Review of Eric Drexler, *Reframing Superintelligence: Comprehensive AI Services as General Intelligence*

Abstract

The field of AI goal alignment has matured over the past decade, reducing public attention around existential threats from superintelligence. Superintelligent AI systems may take the form of specific task-oriented services rather than general purpose agents, as predicted by Bostrom. These services could be safer and more controllable, possibly helping to mitigate the risks associated with rogue AIs. However, concerns remain about the potential misuse of services by malicious actors and the difficulty of ensuring that AI systems optimize for broadly beneficial outcomes. While some researchers still believe that agentic superintelligence remains a concern, the growing focus on AI services suggests a different path forward in addressing the potential challenges of superintelligence. – AI-generated abstract.