Among tens of thousands of papyri that have survived in the sands of the Nile Valley, hundreds provide detailed evidence for a centralized system of public archives in the Roman province of Egypt. This project conducts a focused investigation of these institutions, examining their structure, function, role in administration and impact on documentary culture. The project aims to elucidate the specific characteristics of Roman archiving, posing questions of continuity and change with regard to the Ptolemaic period, and placing papyri in dialogue with other sources to test whether the archival institutions documented in Roman Egypt were generally present throughout the empire. 

The creation of records, their storage in archives, and the ability to retrieve them when needed, are three distinct phenomena that do not constitute a self-evident development, nor does the mere existence of archives necessarily signify their operational importance. Accordingly, this project investigates the dynamics of Roman archival institutions and the ability of institutional actors and private individuals to operationalize archival records. 

Hundreds of thousands of papyri and other documents on organic materials have survived from the ancient Mediterranean, but only 6% of this material has been deciphered so far, due to the difficult and time-consuming nature of this task. This project aims to channel the power of advanced AI to accelerate this process and unlock vast historical archives from antiquity.

In partnership with leading AI developers Mistral AI and SAIL Reply, the project aims to build cutting-edge multimodal Large Language Models (LLMs) for ancient languages and LLM-powered Vision-Language Models to propel the decipherment of ancient documents. We have started with Greek and created Apollo, the first LLM for Ancient Greek, trained on 600 million words of historical Greek and optimized for text restoration. The next step will be to train a Vision-Language model to transcribe masses of unread Greek papyri, of which the Austrian National Library holds one of the world's largest collections. We also intend to develop advanced AI systems for Latin and other ancient languages.

In the past decade, generative AI has been revolutionized by the rise of advanced Large Language Models (LLMs) based on a transformer architecture that have achieved high performance across a wide range of natural language tasks. In the last 3 years, LLMs have acquired multimodal capabilities to process visual inputs and texts embedded in images, including the decipherment of handwriting, opening new horizons for work with historical documents. 

Exploiting these breakthroughs in AI, this project builds the first state-of-the-art multimodal LLM for Ancient Greek. The model has been trained on a comprehensive dataset comprising more than 600 million words of premodern and early modern Greek, the largest machine-actionable corpus of historical Greek to date, sourced from the public domain. The AI system has been trained on chronologically ordered datasets to improve chronological awareness and has been optimized for text restoration, a key research task in working with damaged and fragmentary texts surviving from antiquity. 

Apollo is envisioned as an open digital research infrastructure capable of being improved through additional training and developed in multiple directions by the research community, including systems of Retrieval-Augmented-Generation (RAG) and Vision-Language Models (VLMs) for deciphering ancient documents, including as many as one million unread Ancient Greek papyri.