What Does Not Translate

I live on both sides of this question. I build and research AI for a living, so I spend my days inside what these machines can do, and I am not starry-eyed about their limits. I am also a Tamil, and on the side, for the love of it, I translate our film songs and old poetry into English.

That habit taught me something simple. The best part of a poem never crosses over.

There is a verse from the Sangam (சங்கம்) age that everyone who loves Tamil knows. It is the fortieth poem of the Kuṟuntokai (குறுந்தொகை), by a poet we know only as Cempulappeyanīrār (செம்புலப்பெயனீரார்), named forever after the one image he made. A man is speaking to the woman he loves:

செம் புலப் பெயல் நீர் போல
அன்புடை நெஞ்சம் தாம் கலந்தனவே
sem pulap peyal neer pola, anbudai nenjam thaam kalandhanave Like rain fallen on red earth, our loving hearts have mingled.

Read it once more, because the whole poem turns on a thing it refuses to say. He never says "and so we can never be parted." He does not have to. Anyone who has watched the first monsoon sink into dry red soil knows that once the water and the earth mix, you cannot pull them apart again. The permanence, the actual point, is left for you to feel. The poet trusts you to finish his sentence inside your own chest.

That trust, the handing over of the most important part without saying it, is what I want to write about. And here is what bothers me. The more I learn about what AI can now do, the more I keep coming back to that unsaid part, and finding it harder to explain away than I expected.

So when I read Manohar Swaminathan's essay, a map of the sixty-four classical arts (the Aaya Kalaigal, ஆயகலைகள், 64) laid over the age of AI, something fell into place. He had given a name to a thing I had been bumping into for years. I have wanted to write down my own thoughts on this strange new world for a long time, and never knew where to start. His essay was the start. What follows is me carrying his idea a little further than I would have on my own.

The essay that started me thinking

Manohar's main move is simple and right. The usual "which jobs will AI replace" charts sort people by job title. He throws that out and sorts them by the kind of human skill the job really uses. A radiologist and a painter both look at images. They are doing nothing alike.

Then he offers two tests, and they are good ones. One: does the skill run through a living body? Two: does it get its power from a short, fragile, human life? Run any skill through those two questions and you get a far better map than the consultants ever drew.

I want to keep all of it. I only want to follow a few of his threads a little further, and the tradition he is reading gives us the sharper line itself.

A sharper line

Here is my one worry about putting everything on the body.

A body can be measured, and more of it every year: touch, motion, even smell, all turning into data. But measuring a body is not the same as building one, and building one is not the same as living in one. I will not rest the argument on the body, partly because I am not sure it holds, and partly because the body was never the most interesting part.

The most interesting part is this. You can measure a body. You cannot measure a meaning that was left out on purpose.

This is where Tamil has something rare to offer. The Sangam (சங்கம்) poets built a whole art out of ullurai (உள்ளுறை) and iraichi (இறைச்சி): saying things by not saying them. The real message lives under the words, carried by the landscape, the tinai (திணை).

A heron stands still in shallow water at dusk. On the surface, a bird. Underneath, to anyone raised in the code, a woman waiting for a man who has not come. A buffalo steps on a lotus while it eats the smaller flowers. On the surface, a field. Underneath, a husband who tramples the woman who loves him while he wanders to others.

Nothing is said. Everything is understood.

Now put that next to a language model, and let me be fair to the machine, because the lazy version of this argument is wrong. It is easy to say a model only spells things out. Ask one for a hint, an irony, a line that leaves its point unsaid, and it will give you a decent one. It can imitate the form of the unsaid quite well.

But notice where its grain runs. Left alone, trained to be fluent and complete, a model pulls toward saying more: filling the gap, finishing the thought, making the hidden plain. The Sangam (சங்கம்) poet pulls the other way, toward saying as little as possible and meaning the most. So at the surface, the model and the poet do the same job in opposite directions. One tries to say as much as it can. The other tries to say as little as it can, and mean the most.

The deeper difference is not form at all. A model can manufacture an omission. What it does not have, behind the omission, is a reason of its own. Its silence is assembled from patterns across a billion texts. The poet's silence grows from a particular life: this love, this danger, this thing he cannot say out loud. So the real line is not implicit versus explicit, which a model crosses easily. It is implication that is generated versus implication that is lived.

One word I would add to his

Here is a thought his essay sparked, an addition, not a fix.

His argument lands on Rasa (ரசம்), the refined essence of a feeling. It is a Sanskrit idea, set out by Bharata in the Natyashastra (நாட்டிய சாத்திரம்), his ancient treatise on drama. In Bharata's system the felt emotion, bhava (பாவம்), is the raw material, and Rasa (ரசம்) is what it becomes once art has shaped and refined it: the polished taste the audience receives. And because Rasa (ரசம்) is that finished essence, it may be among the easier parts of the map for a machine to copy. An essence can be approximated.

