Inside the AI-Powered Race to Decode Ancient Roman Scrolls

  • last year
When Mount Vesuvius erupted in A.D. 79, it buried Herculaneum under 65 feet of hot mud and gas. Now, AI tools are making it possible to read some of the carbonized scrolls that remain behind.
Transcript
00:00 I was thinking of the Roman Empire, as one does occasionally.
00:03 I found a book called "24 Hours in Ancient Rome,"
00:06 and I started reading it.
00:08 Turned out the book, I kind of realized halfway through,
00:10 it was actually written for eighth graders.
00:12 But it was a great book.
00:14 I highly recommend it.
00:15 And it kind of just threw me down
00:16 this ancient Rome rabbit hole.
00:18 And eventually I stumbled upon the story of this villa
00:22 that had been covered by the Vesuvius eruption in 79 AD.
00:27 It had been the home to an enormous library
00:30 of papyrus scrolls.
00:32 And that in fact, this was the only library
00:34 of papyrus scrolls that in some way
00:37 had survived since antiquity.
00:39 Knapfriedman is one of many people who are obsessed
00:42 with a collection of carbonized papyrus scrolls,
00:45 discovered inside the ruins
00:46 of a luxurious Herculaneum villa in 1752.
00:50 Since then, there have been many attempts
00:52 to unravel and read the scrolls,
00:54 which unfortunately ended up destroying or damaging them.
00:57 Now, artificial intelligence tools provide a new opportunity
01:00 to decode the content of the scrolls.
01:02 I'm a computer scientist,
01:12 not a manuscript scholar or an historian,
01:15 but I did have this engineering idea
01:17 that we could do digital restoration
01:20 of pages that are not flat.
01:21 We could make them flat.
01:23 And ultimately of things that are wrapped up
01:25 that we could totally unwrap.
01:26 Frank Seals has spent much of his career
01:28 thinking about how to read ancient scrolls.
01:30 In 2002, that's when we had our first version
01:33 of the software that did virtual unwrapping.
01:35 And we had examples we had created in the lab
01:38 that showed the concept from start to finish.
01:40 And I presented that at the Society of American Archivists
01:43 and I actually received from the audience,
01:46 audible gasps when we showed the result.
01:48 And that was when I knew this is powerful.
01:51 In the years that followed,
01:52 Seals was able to read ancient texts
01:54 like the Book of Ecclesiastes
01:56 and the En-Gedi scroll.
01:57 He learned about the Herculaneum scrolls in 2004
02:02 and immediately began trying to virtually unwrap them.
02:04 In early 2023, Seals and his team had a breakthrough
02:09 when one of his PhD students
02:11 confirmed that machine learning
02:12 could detect ink from the scrolls.
02:14 I found an article about this computer science professor
02:20 at the University of Kentucky, Brent Seals,
02:22 who had this wild plan.
02:25 Tech investor, Nat Friedman,
02:26 had become interested in Seals' work in 2020
02:29 when he fell down an ancient Rome rabbit hole
02:31 during COVID lockdown.
02:33 This is one of the coolest projects
02:34 that I've ever heard about
02:35 and I want to follow it and see what happens.
02:38 It reached out to Brent.
02:39 What if we open up this problem to the world
02:42 and we engage all the bright minds
02:45 who I think should be as interested in this as you are
02:47 and as I feel like I am now.
02:49 We raised over a million dollars, a million two,
02:52 and suddenly we had a $1.4 million archeology prize.
02:57 On March 15th, 2023, the Vesuvius Challenge launched.
03:01 The grand prize to be awarded to the first person
03:04 to decode the scrolls is $700,000.
03:08 The organizers have since awarded 36 intermediate prizes
03:11 for milestones along the way.
03:13 And so we put up this first letters prize
03:16 and the idea was to find 10 identifiable letters
03:19 in a four square centimeter region of the scroll.
03:23 And when we put it up,
03:25 we honestly didn't know if it was possible.
03:27 I was doing an internship at SpaceX
03:31 and I was driving to work in South Texas
03:34 and I listened to a Nat Friedman interview.
03:37 Nat just kind of explains
03:39 the whole Herculaneum Scrolls thing.
03:41 He explains the challenge and at the end he says,
03:44 "If you're listening, there's a chance you could win."
03:46 And I was like, "Holy cow.
03:48 I have to give this a go."
03:50 When I started in March,
03:51 I just wanted to get a good handle
03:53 on kind of how the challenge works,
03:56 what the data looks like
03:57 and just develop good intuitions there.
03:59 There was a rare occurrence where I attended a party
04:02 and I was just kind of sitting in the corner
04:04 and I get a text from one of the people
04:07 on the segmentation team, like,
04:08 "Hey, just uploaded this new piece of flattened papyrus.
04:12 Access my computer from my cell phone
04:13 and kind of set it to run on this new piece."
04:16 And then I just put my phone away,
04:18 don't think about it again.
04:20 And as I'm walking out of the parking garage
04:21 to return to my college dorm room,
04:23 I just pull out my phone again.
04:24 Like, you know, I wonder how that, you know,
04:27 wonder how that new segment's doing.
04:29 And I look at my phone
04:30 and there were like three Greek letters there.
04:32 It was amazing.
04:34 I completely freaked out.
04:36 - Luke Ferreter, who is winning the first letters prize
04:41 on behalf of the Vesuvius Challenge.
04:43 - Indeed, thank you.
04:45 - Those letters he managed to uncover
04:47 show a word translated to mean purple.
04:49 - I just want to read the scrolls.
04:51 We all just want to read the scrolls
04:52 as a common catchphrase in the community.
04:55 - Scholars estimate there may be over 30 volumes of text
04:58 in the unopened Herculaneum Scrolls,
05:00 covering a wide range of authors and genres.
05:03 - It's pretty intoxicating to sit down
05:06 and see writing from someone 2,000 years ago
05:11 that we can just read.
05:12 I mean, that is just a fantastic feeling.
05:15 (upbeat music)
05:18 (upbeat music)
05:21 (upbeat music)
05:23 (upbeat music)
05:26 [BLANK_AUDIO]

Recommended