the project to decipher the library of crushed and carbonized papyrus scrolls from Herculaneum
Progress has been quite amazing here.
For reference, what one of these scrolls looks like,
it's a tightly packed, compressed charred mess and attempting to unroll it mechanically however carefully ...
... leaves behind crumbling flakes and not much more.
But somewhere inside there are remains of text.
And there are hundreds of these, and perhaps thousands more still buried.
Since autumn, the 'Vesuvius challenge' have gone from this ... a single word deciphered out of a rolled up crushed carbonized scroll
to entire paragraphs that while not perfectly sharp, are decipherable
This makes it plausible that in fact the library of scrolls will be accessible.
There's a compact write-up from the project here
https://scrollprize.org/grandprize
and a more journalistic treatment
https://www.bloomberg.com/features/2024-ai-unlock-ancient-world-secrets/
This I'd say is one area where machine learning aka AI can deliver spectacular results, and an important thing to consider is, there's no language model involved.
That's important because there is the obvious caveat, if humans can't really see the writing, and can't look inside the scroll, how do we know the AI isn't just making something up that seems plausible?
The algorithms, of which there are competing ones that are all open sourced, are basically looking for the presence or not of ink traces, or at best individual characters, but don't know anything about Greek vocabulary, grammar, the topics of Epicurean philosophy etc.
The entire approach is much more related to medical imagery analysis than ChatGPT et al. and that approach is how the entire Herculaneum scroll decoding effort started (CT scans, segmentation, this is all familiar from medical 3D work as it has existed for decades). And to answer the chicken-egg question ... yes there are some flakes of scrolls where the human eye could recognize letters, or patterns of ink residue, and that's what enabled the first training iterations.
And anyways the results the machine-learning nerds come up with get piped right to teams of established classicists who can contextualize.
If they were trying to decipher unknown writing systems or languages things indeed would be more difficult to verify...