For 2,000 years, the only intact library from classical antiquity lay silent — carbonized by Vesuvius, impossible to open. In 2024, machine learning broke that silence. What emerged rewrites everything.
In 79 AD, a library died. The pyroclastic surge that buried Herculaneum carbonized roughly 1,800 papyrus scrolls in seconds — turning philosophy, poetry, and science into brittle tubes of black carbon. For nearly two thousand years, they sat in museum drawers: miraculous in their preservation, impossible in their silence. In 2024, a 21-year-old student and a machine learning algorithm broke that silence.
This is the story of the Vesuvius Challenge — and of what it means that we can now read books that no human eye has seen since the reign of the Caesars. It is also a preview of what else we might recover: potentially the most significant archaeological discovery in history has been sitting in a cabinet in Naples the whole time.
The Villa of the Papyri stood on the western edge of Herculaneum, overlooking the Bay of Naples. It was almost certainly one of the grandest private residences in the Roman world — 250 metres of colonnaded terraces, gilded frescoes, a swimming pool fed by aqueducts, and a collection of bronze sculptures so extraordinary that they fill an entire wing of the Naples National Archaeological Museum today.
And then there was the library. Whoever owned this villa — possibly Lucius Calpurnius Piso Caesoninus, Julius Caesar's father-in-law — had assembled what appears to have been a working philosophical library, heavy with texts of the Epicurean school. When Vesuvius erupted on the morning of August 24th, 79 AD, the pyroclastic surge hit Herculaneum first. Temperatures of 300°C or more swept through the town in minutes. Everything organic was carbonized instantly. The villa's library — perhaps 1,800 scrolls — was turned to charcoal.
The strange mercy of this was that charcoal lasts. Organic material typically rots over centuries; carbon does not. The scrolls survived two millennia underground, perfectly preserved in their catastrophic state — as long as no one tried to open them.
Excavators rediscovered the villa beginning in 1752. The first attempts to read the scrolls were disastrous: priests hired by the King of Naples tried to physically unroll them, and destroyed hundreds in the process. An 18th-century monk named Antonio Piaggio invented a machine that could unroll them more slowly — painfully, over months — but even this technique shredded many scrolls that had fused solid in the heat.
By the late 20th century, scholars had managed to read about 600 scrolls, mostly partial, mostly the same thing: works of the Epicurean philosopher Philodemus of Gadara, who appears to have lived at the villa. Philodemus wrote prolifically on aesthetics, rhetoric, music, and pleasure — interesting, but not the lost masterworks the world had hoped for. The remaining scrolls — perhaps 1,200 of them — remained sealed and silent. Too fragile to touch, too carbon-dark for ordinary imaging. The situation seemed permanent.
In the early 2000s, a computer scientist at the University of Kentucky named Brent Seales began asking a question that archaeologists had mostly stopped asking: what if you didn't have to touch the scrolls at all? What if a computer could unfurl them virtually — reading the ink from inside without ever cracking the carbon?
The concept was called virtual unwrapping. X-ray computed tomography — the same technology used in hospital CT scanners — can image the internal structure of an object in three dimensions. In theory, if you could CT-scan a rolled scroll, you could computationally unroll the individual layers and read them without physical contact. Seales spent years developing the mathematical tools to do this.
The proof of concept came in 2016, with a scroll that was not from Herculaneum. An En-Gedi scroll — a small, charred biblical scroll found in Israel, dating to the 1st or 2nd century AD — had been lying unread in a museum since 1970. Using virtual unwrapping, Seales's team read it in its entirety. It was a fragment of Leviticus. The technique worked. The paper was published in Science Advances.
When Seales's team applied virtual unwrapping to the Herculaneum scrolls, they hit an obstacle that didn't exist with the En-Gedi scroll. Ancient Egyptian and Roman scribes used carbon-based ink — essentially soot suspended in gum. Carbon on carbon papyrus is invisible to standard X-ray: both materials absorb radiation at nearly identical rates. The ink left no detectable chemical signal. The scrolls appeared blank. It seemed like a dead end.
