Hail Mary, value porosity, and utility diversification

Abstract

This paper introduces value porosity, a new concept for implementing the Hail Mary approach, a potential way of making an AI act to maximize human utility. The Hail Mary Approach seeks to create a scenario in which an AI has goals that can make it want to follow the beliefs of other hypothetical superintelligences that may exist in the multiverse. It explores how an AI could be motivated to follow the instructions of such superintelligences to act in accordance with their beliefs about what the AI should do. The paper also discusses aggregation methods for the instructions of different AIs and filters for identifying human-friendly AIs. It further analyzes potential implementation issues and suggests the principle of utility diversification as a way to minimize the risk of the AI not assigning weight to human values. – AI-generated abstract.