[numitus]

From my perspective it may be just fitting to the answer. We try to find symbols, without confidence, that the paper still contains information, and any text for trainig. With "rights" network you can achieve any possibile result. It is remember me a russian freak-scientist, which try to read words and texts on the detailed sun surface photos.

[lolc]

I wondered about that. My understanding is that the models were trained to look for letter shapes, not words. And that the models couldn't produce known words unless they were trained on the language. If it wasn't trained on a substantial text body, a model producing letter sequences that form known words means it found something and didn't hallucinate.

[dom96]

This was my first thought. How sure are we that their model isn't just over-fitting and finding letters that aren't really there?

[numitus]

I think they need to make similar scroll, write any text and make the similar damages to check is the approach worked at all