A deadly eruption of the famed Mount Vesuvius 2,000 years ago destroyed countless ancient artifacts. But more than 1,800 papyrus scrolls survived the inferno, discovered in the 18th century in the village of Herculaneum—situated in what is Campania, Italy today.

Archaeologists believe the scrolls contain stories and information long thought lost forever, and scientists have been trying to find a way to access the writing without being able to unfurl them. Previous attempts left the scrolls disintegrated into dust, or caused the writing to immediately vanish when exposed to the elements. Thus, the Vesuvius Challenge was born—a competition to access the writings rolled up in ancient paper.

This year, the winners of the 2023 challenge were a group of researchers that used artificial intelligence, nondestructive CT scans, and other specialized techniques. The team estimates they were able to scan and translate about 5 percent of one scroll.

"Ten months ago, we launched the Vesuvius Challenge to solve the ancient problem of the Herculaneum Papyri, a library of scrolls that were flash-fried by the eruption of Mount Vesuvius in 79 AD,” said Vesuvius Challenge cofounder Nat Friedman, former CEO of Github. “Today we are overjoyed to announce that our crazy project has succeeded. After 2.000 years, we can finally read the scrolls.”


Due to their delicate carbonized state, the scrolls represented a formidable challenge to scholars. Unveiling their contents required the convergence of expertise in CT scan technology, computer vision, and machine learning, initiated by Friedman, Daniel Gross, and Brent Seales. Gross is a tech investor who previously served as a partner at Y-Combinator and led artificial intelligence efforts at Apple, and Seals is a University of Kentucky science chair and data science professor.

Vesuvius Challenge
Reconstructed scrolls. Image: Vesuvius Challenge

The winning team was led by Luke Farritor, a 21-year-old University of Nebraska-Lincoln student and SpaceX intern, who developed a machine learning algorithm that managed to discern subtle textual nuances within the sealed scrolls—a feat that had previously eluded scholars. Farritor's neural network meticulously differentiated between inked and blank areas of the papyrus, bringing to light over ten Greek characters, including "porphyras," meaning "purple."

This achievement won Farritor the "first letters" prize in the Vesuvius Challenge contest, setting the stage for a more comprehensive deciphering of the scrolls.

Two other scholars shared the prize with Farritor: Youssef Nader, a PhD student, and Julian Schilliger, who developed a segmentation technique that made it possible to map the papyrus.


Together, they used three distinct model architectures to leverage AI while also avoiding the common problem of “overfitting”—reproducing the same input over and over again—and hallucination. Their rigorous approach allowed scholars to access thoughts and wisdom preserved for two millennia.

The trio got $700,000 for their achievement.

The scroll, now partially deciphered, is a glimpse into ancient philosophical discourse, specifically reflecting the Epicurean belief in pleasure as the highest good. The scroll delves into the intricate relationship between the availability of goods, such as food, and the pleasure they impart. The ancient author refutes the notion that scarcity enhances pleasure. In an excerpt from the text, the author asserts, “As too in the case of food, we do not right away believe things that are scarce to be absolutely more pleasant than those which are abundant.”

Pure Greek philosophy brought forward two millennia through the power of AI.

The whole operation was also open sourced, so the work of Nader, Farritor, and Schilliger in the Vesuvius Challenge could mark a new era in the preservation and interpretation of ancient texts.

Next up? Scanning and deciphering up to 90% of the scrolls —and whoever does it will win $100,000.

Edited by Ryan Ozawa.

Generally Intelligent Newsletter

A weekly AI journey narrated by Gen, a generative AI model.