Nat Friedman (Github CEO) — Reading ancient scrolls, open source, & AI

Dwarkesh Patel introduces Nat Friedman as GitHub's CEO from 2018 to 2021, the founder and seller of two earlier companies (Ximian and Xamarin), the founder of AI Grant and California YIMBY, and now the organizer of the Scroll Prize. The conversation opens with that prize, which Friedman calls the Vesuvius Challenge.

Friedman says he stumbled into the subject during the COVID lockdown, reading about the AD 79 eruption of Vesuvius. Besides Pompeii, the eruption buried Herculaneum, which he describes as sort of like the Beverly Hills to Pompeii. One enormous villa there had once belonged to Julius Caesar's father-in-law. It held statues, marbles and art, and also a large library of papyrus scrolls. Mud and ash buried the town under something like 20 meters of material, and nobody knew exactly where it was until the 1700s. In 1750 a farm worker digging a well, about 60 feet down, hit a marble paving stone from a path at the villa. A Swiss engineer then dug tunnels out from the well shaft. Looting was the spirit of the time: walls were busted through and surviving bronze statues carried off. Along the way the diggers kept meeting lumps that looked like charcoal, and many were apparently thrown away until someone noticed writing on one. Only then did they realize they were papyrus scrolls, hundreds or even thousands of them.

Friedman presents this as the only library to survive from antiquity in any form, though badly damaged, carbonized and very fragile. His contrast is with normal survival: in a Mediterranean climate papyrus rots quickly and had to be recopied by monks every 100 years or so, perhaps less. He says it is estimated that less than 1% of all the writing from that period survives.

The first attempts to read the scrolls were well meant but destructive. The scrolls were so fragile that they turned to ash in people's hands, and hundreds were destroyed. Some were cut down the middle with daggers so that small pieces flaked off, from which people tried to get a few letters. An Italian monk named Piaggio later devised a machine, under the care of the Vatican, that unrolled scrolls very slowly, about half a centimeter a day. A typical scroll might be 15 to 30 feet long. A few were unrolled successfully and contained Greek philosophical texts in the Epicurean tradition by a little-known philosopher, Philodemus. That counted as new text from antiquity, which Friedman notes does not happen often. Physical unrolling was eventually abandoned because so many scrolls were lost, although he says some attempts continued even into "the 200s" (as transcribed) and the scrolls were destroyed. The present situation, in his account, is roughly 600 or more intact scrolls that nobody can open.

Friedman learned of Brent Seales, a professor at the University of Kentucky, who had been scanning the scrolls with increasingly advanced imaging and then using computer vision and machine learning to unroll them virtually, without ever opening them. Seales's most recent attempt, in 2019, took the scrolls to a particle accelerator in Oxford, the Diamond Light Source, for what Friedman calls essentially an incredibly high-resolution 3D X-ray scan. It needed very high-energy photons and produced scans at eight microns, tiny voxels that the team thought would be sufficient.

Friedman waited for news of a successful decode, but COVID slowed everyone down. When he reached out the following year, he found the team had made good progress: machine learning models had started to identify ink inside the scrolls, but no words or passages had yet been extracted, which he calls very challenging. He invited Seales to California, and to his surprise Seales came. They decided to team up, and the approach they settled on was an open competition. Friedman says they did a great deal of work with Seales's team to put the data, tools, techniques and general understanding of the materials into a form in which smart people could become productive quickly. With Daniel Gross, Friedman is putting up a prize, in the manner of an X PRIZE, for the first person or team to read substantial amounts of real text from a scroll without opening it.

The stakes, as Friedman reports them from historians, are large. If the technique worked and generalized to all of the six to eight hundred scrolls, it would approximately double the total texts that survive from antiquity. He says the unit he has heard used is multiple Shakespeares. What is inside is unknown: a few Philodemus texts are of some interest, but there could be lost epic poems or, in his words, God knows what. He puts the chance that someone gets nerd sniped by the opportunity and solves it this year at about 50%.

Patel adds an anecdote: Seales told him that his first reaction to the emails was that they must be spam, since Nat Friedman could not plausibly be reaching out. Friedman confirms that Seales was hard to reach. He emailed a couple of times without a reply, and then had his admin, Emily, phone the secretary of Seales's department before Seales realized something real was going on. They then spoke on Zoom, and Friedman says they now think alike about the project. His assessment of the work so far is that the team has taken it right up to the one yard line. Every key component has been demonstrated. Putting the pieces together, improving quality and doing it at the scale of a whole scroll remains very hard work, and an open competition seems to him the most efficient way to get it done.

Asked by Patel whether the roughly thousand-plus scrolls in question are all from the Philodemus layer, Friedman explains that the villa was never dug out and exposed to air. Most exploration happened in the 1700s through the Swiss engineer's tunnels, 50 to 60 feet down, and the diggers were looking for treasure rather than doing a full archaeological exploration. Excavations at the edge of the villa in the 1990s established two things. It was a seaside villa facing the water, though the eruption deposited so much mud that the shoreline moved outward (Friedman says he has never been). And there were two additional floors that the tunnels had apparently never excavated.

From this Friedman concludes that at most a third of the villa has been excavated. Most of the scrolls found came from one small room and were mostly Philodemus texts, with several revisions of some texts. The hypothesis is that this was Philodemus's working library. In the hallways, however, other scrolls turned up, including crates of them. Historians told him they think the main library of the villa has probably not been excavated and may be a Latin library with literary and historical texts, possibly much larger. He is careful to say that he is no expert and does not know how prone classicists are to wishful thinking, calling it a romantic idea. His evidence for it is the partly evacuated scrolls found in hallways, and he has since read firsthand excavation accounts. Those include heartbreaking descriptions of an entire case of Latin scrolls, perhaps 30, accidentally destroyed while being pulled out of the mud.

Patel notes that Friedman made some scrolls. Friedman explains how papyrus is made: the outer rind of the Nile reed is removed, the inner core is cut into strips, the strips are laid parallel with a second layer at 90 degrees, and the whole is pressed and dried. Sheets are then glued with a paste, usually of flour, into a long scroll. He bought some on Amazon and says it has a lot of texture that you feel when writing.