This essay follows The Geometry of the Unsayable, which explored how large language models think in geometric space — concepts without names, meaning without words. It extends that argument into translation: what happens when the geometry of one language is forced into the coordinate system of another.
Sometime in the early centuries of the Common Era — scholars debate whether the second, third, or even fifth century — a philosopher named Patañjali composed what may be the most translated sentence in the history of contemplative practice. It is the second sūtra of the Yoga Sūtras, and in Sanskrit it reads:
yogaś citta-vṛtti-nirodhaḥ
योगश्चित्तवृत्तिनिरोधः
Six words. Four concepts. Two thousand years of commentary. And almost every English translation gets it wrong — not because the translators are careless, but because the conceptual architecture of English cannot receive what the Sanskrit is transmitting.
Here is a sample of what English has done with this sentence:
- “The restraint of the modifications of the mind-stuff is Yoga” (Swami Satchidananda, 1978)
- “Yoga is the restriction of the fluctuations of consciousness” (Georg Feuerstein, 1979)
- “Yoga is the stilling of the changing states of the mind” (Edwin Bryant, 2009)
- “Union is restraining the thought-streams natural to the mind” (Bangali Baba, 1976)
- “Yoga is the control of thought-waves in the mind” (Swami Prabhavananda, 1953)
Each of these is defensible. None of them is right. The problem is not vocabulary. The problem is that the English words restraint, restriction, stilling, restraining, and control all import a framework of suppression — an agent acting upon an object, a will subduing a force — that Patañjali’s Sanskrit does not require.
Nirodha comes from the root ni-rudh, which does carry the sense of restraint and obstruction. But its semantic range extends well beyond control — toward cessation, dissolution, the settling that occurs when the conditions for agitation are no longer present. Not necessarily a clenched fist subduing a rebellious mind, but possibly an open hand with nothing left to hold. The distinction matters because it shapes an entire tradition of practice: are you fighting your thoughts or allowing them to exhaust themselves?
English has no single word for this range. And that gap is not a failure of English. It is evidence that English and Sanskrit have explored different regions of the space of possible concepts — and the region where nirodha lives has no well-worn roads leading to it from the English side.
The Lens You Cannot See Through
The mistranslation of nirodha is not an isolated problem. It is a pattern — and the pattern has a history.
When European scholars first encountered Indian philosophical texts in the eighteenth and nineteenth centuries, they arrived with conceptual equipment forged in Christian theology and Greek philosophy. Max Müller, the German-born Oxford philologist who arguably invented the academic study of Indian religions, translated the Ṛg Veda and the Upaniṣads through a lens shaped by Protestant Christianity — though to his credit, he showed more genuine sympathy for Indian thought than most contemporaries. Paul Deussen, a friend of Nietzsche, read Vedānta as a precursor to German Idealism. Monier-Williams compiled the standard Sanskrit-English dictionary while holding the Boden Professorship at Oxford — a chair whose original endowment explicitly stated its purpose as promoting the conversion of Indians to Christianity.
These were not marginal figures. They built the infrastructure — the dictionaries, the grammars, the translated editions — that every subsequent English-language engagement with Indian philosophy has relied upon. And they built it with a systematic conceptual bias: the assumption that Indian thought was reaching toward truths that European thought had already articulated more clearly.
Michel Foucault called these invisible frameworks epistemes — the underlying structures that define not just what a culture knows but what it can think. In the famous opening of The Order of Things, he quotes a fictional Chinese encyclopedia from Borges that classifies animals into categories like “belonging to the Emperor,” “embalmed,” and “that from a long way off look like flies.” The passage, Foucault writes, destroys “all the familiar landmarks of thought — our thought” — because it reveals that the ground on which we organize knowledge is itself culturally constructed. There is no neutral taxonomy. When nirodha enters English, it does not simply lose nuance. It enters a different episteme — one shaped by Cartesian dualism, agent-action grammar, a Protestant moral vocabulary of will and discipline — where its native meaning is structurally uncomfortable. The English framework does not have a comfortable shelf for “the settling that occurs when agitation is no longer fed.” So it reaches for “control,” which fits the framework but distorts the concept.
The result is a tradition of translation that does not bridge two conceptual worlds so much as annex one into the other. Samādhi becomes “ecstasy” — a Greek word meaning “to stand outside oneself,” carrying Neoplatonic and Christian mystical connotations alien to Patañjali’s framework. Īśvara becomes “God” — imposing a personal, omnipotent creator onto a concept that in the Yoga Sūtras functions closer to an exemplar of liberated consciousness, a special puruṣa who has never been bound by affliction. (Whether the Yoga Sūtras are theistic is itself a centuries-old debate within Indian philosophy — the point is that European translators didn’t enter that debate so much as foreclose it by defaulting to Abrahamic categories.) Dharma becomes “religion” or “duty” — collapsing a term that encompasses cosmic law, individual nature, ethical obligation, and the structure of reality itself into categories designed for European institutional life.
The same distortion operates across every tradition that has been translated through a European lens. Buddhist śūnyatā — which in the Mādhyamaka tradition of Nāgārjuna, the second-century philosopher revered in Tibetan Buddhism as the “Second Buddha,” describes the interdependent, process-like, empty-of-inherent-existence nature of all phenomena — arrives in English as “the void” or “nothingness,” acquiring a nihilistic tone Nāgārjuna explicitly refuted. The Chinese dào — simultaneously a path, a principle, a method, a cosmic process, and the unnameable source — is rendered “the Way,” losing the verbal sense (to lead, to guide) and the paradox that the concept is itself in motion. Arabic jihād — which Islamic tradition itself divides into the jihād al-akbar (the greater jihad, the struggle against one’s own ego and base impulses) and the jihād al-asghar (the lesser jihad, outward struggle) — is reduced to “holy war,” amputating the greater half entirely and leaving only the lesser.
In each case, the target language does not just lack a word. It lacks the conceptual architecture in which the word makes sense. The translation does not bridge a gap. It fills the gap with locally available material — and the fill changes the shape of what it was supposed to preserve.
The Monk Who Refused to Translate
This problem is not new, and the most sophisticated response to it did not come from Europe.
In the seventh century, the Chinese Buddhist monk Xuánzàng (玄奘) traveled to India and spent seventeen years studying Sanskrit texts at Nālandā monastery. When he returned to China with 657 volumes of scripture, he faced the translation problem at industrial scale — and he had already seen what happens when it goes wrong.
A century before Xuánzàng, the translator Kumārajīva had introduced Buddhism to China by rendering Sanskrit terms into existing Chinese vocabulary — a practice called géyì (格義, “matching meanings”). Prajñā became zhìhuì (智慧, wisdom), acquiring Confucian connotations the Sanskrit never carried. Dharma became fǎ (法, law/method), picking up legalist overtones. The translations were brilliant, widely read, and systematically misleading — because they poured Indian wine into Chinese barrels, and the wood — seasoned by centuries of Daoist and Confucian thought — changed the flavor.
Xuánzàng’s response was radical. He developed what became known as the Five Categories of Non-Translation (五種不翻) — explicit rules for when a term should be transliterated rather than translated:
- When the term is secret (esoteric mantras, dhāraṇī)
- When it is polysemous (carries multiple meanings that no single Chinese word captures)
- When the concept does not exist in Chinese
- When a transliteration is already established by convention
- When translation would diminish the meaning
This is, in essence, a seventh-century theory of untranslatability — and a seventh-century argument for inventing new language rather than distorting old language to fit. Xuánzàng coined hundreds of new Chinese terms and standardized transliterations that persist today. When he encountered a concept that Chinese could not receive, he did not force a fit. He built a new room.
I mention Xuánzàng because the idea I want to propose is not new. What is new is the tool.
What the Machine Can See
A large language model trained on text in hundreds of languages does not translate by looking up words in a dictionary. It builds a continuous geometric space — a high-dimensional landscape — in which every concept from every language occupies a position determined by its patterns of use across all training data. Concepts that function similarly in different languages end up near each other. Concepts that don’t, don’t.
