According to Wired, academic libraries around the world contain hundreds of thousands of fragments of ancient Greek papyri. Although many are so damaged that their meaning is likely lost, scientists are able to reconstruct the remaining pieces by methodically filling in the gaps in words or phrases. To speed up this labor-intensive process, researchers turned to artificial intelligence. On Wednesday, the Austrian Academy of Sciences will unveil the "world's first advanced language model for Ancient Greek", developed in collaboration with French AI laboratory Mistral and technology company Sail Reply. The model, called Apollo, is trained on approximately 600 million historical Greek words collected from manuscripts, papyri and inscriptions. The model will be available to scientists through the chatbot interface. The goal is to help researchers more quickly locate papyri fragments relevant to their specific subdisciplines, as well as open up new avenues for research. Where documents are worn and torn, Apollo is able to fill in the gaps with the most likely words or phrases, potentially revealing hidden details about historical events and practices. Dimitris Vlitas, partner
company Sail Reply, claims that discovering knowledge in this way “would have been unthinkable a year ago.” Until now, restoring damaged papyrus has required a skilled scholar to first identify word divisions—there are no gaps in ancient Greek writing—then accurately date the document, consider relevant sociopolitical contexts, and consult reference materials to select appropriate words to fill the gaps. “There are very few people in the world who know Greek history so well,” says Stephen Colvin, professor of classical philology and historical linguistics at University College London. However, all of this specialized knowledge is built into Apollo. “When he sees Homer, he substitutes Homeric Greek. When he sees an inscription in Doric, he uses Doric,” says Anna Dolganov, a historian and papyrologist at the Austrian Academy of Sciences. Scientists who face difficulties in the painstaking work of restoration expect Apollo to speed up the process, allowing them to focus on the meaning of historical documents rather than what they say. "I think it's very exciting," says Armand D'Angour, professional
r of Classical Languages and Literatures at the University of Oxford, home to the world's largest collection of ancient papyri. “If I had a machine that told me, ‘Here are three possible words that could fit in this gap,’ that would speed up the process significantly.” Apollo is unlikely to change the general understanding of the ancient world; many papyri have not yet been restored precisely because they are ordinary - personal letters, marriage contracts, civil service papers. “If you were an outsider, you might think that all of a sudden we're going to get some new plays by Sophocles, but that's not going to happen,” Colvin says. However, the model may help reveal new details about life in antiquity and confirm existing scientific assumptions. "Every time something is created, it adds a tiny element of knowledge about the ancient world," says D'Angur. If Apollo is successful, Vlitas says, the same technique could easily be applied to other ancient languages—Latin or Egyptian, for example—or to any other academic discipline that would benefit from the distillation and indexing of a large corpus of material. Artificial intelligence has already made significant progress in some
ory areas; OpenAI recently reported that its AI models solved a 200-year-old mathematical problem, while Google DeepMind released an extensive AI dataset that maps how genetic mutations affect molecular biology. One problem may be that relying on a language model—which deals with probabilities—to fill in gaps in ancient documents risks polluting the historical record with errors. But to avoid this problem, Apollo offers a choice of words for scientists to choose from. “The key point is that human competence must remain,” says Dolganov. “If we become completely dependent on AI-based transcriptions and interpretations of historical material, that’s when the problems begin. The material was prepared by a neural network based on the original publication. AI may make mistakes in details (numbers, names, dates, quotes) - for important information, check with the original source. How to speed up your local language model: tips and tricks MIT Technology Review analyzes the summer hype around AI Kyutai has released a Voice of Reason model for verbally solving math problems. Nokia opens access
package for AnyJev, a library for decision making without training. OpenAI introduced GPT-6 Sol and Luna with new API pricing, reducing the cost by 50%.
