Plain text: 07-sec04-prescriptive-descriptive.md · on GitHub

§4 The AI-services model includes both descriptive and prescriptive aspects

Reframing Superintelligence: Comprehensive AI Services as General Intelligence — K. Eric Drexler, FHI Technical Report #2019-1 (Markdown edition). Section 4 of 40. Source label: prescriptive-descriptive. Rendered PDF: https://ora.ox.ac.uk/objects/uuid:9c05427a-6390-4b42-9c55-ee45f73a26ad

Thesis. The AI-services model both describes the architecture of current and prospective high-level AI applications, and prescribes patterns of development that can foster safety without impeding the speed and efficiency of AI development.

Cites sections: §1 R&D automation provides the most direct path to an intelligence explosion; §3 To understand AI prospects, focus on services, not implementations; §6 A system of AI services is not equivalent to a utility maximizing agent; §10 R&D automation dissociates recursive improvement from AI agency; §11 Potential AGI-enabling technologies also enable comprehensive AI services; §12 AGI agents offer no compelling value; §14 The AI-services model brings ample risks; §20 Collusion among superintelligent oracles can readily be avoided; §24 Human oversight need not impede fast, recursive AI technology improvement

Cited by sections: §II

Summary

Does the AI-services model describe prospects for high-level AI application development, or prescribe strategies for avoiding classic AGI-agent risks? Description and prescription are closely aligned: The services model describes current and emerging patterns of AI application development, notes these patterns are accessible, scalable, and align with AI safety, and accordingly prescribes deliberate adherence to these patterns. The alignment between descriptive and prescriptive aspects of the services model is fortunate, because strategies for AI safety will be more readily adopted if they align with, rather than impede, the momentum of AI development.

The AI-services model describes current AI development

AI technology today advances through increasingly automated AI research and development [fn: §10 R&D automation dissociates recursive improvement from AI agency.], and produces applications that provide services [fn: §3 To understand AI prospects, focus on services, not implementations.], performing tasks such as translating languages, steering cars, recognizing faces, and beating Go masters. AI development itself employs a growing range of AI services, including architecture search, hyperparameter search, and training set development.

AI-service development scales to comprehensive, SI-level services

The “AI services” concept scales to sets of services that perform an asymptotically-comprehensive range of tasks, while AI-supported automation of AI R&D automation scales to asymptotically-recursive, potentially swift technology improvement. Because systems based on AI services (including service-development services) scale to a superintelligent level [fn: §1 R&D automation provides the most direct path to an intelligence explosion.], the potential scope of AI services subsumes the instrumental functionality [fn: §12 AGI agents offer no compelling value.] that might otherwise motivate the development of AGI agents. [fn: §11 Potential AGI-enabling technologies also enable comprehensive AI services.]

Adherence to the AI-services model aligns with AI safety

Because the AI-services model naturally employs diversity, competition, and adversarial goals [fn: §20 Collusion among superintelligent oracles can readily be avoided.] (e.g., proposers vs. critics) among service-providers, architectures that adhere to the (extraordinarily flexible) AI-services model can readily avoid classic risks associated with superintelligent, self-modifying, utility-maximizing agents.

Adherence to the AI-services model seems desirable, natural, and practical

There need be no technological discontinuities on the way to thorough AI R&D automation and comprehensive AI services, and continued adherence to this model is compatible with efficient development and application of AI capabilities. [fn: §24 Human oversight need not impede fast, recursive AI technology improvement.] Traditional models of general AI capabilities, centered on AGI agents, seem more difficult to implement, more risky, and no more valuable in application. Accordingly, guidelines that prescribe adherence to the AI-services model [fn: §24 Human oversight need not impede fast, recursive AI technology improvement.] could improve [fn: §14 The AI-services model brings ample risks.] prospects for a safe path to superintelligent-level AI without seeking to impede the momentum of competitive AI development.