The Architecture of Human-Scale AI
Year: 2026
Type: preprint
Abstract: Artificial intelligence has been developed primarily through a strategy of centralized scale: larger models, more computation, more data, larger data centers, greater energy consumption, and increasingly concentrated infrastructure. That strategy has produced extraordinary advances in machine capability, but it is also generating increasingly visible constraints involving energy, infrastructure, latency, dependency, liability, resource competition, and the economics of further generalized capability. This paper argues that some of these constraints may be architectural rather than merely technological or regulatory. A centralized system attempting to provide increasingly generalized intelligence must absorb an expanding range of environments, edge cases, and consequential contexts into a common abstraction. As the marginal value of additional generalized capability declines relative to the resources required to obtain it, a different topology may become economically and operationally advantageous. The alternative proposed here is human-scale AI: a polynodal architecture in which intelligence is placed near the function it serves, bounded by externally defined operating constraints, and connected to humans at the point where consequential decisions occur. Central intelligence remains valuable where aggregation and scale create genuine surplus, but it no longer functions as a universal authority. Local systems retain contextual and epistemic authority while escalating information or problems when local resolution is insufficient. The paper develops several associated principles: human-local loci, bounded intelligence, consequence attenuation, adaptive homeostasis, local epistemic sovereignty, information movement according to function, confidence, and consequence, and AI as compensatory surplus attached to human capability. It proposes the Human Prime Directive from thermoeconomics as a descriptive first-order criterion for evaluating whether an AI architecture preserves or compresses human future option-space. 1The central proposition is simple: the next competitive advantage in artificial intelligence may come less from making one system universally capable than from designing intelligence around the functions, environments, and humans in which capability actually creates value.
Author: Leonard, Paul R.
Institution: Cascade Catalysts, Inc. (Canada)
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213316629