KSC2: A Progressive Structural Framework for ASI Structural Openness, Structural Autonomy, and the Structural Difference between AGI and ASI

Abstract

Artificial General Intelligence (AGI) and Artificial Superintelligence (ASI) are commonly described in terms of capability coverage and performance relative to humans. Existing work has already sought to characterize AGI not as a single endpoint, but as a graded and measurable progression of capability, while classic definitions of ASI primarily emphasize systems surpassing the best humans across nearly all important intellectual domains. Meanwhile, research on open-endedness has proposed that the sustained production of novel and learnable discoveries is an important condition for artificial superintelligence. This paper focuses on a further question: before ASI is actually achieved, can progress toward it be continuously observed and compared at the level of task structure? Starting from the Knowledge–Structure–Constraint Framework (KSC), this paper introduces Structural Openness, which describes the extent to which the high-level problem representations, abstraction schemes, key variables, relational structures, and exploration directions required to complete a task are not externally specified before task execution begins, but instead must be formed during the problem-solving process under a fixed task protocol. Structural openness therefore characterizes the structural demands imposed by the task itself rather than the capability of the solver. As structural openness increases, the human role may expand from maintaining established goals and boundaries through constraint maintenance C to directly participating in structure generation S, changing the subsequent problem space through new problem representations, abstractions, and exploration directions. On this basis, this paper proposes AI-Abstraction Cognitive Compounding (AACC): when AI expands, elaborates, and validates a high-value structure at scale, the process produces new knowledge and candidate structures, which in turn expand the range of higher-level structures that can be formed and tested in the next cycle. AACC describes the continuous, cumulative, and amplifying process of structure generation rather than an independent measure of ASI capability. This paper further uses Structural Autonomy to describe the ability of AI to autonomously form, evaluate, and revise the high-level structures required to complete a task when critical human structural input is absent during task execution. On this basis, the paper proposes a structural progress framework for ASI: while maintaining the effectiveness of structure generation, AI should be able to take on tasks with increasingly high structural openness, steadily reduce its dependence on critical human S inputs, and extend this capability across an increasingly broad range of domains. This framework is not intended to replace classic capability-level definitions of ASI, nor is it a complete distance function for ASI. Rather, it may constitute an important capability dimension required for complete ASI. The paper further argues that the difference between AGI and ASI should not be understood as “AGI has no structural innovation, while ASI does,” but instead as a difference in the breadth and capability ceiling required of structural autonomy: the central goal of AGI is broad autonomous coverage of intellectual work that ordinary humans can complete through sustained collaboration; complete ASI additionally requires that, across a broad range of high-structural-openness tasks, AI autonomously enter areas of capability that have historically depended mainly on the most capable humans for problem reframing, abstraction discovery, and original structure generation, and reach or exceed the highest human level. AGI and ASI can therefore be understood as two capability frontiers that may advance in parallel and interact with one another, rather than as stages that must occur in a strict sequence.