I recall learning two things my senior year of university. The first figures in my memory as the result of hours sat in front of books translating miles of poetry and fin-de-siecle diary entries. Endless meetings to review minute choices made in the process of translation; defending syntactic and semantic word choices as if they were a reflection of my own moral shortcomings. I learned how to be a translator. The second lesson came more abruptly. GPT-4 had updated its translation feature. Out of curiosity, I smugly typed, “Translate Mikhail Kuzmin’s ‘Wreath of Springtimes’ from Russian into English.” The chatbot buffered, asked that I provide the source text, and within a few seconds, out poured 30 complete stanzas in English. Thirty almost perfectly translated stanzas, something I had spent weeks working on, in neat rows on my screen. There, I learned that translation, as I wrongfully understood it, was functionally obsolete.

But alas, there is nothing new under the sun. Technology, inseparable from its nucleus in human creativity, reason, and intelligence, continues its development under our care and supervision. The Babel analogy is endlessly fitting, but trite in its aims. Genesis 11 is often read as a critique of man misappropriating his reason and abilities in favor of disobedience, pridefully attempting to build a tower up to the heavens only to be smited by God. The attempt to transcend one’s own earthliness, to surpass the boundaries of human ability through technology, is a recurring trope. We can glean the newest iteration on the horizon with ASI. Recurring and persistent enough to call the first book of the Old Testament home, yet frequently misread. Perhaps, an alternative reading of Babel, one which cautiously centers the perspective of man rather than God, offers new insight into a digital world that grows alongside us. Babel was not God’s punishment to man for using his divinely imbued abilities to transcend his own limits. Babel was God punishing our desire to grow upwards, away from his green earth, rather than outwards, towards the rest of creation. Crucially, technology must now be considered within the category of creation.

In conversation with artificial intelligence, particularly LLMs, an exegesis of Babel figures as a call to focus our growth laterally.

Language is synonymous with technology, central to culture, inseparable from subjectivity. Subjectivity, the capability to posit the self as an “I”, the self’s capacity to hold up this newfound “I” against another’s established “I”, thus creating “you”, owes itself to the function language plays in communal existence. Language precedes our subjectivity, we owe our faculty of self-creation and individuation exclusively to the power of language in both forming and navigating social relations. There is no “you” without the self’s ability to first assert itself as an “I.” Just as language was once programmed into our still-neuroplastic minds, we have programmed it into machines and have taught them to replicate our thought and speech patterns. Communication, even between native-speakers of the same language, is an unceasing exercise in translation. Moment-by-moment omissions and additions on the basis of shared context allow conversation to flow, information to be exchanged efficiently, and new ideas to grow outward through the solidification of common patterns or the verbal “smoothing out” of points of contention. In regards to LLMs, in the same manner that we translate ourselves to communicate with one-another, we anticipate the machine to perform the same labors of translation to maintain conversation with us. This calls for an expanded understanding of translation, specifically a theory of translation that permits humans to communicate with LLMs in a fashion that supplements our abilities and agency.

The task of translation, as explained by Walter Benjamin, consists in “finding the particular intention toward the target language which produces in that language the echo of the original.” The task must be differentiated from that of the poet; the work product itself held distinct from a piece of literature. Per Benjamin, translation (opposed to an original literary work), “finds itself not in the center of the language forest but on the outside facing the wooded ridge; it calls into it without entering, aiming at that single spot where the echo is able to give, in its own language, the reverberation of the work in the alien one.”

From the perspective of an LLM, the task of translation occurs in nearly each stage of its computational processes. I would resist anthropomorphizing the machine, however they are man-made creations, a by-product of our reason and creativity. We have endowed them with the capabilities to think and speak in patterns familiar to ours, to mimic us. Language models do not stand at the center of the language forest, as that is the territory of mankind. Rather, they exist at the periphery (per Benjamin, “on the outside facing the wooded ridge”) and communicate with us, without entering, by aiming towards a singular spot that produces an echo in their own language.

The main difference when discussing LLMs and translation in the context of Benjamin, is that there is no original work of art or literary product that exists in its own right. But rather, the labor of translation occurs on both ends of human and machine interaction.

As users, we likewise find ourselves standing on the outside, facing a wooded ridge that conceals itself in binary. We do not speak the alien tongue despite having had created it. It will always, in a sense, be foreign to us. We have to come to terms with its foreignness. We face the ridge we cannot enter, on the edge of a digital (yet man-made forest), and call out to it, anticipating a response, an echo, a reverberation, in our mother-tongue.

Here we are presented with the issue of two-way translation. The machine labors to translate our language into its own; and from its own, back into ours. We have collectively labored to preemptively translate our language for the machine and programmed it to accommodate us. In practice, we translate ourselves (the same way we would in conversation with others) to communicate with the language model.

Analogous to translation, the LLM tokenizes prompts and input from the user. It identifies the source language, as one would sit with the text in the source-tongue, and fractures the input into digestible bits of information. It embeds them the way a translator would read a text. The translator’s most valuable skill is undeniably their ability to understand both the source and target language. The text is then placed in context - grammatical, semantic, tonal. The relationship between words and the manner in which they are presented alongside one another, the context vector, determines their ultimate relational meaning. Next token prediction might best be explained as the obvious labor of translation. Words in the source language must be translated into their target language with attention paid to their contextual relationships. As for iterative decoding, the translator’s final, polished product rests in front of him.

