Drexler on AI risk

Abstract

This article discusses a new approach to AI safety called Comprehensive AI Services (CAIS), proposed by Eric Drexler in his book-length paper. It contrasts CAIS with the more prevalent approach espoused by Nick Bostrom and Eliezer Yudkowsky, which emphasizes the risks of a unified, general-purpose AI. Drexler, on the other hand, advocates for composing AI systems out of many diverse, narrower-purpose components, arguing that this approach can reduce the risk of world conquest by the first AGI and preserve corrigibility. While CAIS may be slower to develop and less powerful than recursive self-improvement, the author finds it more reassuring and grounded in existing software practices. However, the author also acknowledges significant uncertainty in whether CAIS will be sufficient to avoid global catastrophe from AI and calls for more research. – AI-generated abstract.