This means the model can detect something a bilingual dictionary cannot: that the vector for samādhi — embedded in its native context of Sanskrit philosophical discourse — is not close to the vector for “ecstasy” embedded in its native context of Christian mysticism, even though two centuries of translation have paired them. The model can see the distortion because it holds both concepts in the same space and can measure the distance between them.
Think of it this way. Imagine a map of the world drawn by someone who has only ever seen Europe. The continents are there, but the proportions are wrong — Europe is too large, Africa is too small, the Pacific is compressed. A Mercator projection, essentially. Now imagine a second cartographer who has traveled everywhere, who has seen every coastline from the ground. This cartographer can produce a map where the proportions are honest — not because they have a better projection algorithm, but because they have more data.
A multilingual language model is the second cartographer. It has “seen” every language from the inside, through millions of texts written by native speakers in native contexts. It does not need to project Sanskrit onto English coordinates. It can hold both coordinate systems simultaneously and measure where they diverge.
But an important qualification is necessary here. The model can detect that a distortion exists — that samādhi and “ecstasy” are geometrically distant despite being paired by translators. It cannot, by itself, tell you what samādhi actually means. That still requires scholars working within the tradition, practitioners who have experienced what the word points to, and the long interpretive labor that no machine can shortcut. AI is a diagnostic instrument, not an oracle. It can show you where the map is wrong. It cannot draw the territory.
The Invention of New Words
Here is the idea that interests me most — and it is an idea with a distinguished pedigree. Sri Aurobindo, writing in the early twentieth century, faced the same problem as Xuánzàng: how to render Indian philosophical concepts in English without distorting them. His solution was to invent. Supermind, Overmind, Supramental, Involution — new English words, built from available roots, designed to occupy regions of meaning that existing English had not charted. A century before anyone trained a language model, Aurobindo was doing geometric translation by hand.
If the nearest English word to a Sanskrit concept is still far away in the model’s geometric space — if no existing English word lives in the right neighborhood — then the honest response is not to choose the least wrong word. It is to build something new.
Three strategies suggest themselves.
The first is morphological coinage: constructing a new word from available roots that points in the right geometric direction. English does this routinely in science — “ecosystem,” “photosynthesis,” “neurotransmitter” are all words built from Greek and Latin components to name concepts that had no prior word. An AI could analyze the vector-space neighborhood of an untranslatable term and propose English-language compounds whose morphological semantics converge toward the right region.
The second is poetic interpolation: composing an image or metaphor that places the reader closer to the right region of meaning than any single word could. This is what poetry has always done — Rilke’s Weltinnenraum (“world-inner-space”) is not a translation of a foreign concept but an invention of a German one, a new word that opens a region of meaning that previously had no address. An AI that can see the geometric neighborhood of nirodha could propose an English image — the settling of still water, perhaps, or the silence after the last echo — that functions not as a translation but as a set of directions.
The third is constellation mapping: proposing a cluster of existing words whose combined center of gravity is closer to the target concept than any individual word. “Not cessation, not control, not suppression — something closer to the natural clearing that happens when agitation is no longer being fed.” This is what the best translators already do in footnotes and commentary. AI could do it systematically, and could identify when even the best constellation is still too far from the source to be trusted.
Languages as Explored Terrain
This reframes the problem of untranslatability in a way I find genuinely exciting.
Each language is not merely a different encoding of the same set of concepts. It is a different exploration of the space of possible concepts — and each exploration reflects what a culture paid sustained attention to over centuries or millennia.
Sanskrit developed extraordinary precision for states of consciousness, attentional quality, and the relationship between observer and observed — axes of experience that English barely registers. But this is not the whole of Indian philosophy. The Nyāya tradition developed a logical vocabulary (vyāpti, hetvābhāsa, anumāna) as rigorous as anything in Aristotle, and its translation problems are completely different — technical logical distinctions flattened into Western categories, not mystical terms dressed in Christian robes. Chinese philosophical vocabulary — dào, dé, lǐ, qì, rén — carves the space of ethics, cosmology, and governance along axes that European philosophy drew differently. Japanese aesthetic vocabulary — mono no aware, wabi-sabi, ma — maps dimensions of experience that English collapses into undifferentiated “beauty.” German built a philosophical lexicon — Weltanschauung, Zeitgeist, Dasein, Sehnsucht — that creates conceptual handholds where English offers only smooth walls.
