In her article titled Mean Images, penned for the New Left Review, Hito Steyerl recalls Ted Chiang’s likening of ChatGPT to a “blurry JPEG of the web” and points out that generative AI models don’t really work like the lossy compression algorithm of Xerox photocopiers because they are fundamentally statistical.¹ That is to say, the generated material isn’t an inferior copy of some original because the model acts as a sort of giant blender, taking in heaps of material and churning out a flattened, average, “mean” output.

The somewhat misleading image of a large language model as a photocopier churning out increasingly lousy copies of a copy does, however, raise an interesting question: what exactly is the original that is being copied so poorly? The common answer, and the one given by Chiang, is that it is the “internet.” After all, 85% of the corpus used to train GPT-3 – OpenAI’s third-generation language model on which the first version of ChatGPT (GPT-3.5) was largely based – originated from the web (60% Common Crawl, 22% WebText2, 3% Wikipedia).²

But both ideas – the idea that LLMs churn out low-res lossy compressions of the web, as well as the idea that they churn out material by means of some stochastic parlor trick – somewhat conveniently overlook the simple aspect that the range of output generated by LLMs doesn’t necessarily equate to the sum of their training data.

Many researchers have pointed out that the ability of LLMs to perform tasks they weren’t trained for and to display knowledge that wasn’t part of their training data, seems attributable to their ability to construct idealized internal models of a given object. This could be the headspace of a human agent manifested through writing and provided as input text,³ or the rules of a game that they weren’t previously exposed to.⁴ A simple yet pertinent example would be translation – a task that transformer-based generative language models aren’t explicitly trained for, but somehow excel at.

In an article titled Can Machines Learn how to Behave?, Blaise Agüera y Arcas explains the process in which transformers acquire necessary skills in order to translate.⁵ Unlike neural machine translation (NMT) models that learn how to translate by studying directly matching pairs of correctly translated sentences, transformers approach translation through analogical reasoning skills acquired during their pre-training stage via seemingly simple tasks such as predicting blanked out words from Wikipedia articles. Jumping back and forth between articles in multiple languages and learning which words correspond to what, transformers eventually learn how to translate sentences that weren’t part of their corpus.⁶

Needless to say, translation requires more than just an index of corresponding words across languages. Agüera y Arcas notes that while this explains how a model can translate phrases that weren’t present in its training data, how they learn the correlation between different languages isn’t entirely obvious. As fascinating as this is, we won’t try to go into it here. The key takeaway for us is that these models aren’t entirely tied to the content they’ve been trained on.

Translational Fluidities

Since we are focusing on the “original” that is being transferred from one context to another, we might approach this as a translation problem. Such an approach might be useful, not only because the ability to translate is a prominent emergent ability displayed by LLMs, but also “precisely because translation is an activity that immediately problematizes the ontological hierarchy of languages – “which is primary and which is secondary?” – it is also the place where the oldest prejudices about origins and derivations come into play most forcefully.”⁷ This is according to Rey Chow, who also utilized the framework of translation to imagine a more horizontal and porous relationship between the supposedly anterior “original” meaning and the derivative “translated” material in the context of cultural translation.⁸

Chow starts her analysis on the relationship between the original and the translation by referencing Walter Benjamin’s peculiar view on translation, as detailed in his lengthy translator’s preface for Charles Baudelaire, widely known as The Task of the Translator. Here, Benjamin questions the traditional notion of translation as a simple transfer of meaning from source language to target language. What should be translated, he argues, isn’t meaning per se, but the fundamental intention inherent within the original text, which he refers to as “the great longing for linguistic complementation.”⁹

The interpretation of this somewhat esoteric phrase gave rise to various readings of Benjamin’s theory of translation. Deconstructionist critics, for their part, tended to equate this to the inherent negativity of writing, seeing the original as an already failed translation that perpetuates the production of differences.¹⁰ Chow, however, supplements such readings by pointing out that this “linguistic complementation”, this process of “putting together”, is “also a process of “literalness” that displays how the “original” itself was put together.”¹¹ In other words, it could be seen as the process or the desire to construct a reasonably coherent whole from disparate, but nonetheless relational individual parts. But what exactly are the parts that are being “put together”?

What needs to be translated from the original, he writes, is not a kind of truth or meaning but the way in which “the original” is put together in the basic elements of human language – words. Hence it is words – in their wordness, their literality – rather than sentences, that matter the most in translation. A real translation, Benjamin writes, “may be achieved, above all, by a literal rendering of the syntax which proves words rather than sentences to be the primary element of the translation. For if the sentence is the wall before the language of the original, literalness is the arcade.”¹²

It would be important to note here that the German word Wörtlichkeit, translated as “literal” or “literalness” in the above passage, Chow prefers to translate – more literally – as “word-by-word–ness.”¹³ So, for Benjamin, words put together in their word-by-word–ness is a plausible, materialist description of the process of writing. With hindsight, we could add that in such a process, likelihood, or a certain associative sensibility that accounts for the contextual coherence of a whole, made possible through effectively retaining what came before, would be key.

