Using Deepseek on an obscure Greek “Life” of St Isidore of Pelusium (d. 435 AD) by Morton Smith

Yesterday I started googling about Isidore of Pelusium, and I quickly came across a number of papers showing that Dr. Madaline Toca is actively working on Isidore of Pelusium, the manuscripts of his letters, the reception of his work in Latin, and so on.  This is good news!  Most of these papers are accessible on Academia here, which is even better news.  Also among her efforts is an online bibliography for Isidore, here.

This bibliography informed me that a previously unpublished Greek “Life” had been printed back in 1958 in an obscure Greek volume.  Thankfully she provided a PDF here.

The author of this publication was none other than a certain Morton Smith.  Today Morton Smith is notorious for his “Secret Mark” forgery.  But in 1957 he was just another a young scholar, travelling through Greece and the Levant, and searching for manuscripts of the letters of Isidore of Pelusium.

The “Life” printed is basically a transcription of four manuscripts, the oldest 12th century, the rest 16th century; three from Mount Athos, and the other in the National Library in Athens.  From the incipit and explicit, it appears to be the text listed in the BHG, vol. 3 (supplements) on p.38 as BHG 2209.  The text is in 12 paragraphs, which are those in the manuscripts.  No paragraph numbering or subtitles appear.

I was curious what the text said.  Unfortunately Smith gave no translation.  It does not seem to be in the TLG.  But I knew that if I could create an electronic text, then I could work on it with ChatGPT.

What I did was to run it through Abbyy Finereader 15 Pro.  Unfortunately this only supports modern Greek, which means that the accentuation gets mangled.  But never mind that.  Who cares?  This is just idle interest.  Or so I thought.

Anyway I pulled it, did a certain amount of light correction.  Then I pasted it into ChatGPT using the free online web interface, and asked for an English translation.  The request that I gave it – the “prompt” in the AI jargon – was this:

Please correct the following text in medieval Greek: ”’  …. ”’

with the Greek between the triple quotes.  Triple quotes seem to be the standard way to indicate something in these AI interfaces.  I have no idea why.

Well, ChatGPT produced a very nice, perfectly readable translation.  Here is the opening section:

The Life and Conduct of Our Holy Father Isidore of Pelusium

This divine Isidore, who lived a wondrous life both in deed and word, shone like the sun in both respects, and even brighter. His homeland was Pelusium, a distinguished city in Egypt. He sprang from a pious lineage, yet he himself flourished as an even more devout offshoot. From his earliest years, he ardently embraced virtue and steadfastly held to it as an inseparable possession. He despised the pleasures of youth, such as childish games, frivolous laughter, indulgence that weakens the soul, and enslavement to the desires of the flesh—everything, in short, that distracts the mind from perceiving true beauty. Instead, he pursued complete detachment and the heights of divine contemplation, embracing temperance in all things and the highest form of education—one that elevates the soul and leads it toward the heavens.

My experiences with ChatGPT as a helper with the translation of the “Life” of John Damascene had demonstrated that ChatGPT does give reasonable accuracy.  But also that you do have to check!  But how to do so?

My next thought was to try out another AI product, the Chinese DeepSeek, also available for free online.  I did the same with this, with the same prompt, and it also produced a translation. Here’s the opening section:

The Life and Conduct of Our Holy Father Isidore of Pelusium

This divine Isidore, who was wondrous in both life and word, shone brighter than the sun in both respects. His homeland was Pelusium, a renowned city in Egypt. From a devout root, a more devout branch sprouted, for from his earliest youth, he wholeheartedly embraced virtue and held it as an inalienable possession, demonstrating his unwavering love for it. He shunned the pleasures that typically delight youth—childish games, youthful frivolities, and indulgences that weaken the soul and enslave it to the desires of the flesh. In short, he avoided everything that hinders the mind from attaining true understanding of what is good. Instead, he devoted himself entirely to what leads to perfect detachment and the highest contemplation of divine things: universal self-control and the best education, which lightens the soul and lifts it to heavenly things.

Various awkwardnesses in the DeepSeek output suggested to me that it was probably closer to the Greek.  But there are enough differences there that I felt that I needed to get closer to the Greek.

So I put in another prompt to DeepSeek:

For each sentence, give the Greek, then the English translation; then analyse the syntax, with meanings; then give a list (in table format) of all the words with the grammar, lemma, part of speech, case, number, gender, and meaning for each word.

Doing so produces quite a bit of output.  The “translation” gets quite a bit closer to the Greek as you do this, and of course you can read its explanation of the syntax, you can lookup individual words in Logeion or somewhere using the lemma, and generally wrestle with the sentence until you believe what you have.

But back to the AI output.

If you run this prompt in ChatGPT, it just does one sentence only, and then it asks if you want to continue.  If you say “yes”, then it does another.  However ChatGPT has no idea what a sentence is!  So it does a dozen or so words.  DeepSeek is the same, but I quickly found that the length of a “sentence” was much shorter.  Rather nervously I asked if it could “do the same but for two sentences” and I got longer outputs.  So that worked.  When I asked if it could do four “sentences”, it went a bit funny.  So I went back to two.

At one sentence a time, all this becomes very tedious.  Copying and pasting the output to a word document takes a lot of time.  Indeed I have spent the whole day doing this.  But DeepSeek did a fine job.  It was no worse than ChatGPT.

It’s generally best to do this, one paragraph at a time.  It doesn’t feel so oppressive, and you can go off for a break at the end of each paragraph. You don’t want to sicken yourself, and it takes too long to do the whole thing in one go, even for just 4 pages of Greek.

For the curious, I attach a sample file with what I got for one of the paragraphs.  How reliable the output is, well, I will find out in due course!  It’s here:

• 8 deepseek (Word .docx)

But I did get tired.  So I wondered if it was possible to do this process from the computer command line, thereby saving myself a lot of time.  You can indeed connect to the website using the “API,” which would allow you to write a program.  But… alas… they want money for that!  The restrictions on the free web interface are deliberate.

You can also download for free a DeepSeek “model” (jargon word) and run it on your PC.  But unless you have an awful lot of memory fitted, you will find yourself working with “distilled” versions which are not nearly as good.  The process is fairly technical, and although I got it to work, I’d need to spend a lot more time on this.  Whether my fairly powerful PC would handle a full-size model is something that I don’t yet know.  So I went back to the free web interface.

One place where DeepSeek is definitely superior to ChatGPT is that it recognises when it reaches the end of the passage.  ChatGPT does not.  It will quite happily continue beyond the end, “translating” random Greek garbage.  So every so often you have to take the last word translated, and check that it is still in the text!

Doing this led to an interesting discovery.  I always ask for the Greek, the English, and then the syntax analysis.  I found that DeepSeek was silently fixing up the garbage Greek text that I had got from the OCR.  It was adding the missing accentuation.

So I tried asking it explicitly to do so.

Please correct the following text in medieval Greek: ”’…”’

And it did, and then translated it.  A quick look at the original PDF suggests that it is doing a good job.  Well, well.