On the front page of Nature's official website, the latest publication is the latest archaeological results, which are sensationalizing the world - after 2000 years, the mystery of the Elder Scrolls (Scrolls of Herculaneum) has finally been revealed! Moreover, this research was done by AI, and behind it was a 21-year-old genius, a SpaceX intern whom Musk liked.

Different from ordinary code cracking, the difficulty of analyzing this scroll is tantamount to deciphering a "wordless heavenly book"——

Because in AD 79, this set of scrolls was burned to carbon in a volcanic eruption, and was only unearthed in 1752. As long as you are not careful, the fragile scroll will become fragments in minutes. It is difficult to unfold it through physical means.

Today, three college students have finally used the power of AI to successfully decipher the passages, opening up a new archaeological research paradigm.

They revealed more than 2,000 characters in the scroll by taking CT images of the scroll and using machine learning technology to segment and recognize them.

This news quickly caused a global sensation. So far, it has been viewed more than 15 million times and has nearly 60,000 likes.

Musk also gave a thumbs up to the three-person team, saying it was really cool.

Demis, the founder of DeepMind, was also very impressed and said that he could not wait to read this mysterious text.

So, how was this mysterious scroll deciphered step by step?

You can decipher the content without opening it

The person who handed over this result was a team of three students. They are:

Youssef Nader, from Egypt, studying for a PhD at a university in Berlin, Germany

Luke Farritor, SpaceX intern, undergraduate student in computer science at the University of Nebraska-Lincoln

Julian Schilliger, Master of Robotics, ETH Zurich, Switzerland

It is almost impossible to decipher this scroll that has been burned into carbon without causing any damage. It is almost impossible to open it and look at it.

In 1865, someone tried to physically open the scroll, but it failed and part of the scroll fell into pieces.

After nearly two hundred years, no one dared to try physical expansion again.

Until 2019, Professor Brent Seales of the EduceLab of the University of Kentucky used a particle accelerator to take three-dimensional CT images of the scroll with a resolution of up to 4 μm.

Stephen Parsons, a doctoral student of Professor Seales, has long been working on using machine learning models to detect ink in CT images. He established a ground-truth data set using partially unfolded scroll fragments with visible ink.

Netizens started a heated discussion about this matter, and it reached Luke's ears through a podcast.

This aroused great interest in him, and he began to use his spare time to train machine learning models to analyze ink marks.

In the end, Luke's model successfully analyzed the invisible symbols in the scroll and discovered a complete word.

The word is the Greek "ΠΟΡΦΥΡΑ?" (lowercase: πορφυ?ρ?α??, Latin transliteration: porphyras), which means "purple".

This discovery made him the first person in history to read an entire word from the interior of the Herculaneum scroll.

Youssef was the second person to decipher the word, using a different method but getting the same, clearer result.

It was after seeing the results submitted by Youssef that the experts confirmed Luke's discovery.

Julian participated in the development of EduceLab's 3D model segmentation tool VolumeCartographer.

VolumeCartographer unfolded the three-dimensional CT image along the scroll and mapped it to the two-dimensional plane, laying the foundation for Luke and Youssef's deciphering work.

Since then, Julian has proposed a new segmentation method, ThaumatoAnakalyptor, to verify the cracked area through re-segmentation and explore more unknown fragments.

After that, in order to decipher more content faster, the three formed a super team and finally successfully decoded more than 2,000 characters.

They used three different model architectures for analysis, and the results can be confirmed by each other. The ink detection program code is also open sourced directly on GitHub.

The more than two thousand characters they analyzed are located at the end of one of the volumes, accounting for about 5% of the volume, and the content is considered to be related to Epicurean philosophy. Epicurus was an ancient Greek philosopher who mainly focused on personal happiness and peace of mind. However, he also emphasized rationality and proposed atomism in the field of natural science, which had an important impact on the development of ethics and psychology in later generations.

In two consecutive sections of the scroll, the author addresses whether and how the availability of goods (such as food) affects the pleasure they can bring.

Therefore, this archaeological discovery may actually help further understand the ancient people's understanding and exploration of the world.

A challenge, three students, 5 million prize money

What brought these three students together was a challenge called "Vesuvius".

After Stephen built the data set, Silicon Valley entrepreneur and former GitHub CEO Nat Friedman became very interested in it and sponsored the competition.

The organizers announced the selection criteria in March last year. Participants need to decipher at least 4 paragraphs, each paragraph has 140 characters, of which at least 85% of the characters can be recovered.

At that time, the expert group generally believed that the probability of someone successfully challenging the challenge within a year was less than 30%, but the result was unexpected -

According to reports, they received many works after the news of the competition was released, and there were still teams submitting their results a few minutes before ddl.

After a month-long review, the 2,000-character text was deemed successfully deciphered.

The three winners will share a prize of US$700,000, equivalent to more than 5 million yuan.

In addition, three teams won the second prize for submitting results that were significantly readable. These three teams will each receive US$50,000.

Among them is the team led by Google Chinese engineer Shao-QianMah.

Next, the organizers of the competition hope to decipher 90% of the four sets of scrolls within one year, and subsequent competition information will also be announced in March.

Of course, such challenge prizes are only one aspect. More importantly, it shows the possibility and feasibility of using AI to solve challenges in all walks of life, including subjects such as archeology.

Perhaps you have also heard that AI research on Oracle began earlier. Do you want to know about other projects? Also welcome to share~~

Reference links:

[1]https://scrollprize.org/grandprize

[2]https://twitter.com/natfriedman/status/1754519304471814555