The Library That Came Back

01 — The BurialThe Only Library

On October 24, 79 AD, Mount Vesuvius erupted. A pyroclastic flow traveling at hundreds of kilometers per hour engulfed the Roman town of Herculaneum, burying it under twenty meters of volcanic ash and debris. The heat was so intense that organic material carbonized on contact. Wood became charcoal. Papyrus became black cylinders of compressed carbon, indistinguishable from lumps of coal.

Inside a seaside villa now known as the Villa dei Papiri, roughly 1,800 papyrus scrolls sat in shelves and cases. Some had been packed for transport — their owner may have been preparing to evacuate. The eruption was faster. The scrolls carbonized in place.

In the 1750s, workers tunneling beneath the hardened volcanic rock rediscovered the villa and extracted hundreds of scrolls. They were the only intact library surviving from the ancient Greco-Roman world. Every other library from antiquity — Alexandria, Pergamum, the great Roman collections — had been destroyed by fire, war, or decay and left nothing behind. Herculaneum's library survived because the same catastrophe that should have destroyed it also preserved it.

The problem was reading them. Early attempts to unroll the carbonized scrolls destroyed many. An eighteenth-century machine designed by Antonio Piaggio could unwind some partially, but the process took years per scroll and shattered fragile passages. By the twentieth century, scholars had extracted text from roughly 800 partially opened fragments, mostly works by the Epicurean philosopher Philodemus. But hundreds of scrolls remained sealed — too fragile to open, too carbonized to read by any known method. The text inside was intact. It was also invisible.

02 — The WaitTwenty Years of Scanning

Brent Seales, a computer scientist at the University of Kentucky, spent more than twenty years working on the problem. His premise was simple: if you cannot physically unroll a scroll without destroying it, perhaps you can virtually unroll it instead.

In 2009, Seales and his team traveled to the Institut de France in Paris — which holds several Herculaneum scrolls — and performed the first-ever micro-CT scans of completely sealed scrolls. The scans produced detailed 3D images of the scrolls' internal structure, revealing the tightly wound layers of papyrus. They could see the layers. They could not see the ink.

The carbon-based ink used by ancient scribes was nearly identical in density to the carbonized papyrus itself. In a CT scan, ink on carbonized papyrus was like black ink on black paper. The writing was physically present. It was radiologically invisible.

Seales' team developed software called the Volume Cartographer — a program that could map 2D surfaces within a 3D CT volume and computationally "flatten" them. In 2015, they proved the approach could work: they virtually unwrapped the En-Gedi Scroll, a badly charred Hebrew parchment from ancient Israel, revealing text from the Book of Leviticus. That scroll used metal-based ink, which showed up clearly on CT scans.

Herculaneum's carbon ink did not cooperate. The technology could unwrap the scrolls. It could not read them. That gap persisted for years.

03 — The ChallengeOpen Source, Open Data, Open Prize

In March 2023, Seales partnered with Nat Friedman, the former CEO of GitHub, and Daniel Gross, a Silicon Valley investor and AI entrepreneur. Together they launched the Vesuvius Challenge: a machine learning competition to read the Herculaneum scrolls. Seales released his team's CT scans and Volume Cartographer software to the public. The three organizers put up prize money. Donors followed — more than thirty individuals contributed between $5,000 and $450,000, bringing the total prize pool to over $1 million.

The structure was deliberate. Rather than funding a single lab or hiring a team, they created an open competition. Anyone in the world could download the data, train a model, and submit results. The scans were enormous — terabytes of micro-CT volumes. The problem was precise: detect the ink that CT scans could not distinguish from the papyrus.

The grand prize required participants to decipher at least 85 percent of four passages from a sealed scroll, each at least 140 characters long. Smaller prizes rewarded intermediate progress. The prize structure was designed to create a ratchet: each breakthrough — a technique, a model, a labeled dataset — would be shared publicly, so the next participant started from higher ground.

04 — The CrackleSeeing What CT Could Not

In early August 2023, five months after the competition launched, a contestant named Casey Handmer found the first crack in the problem. Handmer — an ex-JPL engineer and startup founder — had been staring at the segmented CT scans for hours. He noticed something no one had documented before: a faint, distinctive texture on the papyrus surface where ink had been applied. The carbon ink had left no density contrast detectable by the CT scanner. But it had left a physical trace — a subtle disruption in the surface structure of the papyrus fibers that he called a "crackle pattern."

