I’ve now trudged all the way through the remainder of the Life of Abba Garima, taking each paragraph and getting Gemini AI to translate it into English. Why Gemini? Well it was there on my list of AI sites, and I used it at random. For the immense chapter 7, I broke this down into smaller paragraphs, and this worked.
I tend to believe that what I am getting out of Gemini is indeed what the Ethiopic text says, more or less, although I have no way to be certain. At one point the name of a monarch came up, and I asked for a transcription into Roman letters of the passage. The result verified the presence of the proper name. That’s encouraging.
Gemini was often reluctant to translate the text, claiming “I’m just a poor LLM, I only know a subset of languages.” I responded, “Try again” and invariably it then decided that it did know Ge`ez after all. What this means, stripped of the fake “AI” stuff, is that the search simply failed first time, and worked the second time.
This seems to be what AI is. It’s just a search engine. Nothing more. The talky bit on the front is just a chat engine, such as banks use on websites to demoralise customers who need help and make them give up. The LLM is a big flat file containing the database that it is searching. That database is composed of a vast amount of data, including a large quantity of pirated books. The whole “AI” stuff is just nonsense, to camouflage that it’s engaged in massive copyright violation. The search brings back stuff from pirated grammars etc, and assembles them into an English narrative. As with every search engine, sometimes it guesses wrong, or gets unreliable data. But “AI” has to choose something in order to pretend to be human. So you get bogus stuff sometimes. It would be interesting to know how well AI works in other languages, like French, German, etc, where there may be much less pirated data in the database.
Back to Abba Garima. I realise that Rossini appended notes to his edition, explaining various points in the text. The philological notes are of no use to me. But I have now decided to OCR his Italian and see what else he says about the text. I think it may be of interest. We’ll see.
Roger, I recently translated an 800-page somewhat technical history book from Italian to English with the help of DeepL translator (I know Italian well). My experience was that on average about one in three sentences was good, even sometimes elegant. A second of three needed light correction or rewriting, especially for word order or mistranslation of technical concepts. The third of three was seriously wrong and had to be translated from scratch. I considered this a good result and it certainly streamlined the task. Of course, it’s still far from perfect and the text still had to be checked sentence by sentence.
That is very interesting, from someone who knows the language. It feels right. These things are not translators, but database search tools plus an English language text generator.