I insert a query like this:
translate into English: ὑφ᾽ ὧν ἐκεῖνοι παρὰ πᾶσαν ἐπιβουλευόμενοι τὴν ὁδὸν
With both google ai and Chatgpt my query starts a conversation. If ai asks for the source, I just type in, for example, “Isocrates 4.148”. Both ai-sites can work with that kind of reference. ai can also handle short questions, for example “what about ὑφ᾽ ὧν?”, and "isn’t this an allusion to Xenophon’s Anabasis?
I ask whatever I’m wondering about. I ask questions like this: “When you parse, are you consulting commentaries, or can you actually read Greek?”
But what about ai making mistakes? I ask ai for references to Smyth and to dictionaries. I ask, “How do we know that the antecedent of X is Y?” So even though ai parsings may elude my understanding at first, I can compare them with standard references.
2 Likes
90% of the time, AI will just remind you of rules that you’ve forgotten. If it cites a new rule, then ask where it got that rule from. About 6 months ago, it had trouble giving the exact numbers for reference but now it’s doing much better, at least Gemini is, which is the better AI. I learned in 1 month as much about Ancient Greek morphology using AI than I would 4 years using traditional methods, κατα την εμην γνωμην
hello kylefoley, that’s interesting about using AI for morphology.
Liberation Philology Ancient Greek: this app lets one use wait time in offices, etc., for practicing forms.
I saw your earlier discussion on this. They don’t actually have a set of grammatical rules they consult, and they don’t work by searching a giant database of translations and commentaries.
What happens is that you feed an artificial neural network enormous amounts of text, and during training it learns statistical patterns in that text.
So if you give it a particular phrase, it may recognise that the phrase is strongly associated with a particular author and produce the author’s name. Or if you give it the beginning of a sentence, it predicts what is most likely to come next, then what is most likely to come after that, and so on.
It is not fundamentally different in principle from the old auto complete systems on mobile phones. If you typed “hel”, the phone might offer “hello” or “help”. The more surrounding context you gave it, the better the prediction became. “Daddy’s home. Go and say hel…” is much more likely to end in “hello” than “help”.
A large language model is essentially that principle taken to an absurd level in terms of the data it is dealing with. It doesn’t have a dictionary, and it doesn’t have any rules. It has learned extremely intricate statistical relationships from a vast number of examples, and collecting vast numbers of examples of what goes with what..
They are not programmed with code to follow rules based on data, it’s like if you throw enough regular sequences of words at it as in millions and millions of books and the entire internet, it’s going to quite reliably be able to map the string of words that you’ve said to it to the kind of reply you are likely expecting. And they are fed grammar books as well, as well as dictionaries, by the thousands, and they are fed them several times over.
Very interesting description, James, with which I find no fault. Here is a recent snip from google ai:
“The double construction of ἀξιόω (taking both a genitive of the thing and an accusative of the person) is particularly notorious for throwing readers off balance when the words are separated like they are here. You should be proud that your trial analysis got you to the right meaning!”
I had given google ai a query on this clause from Lysias 12.
τοιούτων ἠξίωσαν οὐχ ὁμοίως μετοικοῦντας ὥσπερ αὐτοὶ ἐπολιτεύοντο.
The commentary by google ai seems apt, and it shows that however it learns language, for users it can employ the categories of first-year Greek textbooks.
1 Like
fascinating it is the same in Russian.
Они сочли нас достойными подобного, хотя мы, как метеки, вели себя не так, как они вели себя в качестве граждан
нас :accusative.
подобного: genitive.
I found out I could read Greek by accident after studying Russian, Cyrillic is based on Greek
And the grammar is similar.
Russian perfective and imperfective map onto the Greek aorist and present/imperfect
respectively.
(ὁμοίως … ὥσπερ also maps cleanly onto не так, как)
(My Russian is dreadful, and my Greek is worse)
Here is Gemini chipping in, just to bring us full circle
That parallel is striking. The exact alignment between Greek ἀξιόω and Russian счесть (достойным) is a perfect example of how Proto-Indo-European (PIE) case syntax and valuation semantics persist across branches.
Syntax Comparison
Why this happens
-
The Genitive of Value/Price (Genitivus Pretii): In both Greek and Slavic (and Latin: aestimare + genitive), verbs of valuing, assessing, or deeming worthy inherently require a genitive to mark the degree of worth or price. The genitive answers “of what amount/quality?”
-
Adjectival Inheritance: Even though Russian inserts the adjective достойный (governed in the predicate instrumental by сочли), достойный itself inherits that exact same genitive requirement (достойный чего? $\rightarrow$ подобного). Ancient Greek collapses this into the verb verb directly: ἀξιόω built from ἄξιος (“worthy of [genitive]”).
It shows how deeply rooted the genitive of specification/value is when expressing judgment or worth across Indo-European languages.