This notion of “word-by-word–ness” could illuminate a new understanding of language models.¹⁴ Unlike a human translator who might take liberties with words for the sake of meaning or flow, a language model is inherently “literal.” This way, the output isn’t a lossy copy but a linguistic mosaic, composed from countless fragments of language data. This mosaic isn’t inferior; it’s just different, and it can offer us a unique perspective on language itself.

If we can generate plausible, meaningful, and creative outputs from a fundamentally statistical process, does this challenge our understanding of human cognition and consciousness? If consciousness is perceived as a continual generation of words, how does this differ from the language model’s process of generating words?

The parallel between the language model’s “word-by-word–ness” and human cognition invites us to ponder the nature of our own thoughts. Could our thought patterns also be seen as a form of “word-by-word–ness” where each thought builds upon the last, influenced by the vast web of experiences, memories, and knowledge we’ve accumulated over a lifetime? In this light, the language model could serve as a microscope, revealing the architecture of our own thought processes.

Interestingly, such a conception also has implications for the way we perceive our interactions with artificial intelligence. If we approach conversations with language models as a kind of translation, where the input is rendered into a contextual response based on specific patterns, we can think of these interactions as less of a mimicry or copy, and more of a kind of conversation across boundaries. This is a conversation not just between human and AI, but also between two fundamentally different ways of processing and understanding language.

If we take this concept further, we may even start to question the idea of authority in conversation. In a dialogue with an AI model, who holds the upper hand? Is it the human who gives the prompt, or the AI that provides the response? Perhaps the power dynamic is fluid, shifting with each exchange, likely leading us to rethink our assumptions about conversation dynamics, hierarchy, and control.

Neither the human input nor the AI output can so easily claim the mantle of the “original”. The conversation becomes a fluid exchange where each response is a translation of a previous input, creating a cascading flow of interpretations. This fluidity destabilizes the conventional binary of “original” versus “translated” and invites us to reconsider the inherent value we place on originals and the potential creativity embedded in the act of translation.

Word-to-Vibe

But what would this “linguistic complementation” – the “putting together” of words in their “word-by-word–ness” – amount to, if not “meaning”? Again, what is it that’s being transferred through the passageway of the arcade that is “word-by-word–ness”? It must be a certain sense of unity, a constructed coherence that has been “put together” through a sequential yet relational sign system; a certain sensibility that slips through the grasp of language; a certain feeling; a “vibe”.

The “word-by-word–ness” in this linguistic architecture parallels the neural networks in the human brain. Each word, like a neuron, contributes to a larger thought, feeling, or concept – a “vibe”. This raises intriguing questions about the essence of cognition. Are our thoughts, much like language models’ outputs, an assemblage of synaptic words, creating a “vibe” that constitutes our conscious experience?

Recognizing the “vibe” of a text as a product of intricate “word-by-word–ness”, akin to the natural and artificial neural network operations, extends our understanding of creativity. It seems we have entered a new paradigm where creativity is no longer a strictly human phenomenon. Language models, too, through their artificial neural networks, are capable of producing a novel and striking assemblage of words. This prompts a redefinition of creativity as a function of networked information.

When we perceive creativity as a function of networked information, how does this influence our approach to literary criticism? When analyzing a text produced by an AI, we may need to shift from traditional hermeneutics to a computational critique, focusing on the intricate network connections that generate the text’s “vibe”. This could offer a new vista on literary interpretation – one grounded in the science of neural networks.

Peli Grietzer’s research on “vibes”, conducted at the interstices of mathematics and literary theory, demonstrates how the Modernist, avant-garde notion that a literary work compresses a truthful world view has strong resonances with the operations of an autoencoder.¹⁵ An autoencoder is an artificial neural network used to compress and reconstruct data. But in order to compress, that is, to create idealized representations of data, the neural network extracts and amplifies what is best described as the “vibe” (i.e. the sense of unity) of a given data.

The meaning of a literary work like Dante’s “Inferno,” Beckett’s “Waiting for Godot,” or Stein’s “Tender Buttons”, we would like to say, lies at least partly in an aesthetic ‘vibe’ or a ‘style’ that we can sense when we consider all the myriad objects and phenomena that make up the imaginative landscape of the work as a kind of curated set. The meaning of Dante’s “Inferno,” let us say, lies in part in that certain je ne sais quoi that makes every soul, demon, and machine in Dante’s vision of hell a good fit for Dante’s vision of hell. Similarly, the meaning of Beckett’s “Waiting for Godot” lies partly in what limits our space of thinkable things for Vladimir and Estragon to say and do to a small set of possibilities the play nearly exhausts. Part of the meaning of Stein’s “Tender Buttons” lies in the set of (possibly inherently linguistic) ‘tender buttons’ – conforming objects and phenomena.¹⁶

According to Grietzer, much like how an artist amplifies an elusive feeling by piecing together disparate, disjointed objects and phenomena of the real world, an autoencoder amplifies the “loose vibe” perceived amongst real-world objects and phenomena into an idealized “dense vibe”. Our senses provide us with a vast array of information that our brains process and piece together into a coherent narrative of reality. But much like a language model’s output, our perception of the world might not be a lossy compression of an objective reality, but rather a “vibe” constructed from a vast array of sensory data.