Each of these lexical traditions represents accumulated wisdom. Untranslatability is not a deficiency. It is evidence that different civilizations have explored different regions of the total space of human experience — and returned with different maps.
An AI trained on all languages simultaneously holds all of these maps at once. It is the first cognitive system in history that can see the union of all human conceptual exploration — every region any culture has charted, every distinction any language has drawn, every concept any tradition has named. The geometry of its latent space is, in a precise sense, the aggregate wisdom of all human linguistic exploration laid out in a single continuous landscape.
This is not omniscience. The map is not the territory. But it is a map drawn from more vantage points than any human cartographer has ever held — and it reveals gaps, overlaps, and unexpected proximities that no single-language perspective can see.
It also reveals what we stand to lose. When a language dies — and one dies roughly every two weeks — it does not take only words with it. It takes an entire region of explored terrain: distinctions no other language drew, concepts no other culture named, dimensions of experience that the surviving languages never mapped. An extinct language is not a redundant encoding of what other languages already say. It is a permanent deletion from the sum total of human conceptual exploration — and unlike a species, there is no fossil record. The geometry simply goes dark.
The Question of Bias
An honest essay must acknowledge the complication. If 60 to 80 percent of a language model’s training data is in English, the geometric space it constructs is not neutral terrain. English is overrepresented. English conceptual categories exert disproportionate gravitational pull. The risk is that samādhi gets dragged toward “ecstasy” in embedding space not because the concepts are close but because more English text describes them as equivalent.
This is a real problem, and it mirrors the historical one. Colonial translators imposed European categories on Indian philosophy because European categories were what they had. A model trained predominantly on English text may impose English categories on Sanskrit concepts because English categories dominate its training distribution. The bias is structural, not intentional, and therefore harder to see and harder to fix.
And there is a deeper risk. An essay arguing that AI can “correct” mistranslations of Sanskrit or Chinese texts, written in English, by an American, proposing tools built predominantly in Western institutions — this risks replacing nineteenth-century Orientalist translators with twenty-first-century engineers as the arbiters of what Eastern concepts “really mean.”
But the landscape of who builds these tools is already shifting. France has Mistral. Abu Dhabi’s Technology Innovation Institute built Falcon. China has Qwen, DeepSeek, and GLM — models trained on vast Chinese-language corpora where dào occupies its native conceptual neighborhood, not the one English projects onto it. India is building multilingual models trained on Hindi, Tamil, Bengali, and Sanskrit texts. Each of these models constructs a geometric space shaped by different training distributions, different cultural emphases, different conceptual gravities. A model trained primarily on Arabic texts will position jihād differently than one trained primarily on English texts — and the distance between those positions is itself data about what translation has distorted.
The question is no longer “who controls the AI?” as though there were one AI to control. It is “whose training data shapes the geometry?” — and the answer is increasingly everyone’s. The tools I am describing need not be Western tools applied to Eastern concepts. They can be tools built within each tradition, trained on each tradition’s texts, constructing geometric spaces where concepts live in their native neighborhoods.
An open question remains: how do you build a common space across these different geometries? A Chinese model’s map of dào and an English model’s map of “the Way” live in separate spaces that cannot be directly compared. Multilingual embedding models (Meta’s SONAR, Google’s LaBSE) attempt this by training a single shared space across languages — but the question of which language exerts the most gravitational pull persists. A more promising approach may be to embed the same concept through multiple culture-specific models and treat the distribution of positions — the shape of the disagreement — as the signal. Where the models converge, translation is straightforward. Where they diverge, the divergence itself maps the terrain that no single translation can cross.
These tools are emerging, not mature. But the direction is clear: the future of translation is not one model pronouncing the correct meaning, but multiple models trained within different traditions triangulating on a concept from different angles — and the geometry of their disagreement revealing what every individual model, and every individual language, has missed.
