On the evening of September 24–25, historian Ben Brinn's piece made the rounds about how large language models are tackling old unsolved problems: deciphering John Dee's grimoire Liber Loagaeth, a Wehrmacht message from July 10, 1941, a WWI radiogram, hunting down the source of Newton's Latin translation from a French alchemical treatise. Charming stuff, but what hooked me wasn't the 16th-century mysticism — Enochian and Dee have been covered separately. What hooked me was the phrasing about "old open problems": the machine arriving where decades of work relied on manual skill.
That same day, another item hung in the feed that you'd normally scroll past: PHerc. 1667, a scroll from Herculaneum, charred in 79 CE, read from beginning to end — never once unrolling it. Two stories about the same thing.
Then came the engineering question that finished me off. How do you read black on black? Soot-based ink on papyrus that itself turned to soot. The answer turned out more interesting than the discovery itself.
The Villa dei Papiri library is the only ancient library that survived intact, shelves and all. The villa was discovered by accident: in 1750, a peasant digging a well hit marble flooring. Two years later, black lumpy cylinders started turning up in the underground passages — at first they were mistaken for charred branches, some were simply thrown out, and others burned for warmth: there was nothing to heat the tunnels with.
Then the destruction machine ran for a hundred and fifty years. The first scroll took four years to unroll, millimeters per day, using silk threads glued to the papyrus edge — the device was invented by Genoese priest Antonio Piaggio, sent from the Vatican Library in 1753. Then came Humphry Davy's chlorine, then ether, then mercury, then in the 1980s a mixture of gelatin and acetic acid. By the mid-20th century, 585 fragments had been fully unrolled from the collection, another 209 partially; roughly two hundred texts were deciphered and published.
Against that backdrop, read today's triumph. PHerc. 1667 today is only what remains of the scroll after three opening attempts: in the 19th century, in 1969, and in the 1980s. Those attempts destroyed the outer layers, and only the dense core survived — about eight centimeters of the original 19–24. In the 1980s, the scroll was declared completely illegible. When they call it "completely" read, they mean within these boundaries: the surviving part, 22 columns of Greek, nearly a meter and a half of text.
The content wasn't what they expected. An ethical treatise on human nature, impulse, and moral progress, with terms like ὁρμή ("impulse," drive to action) and φρόνησις ("practical wisdom"), closest to the Stoics. In the last surviving column — the name Aristocreon, nephew and student of Chrysippus. Based on handwriting, the text was dated to the 2nd, possibly late 3rd century BCE, meaning the author is most likely not Philodemus, and this is one of the oldest scrolls in the collection. A Stoic in an Epicurean library isn't odd: the schools debated openly, and the villa's owner kept works by those his circle criticized.
For honesty's sake: in press releases and on the project site, the number of columns in PHerc. 1667 varies — 22 columns from one team, around 20 from another, even though they're talking about the same scroll. The reason is likely that some columns are fragmentary and get counted or not depending on the individual papyrologist's decision. I can't settle on a single number.
To understand the main obstacle here, forget the hospital picture. A medical CT scanner distinguishes densities: bone is brighter than soft tissue. In a Herculaneum scroll, there's nothing to distinguish. The ink was made from soot, the papyrus charred into soot — in an X-ray volume scan, letters aren't invisible because they're small, but because chemically they're the same substance.
The industry lived with a wrong assumption for quite a while. Brent Seales's 2019 paper in PLOS ONE was titled exactly that — "From Invisibility to Readability," and its first conclusion reads like a confession: it was believed that micro-CT contained no information about carbon-based inks in principle. The information was there, just weak — it lay in the wave phase and surface microtopography, not in brightness. The equipment needed for this is accordingly specialized: a synchrotron, accelerating electrons to speeds that produce a narrow and extremely powerful beam.
The cost of resolution scales by the cube. Cut the voxel in half along each axis — data volume grows eightfold. Going from 8 micrometers to 2 gives roughly 64 times growth. Recent scans of complete scrolls on beamline BM18 in Grenoble run at 2.4 micrometers, propagation distance 22 centimeters, and average energy around 78 keV. One scroll — up to 300 terabytes, the largest dataset in the history of the European Synchrotron Radiation Facility.
Here too, sources don't align. The project's technical breakdown talks about "tens of terabytes" per large scroll, the synchrotron's press release says 300 terabytes per scroll. That's an order of magnitude difference, and I found no explanation; I suspect some count the scan volume itself while others count all intermediate processing products, but there's nothing to confirm this.
The breakthrough didn't come from physics. In August 2023, physicist Casey Handmer published an observation on his personal blog that papyrologists had missed: on flattened surfaces, some areas look like cracked mud — raised scales 0.1–0.5 mm, separated by narrow dark channels, and these areas run in 2–4 mm bands, neatly tracing letter shapes. The ink layer dried differently from the papyrus fibers and left relief. Handmer got $10,000 for that observation.
