometimes directly reproduce specific passages. The models themselves were created as extensions of the original medium, initially through simple predictions of what would likely come next.
This creates a dramatic shift in the relationship between writers, readers, history, and the present moment. Historically separate authors and their many disputes and arguments now contribute to learned associations, patterns that can be activated together within a machine response, almost indistinguishable from the original text. Complex queries can produce resulting passages that may never have existed. The model doesn’t contain every unwritten sentence waiting to be retrieved. Latent writing instead suggests the practical reality that historical language has been transformed into a capacity to generate new, context-sensitive language. A new process of mediation with the past, a powerful new way to interact with language, an artificial construct that has clear utility, uncertain accuracy, and an unknown ability to intuit genuinely new ideas.
The user, the machine, and the historical Other
What is the actual relationship between authors and readers, and how does it relate to this new working relationship with models? Is this process a new kind of dialectic, a new machinic process itself? Currently, excitement about advances in AI is matched by worry and skepticism, a split between its potential to accelerate most fields of knowledge and work, and its threat to those within those fields and even our actual footing within the historical process itself. If writing and language models mediate our relationship with historically accumulated thought, they also transform relations of dependence. We depend on what others have written to develop our own understanding. But what changes when the medium itself begins to answer? Hegel’s work might provide an account of recognition, dependence, and labor that offers one possible way of approaching this question.
Hegel produced an incredibly influential description of history through his approach to the master–slave dialectic. Hegel’s account of the continual relations between masters and those bonded to them in the Phenomenology of Spirit focuses on desire and recognition. The master wins a kind of competition of animal desires, seemingly becoming “more” human through a willingness to sacrifice base existence for a kind of meta-desire for recognition. This is taken as a kind of extreme view of competition that, for Hegel, explains human social relations and subsequently historical development. Most importantly, though, the new lord seems independent, but the recognition he obtains from his subject is compromised because the slave’s independence is denied. The servant, meanwhile, encounters resistance through labor and gains a transformed relation to the world and to himself. The servant no longer provides sufficient recognition to the master, and the master paradoxically becomes a slave to his own needs. Meanwhile, the servant gains meaning through enacting actual change in the world. The familiar reversal is not simply that a servant becomes stronger; dependence and independence are more complicated than they initially appear, and this contradiction drives historical change.
This general relationship can be interpreted in many relational contexts, for example, author-as-master and reader-as-slave. No longer strictly dialectical in sequence, but still unfolding in time, the master sacrifices to produce a more ‘human’ artifact, a work to be preserved. A future reader, initially dependent on the work, may experience a further reversal as they gain mastery through interpreting and transforming it. Extending this to the human–LLM relationship is provocative, though not exactly equivalent. The machine strictly serves the human, yet that is incidental to this historical perspective. The model is also conditioned on the old masters, interpolating between them. The key relation is therefore not only user ↔︎ machine. It is also user ↔︎ historically accumulated human thought, mediated by machine. Hegel’s slave is a self-conscious agent involved in a process of desire, struggle, labor, and recognition. Models lack these from the bottom up, existing only through training and reinforcement. Nor are they able to undergo the process of mutual recognition, which can only be achieved in a kind of simulacrum. This analogy would arise from the structure of mediation and dependence, not from an established identity between artificial generation and human self-consciousness.
The potential, or danger, lies in how disruptive this is to the dialectic of master–slave relations themselves. What happens when we strip desire and recognition from the process? The relationship (i.e., being a slave) should not be taken too literally, but understood in the sense that we are all servants undergoing a process of mastery. Do these new tools help or hinder that process?
Hegelian unity or Deleuzian multiplicity?
In summary, that interpretation sees the language model as a new kind of tool through which historically collected thought emerges from a process to participate in our self-understanding. A new stage in a long movement of self-externalization and mediated return.
A contrasting, though not mutually exclusive, view would be inspired by more ‘postmodern’ thinkers like Deleuze and Guattari. Here, we don’t simply have a unified historical subject and dialectical process, but rather an assemblage of heterogeneous processes: texts, languages, institutions, training procedures, people, platforms, and so on. The apparently unified voice that we identify in the billions of parameters in a model is a precise and localized outcome of this complex arrangement. This can make previously distant ideas connect, or otherwise flatten their differences into a familiar pattern, with reality and hallucinations mixing in this new landscape.
In some ways, this is a more comfortable picture of the normal ‘machine learning’ space, more ambivalent toward the dynamics and relationships and allowing their multiplicity to emerge from their positions. Naturally machinic, less obviously human. Latent writing here is not directly analogous to the Deleuzian ideal of the virtual, the metaphysics of difference and becoming, but perhaps the latent space provides a mechanism for real structural emergence. These approaches offer different ways of organizing the history of ideas, and they lead to different questions. Does this technique deepen a subject’s encounter with its own historical world? What heterogeneous processes make that encounter possible, and what differences survive it?
Changes to the scope of forms of thought
Historical techniques, teachers, commentaries, and encyclopedias all respond to human intellectual needs. New generative systems did not invent mediation or the recombination of ideas. The principal difference lies in how quickly it happens, the scope of information within reach, and how much of it can happen at once. The discontinuity may therefore be operational rather than metaphysical. LLMs combine older operations at a scale and speed that may reorganize the practice of reflection itself, even as the underlying dialectical and machinic processes persist.
You could describe a single conversation as a process with feedback. Humans pose questions to other humans, to books, and to models. The deeper process unfolds over years. People read generated writing, incorporate it into further writing, and alter the cultural record from which future models may learn. The technological medium then helps produce some of the history that later versions of the medium encounter. There may be a change in degree with this new technology great enough to change the conditions of intellectual life. The most important question becomes not whether the machine possesses the Spirit of history. It is what kind of self-relation becomes possible when we encounter our inherited intellectual world through a medium that can compose a reply, and what kinds of thought may be lost if we confuse that reply with understanding itself. The critical point is how new relations alter the dynamics and affect future generations of this process. A flattening of our intellectual history could cause us to be trapped within this latent space, content to forever inhabit the hallucinations of the past.