This presentation focuses on the issue of artificial intelligence in the information society through the conceptual framework of the database. It introduces several different notions of databases — including the encyclopedic databases of the colonial era, material-based databases in archaeology, the technological databases of film history, and the databases of artificial intelligence — and examines these through the lens of the artist’s own research and creative practice. By tracing the historical development of databases, the presentation aims to explore the contemporary possibilities of the database as both concept and methodology.

A central component of the research is the Eye Filmmuseum’s archive of approximately 1,500 documentaries from the Dutch East Indies, produced between 1913 and 1945. These materials—ranging from family films to corporate and governmental records—provide diverse perspectives on filmmaking history, colonial life, and natural science.

From this archive, Hsu selected 250 copyright-free films and worked with engineers to train an AI model on this material. The AI segmented the footage into more than 30,000 individual shots and generated textual descriptions for each. These images and descriptions were then assembled into a dedicated database used to train a new model—an AI that “learns” from the visual culture of the Dutch East Indies.

This project approaches AI critically. The research also delves into the microstructure of AI hardware, particularly the properties of the CoWoS chip, a key technology in modern AI computing that enables chip stacking. Through this technology, the study proposes a new concept: the “stacked database.”

Beyond the physical stacking of chips, AI computation operates within an abstract high-dimensional mathematical space, in which all entities are decomposed into smaller units called tokens. Through a process known as embedding, these tokens are transformed into complex numerical vectors in multidimensional space. From this moment, the meanings and relationships of language are translated into positions and distances within mathematical space. The interpretive frameworks once structured by human keywords are thus dismantled, allowing heterogeneous databases to collide, merge, confront, and coexist — leading to a networked perspective on the multiplicity of databases.