Luke Ferritor: the DOGE Zoomer who first cracked ancient Roman scrolls using ML on a hand-built PC

Summary

Luke Ferritor: Ancient Scrolls to Government Data

Luke Ferritor is a 22-year-old from Nebraska who won a portion of a $700,000 prize for developing machine learning software that reads carbonized ancient Roman scrolls buried by Mount Vesuvius in 79 AD. He has since joined DOGE, where he has read access to US Treasury payments data and is applying similar data science techniques to understand where government money flows.

The scrolls story began when Nat Friedman, former CEO of GitHub, became obsessed with the Herculaneum papyri during pandemic lockdown. He learned that roughly 800 scrolls have been recovered from a villa thought to belong to Julius Caesar's father-in-law, but thousands more likely remain buried. Conventional wisdom holds that reading them could multiply the known supply of ancient Greek and Roman literature by five to ten times. The obstacle was technical: the scrolls are brittle, and even careful unrolling destroys them. Three-dimensional CT scans offered a path forward, but humans could only glimpse snippets of ink.

Friedman's gamble was to run a contest. After pitching the idea at a Silicon Valley event where "no one bit," he and his investing partner Daniel Gross funded the Vesuvius Challenge, offering $1 million in prizes for software capable of reading passages from a single scroll. The strategy worked. An Australian mathematician named Casey Hanmer noticed patterns he called "crackle"—faint networks of cracks that resembled dried ink lifted from the page. Ferritor, then a summer intern at SpaceX, saw Hanmer's post in the contest Discord and began training a machine learning model on crackle data, adding newly identified letters back into the training set iteratively. Unlike large language models, his system was lightweight enough to run on a hand-built personal computer and required only binary classification per pixel: is there ink here or not?

In August, while at a house party in Nebraska, Ferritor remotely ran a new scroll image through his model and found three Greek letters. By 2 AM, he texted Friedman and other contestants "fighting back tears of joy." He eventually found 10 letters and won $40,000 from a progress prize. A classicist confirmed he had found the Greek word for purple. Ferritor later formed an alliance with two other contestants, Yousef Nader and Julian Schillinger, agreeing to combine their technology and share prize money. Their team won the grand prize, stitching together more than a dozen columns of text and entire paragraphs. The winning submission revealed a work by Philodemus centered on the pleasures of music and food.

The project yielded about 5% of one scroll. Friedman says a new contest might reach 85%. He is now contemplating buying scanners to place at the villa for parallel scanning and has become one of the few living people to explore the villa tunnels.

Ferritor's work at DOGE applies the same sensibility: treat government data like an archaeological problem. One early DOGE project involved analyzing a dump of federal payments to identify anomalies—why, for instance, were tax dollars funding something called "Ford Raptor LLC"? At one California university receiving roughly $700 million annually in federal funds, the team found that 40% of the budget, about $300 million per year, goes to administrative costs. The leverage is in pattern recognition across massive, messy datasets.

What distinguishes this story is the earnestness. Everyone involved—Friedman, Gross, Seals, Ferritor, Hanmer—worked without cynicism toward a shared goal. Ferritor's public profile since joining DOGE has drawn aggressive criticism, with some calling for his removal and his account being pressured into private mode. But the scrolls work shows someone genuinely excited by hard problems and willing to pursue them across disciplines: from machine learning on hand-built hardware to reading 2,000-year-old texts to auditing federal spending. The tension between celebration and controversy reflects a broader pattern: by the time most observers form an opinion on a person or project, the person has usually moved on to something else entirely.