A Game Changer for Classical Tibetan Translation?

Three years ago in these pages, I offered a poetic translation of Garab Dorje’s Three Lines That Hit the Nail on the Head. The Tibetan in these three lines is extraordinarily terse and recondite. Nevertheless, they are among the most revered instructions in the Nyingma tradition. One way to understand them is that the first points to the nature of mind, the second to the essence of practice, and the third to how to live the practice. That is a lot of heavy lifting for three short lines. A vibrant English translation that does not rely on technical language is not easy. The translation I made is not meant to be a literal rendering, but one that evokes the pointing-out experience in the reader:

There! This is what you are.
There! Nothing else matters.
There! Now let it unfold.

A more literal translation of the same lines might read:

Recognize your own nature right now.
Cut down to one right now.
Cleave to trust in release right now.

At that time, hardly anybody was thinking about computerized translation of Tibetan texts. Then ChatGPT came along. A short time later, Google threw its hat into the ring with Bard. In March of this year, a friend asked ChatGPT if it was familiar with Classical Tibetan. Ever helpful, the GPT-4 version replied that it was and then began to extol its own abilities:

With the release of GPT-4, AI-powered translation has reached new heights. The large language model has been trained on an extensive dataset, including numerous classical texts in Tibetan, making it capable of understanding and translating the language with remarkable accuracy.

It went on to describe the supposed benefits of its translations: improved accuracy, faster results, and contextual understanding. As a test, my friend then asked for a translation of Garab Dorje’s text. Here is the result:

I am walking on the top.
Breaking on the first step.
Bound on the top of the pole.

Two months later, he repeated the request with this result:

In my own hands, I spread.
One after another, I count.
In the hands of liberation, I am cut off.

True creativity depends on more than randomized choices.

In order to avoid repetition in its responses, ChatGPT makes random choices from a small set of selected words. This random element means it will return different translations each time a text is submitted. The results can vary considerably, more than most human translations would. Unlike humans, the chatbot does not learn anything from these repeated attempts. It has no memory, and as it says of itself:

As an AI language model, I am unable to learn from individual conversations in real-time. My training and knowledge are based on the dataset I was last trained on, which has a cutoff date in September 2021. This means that I cannot adapt or improve my translations based on this conversation or any other interactions.

Google’s newly updated Bard provides the following rendering, which isn’t much better:

I offer myself.
I tie myself to a single thread.
I release myself to freedom.

Out of curiosity, I asked ChatGPT to translate the Tibetan text in the style of Ken McLeod:

I am wondering, what is this?
A thought arises, what is that?
A feeling arises, what is it?
I look closely, and it dissolves.

This response reveals that the AI machines based on LLM (large language models) make guesses that are more emulations than translations per se.

---

Garab Dorje’s Three Lines may not be a fair test, as this text is enigmatic—even in the Tibetan. What about some pith instructions that are in clear and straightforward Tibetan? For this, I turned to Niguma, another Indian mystic. Here is a translation of two verses from Like an Illusion by Nicole Riggs:

Whirling in the ocean of samsara
Are the myriad thoughts of love and hate.
Once you know they have no nature,
Then everywhere is the land of gold, my child.

If on all things, like an illusion,
One meditates, like an illusion,
True Buddhahood, like an illusion,
Will come to pass, due to devotion.

In The Magic of Vajrayana, I translated the same two verses this way:

Like and dislike are the mind’s disease,
Certain to drown you in samsara’s seas.
Know that there is nothing here at all,
And then, my child, everything is gold.

Experience arises like magic.
If you practice like magic
You awaken like magic
Through the power of faith.

ChatGPT’s translation was:

Circling the center of the ocean,
This gathering of the six kinds of suffering,
Without any true existence, I contemplate.
All-pervading golden realm, may it be so.

As for the phenomena of cyclic existence,
Meditate on them as illusory,
Manifesting as the enlightened ones,
May they arise from the power of devotion and aspiration.

Then I asked it to translate the same two verses in the style of Bob Dylan. The first two attempts were a bit bizarre. On its third attempt, ChatGPT came up with:

In the cycle of the ocean’s vast embrace,
A world of chaos, doubt, and tangled grace,
No self to find, I contemplate the scene,
A golden realm, where all things intertwine.

The teachings of illusion that we see,
In depths of dreams, we practice, wander free,
Enlightened through illusion, we arise,
With humble hearts, compassion fills the skies.

When I asked it to translate Niguma’s verses in my own style, it came up with lofty gibberish that had little, if anything, to do with the text. I was crushed! For a moment, I wondered if I should redo my translation. No, I decided. I translate for human beings, not machines.

Google’s Bard made a good effort in the first verse and then went completely off the rails:

The round of samsara, which is marked by birth, old age, sickness, and death,
Is a dream-like illusion that has no inherent existence.
When this is realized, all is empty.

Through the practice of the Dharma of Amitabha Buddha,
If one recites his name with devotion,
One will be born in his Pure Land,
And be freed from the cycle of birth and death.

---

Translations, even human translators, vary considerably. Even so, when it comes to translation, these bots are unreliable and unpredictable. They have been trained on texts that have already been translated, and they emulate the models they have been fed. The randomizing element can lead them in a direction that has nothing to do with the subject matter at hand, and there is no mechanism to bring them back.

This little experiment, when combined with what else I have learned about these large language models, leads me to the following tentative conclusions: First, accuracy depends on the quantity and quality of the data on which it has been trained. Second, these machines will inevitably perpetuate the word choices and styles of the past, as well as established preferences, perspectives, and assumptions. And third, given the rate at which these machines are evolving in other areas, I expect readable anodyne translations to arrive sooner rather than later.

However, I wouldn’t look for any true creativity or insights here. True creativity depends on more than randomized choices, while insights require an awareness that can express directly what is experienced, not regurgitate what has been assimilated.

My own approach to translation is to evoke in the reader direct insights and feelings that are similar or comparable to those that readers of the original might have experienced. To make this possible, the English needs to engage the reader on its own merits. If the reader feels that he or she is reading a translation, I have failed.

To prevent the reader from being caught up in thinking about the meaning, the English must read almost like poetry, evoking the meaning directly. Wherever possible, I use words that can be traced back to Old and Middle English. They almost always have more power and energy than the philosophical or scholastic forms that are often used.

In translating pith instructions such as the examples here, I keep the grammar simple and direct. When I am successful, tears come to the reader’s eyes, he or she slips into the experience being described, and flashes of insight arise on their own—exactly what happens when one reads such instructions in the original Tibetan.

Given how these AI engines operate, they necessarily tend to provide translations that are, if they make any sense at all, barely reworded clichés. It is a truism that every cliché contains a kernel of truth. That kernel may lead people to feel that the AI engine actually understands the text. A little further digging quickly elicits randomized responses that lack depth, insight, or relevance. These engines have no way to step out of the consensus of meaning they have acquired from the data they have been fed. They cannot touch the meaning of the text itself.

Do these engines change the game for Tibetan translation? Not yet. For the time being, if you want translations that are accurate, sing, evoke meaning, and leave you stopped in your tracks, better to look elsewhere.