Interesting from Anton Repushko:
“A lengthy text about how I’ve been going insane from despair over the last 8 months and started learning Ancient Greek.
Posts about incomprehensible white letters and jokes about lazy ancient Greek scribes have already flickered above. It’s time to put everything together into a coherent picture and explain myself. I can’t publish and explain the technical solution in detail due to the rules (it’s under review), but I’ll share various other things.
The competition is called the Vesuvius Challenge. When Vesuvius erupted and destroyed Pompeii, it also destroyed the city of Herculaneum at its base. We are interested in Herculaneum because there they excavated a villa that had a very rich library. And in this library, they have already found/expect to find about 4-5 thousand scrolls. Important clarification: from Antiquity, very few texts have survived to us: the Odyssey, the Iliad, all philosophy – just a small percentage of all texts that existed back then. Imagine how much knowledge and history are stored in these thousands of scrolls. But of course, there is a problem: they look like potatoes charred on coals due to conservation in lava and ash (but thanks to this they lay in the ground for 2000 years).
Enter Professor from Kentucky, Dr. Brent Seales, who spent his entire career pushing the following method: let’s make 3D CT scans of these scrolls and then try to cleverly restore and read them. The plan is reliable, like a Swiss watch, if it weren’t for one thing: the inks there are carbon-based and are invisible on MRI results. There were also attempts to scan the same scroll at different beam phases and by subtracting the results of this get some letters. The letters were enough for an article in Nature, but it looks pretty awful. What to do with this next?
Enter ex-CEO of Github, Nat Friedman: a philanthropist, organizes various AI grants, and invests in AI startups. A very nice fellow who helped organize this competition and stuffed the prize fund with money. Here’s his private fund and so far, to me, he seems like a healthy-person’s Elon Musk.
The competition itself started last February with a prize fund of 1M USD. It began on Kaggle, and then it all continued on Discord. There are 3D CT scans of the scrolls and the text needs to be read from there. The process of virtual unfolding of text proposed by the organizers is quite complicated and tricky, and now there’s about 7-8Tb of data, which sets a pretty high technical entry barrier.
After the Kaggle competition, things quieted down until in August, two students (one from the states, the other from Berlin) independently found some unclear reliefs with their eyes (!) that confirmed the initial idea: despite the ink being invisible, the paper remained deformed at the place of the written letters. Thus, the task became trying to programmatically learn to identify these deformed places (sometimes invisible to the eye) and read the letters. Around early August, that’s when I dived head-first into the competition.
So many sleepless nights, worries, and everything else throughout these months. I came and pestered anyone who could help me with advice/ideas (thank you all, friends. I couldn’t have done it without you). When discussing news and what happened during the day with my girlfriend, I thought dozens of times that nothing at all happened to me: I was solving the scrolls, just like yesterday, and will continue tomorrow.
And despite not meeting the official criteria for the main prize (four texts with 140 readable characters each), I think no one did. According to the organizers, they received about a dozen submissions. And for about a month now, we and the other participants have been waiting for the results of the papyrologists’ analysis. During these six months, I’ve met a lot of interesting people, won an intermediate prize of 5200 USD, started learning ancient Greek and greatly improved in solving such unclear problems.
Interestingly, during the competition, I felt incredibly fulfilled and it’s a very good alternative to all the races with LLM-research: you have a complex unclear problem, which intuitively seems solvable with current means. And the whole history of these scrolls (which were passed from hand to hand for decades until they could be read with new technical means) is a great example of human cooperation for something that can’t be measured in money. I really like it.
The organizers have already announced a new stage of the competition sometime in February, so there will be even more posts about it. Below is a picture with roughly the final results achieved: individual words and prepositions are readable, but without knowing Ancient Greek, it’s hard to read more. Let’s see what happens next.”
Hi, I'm Rauf Aliev. 