Letters from the Past AI read messages that are 5,000 years old

For centuries around the world, texts have been found that no one could read. Clay tablets with mysterious symbols, stone slabs covered in inscriptions, not resembling any known alphabet. Scrolls written in languages that died thousands of years ago and left no descendants. Generations of linguists, historians, and cryptographers struggled with these texts and suffered defeat. The human brain, for all its power, was unable to cope. Too little data, too many possibilities, the key to the solution buried too deeply in the past. And then neural networks appeared. Machines capable of processing millions of options per second to detect patterns invisible to the human eye and propose solutions that no scientist could have conceived in an entire lifetime. In the past ten years, artificial intelligence has made breakthroughs in deciphering ancient texts. which seemed hopeless. The first story, and perhaps the most impressive one, concerns the charred scrolls from Herculaneum. In 79 CE, the eruption of Mount Vesuvius destroyed not only Pompeii but also the nearby city of Herculaneum. A hot flow of gas and ash, a pyroclastic surge Overwhelmed the city in mere minutes, killing all its inhabitants. and burying the structure under a 20-meter-thick layer of volcanic material. But unlike Pompeii, where the structures were buried under ash, the temperature of the flow in Herculaneum was so high, that the wooden objects did not burn but carbonized turned into fragile black carbon, preserving their shape. In the 18th century, when the excavations of Herculaneum began, Workers discovered a luxurious villa, known as the Villa of the Papyri. in which the only surviving library of the ancient world was located. About 1,800 papyrus scrolls, charred to blackness, clumped together, fragile as dry leaves. Each scroll is a potentially lost work. Greek or Roman literature, philosophy, science. Texts that existed in a single copy and were considered lost forever. The problem was that they were impossible to read. The scrolls were so fragile, Any attempt to deploy them always ended in catastrophe. The papyrus crumbled into dust, carrying the text away with it. For 300 years, scientists tried everything – chemical reagents, humidity, heat, mechanical devices. A few scrolls have managed to be partially unrolled. discovering fragments of philosophical texts in the Greek language. But most remained unreadable black blobs. A miracle occurred in 2023. A team of researchers, physicists, programmers, and papyrologists joined the Vesuvius Challenge project. The idea was insane—illuminating the scrolls with an X-ray tomograph. How does a doctor scan a patient's body? and obtain a three-dimensional image of the internal structure without unfolding them. The problem is that the ink used by ancient Greeks was carbon-based and black. And the charred papyrus is also carbon-based and also black. On the X-ray photograph, the ink and the papyrus looked the same. as if you were trying to read black text on black paper. The human eye could not detect the difference, but the neural network could. Researchers trained a machine learning algorithm on those few fragments where the text was visible. Taught him to distinguish microscopic textural differences Between the surface of the papyrus with ink and without. The differences were negligible. A difference of a few microns in the topography, invisible to any instrument except a synchrotron tomograph with a resolution of units of microns. The neural network processed terabytes of data and began to see letters. Greek letters on the surface of a scroll that had not been unrolled. A text that no human has read for two thousand years. The first word read was porphiros. purple in Ancient Greek One word and the hall where the results were announced exploded into applause. Because one word meant that the method worked. What can be read in the entire scroll? What can be read in all 1,800 scrolls? the lost texts of antiquity, works that we had considered lost forever Perhaps there still exist others, encased in charred carbon, and the neural network is capable of extracting them. By the beginning of 2024, several columns of text had been read. It turned out that the scroll contained a philosophical treatise from the Epicurean school. Reflection on pleasure, taste, and the enjoyment of food The author is presumably Philademos of Gadara. a philosopher of the first century BCE, Their works were known only through fragments. This is neither Aristotle nor Plato. But this is a text that the world has not seen for two millennia. And this is only the beginning. More than 1,700 scrolls are waiting in line. Among them may be lost tragedies of Sophocles, Unknown Dialogues of Aristotle Lost books of Titus Livius Or nothing of this. We won't know until the neural network reads them all. The scale of the potential discovery is difficult to overestimate. Of all ancient Greek literature that has come down to us, Less than 1%, according to various estimates. Ninety-nine percent of what the Greeks and Romans wrote has been lost. Of the 123 plays by Sophocles, seven have survived. About 7 out of approximately 90 plays by Aeschylus. Everything else perished in the fires of the libraries, in wars, invasions, and simply over time. The Library of the Villa of the Papyri the only surviving library of the ancient world, may recover part of these losses for us. Or cannot. We don't know what's in