Inner alignment in artificial intelligence systems

• Grant number: 23/15356-7
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: March 01, 2024
• End date: August 31, 2024
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Jacques Wainer
• Grantee: Gabriel Antunes Rodrigues
• Host Institution: Instituto de Computação (IC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil

Abstract

In this work, we investigate the recently formalised inner alignment problem. In broad terms, to align an artificial intelligence is to construct or adjust it in such a way that its outputs are in accordance with human preferences. Internal alignment is a subtask within this exercise, in which the system is treated as an optimisation mechanism which is in turn optimised by some other optimiser. We aim to assess environments where inner alignment failures occur, as well as identify potential causes of this phenomenon. Our intention is to elucidate the nature of the problem, which we regard as important for the future of the field.