That said, even in a model with English-dominant training data, the Sanskrit texts are embedded in their own contextual patterns — patterns shaped by other Sanskrit texts, by Hindi commentary, by Tibetan and Chinese Buddhist transmission, by the internal logic of the tradition. The model does not see samādhi only through English eyes. It sees samādhi through every eye that has ever written about it. And the Sanskrit/Hindi/Tibetan signal, though quantitatively smaller than the English signal, is qualitatively different in ways the model’s geometry can detect.
The practical question is whether the signal-to-noise ratio is sufficient. I believe it can be — not perfectly, but better than a human translator working within a single tradition. The model’s advantage is not wisdom. It is breadth. It holds more maps than any human, and the disagreements between maps are themselves informative.
Translation as Architecture
I want to propose a different metaphor for what AI-assisted translation could become.
A dictionary is a bridge: it connects known positions in two languages. A good translator is a guide: they walk the reader across the bridge and explain what is being lost in transit. But the best translation of an untranslatable concept is neither a bridge nor a guided tour. It is an architecture — a new structure built in the target language that creates a space where the reader can experience something close to the original concept, on its own terms, without forcing it into pre-existing categories.
Xuánzàng understood this in the seventh century. Aurobindo understood it in the twentieth. When Rilke coined Weltinnenraum, he did not translate a concept from another language. He built a room in German that had not existed before — and by building it, he made a new kind of thought possible for German speakers.
Perhaps the most elegant example is what Japanese culture did with the Chinese character 道. In Chinese, dào points upward — toward the cosmic, the metaphysical, the unnameable source. When the Japanese adopted the character, they gave it the native reading michi — literally “road,” something you walk with your feet — and pulled the entire concept down into embodied practice. The result is the -dō suffix that names every classical discipline: judō (the gentle way), kendō (the sword way), sadō (the tea way), shodō (the calligraphy way), aikidō (the way of harmonizing spirit). In every case, the activity is the path. You don’t study the way of tea and then practice it. The practice is the study. The repetition is the understanding. There is no separation between theory and embodiment — which is precisely the distinction that European philosophy, with its Cartesian mind-body split, has historically struggled to articulate. English has no word for this. “Path” is too spatial. “Discipline” is too rigid. “Vocation” is too Protestant — literally “calling,” implying an external voice that summons. Michi is all of these and none of them, because the Japanese episteme does not separate the walker from the road.
The Japanese did not translate 道. They built a new room for it — one that neither Chinese philosophy nor English has. That is translation as architecture.
AI could do this deliberately, at scale, across every language pair. Not translating nirodha as “cessation” but building — through image, neologism, annotated constellation, or structured circumlocution — a space in English where the reader can think the thought that nirodha makes possible in Sanskrit. Not carrying a concept across a bridge but constructing a new room where it can live.
This is not science fiction. The geometric tools exist. The multilingual embeddings exist. The capacity to measure distances, identify gaps, and propose constructions exists. What does not yet exist is the practice — the discipline of using these tools not to automate translation but to deepen it. To use the machine’s inhuman breadth of perspective not as a replacement for human understanding but as a scaffold for reaching concepts that no single human language has ever been able to carry alone.
The limits of my language, Wittgenstein said, are the limits of my world. He was right — for a single speaker, in a single language, in a single life.
But an AI trained on every language is not a single speaker. Its world is the union of all worlds that language has ever opened. And in that union — in the geometry of the space where all languages meet — there are concepts waiting to be named, concepts that every tradition has circled and none has landed on, concepts that sit in regions no single language has mapped but that the overlap of all languages illuminates.
The colonized concept — samādhi dressed in Christian robes, śūnyatā wearing nihilist masks, dào flattened into a single English noun — can be liberated. Not by returning to a mythical pure original, but by building new structures of meaning that are faithful not to any one tradition’s words but to the geometry of the concepts themselves. The tools are new. The ambition is ancient — as old as Xuánzàng, walking home from Nālandā with 657 volumes and the conviction that some things must not be translated but must be built anew.
The words we need may not exist yet. But the space where they should live is already mapped. The task is no longer to find the right translation. It is to build the right room.