果冻传媒

Topic 

The varieties of value alignment

Abstract

There are two approaches to the value alignment problem, namely, the problem of how to guarantee that the actions and decisions of artificial agents will align with our values. I first examine the most general version of this problem, and show that although humans are never perfectly aligned鈥攐therwise, we would not be unique or free鈥攖here is enough alignment to allow for reliable communication and sufficiently coherent ethical guidance. This fortunate combination of alignment and diversity is explained by the resemblance between our representational and emotional interests and needs, which we must satisfy in similar ways.

The role of joint attention is key in explaining how we end up sufficiently aligned in a non-lucky and non-arbitrary way. I then explore two ways of modelling alignment in AI agents. One of them is by predicting patterns by data compression, which is our current model. The other focuses on joint routines to create habits of attention that satisfy urgent needs and make the agent vulnerable. I argue that the second model is the only realistic model of value alignment, and that contemporary models should acknowledge this limitation as we look for better models. In fact, current models make value alignment, when it happens, either lucky or arbitrary.

About the speaker

Carlos Montemayor is a Professor of Philosophy at San Francisco State University. With a background in Law, Philosophy, and Cognitive Science, he conducts highly interdisciplinary research at the intersection of Legal Epistemology, Philosophy of Mind, Cognitive Psychology, and Artificial Intelligence. His recent work explores the notions and prospects of AI consciousness and agency. In his latest book, The Prospect of a Humanitarian Artificial Intelligence, he integrates insights from these topics to offer a comprehensive analysis of the alignment problem in AI from both epistemic and ethical perspectives.

Find more info on 

Your host

Carlos Zednik, Assistent Professor (tenured) for Philosophy of AI, will host Professor Carlos Montemayor of San Francisco State University.

Registration is required but free of charge.