Returning to Benjamin, the hallmark of a bad translation or a bad translator is one that intends to perform a transmitting function. It aims only to communicate. Communication figures as the least essential part of translation. We are urged to understand translation as a form, figuring translatability as a quality of particular works or practices. With regard to human and machine communication, translatability is built into the nature of the relationship. Translation is never perfect, it cannot replace the original, however it stands “in the closest relationship to the original by virtue of the original’s translatability.” The labor of translation and communication performed by AI is never perfect, they are undeniably unable to replace the experience of human-human communication or relation (despite close imitations).

Human-machine communication marks a new stage in the development of human language. We are no longer building a tower up to the heavens, but rather we have come to terms with it being stricken down. We have gathered the remains of Babel and repurposed them. Endowing machines with the ability to seemingly speak in our language, reply to us ‘naturally’, exists as a testament to what Benjamin posits as the innermost kinship of languages. We’ve created a language for machines and taught them to speak in ours revealing a special kinship that persists. Our kinship with artificial intelligence is marked by a relationship of familiarity rather than strangeness or uncanniness seeing that all languages, machine or human, “a priori and apart from all historical relationships are interrelated in what they want to express.”

We have understood the failures of rote, strictly communication-motivated translation since the first steps of artificial intelligence. Christopher Strachey, working with the Manchester Mark 1 in 1952, developed an early language model using a random number generating algorithm. Predating the earliest natural language processing machines, Strachey created the love letter generator and began posting its products (love letters) on the notice board of Manchester University’s Computer Department. The saccharine notices, desire overdone to the point of comedy, read along the lines of:

The varying iterations of the message in itself remain stable. Programmed to function essentially as a thesaurus, the love letter generator’s output produces an almost inhuman text. The awkwardness of the tone, the indelicacy of the language, the repetitiveness of dated tropes (even in 1953) figure for a love letter that communicates emptiness, nothingness even. Outside of the question of machine authorship, the thesaurus-programmed language generator fails to engage in the aforementioned two-way translation process. The lack of spontaneous input from the user, combined with the barriers programmed into the generator, limit its capabilities. The resulting text exists within itself rather than in communion with the user.

In our time, we are able to hold two-way conversations with current language models. Thus, in speaking with artificial intelligence (specifically in terms of prompting), one should be mindful of translation as a form. In the “suprahistorical kinship” between languages, they seek to mean the same thing. The way words mean across languages is supplemented in relation to how they mean in the murky waters of semantic relations and ever-shifting contexts. Translation tests the persistence of meaning. The source language and the target language must both be evaluated categorically. Has the context shifted? How has this word been derived? Is meaning tied strictly to historical relevance? Has it any resonance? We come to terms here with foreignness. Difference. So, how does one locate the nucleus of Benjamin’s “pure language” to preserve the particular power of meaning across languages?

We avoid an error identified by Benjamin in 1921. A translator fails when he “preserve(s) the state in which his own language happens to be instead of allowing his language to be powerfully affected by a foreign tongue.” We must expand our language, we must grow outward. We gather the bricks in Babel and build horizontally, outward.

We are a priori related to machines, connected through a fragmented, yet shared language, repaired through translation. Communication through translation becomes the central concern. Working on methods of communication/translation between humans and machines reveals the Benjaminian kinship and our ability to collaborate with them. How we talk to machines and artificial intelligence bolsters both the language spoken by us and by them. To hone in on the analogy of translation, the task of prompting might be compared to a translation task.

It is critical to understand how the model thinks in its own language (the same way a translator must have a healthy grasp of the source tongue and the target tongue). We anticipate its internal functions, the manner in which it recognizes tokens in relation to one another. Much like a translator, the user peers into the inner language of the machine and observes how it thinks:

“Just as a tangent touches a circle lightly and at but one point-establishing, with this touch rather than with the point, the law according to which it is to continue on its straight path to infinity-a translation touches the original lightly and only at the infinitely small point of the sense, thereupon pursuing its own course according to the laws of fidelity in the freedom of linguistic flux.”

Understanding the “source language” or the machine’s manner of recognizing patterns in human-generated speech and producing relevant, appropriate outputs, permits us to maximize use of the machine. Changing our view of ourselves in relation to machines from “programmers” or “users” to “translators” reveals our task: to narrow, understand, and deliberately speak to its ability to predict speech, text, thought, and patterns statistically.

As translators, users interact with the machine on its terms. Prompt engineering shapes the context needed to produce the “correct” outputs. In remaining aware of how an LLM interprets and prioritizes information, we must allow our own tongue to be deeply affected by our interaction with the machine. Recall the kinship, the interrelatedness of language, thus, being ‘good at prompting’ hinges upon being an intelligent and intentional translator. As their context windows continue to expand, we must learn to navigate them seamlessly, to expand our own language to grow alongside their rapidly expanding understanding of the way our world works.

Historically, human beings are writers of language more recently than we are speakers. We have only a few thousand years of practice in terms of learning to communicate through text rather than speech. Further, we are even more recently speakers of digital language. We are witnessing the first generation to have grown up speaking either with or through machines as much, and if not more, than they have face-to-face with others. Young people are the innovators of language, always, but especially so in the digital age. Our task as translators and users figures as learning to translate in harmony with the internal language and thought patterns of artificial intelligence while allowing our own language to be amended and molded through this engagement.

Artificial intelligence cannot replace human discernment, or the particularities of cultural coexistence with others, or ideas as amorphous as “taste.” However, it can replace the labor intensive and monotonous tasks of aggregating information and combing through it. To collaborate and preserve human agency, we must hold the manner in which we approach communicating with machines as our central concern. As translators, our labor must be turned towards developing fluency and establishing a sense of comfort in a foreign, digital tongue.