these scrolls yet. Perhaps these are duplicates of already known texts. Perhaps philosophical treatises, Interesting only to specialists. Or maybe just one line from Sofo, A scene from a lost tragedy, One idea that will revolutionize our understanding of antiquity. The neural network does not distinguish between the great and the mundane. She's just reading. As for deciding what is important, we will still do it ourselves. Organizers of the Vesuvius Challenge A prize of 700 thousand dollars has been announced. for the first few lines read and received over 1,600 applications from programmers, physicists, and enthusiasts around the world. The winners are students from Nebraska. and two researchers from Switzerland and Egypt, worked independently of each other and arrived at the same results. This confirmed the reliability of the method. Three people, three approaches, the same text. The second story takes us back 4,000 years. one of the oldest undeciphered writing systems in the world. Proto-Elamite script was used on the territory of modern-day Iran around the 3rd millennium BCE long before the rise of the Persian Empire, Before Cyrus the Great, before Zoroastrianism. About 1,600 tablets with proto-Elamite signs were found during excavations in Susa and other ancient cities, And for more than a hundred years, no one could read them. Writing was not associated with any known language, did not have a bilingual inscription which would allow establishing correspondence, and contained over a thousand unique characters, Most of whom had met only a few times. In 2020, a group of researchers from Oxford and several technical universities applied to proto-Elamite tablets clustering algorithms and neural networks The idea was as follows – even if we do not know what each symbol means We can analyze the structure of the texts. Which signs appear nearby? In what order, with what frequency, and based on these regularities, determine Which signs are numbers? which — by names, which — by product names. Most proto-Elamite tablets are administrative records. Grain, livestock, and oil accounting. This is neither poetry nor philosophy, This is accounting from the Bronze Age. And accounting, unlike literature, It is subject to strict rules. The numbers are in specific positions. units of measurement in others, Names — in third place. The neural network has processed all 1,600 cards. and highlighted groups of symbols, which are highly likely denoted numerals and units of measurement. This allowed partially reading the signs not in the sense of translating into modern language, and in terms of understanding the structure. Here it is written how many, Here is the product name, Here is the official's name. Complete transcription of proto-Elamite script Still far away, but for the first time in a hundred years, scientists can say, What is meant by the logic of the text. And this is thanks to machine learning. The third story — one of the most beautiful ones in the history of linguistics In 1908, Italian archaeologists, those who dug on the island of Crete, We found a clay disc with a diameter of about 16 cm. spiraled on both sides from mysterious symbols. The Phaistos Disc, named after the place of its discovery to the palace in Feste, became one of the most famous and the most mysterious archaeological discoveries 20th century. 242 characters, 45 unique characters not even a hint about their meaning. The Phaistos Disc is further intriguing because How was it made. The symbols were not scratched into the clay, They were pressed in using stamps. Each character is a separate stamp, pressed into the soft clay like a seal into wax. This makes the Festive Disc, perhaps the first printed text in the world 3,000 years before Gutenberg. If so, then somewhere on Crete There was a set of stamps, a kind of typographic font from the Bronze Age which has not reached us. Neural networks attacked the Festsky disc from several sides Pattern recognition algorithms classified the symbols, identifying patterns in their distribution. Certain signs occur more frequently some at the beginning, others at the end thirds – only in pairs. Statistical analysis showed that What is the structure of the disk text? characteristic of syllabic writing writing, in which each sign denotes a syllable, not an individual sound. This narrowed down the search field. If the disk is written in syllabic script, his tongue should have a certain phonetic structure, and the number of possible languages of candidates decreases sharply. There is still no complete decryption. and perhaps it won't. 242 characters is an excessively small text body for unambiguous decryption. But neural networks have done that, What a human could not do. They examined all existing hypotheses. and more than 20 of them, ranging from Minoan to Etruscan from Luwian to Semitic and assessed the probability of each based on mathematical analysis text structures The most likely candidate turned out to be a language related to the Minoan Linear A Another undeciphered writing system, but about which we know a little more, about the symbols of the Phaistos Disc The fourth story takes us to Mesopotamia – cradle of writing the land where 5,000 years ago People first began recording their thoughts on clay tablets Cuneiform – the writing system of the ancient Sumerians, Babylonians and Assyrians – was deciphered in the 19th century, And today specialists can read Cuneiform texts are almost free. The problem is another one—the