Optimal Brain Damage

Pieter Huys, A Surgeon Extracting the Stone of Folly (1561). Source: Wellcome Library, London. Licence CC BY 4.0.

The art critic Craig Owens once proposed indexing the emergence of postmodernism to artists’ writings of the 1960s and the peculiar way they seemed to erupt into the aesthetic field.1 Whereas the texts of modernist artists had constructed and confirmed the boundaries of art as a visual practice, these new writings marked the rise of a textual condition that punctured aesthetic autonomy, distributing the work across a chain of visual and verbal signifiers that undid the notions of a single object, author, or meaning. Language transformed from a vehicle of expression into a system of hidden codes governed by feedback, allegory, and recursive logic. Art’s discursive turn thus foretold a larger shift. The emergence of postmodernism was not simply an artistic or theoretical juncture but the cultural expression of a broader epistemic rupture catalyzed by the convergence of cybernetics and structuralism in postwar think tanks.2

The shared intellectual genealogy of these fields resurfaces today in the statistical operations of machine learning, where the ascendancy of technical codes and the arbitrariness of the sign appear as two sides of the same coin. Text can do more things than ever before, yet those things no longer correspond to the historical structure of literary culture and its operating system of narrative time. The course of technological progress charted by literary culture has come to subvert its own core function, optimizing text for computation rather than consciousness. As it was for the artists of the sixties, the “written word” is now technically interoperable with all manner of nonlinguistic signals. The recursive textuality Owens identified—the endless deferral of meaning across signifiers—has metastasized into the circuits of synthetic cognition, which extend the spiral of possible meanings indefinitely, and beyond the horizon of human rationality.

Text today circulates predominantly as a mode of technical coordination—between servers, sensors, models, platforms, and users—with machine-generated outputs synchronizing behaviors across computational systems. This is an epochal shift in the function of language: algorithmic processing substitutes for human understanding, logical operations for reading. Large language models (LLMs), such as those driving most public AI interfaces, condense this mutation, ingesting writing as raw material for pattern recognition.3 The texts consumed and authored by machines remain meaningful, but their meaning resides in the deeper patterns of linguistic usage they index and unfurl. If a human sentence draws nuance and meaning from its author’s experience in language, which constrains but does not foreclose syntactical invention, a machine sentence reduces this background data to a matrix of linguistic possibility. The text emitted by a model represents a walk through a sequence of probabilities learned from a vast corpus of unrelated writings; the connective tissue between words is no more than a calculated likelihood represented in language for the benefit of its human users. These texts’ mathematical transparency is the obverse of their semantic impenetrability.4 On the one hand, the rationality of the text generated by an LLM—its statistical coherence—destabilizes its capacity for human meaning. On the other, the rationality that gives rise to an LLM—the reworking of language into technical codes capable of generating texts—destabilizes the possibility of meaning between humans.

This opacity marks a broader cultural restructuring of language from a mode of introspection and historical reasoning into a behavioral interface. Whereas the twentieth century identified language with desire and repression, the twenty first wires it to neural networks and user data. In contexts ranging from haptic technologies to social platforms to Spotify mood filters to object recognition models, text seems to index everything except its own semantic meaning. Language is simply another facet of technical infrastructure, mapped to internal states, biometric cues, sentiment scores, and image databases. Within LLMs, this shift becomes especially visible in their treatment of prompts—the texts provided by users to elicit generative outputs. Intermediary techniques like LLMLingua compress prompts internally by learning which tokens (units of text) most influence model behavior, treating language as a manipulable interface rather than a communicative form. Stripped down to its barest statistical signature, human-parseable language is refined into a machinic shorthand, the minimal contours required to activate the model.

