[Vesuvius Challenge]
According to a Wall Street Journal article published on Feb. 9, 2024, the contest was launched by investors and scholars hoping to spur breakthroughs in reading these carbonized scrolls. It ultimately drew more than 3,000 competitors, including college students, software engineers, and AI researchers—all aiming to unearth texts lost for nearly 2,000 years.
Luke Farritor from the University of Nebraska– Lincoln video
Two years after virtually unrolling a scorched Herculaneum scroll along with Youssef Nader and Julian Schilliger, 23-year-old software whiz and Thiel Fellow Luke Farritor is once again in the spotlight—this time for his involvement at DOGE, reportedly receiving access to the Department of Energy’s IT system and listed as an engineer in the DOE’s staff directory.
Applying modern AI to ancient scrolls
Yet despite the intrigue surrounding his new government role, Farritor had a clear interest in applying computer science skills (he was a senior at the University of Nebraska–Lincoln) to unravelling ancient history. “I’ve always been interested in history,” he said in a video interview published by the University of Nebraska–Lincoln from 2024. “Growing up, I learned Latin. I was never that good at it, but I was always kind of into that sort of stuff—read a lot about the Roman Empire and things.”
After catching wind of the “Vesuvius Challenge” from a podcast appearance by tech entrepreneur and former GitHub CEO Nat Friedman in March, Farritor discovered that archaeologists at the University of Kentucky had released high-resolution CT scans of scrolls carbonized by the eruption of Mount Vesuvius in 79 AD. “They uploaded the data on the internet as a competition to see who could find writing inside,” Farritor recalled. “When I heard that, I was just immediately like, ‘Oh my goodness, this is such a cool project. I have to work on this.’” Within days, he was poring over gigabytes of 3D scans, trying to train AI models to spot even the faintest traces of ink. The ink detection code used in the competition is now archived on GitHub.
How text from the Herculaneum Papyri sample was deciphered
A sample of the Herculaneum papyri from the team that won the Vesuvius Challenge in 2023.
First, the scrolls were scanned using X-ray tomography at a particle accelerator (Diamond Light Source). This created a detailed 3D digital model of the rolled-up scroll, showing the internal structure without physically opening it. Next, software (primarily Volume Cartographer) helped trace the individual layers of the papyrus within the 3D scan. This process is called segmentation. Think of it like digitally peeling away each layer of the rolled-up poster, one at a time, and flattening them out. This was a combination of manual effort and automated algorithms. Once the surfaces were “flattened” digitally, machine learning models helped identify where ink was present. This was challenging because the ink and the papyrus were both carbonized, making them very similar in the X-ray scans. Two key breakthroughs enabled the ink detection: the discovery of a “crackle pattern” indicating letter shapes, and the use of domain adaptation to apply models trained on scroll fragments to the intact scrolls. These models identified tiny ink “spots,” and the aggregation creates the words and letters. Finally, papyrologists analyze the text. For more on the methodology, check out the Vesuvius Challenge Page.
Researchers at the University of Kentucky were able to use high-energy X-rays to render the scorched scrolls in 3D. [University of Kentucky]
In the aforementioned video interview on the University of Nebraska website, Farritor credits the University of Nebraska for encouraging him to think boldly—something he says “played a pivotal role in my work on this project.”
An output from the autosegmentation on scrollprize.org