labels are too numerous. According to various estimates, museums around the world stored at half a million up to a million cuneiform tablets, and most of them have never been read. Not because we cannot, Because we don't have enough time. For reading one label An experienced Assyriologist spends from several hours to several days At this rate, it'll take forever to read all the signs. It will take thousands of years. There are specialists in cuneiform in the world – a few hundred Several hundred thousand labels. Mathematics is merciless. In 2023, a group of researchers from the University of Munich and Google DeepMind imagined a neural network, capable of automatically recognizing cuneiform signs on the photographs of tablets and translate them into English. The system was trained on a corpus from tens of thousands of already deciphered tablets and achieved accuracy, which the developers evaluated comparable to the work a beginner Assyriologist Not an expert — a beginner. But a novice Assyriologist, which operates at a speed of thousands of tablets per hour – This is a revolution. The system has already helped read Several previously unstudied tablets from the collection of the British Museum. Among them were household accounts – Accounting again, lists of employees salary calculations in kind (grain) receipts for oil collection. Nothing sensational. but precisely such boring texts comprise 90% of the cuneiform corpus. And it is precisely they who tell us On everyday life in ancient Mesopotamia more than any epic. How much grain did the worker receive per day? What was the price of the house? How much did the slave cost? The answers to these questions lie in the unread labels. And the neural network can extract them within hours. not for years. But there is also a more ambitious goal. Among hundreds of thousands of tablets Probably conceal texts, that will overturn our understanding of ancient history. Lost myths, unknown laws diplomatic correspondence between kingdoms about which we know nothing. Each unread tablet – a sealed envelope with a letter sent five thousand years ago. A neural network – a postman, which will finally deliver these letters to their recipients. The fifth story about language, which had long been considered untranslatable. The Rongorongo script of Rapanui – One of the most mysterious writing systems on Earth. It was used on Easter Island. in a tiny fragment of Earth in the middle of the Pacific Ocean, known for their giant stone heads Muay. Rongorongo is a series of symbols, carved on wooden plaques depicting people, animals, plants and geometric shapes. 26 tablets have been preserved in museums around the world, And none of them had been read. The Rongorongo problem – in her isolation. Easter Island – One of the most remote inhabited places on the planet. The nearest land is the coast of Chile. 3,700 kilometers away. Writing developed in complete isolation. from the rest of the world. There were no contacts with other civilizations. There were no borrowings, There were no bilinguals. When Europeans arrived on the island in the nineteenth century, They brought diseases and slave owners with them. Epidemics and slave trade destroyed most of the population, and with them all those who could read Rongorongo. By the time scientists became interested in the little tables There was not a single living person left. able to read them. A language locked in 26 pieces of wood not even one key to the lock. In 2024 Researchers applied to Rongorongo methods of computer linguistics which were previously used for analysis genetic sequences. The idea. If the text cannot be translated, At least its structure can be determined. Is Rongorongo genuine? writing system in which symbols transmit the language, or by a mnemonic device something like the Incan knotted writing, which records not words, and ideas. The neural network has analyzed character distribution, frequency of repetitions the length and structure of words sentences. And I came to the conclusion that Rongorongo is highly likely is logosyllabic letter. A system in which one sign denote whole words, while others Syllables. This is crucial. This means that Rongorongo the present writing, not a set of pictures. And this holds out hope for decryption, Not tomorrow, anyway. Sixth and The latest story is perhaps the most ambitious one yet. It is not related to a single text and not just one sign, but with a whole civilization. A civilization Indus Valley. One of the three great civilizations of the Bronze Age. Along with Egypt and Mesopotamia. Five thousand years ago in the territory of modern-day Pakistan and in northwestern India existed cities with populations in the tens of thousands thousands of people. Mohenjo-daro, Harappa, Lothal. Cities with sewage systems, public baths, grain silos and standardized bricks. Cities that exceeded in terms of urban development everything that existed in Europe up to the Roman Empire. And this civilization left us their writing system. Around four thousands of inscriptions on seals, tablets and ceramics. Indus script approximately four hundred unique not deciphered. This the most famous undeciphered Writing in the world. And above it the best linguists of the twentieth century struggled and the twenty-first century. In vain. We don't know what language it is. The residents of Mohenjo-daro said. We don't know what is written on their seals. We do not know how they called them. their cities, their gods, themselves. A whole