An Implementation-Neutral Solution to the Normative ASI Superalignment Problem
Year: 2026
Type: preprint
Abstract: In order to maximise the net benefit of Artificial General Intelligence (AGI), and, in particular, agentic superintelligent AGI, a.k.a. Artificial Superintelligence (ASI), for all humanity, without favouring any subset thereof, we (1) imagine an intrinsically superintelligent agent $S$ that is both superaligned and supervalidated, (2) decompose the superalignment problem into normative and technical components, and (3) formulate a proposed (and, to the best of our knowledge, novel) implementation-neutral solution to the normative ASI superalignment problem. The net effect of our proposed solution (the $\mathbf{TTQ}$+$\mathbf{NAP}$ combination) is to reduce the (seemingly impossible) problem of building a superaligned and supervalidated agentic superintelligence $S$ to the (much easier) problem of building an "$\mathbf{NAP}$-compliant, goal-less, non-affecting superintelligent-agent-under-construction" $S^-$ (where $S$ = $S^-$ + $\mathbf{TTQ}$, and $S$ is supervalidated). Having, in this paper, both defined supervalidation conceptually and addressed normative superalignment, we defer technical superalignment and the details of supervalidation to later papers.
Author: Turner, Aaron
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225732496
Cites:
Cited by:
Related to: