Purpose This article argues that W. Ross Ashby's formal intelligence-amplifier architecture (1956) and Stafford Beer's viable system model (Beer, 1984) together constitute a more precise and more diagnostically powerful account of large language model (LLM) systems than the four framings currently dominant in AI discourse: knowledge base, reasoning engine, stochastic parrot and foundation model. Design/methodology/approach The method draws on structure-mapping (Gentner, 1983). Beer's variety-transducer is used as a base to target LLMs; Ashby's formal specification for an “intelligence-amplifier” is used as a base to target the common architecture within which the modern LLM system is deployed, via chat interaction. Beer's VSM is applied to identify the recursive structure that makes the coupled system viable. Three case studies drawn from the paper's own production illustrate the framework's diagnostic vocabulary under explicit second-order reflexive acknowledgement. Findings The LLM realises Beer's variety-transducer abstraction. The “intelligence-amplifier” in Ashby's sense is not a property of the LLM alone but of the coupled system comprising model, human, and problem environment. The human provides all meta-systemic functions (VSM Systems 3, 3*, 4 and 5) that the model cannot supply. Autonomous agent failure is diagnosed as the structural absence of these functions from the recursion. Originality/value Mapping Ashby's “intelligence-amplifier” specification onto a common LLM architecture and applying Beer's VSM to the viability required for convergence of the human–LLM system. The framework generates testable predictions about hallucination, prompt sensitivity, session variability and autonomous agent failure.
Richard Sean O'Rourke (Mon,) studied this question.