We sorted every translated page in Source Library by length, longest first, to see what was at the top. The single longest “page” in the whole library is page 184 of a nineteenth-century Old Testament in Javanese. The AI’s English translation of that one leaf is 491,418 characters long — about eighty thousand words, longer than The Great Gatsby. It begins like an ordinary genealogy, and then it finds a word it likes:
A genealogy — the generations of Adam, the generations of Noah — asked the machine for the generations, and the machine, ever obliging, generated them. Handed a list of begettings, it begat without end.
And it is always like this, and it is always scripture. Here is the AI translating a seventy-thousand-word Tibetan volume of the Buddhist canon — a liturgy of accomplishment for the benefit of all beings — which it renders, faithfully, into the one thing the liturgy is for:
Here it is inside a commentary on the Zohar, the central text of Kabbalah, where it locks onto one of the divine attributes and will not let go — a mind circling a single mystery, which is more or less what the book is about:
It does this with the Taoist Canon (it comes apart on 98% of the pages), with Syriac homilies, with Ethiopic gospels, with the Mishnah. Hand it Cicero and it translates all day without breaking a sweat. But the texts that undo it are, almost without exception, the ones built out of sacred repetition — mantras meant to be recited ten thousand times, litanies, the names of God, the long chains of who begat whom. Repetition is the whole point: the thing a devotee chants on purpose, for hours, to empty the mind. Feed that to a machine that is already unsure of the script, and it does exactly what the devotee does. It falls into the chant.
The only difference is that the monk chooses to stop.
Excerpts are verbatim from the model’s output. We’ve since flagged the runaway pages and are re-translating them with a hand on the shoulder.