Two months later, Luke Farritor, a 21-year-old student who'd spent the summer interning at SpaceX on the launch complex, heard about the contest on a podcast, trained a model on these cracks in his evenings — and saw the word ΠΟΡΦΥΡΑϹ, "purple." Egyptian robotics PhD student Youssef Nader independently found the same word in the same place, only clearer. Then both, along with Swiss Julian Schilliger, split the $700,000 grand prize of 2023, and the organizers honestly wrote: when they announced the conditions, they themselves gave success less than 30% probability.
What happened next rarely gets described: almost the entire current research team on the project are former contest participants. People came for the prize money and stayed to work.
And here's the most interesting turn, fresh — September 16, 2026. Douglas Seiler from Berkeley SETI went a different route. He bought papyrus and reed pens in Egypt, lampblack in Japan, paid schoolchildren to fill scrolls with lines from Star Wars, Bible quotes, and a phrase from "The Outer Limits," rolled it all up, burned it in a nearly oxygen-free furnace — and got his own ancient scrolls with known content.
The result is simple: if the ink contains lead, it absorbs X-rays up to 25 times more strongly than charred papyrus. The difference is crude and glaring — visible without any math. And you can check this with a cheap handheld XRF scanner, without sending the scroll to a synchrotron: click the device — and you know which scroll will give readable text and which will remain a black smudge.
The most unpleasant part of this work — the footnote: not one of the scrolls that had been virtually unrolled by that point had been systematically checked for lead. Meaning three years of competition optimized neural networks and unwrapping geometry, while the question "which scrolls are even worth scanning first" was solvable with a pistol-sized device. Side note: on Oxford's scroll PHerc. 172 from the Bodleian Library, back in 2025 they noted that the ink was unusually visible — and suggested lead admixture. Nobody followed that thread.
The technical part, roughly speaking, was solved. Then it turned out the bottleneck had moved to a place nobody expected — literally up top, in the city.
The villa sits beneath modern Ercolano. Above the atrium — the historic town hall with garden, next to it a police station and residential houses. To reach the unexplored 2,800 square meters, estimates say they'll need to demolish fully or partially at least seven houses, the lot, and possibly the town hall. Archaeological park director Francesco Sirano speaks about this without pathos: €150 million against the site's annual budget of €12 million, not counting lawsuits with property owners. And he adds a detail that makes the whole discussion concrete: his own father's house was expropriated for previous excavations. Plumber Liberato Raia puts it simpler: "Demolishing our houses is a problem. Where do we go?"
Richard Janko from Michigan insists: "Until now we were told we'd only damage new scrolls and couldn't read them. That objection is removed." Robert Fowler from Bristol adds that 18th-century excavations didn't reach the first floor, meaning the book reserve could be an order of magnitude larger. Andrew Wallace-Hadrill from the British School in Rome counters no less firmly: the Roman villa is rotting — restored roofs are collapsing, tiles lie scattered on mosaic floors, charred wood constantly crumbles. "It would be a scandal to expose it to daylight now, while we can't guarantee it will survive."
A separate detail that finished me: these conversations have been going on since 2004. In the Herculaneum Society bulletin that year, they wrote that technical-economic justification for excavations "is underway, results expected within a year." The first part of the report came out in 2008 — no deadlines, no figures. In 2016, a group of British, American, and French scholars wrote a letter to The Times: no new work since 2009, "excavations must be completed." What's happening with that justification now, I honestly don't know: Italian pages from 2026 report the villa is closed, but for a different reason — accessibility and drainage works, with the wording "opening date not determined."
Announcing the first results in June 2026, the project also announced a new grand prize: $800,000 to whoever fully unrolls and makes readable another scroll from a list of thirteen; total prize fund for the new cycle — a million, plus half a million for "first letters" across ten scrolls, plus 20 thousand monthly for best open-source contribution. Formally this continues last year's story. Actually — a metric shift, far more interesting than the money.
The conditions include a limit: no more than eight hours of human labor. The task isn't "read" but "read almost without humans." The team already knows how to read one scroll by hand, but there are over six hundred unopened ahead, and a year of work each — that's not a plan.
It immediately became clear that even "manual work" itself is now different. The night before the Naples presentation, the team extracted around 140 columns from PHerc. Paris 4, rescanned at 2.4 micrometers: ink became visible directly in the 3D data, and the read text matched exactly, one-to-one, what had been deciphered from cracked mud in 2023. And on another scroll, PHerc. 139, they read the title: "Philodemus. On Gods, Book 8." Before this, only the first book and, presumably, the third were known — the figure "8" proves the existence of at least eight books of a treatise history had been silent about.
Seales framed the transition in one sentence: now they need not so much machine learning specialists as people who can read, edit, and understand what they've read. "Today we're hearing voices silent for two thousand years," he said. "But most importantly, we're beginning to understand them."