The crackle was not ink in the traditional imaging sense. It was the physical deformation that ink had caused on the papyrus surface before carbonization. It was too faint for human eyes to reliably detect at scale, but it was a trainable signal. If you could label enough examples of crackle versus no-crackle, you could train a machine learning model to detect it automatically across the entire scroll.

Handmer won $10,000 for being the first person to find "substantial, convincing evidence of ink within the unopened scrolls." The Vesuvius Challenge organizers described him as the first person in 2,000 years to find ink inside a sealed scroll.

What happened next moved faster than anyone expected.

05 — The WordA 21-Year-Old and a Greek Word for Purple

Luke Farritor was a 21-year-old computer science student at the University of Nebraska-Lincoln and a SpaceX intern. He heard about the Vesuvius Challenge from a podcast and started spending evenings training a machine learning model on Handmer's crackle pattern. The model learned to distinguish papyrus fibers that had been deformed by ink from those that had not.

In August 2023, Farritor became the first person in nearly two millennia to read an entire word from inside a sealed Herculaneum scroll. The word was porphyras — ancient Greek for "purple." He won the $40,000 First Letters Prize. Youssef Nader, an Egyptian biorobotics graduate student in Berlin, placed second in the same competition, having independently found the same word using a different modeling approach.

One word does not make a library. But it proved something fundamental: the ink was detectable by machine learning even when it was invisible to CT scanning and to the human eye. The barrier that had held for 275 years — the radiological invisibility of carbon ink on carbonized papyrus — had been bypassed not by a better scanner, but by a better algorithm.

06 — The Prize2,000 Characters from a Sealed Scroll

The three contestants who had made the individual breakthroughs joined forces. Youssef Nader led the team, with Farritor and Julian Schilliger, a robotics student from Zurich who had won three earlier progress prizes for his work improving the Volume Cartographer's segmentation pipeline. Together, they trained an ensemble of AI models — multiple neural networks whose outputs were combined to increase accuracy — on the crackle pattern signal and submitted their result in late 2023.

On February 5, 2024, the Vesuvius Challenge announced the Grand Prize winners. The team had decoded more than 2,000 characters from fifteen partial columns of text inside scroll PHerc.Paris4, which had never been opened. They split the $700,000 prize.

The decoded text turned out to be a previously unknown philosophical work. Papyrologists identified it as a discussion of pleasure — specifically, whether the scarcity or abundance of goods affects the pleasure they provide. The author was identified as Philodemus, the Epicurean philosopher whose works dominate the Herculaneum collection. This was not a text that had been lost and rediscovered. It was a text no modern scholar had ever seen.

In May 2025, researchers Marcel Roth and Micha Nowak from the University of Wurzburg, along with independent researcher Sean Johnson, decoded the title and author of a different scroll — PHerc.172. It was identified as On Vices, part of Philodemus' ethical treatise On Vices and Their Opposite Virtues. It was the first time a title had been read from a still-sealed Herculaneum scroll.

07 — The Scale5 Percent of One Scroll

The Grand Prize team decoded roughly 5 percent of one scroll. The Herculaneum collection contains approximately 1,800 scrolls, most of them still sealed and unscanned. The 2024 goal was to read 90 percent of four scanned scrolls. As of early 2026, that work continues — though the ink-detection models trained on PHerc.Paris4 do not generalize cleanly to all scrolls, and scanning protocols are being refined for the remaining collection held primarily at the Biblioteca Nazionale di Napoli.

The scale of what remains is difficult to overstate. The scrolls from the Villa dei Papiri represent the largest surviving collection of ancient philosophical texts. Forty-four known works by Philodemus have been identified in the partially opened fragments. But the sealed scrolls may contain works by other authors — Epicurus himself, whose original writings are almost entirely lost, or Greek tragedians and historians whose works survive only in fragments quoted by later writers.

Seales has described the ultimate goal: scan and read all 1,800 scrolls. In July 2025, his team secured a European Research Council grant to continue the work. The methodology is open-source. The CT scans are public. The competition framework has demonstrated that a global community of researchers — students, engineers, physicists, classicists — can collectively solve problems that no single lab could address alone.

The 2023 Vesuvius Challenge went from 0 percent to 5 percent of one scroll. That is a small number. It is also the largest since Piaggio's unwinding machine in the eighteenth century. The constraint is no longer whether the scrolls can be read. It is how fast, and at what resolution, the remaining 95 percent — and the remaining 1,799 scrolls — will yield their text.