He also argues that “learning to sense a system and learning to sense in relation to a system – learning to see a style, and learning to see in relation to a style – are, autoencoders or no autoencoders, more or less one and the same thing.”¹⁷ Whether you accept this proposition or not has major implications.

Through this perspective, we could venture to say that the “vibe” of a text – whether it is produced by a human or an AI – is like the resonance of a system. Just as a system of interacting parts produces a certain pattern of behavior that characterizes the whole system, the individual words of a text, interacting within their linguistic rules and associations, give rise to a distinctive “vibe” that speaks of the text as a whole. This understanding of text as a system that gives off a certain “vibe” could bring a new dimension to our appreciation and understanding of literature. Grietzer expands upon this systematic perspective as follows:

One reason the mathematical-cognitive trope of autoencoding matters, I would argue, is that it describes the bare, first act of treating a collection of objects or phenomena as a set of states of a system rather than a bare collection of objects or phenomena – the minimal, ambient systemization that raises stuff to the level of things, raises things to the level of world, raises one-thing-after-another to the level of experience.¹⁸

What he seems to be trying to convey here is the fundamentally apophenic tendencies of neural networks, whether artificial or natural. Through this lens, the “longing” to “put together” words in their “word-by-word–ness” could be seen as the urge to construct patterns. The desire to contextualize something – oneself, a word, an other, an object, a phenomenon – within a web of relations. Our tendency to elevate the aimless “one-thing-after-another” into beads of “experience” neatly woven on a piece of string; our need for surface-level coherence – that is what artificial neural networks seem to have a propensity to model.

Abstract Web

There’s a subtler facet of language models that often goes overlooked: their influence on human language and thought. As these models become more integrated into our lives, their output starts shaping our communication. Are we seeing the birth of a new form of “AI-tinged” language? How does the rise of language models impact the evolution of human language itself?

Indeed, it is not a one-way street where humans create and machines imitate. The process is rather circular, feeding back into itself. As we converse with language models, we adjust our language to be better understood by them. Subtly and over time, these adjustments could reshape our own language use, prompting us to think more carefully about our wording. This feedback loop, in turn, could drive the evolution of language in unexpected directions, blurring the lines between human creativity and artificial intelligence.

As we adjust our language and thoughts in such a way, we inadvertently engage in a process of perpetual translation, molding our language into new forms that resonate with AI-understandable “vibes”. In essence, we become translators of our own thoughts. This not only redefines our communication but also creates a richer and more complex linguistic tapestry, sewn together with natural and artificial threads.

In this co-created linguistic tapestry, it’s not the words themselves that hold significance but the relational dynamics among them. This “putting together” of words is not a simple act of stringing together letters, but a complex process of building and navigating systems. With AI, this process takes on an additional layer of translation.

This additional layer could, in turn, allow us to conceive of a third layer. This third layer could be a point of introspection—a moment where we step back and critically examine the co-created language. Here, we are not just interacting with AI but also evaluating and modifying our linguistic approaches based on these interactions. It introduces a layer of self-awareness and reflexivity, where we consciously navigate the complex networks of “putting together” words and uncover new ways to communicate in this ever-evolving landscape.

Navigating this layer of introspection, our notion of the “original” undergoes a radical shift. The “original” metamorphoses into an abstract landscape of ideas, losing its rigid identity in favor of a more fluid, flexible form. Just like a river changes course over time, carving out new paths and merging with others, the “original” too is perpetually reshaped by a translational process, evolving into an ever-changing panorama of linguistic possibilities.

The “original”, then, isn’t merely a corpus, and the abstraction of the “original” embodies its translation from a simple given into something that is “put together”. As the “original” goes through the process of abstraction, it evolves from being a well-defined point to a web of potentialities. In this web, each word, each conversation, each input-output, represents a point of potential “original”. As it is the case with abstractions, each and every manifestation of this potential original will only act to transform and delay it.

*The author wrote this article in Korean and English simultaneously using GPT-4 in order to generate texts. Among bilingual texts, you can read the Korean version via this “인공적 원본: 합성 텍스트와 번역적 가변성”.

*This article is produced in conjunction with the 2023 SeMA Agenda Research Project.