Pour une histoire de l’IA #16 : David Berry

La seizième séance du séminaire Pour une histoire de l’IA accueillait David Berry le 01 octobre 2024 pour une séance intitulée Constitutional Alignment: Critical Methods for Machine Learning and Artificial Intelligence. Ci-dessous, un court synopsis de la séance ainsi que la biographie de notre invité :
Constitutional Alignment: Critical Methods for Machine Learning and Artificial Intelligence
“Alignment”, put simply is an attempt to align a computer model with norms and values and towards a defined end. The key question is how to install values into a computer system such that they reflect the norms of a society or “humanity.” Should they be prescribed into the system (creating a set value constellation) or should the system learn the best values (a value constellation that it can adapt)? One approach we might call Forward Alignment (alignment training) produces trained systems based on alignment requirements. The second case we might call Backward Alignment (alignment refinement) ensures the practical alignment of trained systems and revises alignment requirements. In this talk I want to explore these theorisations of alignment as a concept, but also how they are operationalised in practice. I argue that concept formation around the notion of alignment reveals interesting elements of AI system design and potentially provides a critical case study suggesting new methods for critiquing machine learning. Although we are still developing methods for working with these systems at a granular level, we are beginning to have some sense of what good approaches are. AI/ML offers a new challenge for research, not only in the complexity of the stack but also in the seeming absence of an identifiable research-object beyond the prompt interface in and of itself. By developing approaches through the notion of alignment, probes can be used to engender new methods for understanding how these LLM systems are developed, shaped and controlled.
David M. Berry is Professor of Digital Humanities at the University of Sussex. He publishes widely on knowledge formations, algorithms, code and software and his most recent book is Digital Humanities: Knowledge and Critique in a Digital Age (Polity). He is currently working on a book titled Artificial Intelligence and Critical Theory (MUP) and another co-written book on the history of Joseph Weizenbaum’s ELIZA software titled Please Go On: The Invention of ELIZA (MIT Press). His latest research project is a history of the Plate Glass universities.
OpenEdition vous propose de citer ce billet de la manière suivante :
Carla Marand (10 décembre 2024). Pour une histoire de l’IA #16 : David Berry. CulturIA. Consulté le 14 février 2025 à l’adresse https://doi.org/10.58079/12w1y