Abstract
Traditional engineering methods rely on the decomposition and recomposition of systems, a process that assumes components have clearly partitioned inputs and outputs. However, Artificial Intelligence (AI) systems often influence their environments in ways that, in turn, influence the systems themselves. This environmental coupling challenges the assumption of independence required for traditional component-level engineering. This paper proposes two new general systems precepts—the Core and the Periphery—to provide a theoretical foundation for engineering general intelligence. By leveraging abstract systems theory and Ashby’s Law of Requisite Variety, these precepts allow engineers to model AI not as a standalone component, but as a regulatory relationship within a closed system.
The Breakdown of Traditional Decomposition
In classical systems engineering, the "V-model" serves as the standard: a complex problem is decomposed into smaller, independent functional components. These components are built and tested in isolation, then recomposed into a whole. This methodology succeeds because we can typically draw a hard boundary between a component's inputs and its outputs.
Artificial General Intelligence (AGI) breaks this model. Because an AGI is designed to interact dynamically with its surroundings, the boundary between the system and its environment dissolves. This "environmental coupling" creates a feedback loop where the system's outputs eventually become its own inputs. When inputs and outputs cannot be strictly partitioned, traditional decomposition fails. The result is a significant gap in AI engineering: while we can specify the functional requirements of a Deep Neural Network (DNN), we struggle to guarantee specific outcomes because those outcomes are a product of a shifting, complex relationship with the world.
Theoretical Foundations: Requisite Variety
To bridge this gap, the authors look toward Cybernetics and Abstract Systems Theory, specifically Ashby’s Law of Requisite Variety. This law states that for a system to successfully regulate another (such as an AI regulating its environment to achieve a goal), the "variety"—or the internal complexity and range of actions—of the regulator must match or exceed the variety of the disturbances it faces.
Under this lens, intelligence is not viewed as an inherent property of a machine, but as a property of the relation between the system and its environment. Engineering general intelligence, therefore, requires a framework that can account for this relationship without relying on the false assumption of component independence.
The Core and Periphery Framework
The paper introduces two formal precepts to replace traditional open-system decomposition:
- The Core ($C$): These are the parts of a system’s structure that remain invariant over time. The Core represents the stable, "mechanical" aspects of the system where traditional engineering principles still apply.
- The Periphery ($P$): This represents the "residual change"—the parts of the system that shift, adapt, or grow in response to environmental stimuli.
Mathematically, the authors define these based on the set differences of system states over time. If $S$ represents the system and $R$ represents the residual change (adaptation), the Core is defined as $C = \bar{S} \setminus R$, while the Periphery is $P = R$.
This distinction allows engineers to track how a system adapts without needing to assume that its components are independent. It provides a way to analyze "blocking"—how a system uses its internal variety (the Periphery) to cancel out environmental disturbances to maintain the stability of the Core and achieve desired outcomes.
Application to Deep Learning
This framework offers a new way to conceptualize modern AI architectures. In a Deep Neural Network, the Core consists of the universal functions and architectural invariants, such as the backpropagation algorithm or the specific layer-to-layer information passing structures. These are the elements designed by engineers using traditional methods.
The Periphery, conversely, consists of the specific weights and the training data. These elements are environment-dependent and do not follow "universal" engineering rules; they are the result of the system’s adaptation to its surroundings. By separating these two, engineers can better understand which parts of the system can be guaranteed through design and which parts emerge through environmental interaction.
Conclusion and Future Directions
The authors conclude that while traditional engineering is a precept for "open systems," the Core and Periphery are precepts for "closed systems"—where the system and environment are treated as a single, coupled unit. This shift is essential for the safe and predictable development of AGI.
However, several hurdles remain. The framework requires empirical validation to see if it can effectively isolate functions in real-world intelligent systems. Furthermore, while the concepts are formally defined, a more rigorous mathematical elaboration involving specific theorems is still needed. Finally, the authors note a "boundary paradox": even if boundaries dissolve in reality, engineers still need to define temporary boundaries to perform measurements. Future research into the "dynamics of boundaries" will be critical for the practical application of these precepts.
Summary updated