civilization, millions of people who lived and died for a thousand years. And we do not We cannot understand a single word they say. Think about it. We know, What did the inhabitants of Mohenjo-daro build? straight streets intersecting at a right angle four thousand years ago to the New York grid We know that in every house there was sewage system in three thousand years before the Roman sewers. We know that They used standardized weights. and measures. Cubic weights with precision to the nearest tenth of a gram. We know that they traded with Mesopotamia. Their seals have been found in Sumerian contexts. cities. But we don't know how they They called all this themselves. We don't know them. myths, their laws, their prayers. An alien civilization — the most A competent one among the silent. She speaks. We're done. With our own architecture, with their own seals, with their own sewage system And says nothing. Because we cannot read Her tongue. Neural networks are attacking. the inscriptions from the beginning the 2010s. A group of under the guidance of Raja Sharao The University of Washington trained algorithms on the corpus of all known inscribed texts and analyzed the statistical structure of texts The result confirmed what many had already suspected. Linguists suspected. Indian Signs convey language. Not decorative patterns, not random tags, and structured text with grammar, syntax, and lexically. The neural network detected regularities characteristic of so-called agglutinative languages. Languages in which words are formed by affixation of suffixes and prefixes to the root. Dravidian languages of South India — Tamil, Telugu, Canada — are agglutinative. And some Researchers believe that the Indian civilization spoke in an early stage Dravidian language. Neural network did not prove this hypothesis, but Gave her quantitative reinforcement, which had not existed before. The problem is that Indian inscriptions are very short. On average, 5 characters. That's like If only a single work of Russian literature remained 4,000 fragments of five literally. Good, winter, The tsar, the fish. Try based on Restore the Russian language and you will understand the scale of the task. The neural network can analyze the structure, but for complete decryption, a either a bilingual text, or significantly longer inscriptions. Neither has been found yet. But archaeologists continue digging in in the Indus Valley, and each season brings New stamps and plates. Sooner or later, perhaps tomorrow, Perhaps in 100 years, someone will find a long one text or bilingual. And at that moment neural network, already trained on the entire existing hull, Decrypt it in hours. The machine is ready; it is waiting for data. 6 stories, 6 ancient ones messages. Charred scrolls Herculaneum, two millennia silent beneath the ash of Vesuvius, and the neural network, seeing the black letters On a black background. Proto-Elamite signs that had kept the secret 4,000 years and the algorithm the structure found in the chaos of signs accounting book. The Phaistos Disc, challenging a hundred years of linguistics, and the machine that determined the type letters No. 242 cuneiform tablets that humanity does not have time to read and a neural network reading thousands of tablets per hour. Rongorongo from Easter Island, locked in 26 wooden planks, and computer, proving that this is real Writing. Indus Civilization, Whose language has been silent for 5000 years? and an algorithm waiting for data to speak on her behalf. Each of these stories is about the boundary between human and machine. Neural network Does not understand the text, she does not know, What is "purple," doesn't feel the beauty of the verse, does not weep over tragedy. She sees patterns, patterns in data, statistical distributions, correlations between the symbols. But precisely this — the ability to see patterns on a scale inaccessible to human beings to the brain, making her indispensable. A person reads text, a machine reads the text structure. Together They can read what no one else would have read None of them individually. And here's what's surprising — dead languages come alive. The people who wrote these texts thousands of years ago could not Imagine that their words will be read by a machine. They wrote for their contemporaries, for priests, officials, merchants who knew their language and could understand them. They did not think about the future. They did not insert the capsule. time. They simply recorded their life, my own accounts, my own prayers, my own thoughts. And here it is after thousands of years of silence Their voices return, decrypted not a human, but a machine, who doesn't understand a single word But perhaps read each one. We are living in a unique moment. For the first time in history, we have a tool, capable of breaking through the wall of silence that separates us from ancient civilizations. A neural network is not magic. A stick. It won't decipher what what cannot be deciphered and will not be found what is not there. But she can do in seconds, what would take a human It would have taken years. She could see what the eye cannot see. She can Hear what has been silent for a long time. And every time she draws it out words from a charred scroll the number from the clay tablet, from the stone seal's structure, it brings us closer to the people who Lived thousands of years ago and wanted one thing, So they could hear us. They weren't expecting us. They didn't know us. But we came. And we listen.