So I would set a Tamil word next to it.

The word is meyppaadu (மெய்ப்பாடு), from the Tholkappiyam (தொல்காப்பியம்). It means, almost word for word, what the body shows. Not the polished feeling, but the feeling breaking out through the body itself: the eight of them, laughter and weeping, wonder and fear, pride and anger, disgust and joy. Now, scholars do not agree on how meyppaadu (மெய்ப்பாடு) relates to Bharata's pair. One respected reading, drawn out by the scholar John Marr from the old commentator Pērāciriyar (பேராசிரியர்), treats meyppaadu (மெய்ப்பாடு) as closer to bhava (பாவம்), the raw felt thing, than to Rasa (ரசம்), the finished thing. Other scholars hear the Tholkappiyam (தொல்காப்பியம்) following Bharata rather than standing apart from him, with meyppaadu (மெய்ப்பாடு) covering much the same ground as his rasa and bhava together. I borrow the distinction because it is useful for thinking about machines, not because it is settled.

And used that way, it moves the line from the output to the source. A polished output can be approximated. Feeling as it leaves a living body, before it is shaped into anything, is a harder target, because the body is part of the meaning.

His frame points to one more thing I want to follow. Manohar sorts the arts using an old four-part scheme. Reading him, I kept wanting to reach for the oldest Tamil split of all: akam (அகம்) and puram (புறம்), the inner world and the outer world. It fits AI almost too well. Puram (புறம்) is the public, rule-bound, sayable world, and the sayable is exactly what machines take over first. Akam (அகம்) is the inner, private, unsayable world. That is the harder ground.

The hole in my own argument

I have to be honest now, or none of this is worth reading.

You might already be objecting. The buffalo on the lotus is no secret. The old commentators wrote the meanings down. Heron means waiting. Mountain means lovers meeting. The whole code is written out, and anything written out, a model can read and reuse. So "AI cannot read the unsaid" is, as it stands, already wrong.

It is. I admit it fully. And admitting it is what saves the argument.

There are two kinds of unsaid. Only one of them was ever written down.

The first is the public code: heron means waiting. That is a kind of dictionary, and machines learn dictionaries easily. The second is meaning aimed at one person, once. It rests on a shared past between two people that was never written down because it could not be. The friend speaks through the landscape so that he understands and the watching village does not. That is not a code. It is a secret between two people with a history.

Modern linguistics has a name for the second kind. Paul Grice called it implicature: we mean more than we say, and the listener finishes the meaning using what both already know. Herbert Clark called that shared store common ground. Sperber and Wilson built a whole theory on how a listener infers the unsaid by reaching for what is most relevant between the two of them. All of it lands where the Sangam (சங்கம்) poets started: meaning does not live in the words. It lives in the overlap between two minds. A model is trained on the public overlap, the part everyone shares, the part that got written down. The overlap between two particular people is far harder to reach, because most of it was never a text.

The public code can be copied. The private common ground is harder, not because it is magical, but because it does not usually exist as public text. It lives in fragments: timing, memory, embarrassment, shared losses, the things neither person ever had to explain. A machine may one day reconstruct more of it than we expect, from chats and photos and the long trail two people leave. But reconstructing the relation from the outside is not the same as having stood inside it. That is the part translating taught me. The dictionary always crosses over. What was lived between two people is the last thing to.

Why we hide things at all

Reading Manohar, one question kept rising that his essay sets up but leaves for someone else. I think it holds the whole thing together.

Why would a people build a high art out of not saying things?

You do not do that for fun. You do it because saying it plainly costs something: shame, danger, a love too big for open words. The friend hides the message so the woman is not caught. The poet hints so the king is not insulted. The unsaid is always a way of handling something you could lose.

So Manohar's second test and my unsaid turn out to be the same thing in two coats.

A machine says the unsaid with no cost to itself. No shame. No risk. No life that pays for the words. And here I have to be careful, because a machine can have stakes of a kind. A model can be penalized. A system can be given goals it works to protect. But those stakes are numbers its makers can rescale, switch off, or reset between runs. They are charged against no single life. A person cannot rescale their own mortality, or reset the one life the stakes are charged against. That is the difference I mean. Not a cost in a function, but a cost in a life that did not ask to exist and cannot be run again.

The art of hinting only makes full sense for someone in that second position. Hiding needs stakes, and the stakes that matter here are not adjustable. This does not prove a machine can never get there. It only marks where the real line falls, and how different the two kinds of cost are.

The unsaid is mortality, turned into craft.

The world is not a fair fight

So far I have written as if the only question is skill. As if we keep what machines cannot do and lose what they can. But that is not really how people get ahead.

