Nat Friedman - Reading Ancient Scrolls, Open Source, & AI

Wed Mar 22 2023

The Vesuvius Challenge and the Papyrus Scrolls of Herculaneum

There are over 600 tax scrolls from ancient times that cannot be opened, but new techniques may allow them to be read.
Thousands of papyrus scrolls were preserved by the volcanic eruption of Vesuvius in AD 79 in the ruins of Herculaneum.
The score prize, also known as the Vesuvius Challenge, was founded by Nat Friedman to uncover new information from the papyrus scrolls.
Professor Brent Seals at the University of Kentucky has been using advanced imaging techniques and machine learning to virtually unroll the scrolls without damaging them.
In 2019, the scrolls were taken to a particle accelerator in Oxford for a high-resolution CT scan.
The goal is to launch an open competition to read substantial amounts of real texts from one of the scrolls without opening them.

Julius Caesar's father-in-law owned a villa that was buried in mud and ash during the eruption, containing a large library of papyrus scrolls.
The villa has only been partially excavated, with two additional floors remaining unexplored.
Most of the exploration occurred in the 1700s through Swiss engineers' tunnels underground.
The papyrus scrolls found in the villa were mostly philodemus texts, possibly from Philodemus' working library.
There may be a Latin library containing literary and historical texts that has not yet been excavated.

Papyrus scrolls from ancient times were fragile and decayed quickly, leading to only 1% of all writing from that period surviving.
Attempts to physically unroll the scrolls often resulted in their destruction, but a monk named Piagio developed a slow unrolling machine that successfully revealed some Greek philosophical texts.
Professor Seals has made progress in identifying ink inside the scrolls using advanced imaging techniques and machine learning.
The resolution of the scans is good enough for recognition, but it requires a smart or dogged group of people.
The uncertainty lies in whether there is enough data to train a machine learning model for the scrolls.

Nat Friedman and Daniel Gross are offering a prize for the first person or team who can read substantial amounts of real texts from one of the scrolls without opening them.
Reading all the scrolls could double the total text from antiquity, including lost epic poems.
There is a 50% chance that someone will solve this challenge and get the data this year.
If one scroll can be read, it could lead to scanning all remaining scrolls and excavating the rest of the villa.
This challenge could be a catalyst for more funding in archaeology.

Microsoft acquired GitHub for $2 billion and runs it independently like LinkedIn.
Nat Friedman was appointed as CEO to run GitHub independently and worked hard to alleviate users' concerns.
The acquisition caused panic among users when it was leaked before the official announcement.
GitHub profiles and resumes contribute to how people understand a person, but they are not definitive resumes.
Open source tends to excel in infrastructure and standalone components but struggles with user experiences that require editorial control.

GPT-4 is better at certain languages than GPT-3, indicating progress in language models.
Pre-trained models are like unrefined crude oil until they are trained and put out into the world.
There may be wide-scale proliferation of foundation models, with structural forces promoting both proliferation and concentration.
Building a good end user application requires establishing a tight feedback loop with users, which is difficult even in a company with hierarchical control.
The cost of training machine learning models may not be as high as expected due to many available techniques for improving efficiency.
Tinkerers are needed to explore the possibilities and value of AI tools like Copilot.

Acquisitions can save time and money on building distribution and relationships with customers.
Large companies face challenges in doing acquisitions due to different approval processes and cultural buy-in processes.
Open source tends to succeed more where the coordination burden is lower.
Software and new forms of governance can improve coordination and expand the territory for open source success.
Governance dynamics can make coordinating with multiple contributors challenging in open source projects.

Companies should empower individuals with great judgment to make decisions and exercise their expertise.
Consensus environments can lead to average outcomes, while high variance outcomes can result in either excellent or disastrous results.
The speaker believes in cultivating a forward-looking view and doesn't dwell on past achievements.
Capabilities and alignment may not be opposing forces, as both are important in the context of AI's potential benefits and risks.