LLMs are not ends in themselves but the mechanisms that tether text to these myriad other signals. What emerges from these pairings is not meaning in any traditional sense but the semblance of meaning, as when models trained to relate semantic features to the measurement of human brain activity use LLMs to reconstruct what someone is “thinking.”5 Having once made thought visible, the alphabet now transforms it into data, generating plausible text outputs statistically correlated to brain-signal inputs. What speaks here? Not thought, not perception—only the model. From within the shell of the alphabet, LLMs strip language of its traditional functions, decoupling text—the conventional form of embodied thought—from its capacity for expression. In doing so, they unsettle the presumptive historical resolution of the alphabet as the apogee of human communicative reason. To an LLM, text is but one possible way to mediate relationships foreign both to the durational, phonemic structure of human speech and the structures of meaning that speech unfolds. Here, the alphabet confronts its own obsolescence: a world that writes itself without knowing itself.

Nowhere does this inversion manifest more vividly than in the generation of images from text prompts. In these multimodal models, which use LLMs to process user inputs, the prompt’s data character—its openness to multiple interpretations—comes to life. To a text-to-image model, a prompt is not a description but a program, a cluster of coordinates whose latent visual potential is drawn from hundreds of millions of captioned training images. Communicative prose becomes parameterized scene specification, a function input rather than a message, with certain words keyed to spatial and appearance constraints according to their correlations with visual patterns. The prompt is, in effect, already an image, just alphabetically encoded. In this contraction, both the imaginative potential of the image and the syntactic order of the text are subsumed into a single topological zone. A prompt is an image-in-waiting; an image is a prompt obtained.

This displacement of the alphabet’s visual project leaves us without a referent. As text evolves from picturing thought to being pictured by machines, there is no longer a source to trace: a generated image indexes a statistical correlation like a photograph indexes a light event, yet neither model nor lens can recover origin from output.6 As in realist literature, where the accumulation of textual details produces the illusion of the real rather than denoting reality per se, the prompt signifies not by reference but by function.7 Images and texts no longer mirror the world but hallucinate one another into being.

A Minimum Viable Stupidity

The transformation of language from an instrument of human reasoning and proposition into a surface layer of intra-machine communication—the code by which machines generate images, simulate discourse, and coordinate among themselves—also dismantles the inner life it once provisioned. Stripped of its expressive depth, it acts back on its human users, who now write for interfaces rather than interlocutors, editing for uptake rather than understanding. In this shift, the reflective consciousness of linear writing gives way to a culture oriented towards pattern recognition and stimulus response. If postmodernism heralded a reorganization of meaning along the lines of code, culture today is reforming around computational protocols that suppress the search for meaning entirely. Beyond the dawning awareness that every text may now conceal a nonhuman author, unaware of what precisely it has written, the creep of platform culture infects even human texts with the demand not only to appeal to but to exploit the nonmeaningful functions of the human brain. In the same way that machine language expresses a location in mathematical space, human language can now be seen to divulge its own operating system. Call-to-action email subjects, market-tested branding lexicons, clickbait headlines, dog-whistle politics, academic abstracts keyworded to boost citation counts: all picture our evolved neural biases and learned neurochemical reward loops.

As text decouples from natural language, it reverts to the logic of its constituent glyphs, whose visual properties make them intrinsic attention attractors irrespective of their meaning. Since the brain has no dedicated structures for script, text recognition must be learned in individual developmental time, recycling cortical territory specialized for facial and object recognition. Our heightened sensitivity to edge detection and line junctions, for instance, corresponds to the high probability of encountering contours, corners, and object boundaries in natural environments; the human visual cortex mirrors the statistical structure of the world. All human symbol systems redraw these features in proportion to their frequency in the environment, simulating the world in order to exploit our evolved perceptual capacities.8 Tuned to these biases, letters—special forms of glyphs that lack inherent meaning and are unable to be decomposed or explained by other linguistic units—become lures, archaic tools for cognitive hacking.9

Text becomes glyphic when its stimulating visual economy eclipses its linguistic operations. A glyphic culture, then, is one built on the contours of content rather than its substance—human-authored language warped beyond recognition by the protocol imperatives that now sort, hierarchize, display, and distribute it. Like the emergence of postmodernism, the ascension of glyphic culture is rooted in the determinate histories and material contexts of late capitalism.10

The double bind of this culture is that we are illiterate to its logic—which, like machine language, is indexed to signals rather than denotative meaning—so we approach it still clinging to our old interpretive habits. Glyphs, however, show without telling, collapsing thought into stimulus. Written language, once the scaffold of a continuous interiority, now arrives fragmented: text as object, text as cue. The narrative self falters, replaced by a perceptual self shaped by the rhythms of interface rather than introspection.

