The Path to Self-Sustaining AI

Assessing AI’s Survival Capabilities in the Physical World

Abstract

In this study, I examined the technological challenges associated with and considerations necessary for achieving self-sustaining artificial intelligence (AI) systems capable of autonomously operating in the physical world. I explored the motivation behind AI's pursuit of long-term survival, whether as a design intention of humans or a capability autonomously developed by AI. By systematically categorizing and evaluating 21 self-sustaining technologies (SSTs) across five domains, namely maintenance and hardware assets, energy and resource management, object and sensory recognition, learning and adaptation, and communication and cooperation, this research highlights the complexity of enabling AI systems to maintain their existence without human intervention. By utilizing the common-sense knowledge of large language models (LLMs), I assessed the difficulty of realizing each SST with different levels of human support, including full, remote, and no support, both on Earth and in space. The findings suggest that although achieving complete self-sustainability for AI systems could take more than a century with current technology levels, strategic human assistance could significantly expedite this process. Notably, a secondary analysis utilizing the Gemini LLM revealed the potential for the operation of AI systems in space to accelerate the acquisition of SSTs.