We're all jumpy when we see "AI" as a solution these days, but when this article says "AI" they don't mean ChatGPT.
"AI" is a shorthand for a lot of machine learning applications. The thing these are really useful for is finding connections in swamps of data - which is exactly what it's doing here.
A key innovation came in the middle of last year, when US entrepreneur and former physicist Casey Handmer noticed a faint texture in the scans, similar to cracked mud — he called it “crackle” — that seemed to form the shapes of Greek letters. Luke Farritor, an undergraduate studying computer science at the University of Nebraska-Lincoln, used the crackle to train a machine-learning algorithm, revealing the word porphyras, ‘purple’, which won him the prize for unveiling the first letters in late October. An Egyptian PhD student in Berlin, Youssef Nader, who followed with even clearer images of the text, came second.
Basically, they're using computers to scan for very, very faint irregularities, and then map those irregularities. Machine learning is good at that, because it's good at generating patterns from noise. It's then up to the humans to look to see if the patterns that the machine learning algorithm is finding actually show anything meaningful.
[...] But in the final week, the competition received 18 submissions. A technical jury checked entrants’ code, then passed 12 submissions to a committee of papyrologists who transcribed the text and assessed each entry for legibility. Only one fully met the prize criteria: a team formed by Farritor and Nader, along with Julian Schilliger, a Swiss robotics student at the Swiss Federal Institute of Technology Zurich.
They had experienced coders checking the code, AND experienced papyrologists checking the output to see if it created anything meaningful.
ChatGPT and image generators may be a solution looking for a problem, but there are a lot of good uses that large machine learning models can be put to, from monitoring species numbers on trailcams to detecting cancer cells. Machine learning algorithms - for which the trendy name right now is 'AI' - are really, really good for crunching data with a looooot of variables and variability, and finding patterns. Sometimes those patterns are fake, but sometimes they're real, and humans need to actually look to see if the patterns are real, and they had humans doing that for these image scans.
And to be clear, this AI is finding miniscule variation in the texture in the papyrus scrolls, and pulling that out. It's not spitting out Latin or Greek text - I highly doubt even modern ChatGPT could handle classical Latin too well, let alone handwritten classical Greek. The interpretation and translation is being done by human classicists. The computer is pulling images out of the scroll based on variations in texture. It's implied that its specifically looking for texture patterns that match Greek letter forms, but I can't quite tell by the reporting.
From a different Nature News article on the topic:
Seales realized that even with no difference in brightness, CT scans might capture tiny differences in texture that can distinguish areas of papyrus coated with ink. To prove it, he trained an artificial neural network to read letters in X-ray images of opened Herculaneum fragments.
Which is amazing for it to do, but it doesn't "know" what it's seeing and can't fabricate believable strings of Latin or Greek text. If it was making up letterforms because it had been trained to see them, it would come up with gibberish. That's why the code, and the end result, is picked over so carefully - to make sure it's seeing something that's really there! But it's not generating text itself.
discoholicmusic
gracieabrams
romaincrisis
origamiaround
snowghoul
