# Preface

*Reframing Superintelligence: Comprehensive AI Services as General Intelligence — K. Eric Drexler, FHI Technical Report #2019-1 (Markdown edition). Front/back matter. Rendered PDF: https://ora.ox.ac.uk/objects/uuid:9c05427a-6390-4b42-9c55-ee45f73a26ad*

---
The writing of this document was prompted by the growing gap between models that equate advanced AI with powerful agents and the emerging reality of advanced AI as an expanding set of capabilities (here, “services”) in which agency is optional. A service-centered perspective reframes both prospects for superintelligent-level AI and a context for studies of AI safety and strategy.

Taken as a whole, this work suggests that problems centered on *what high-level AI systems might choose to do* are relatively tractable, while implicitly highlighting questions of *what humans might choose to do with their capabilities*. This shift, in turn, highlights the potentially pivotal role of high-level AI in solving problems created by high-level AI technologies themselves.

The text was written and shared as a series of widely-read Google Docs released between December 2016 and November 2018, largely in response to discussions within the AI safety community. The organization of the present document reflects this origin: The sections share a common conceptual framework, yet address diverse, overlapping, and often loosely-coupled topics. The table of contents, titles, subheads, summaries, and internal links are structured to facilitate skimming by readers with different interests. The table of contents primarily of declarative sentences, and has been edited to read as an overview.

Several apologies are in order: A number of topics and examples assume a basic familiarity with deep-learning concepts and jargon, while much of the content assumes familiarity with concerns regarding artificial general intelligence circa 2016–18; some sections directly address concepts and concerns discussed in *Superintelligence* (Bostrom 2014). In this work, I have made little effort to assign proper scholarly credit to ideas: Concepts that seem natural, obvious, or familiar are best treated as latent community knowledge and very likely have uncited antecedents. Ideas that can reasonably be attributed to someone else probably should be. Finally, how I frame and describe basic concepts has shifted over time, and in the interests of early completion, I have made only a modest effort to harmonize terminology across the original documents. I thought it best to share the content without months of further delay.