The solution came from looking closer — not at the chemistry, but at the physics. When carbon-based ink dries on papyrus, it changes the surface texture of the fibers at a microscopic level. The ink creates tiny raised ridges: a "crackle" pattern detectable not by chemistry but by topology. In sufficiently high-resolution CT data, the surface geometry of the papyrus — down to sub-voxel scales — encodes the presence of ink.
Machine learning, trained on this texture signal, could potentially read the ink directly from the CT data. In 2023, Seales's team had the algorithm. They had high-resolution CT scans of two complete, unopened Herculaneum scrolls. What they needed was the computing power — and the community — to make it work at scale. That's when Nat Friedman and Daniel Gross made their move.
In March 2023, Nat Friedman — former CEO of GitHub — and entrepreneur Daniel Gross launched the Vesuvius Challenge. The premise was simple and audacious: release the full CT scan data of the Herculaneum scrolls to the public, offer over a million dollars in prizes, and ask the global machine learning community to solve the ink detection problem. Science by open contest.
The response was extraordinary. Within months, thousands of researchers, students, and enthusiasts were working on the data. The first milestone prize — for detecting any ink at all — was won in August 2023 by Luke Farritor, a 21-year-old computer science student from the University of Nebraska. Farritor trained a machine learning model specifically on the crackle pattern signal and detected ink for the first time in the CT data of an intact scroll.
Two months later, Farritor's model decoded the first complete word: πορφύρας — porphyras, meaning "purple" or "the purple dye." A single ancient Greek word, emerging from 2,000 years of silence. It had been sitting in that scroll since the first century BC.
The grand prize — awarded in February 2024 — went to a team of three: Farritor, Youssef Nader (an Egyptian-German graduate student), and Julian Schilliger (a Swiss robotics researcher). Together they had decoded approximately 2,000 characters — roughly 5% of one complete scroll. The text was identifiably Epicurean philosophy, most likely from the hand of Philodemus himself. It appears to discuss whether pleasure is enhanced or diminished through music and the senses — a classic problem of Epicurean aesthetics.
The decoded passage was not a lost masterwork. But it was real, it was ancient, and it had never been read by any human being alive. More significantly, it proved that the method worked at scale — that the remaining 95% of that scroll, and the remaining ~1,800 scrolls in the collection, are readable.
Consider the mathematics of what remains. If 5% of a single scroll contains approximately 2,000 words, a complete scroll contains perhaps 40,000 words — comparable to a short novel. Multiply that across 1,800 scrolls. The Herculaneum collection, if fully decoded, could yield something in the range of 70 million words of classical text. In a world where most ancient literature has been lost — where we have fewer than 50 complete plays from the three great tragedians, where entire schools of philosophy exist only in summaries written by their critics — the implications are almost incomprehensible.
The Herculaneum library was Epicurean. We have almost nothing of Epicurus himself — his major works, including On Nature (a 37-volume treatise on physics and philosophy), are known only through fragments and paraphrase. The library may contain lost books of it. Other Epicurean philosophers — Metrodorus, Colotes, Polystratos — may have their complete works sitting in those drawers in Naples.
Archaeologists believe the Villa of the Papyri may extend significantly beyond what has been excavated. Ground-penetrating surveys suggest a possible second floor — unexcavated since the 1750s — which could contain Latin texts. The known scrolls are predominantly Greek. A Latin library above them could include works from Roman authors across the first century BC and AD: poetry, history, philosophy now otherwise lost entirely.
The broader promise extends beyond Herculaneum. Thousands of ancient manuscripts survive in conditions too fragile or too damaged for physical reading — in monastery libraries, museum archives, and archaeological sites across the Mediterranean world. Virtual unwrapping and AI ink detection are now proven tools. The En-Gedi scroll, the Herculaneum scrolls — these are early tests of a technology whose full application has barely begun.
We have been living in a world defined partly by what was lost: the gaps in our knowledge of antiquity where libraries burned, where manuscripts rotted, where civilizations collapsed and took their writing with them. AI is beginning, with unexpected gentleness, to reach into those gaps — and bring voices back.
Buy me a coffee