“This side of the movement in the scholarship has, however, been slow because doing it well requires a lot of language expertise: not just Greek and Latin, but perhaps also Babylonian, Old Persian and Egyptian. Not a lot of scholars have that kind of language expertise, both because it’s hard to learn that many languages, but also because it is hard to learn those eastern languages at all because they’re simply not taught in many universities. Indeed, the number shrinks over time, as language programs are often under siege at universities.” — Bret Devereaux, On the Reign of Alexander III of Macedon, the Great?

This translation effort stemmed from two things. First, a desire to share the text with an English speaker (having myself read a French translation years ago) but being unable to find an English translation whose quality I could ascertain¹.

And second, an exercise in using artificial intelligence to bridge the gap between a primary source and myself. A lot of historical work is built upon reading and analyzing primary sources written in the native tongue of their authors (often a dead language for all intents and purposes), but the number of historians able to read those languages is continuously shrinking, and some texts that might be of interest to a particular individual are so obscure that available translators cannot be expected to dedicate time to them.

The solution is obviously to reform universities’ language programs. But, in the meantime, I wondered if artificial intelligence could help us bridge that gap and figured that a text in 17th-century Latin, a language I do not speak², would be a good way to experiment. The goal was to evaluate the feasibility of the task, understand the required intermediate steps, as well as how much manual work would be required on top of the automation afforded by artificial intelligence.

The current result is very much a cyborg translation, something that could not be achieved solely with a human or machine translator. While I have a hard time picturing humans being fully taken out of the loop³, my perception is that the fraction of the task that can be automated is growing every year⁴, and I very much expect to obtain a better end result if I were to redo this work in a couple of years with a new and improved model.

However, this work was only possible due to the existence of a large number of freely and easily accessible scans of the original book. This, as well as establishing a clean reference transcription out of them, is work that can be done once now and benefit all further translation efforts. Furthermore, it is work that can be scaled to large corpora of texts, making them accessible to a wide and diverse array of future scholars who might not have been able to procure a physical copy of the original. I look forward to further progress along those lines, independently of any progress in artificial intelligence, lowering the bar of entry to research.