Social platforms materialize the glyphic turn. As structuralism is to cybernetics, so is an algorithmic feed the other side of an LLM: a place without history for a text without time. Like LLMs, platforms are transacted through texts whose alphabetic surfaces obscure their underlying codes. The commonplace conflation of social media with reality not only mistakes rhetoric for action but, more fundamentally, misreads which signals the text is actually indexed to. The meaning of content, in fact, is only the length of time it captures a glance. Glyphic culture is this principle writ large, the generalization of the glyph’s recursive optimization into the technical forms that today run machines and humans alike as pattern-recognition engines, interlinked via their respective linguistic protocols.11 Through these co-constituting protocols, the tokenization of language by LLMs to generate probabilistically sound texts comes to mirror the taxonomization of behavior by social computation to engineer human engagement. Where computer science made language machinic to build models capable of generating human-like language, we now make our own language machinic to exploit the affordances of platforms refined and powered by those same models.

If the glyph reencodes human culture with the biomorphic logic of machine learning, it also points us back to an older paradigm of model refinement increasingly relevant to the computational demands of LLM training and inference. “Optimal brain damage” is a technique for reducing model size through the calculated deletion of information, first introduced by Yann LeCun (currently Chief AI Scientist at Meta) in 1989. The procedure uses an algorithm to identify and trim extraneous weights from a network after training, selectively forgetting low-probability outcomes: “Compression disproportionately impacts model performance on the underrepresented long-tail of the data distribution.”12 Lightweight models such as DeepSeek achieve further performance gains by training on other models, eating their own words to extend rather than retract them.13 Just as LLMLingua distils a prompt to its minimal tokens, optimal brain damage pares a model to its most efficient parameters.

For humans, too, glyphs are the mark of optimal brain damage: the efficient production of outputs from prompts. Where alphabetic writing was a tool of reflective consciousness, glyphic culture is the rule of unconscious cues that precede comprehension. To sever the wiring of these automatic connections, we might look to the neurological precursor to “optimal brain damage,” synaptic pruning. In this process, the maturing brain replaces the simpler associations learned during childhood with more complex neuronal structures. The linear mapping of letter to sound gives way to the tangled network of orthographic processing, phonological rules, grammar, and semantic meaning that enables each of us to read. Children develop away from the linear, “algorithmic” functions of the early brain, upscaling their abstraction capacities until they can robustly respond to circumstances for which they have never been “trained.” The pruning of a model, by contrast, is only ever a lossy compression towards a fragile mean—“compressed networks [being] far more brittle than non-compressed models to small changes in the distribution that humans are robust to”—a progressive desensitization towards a minimum viable stupidity, which is now turned back on culture to automate a diminished form of cognition.14

The achievement of selfhood is coincident with the capacity gains of synaptic pruning, which allows us to join and contribute to a human community in unpredictable ways. Yet it is precisely the neurological machinery of abstraction that has produced our brain-damaged present through centuries of alphabetically fueled technological development. The suggestion that machine intelligence will expand the human horizon ignores the myriad ways that horizon has already begun to narrow as machines progressively filter out improbable, long-tail events until they have disappeared from the distribution entirely.

As the medium through which modern scientific rationality emerged—and with it, the matrix of capital and empire that continues to shadow consciousness globally—alphabets have shaped the languages of programming, the interfaces of human-computer interaction, and the corpora on which LLMs are trained.15 Today they are themselves the substrate of computation, the grid along which the mind they helped construct—ours—is being rewritten. We no longer live in the reality we were trained for; moreover, the symbolic structures that scaffold our cognition now work against our capacity for abstraction. If a model falters when the world shifts out of alignment with its training data, perhaps the task is not to train better models but to imagine how we might use the ones we have to shift the world instead—a syncretic politics that wields our tools not as oracles but as wedges.

This essay is adapted from Proof of Personhood, published by the Singapore Art Museum and End of Medium and distributed by Library Stack.