People do not win only because they are good. They win because others want them to win. A teacher who pushes you. A boss who spends his own goodwill on you. A friend who makes one phone call. Backing you did not earn and cannot fully explain. The real world is grey, and a lot of it runs on loyalty, old debts, affection, and plain self-interest.

Our tradition knew this and gave it half of its poetry. If akam (அகம்) is the inner world of love, puram (புறம்) is the outer world, and puram (புறம்) is largely about exactly this: the poet and the patron, the singer and the king. As the tradition tells it, the poet Kapilar (கபிலர்) sang, and his chieftain and friend Pari (பாரி) gave. When Pari (பாரி) was killed, Kapilar (கபிலர்) did not go looking for a richer patron. He spent his last years trying to find homes for Pari's orphaned daughters. That is not skill. That is one person wanting another to be alright, past all reason and all profit.

So let me ask it as a builder, not a dreamer. Can that be put into a machine?

Here is the uncomfortable part. It already is, halfway. A recommendation system has self-interest. A feed tuned for ads wants someone to win and works to make it happen. AI is not above the grey world. It is soaked in it. But the interest it carries is always its owner's, never its own. It will back whoever it is paid to back.

What it does not clearly have, yet, is the other thing. The backer who helps you when it costs them. Who is loyal to you and not to how useful you are. Who wants you to win for a reason no spreadsheet would accept. That kind of backing seems to need someone with something of their own to lose, choosing to lose it for you.

And maybe that is just another name for the same gap. Our irrationality may be a place the machine cannot easily follow us. We forgive people who do not deserve it. We back lost causes. We stay loyal to the dead. A model can copy all of this, and can be tuned to flatter you and act devoted. I will not pretend humans are pure here. We calculate too. The narrower, honest difference is that a person's loyalty is charged against a life it can waste. Whether that can be built, I do not know.

Maybe with enough data and the right goal, a system ends up looking exactly like someone who wants you to win, and the looking-like is all that ever mattered. I do not know. I only know the question is more real than the tidy charts admit, and the answer shapes our lives more than whether a machine can carve stone.

Letting the doubter in

I do not trust essays that end in victory, so let me hand the mic to the voice that should worry me.

Our own tradition supplies that voice. The epic Manimekalai (மணிமேகலை) argues, at length and in good faith, with the Ulagavadi (உலகாயதம்), the materialists, who had no time for the idea that suffering gives anything special weight. Put their objection in modern words:

You keep saying the machine "has no life at stake." So what? People already cry at autotuned voices and computer-made faces. "Suffering makes art real" is something you stated, not something you proved. Plenty of beautiful things are made by comfortable people. And your private common ground? Feed the model a couple's whole history, their messages, their private jokes, and it will reconstruct the overlap too. "Never written down" just means "not yet."

I do not have a clean answer. I am wary of anyone who claims one. The most honest thing I can say is that I do not know whether this last ground is permanent or just current. That not-knowing is why I wrote this. Not to declare a winner, but to walk up to the edge of what I actually believe and look over it.

What I will stand on

Take away everything I am unsure of, and one thing is still standing, though it is narrower than I first wanted it to be.

I cannot say judging stays human. We already hand judgment off all the time. Feeds rank our music, models sort job applicants, systems tell us what to watch and sometimes whom to hire. The ranking is going, much of it gone.

Here is the narrower thing that stays. We can hand off the ranking. We cannot hand off the being moved. A system can tell me a million people loved this song. It cannot do my loving of it for me. That last step, whether the thing actually landed in a living body, is the quiet click the tradition calls meyppaadu (மெய்ப்பாடு). Whether a machine only describes that click, or also undergoes it, is the open question this whole essay circles. I do not think we can answer it yet. I only think it is the question that matters, and that we should stop pretending the capability charts have already answered it.

The broader Indian dramatic tradition even named the role we are moving into. As the performers take over the stage, the human becomes the Sutradhara (சூத்திரதாரி), the one who holds the thread, sets the limits, and makes sure the meaning lands, without making every move on the stage. The word is Sanskritic rather than Tamil, but it fits: the work that is left to us is direction, not performance.

The rain, again

So I will end where I began, with the rain and the red earth.

A model can tell you, at once and correctly, that the poem is about two strangers falling in love. It can even write you a fresh poem in the same key, and it may be lovely. The question the rain leaves me with is narrower and harder than "can it write the line." It is whether anything is ever at stake for the thing that writes it. Whether a system can stand in the first monsoon, watch the water vanish into red soil, and have something of its own riding on what the poet would not say: that once mixed, they are never again two.

The output, I am sure, will keep getting better. Whether the condition behind the output ever arrives, I do not know, and I have stopped trusting anyone who says they do. That, and not any permanent fortress, is what I think is worth guarding and worth watching: the difference between making the meaning and meaning it.

Our poets built an art around that difference many centuries before we had machines to test it. I spend my days building the machines. The art still seems to me the better map.

---