Physicist and founder of Terraform Industries Casey Handmer reveals how his efforts to solve The Vesuvius Challenge led to a major breakthrough in the decoding efforts.
Timestamps:
0:00: The library of Herculaneum
1:17: The Vesuvius Challenge
2:30: A unique approach
3:49: Deciphering ‘crackle’
5:01: Solving an ancient puzzle
Transcript:
The library of Herculaneum
In AD 79, Mount Vesuvius erupted in Italy, and that eruption destroyed multiple ancient Roman towns, of which the most famous is Pompeii, and a close second is probably Herculaneum. Herculean was a beach side resort town. The villa next to Herculean was owned by Calpurnius Piso, who was the father-in-law of Julius Caesar. It was full of this ancient library. We know the ancient Romans wrote a lot. They didn't have the printing press, but they had industrialized scribes systems set up. They produced lots and lots of books, and less than 1% of those books have survived to the present day.
Herculaneum is the only ancient library from antiquity that has survived to the present day, but when we say survived, it is not in good shape. These scrolls were buried by superheated mud that crushed them and carbonized them and buried them, and then 1,700 years later, they were dug up and mistreated, and in many cases we've just got these little fragments left, but some hundreds of them are still in the unrolled state. They've been left that way for posterity because philologists knew in the future, humanity would develop technology that might make it possible, and we have. Maybe the AIs can help us.
The Vesuvius Challenge
What it really comes down to is how much time do you spend thinking about the Roman Empire? When I was in grad school, I knew about people working on this problem. I studied Latin for seven years, so I spent quite a lot of time thinking about the Roman Empire. There's a team led by Brent Seales, at the University of Kentucky, which has been working on this problem for 20 years. Nat Friedman, who's a friend of mine, became aware of this problem during COVID.
They basically decided that they were gonna run a competition to try and inspire people to come out and solve this problem using software. The scrolls were flown by private jet from Naples, in Italy, to the UK, scanned at eight micron resolution. So we have these terabyte sized files of data, basically like a CT scan, like a medical scan of these scrolls, and you can look at them and you're like, "Okay, that's a real mess, I can't see any text."
The competition kind of split in two directions. One half was, can we use software to virtually unroll these scrolls, so that instead of just being this crunched up mess, like a roll of toilet paper that's been set on fire, it can be unrolled and we can actually see the surface? And then the second half was to try and detect ink on that surface. I said, "Look, I'll just take a crack at it Casey style."
A unique approach
I like to think that I have a fairly unique approach to solving problems, and sometimes it's just doing the really obvious thing, and the really obvious thing that no one had done apparently with this dataset was just stare at it until they understood what they were looking at. It's the same thing if you're learning to read a medical X-ray, you stare at, you're like, "The white stuff is bones, I guess, but what else am I looking at?"
If you're a trained technician, you can learn to recognize cancers and growths, and foreign bodies and all kinds of stuff that the amateur could not see. I stared at this most evenings for about a week, probably in total 20 or 30 hours, getting a feeling for what sort of textures I was seeing. You can see the texture of the papyrus, which is kind of this crosshatch texture, and it has bright spots in it because the papyrus plant actually has little mineral concretions that it precipitates. And then there were all kinds of artifacts caused by the scanning process and the unrolling process that had kind of red herrings, cracks, crumples, all kinds of mess.
But there was one particular texture that I kept seeing that I couldn't explain away, and it looked like cracked mud. And I was like, "That's interesting. Why don't I kind of scroll around in this segment," which on my computer was about, you know, about a few feet across, but in reality is about the size of postage stamp. Eventually I found an area of this scroll, it had a shape to it that did not appear to be possible by random chance.
Deciphering ‘crackle’
I rotated the image and then I very quickly began to recognize ancient Greek letters, which of course I know from mathematics and physics practice. I saw a pi, a lunate sigma, an iota, a couple of lambdas, a delta. I probably saw 15 letters or 13 letters of which I identified 7 or 8 correctly. I kind of let the Slack group know, "Oh, I've seen something here." We call it crackle, it's like pentagons and hexagons and squares, lighter shade material with dark spaces in between. And in some places it seemed to be kind of curling up at the edges, like cracked mud does in a lake bed. They were justifiably skeptical because plenty of lunatics get involved in this challenge and send in all kinds of stuff, and I was just one lunatic amongst many. I showed that it was solvable, and then other people went and did a lot more work and actually solved it.
Fortunately, Luke Farritor took me seriously, because he was working at SpaceX and I've done a little bit of writing about SpaceX on my blog, and so he kind of saw me as at least moderately credible. He took a look at it and we collaborated for a bit and he very quickly trained himself to see the same thing and then made some fabulous masks, and then he was able to train his AI to see this.
And then in the meantime, another contestant, Youssef Nader, who's based in Germany, pursuing a different technique, but basically automating the computer to do what I'd done by hand, which was just stare at it until you recognize like four or five or six different kinds of texture. His computer system started to be able to recognize the crackle texture as a unique feature as well. Then Luke and Youssef got together with Julian, who was doing the unrolling process.
Solving an ancient puzzle
In about a month and a half or so, we're able to really push this a long, long way. Three students, including a 21-year-old American found the best solution and this is the result. They're able to make clearly visible and easily readable about 2,000 letters of ancient text. ChatGPT can read them in some cases, which is pretty crazy. The papyrologists and the people who can actually read ancient Greek got involved and they read it, produced translations, and then we had the announcement and it was all pretty exciting.
15 columns of text in ancient Greek seen for the first time in almost 2,000 years. The competition is still going, and Elon Musk put the money in this year, so if you think that Elon could do with $2 million less, you should get involved. Once we can read a scroll for like $10,000 instead of a million dollars, which is currently kind of the going rate in terms of labor time required, then I think it makes sense to scan all the extent scrolls and then restart excavations, extract, conserve, scan, and read all the other scrolls, and then go on a hunt for other libraries that were destroyed and preserved by the eruption of Mount Vesuvius, of which we know at least two others. We don't know where they are, but we know they're somewhere under the mountain, so we should go and get those too. And the crazy thing is, all the surviving ancient text from ancient Rome, including translation, you can fit on a modest sized bookcase. In the 1700s, 1800s, there were philologists who, like me, like you, were desperate, just desperate, for even a word, even a sentence, even a book of new work from the ancient world. If we are able to decode and read all the scrolls that we've already excavated, it will double the size of the ancient corpus. That's insane. 40 generations of ancient historians have waited for this, and our generation gets to do this.
Share this post