Rebuilding Ashby's Amplifier

Between 2021 and 2026 the practice of directing large language models was renamed four times (prompt, context, harness, and loop engineering), each announced as superseding the last. This paper argues that the four are not successive techniques but four components of one machine, built out one at a time: the coupled "intelligence-amplifier" W. Ross Ashby specified in 1956. The field that builds these systems is a descendant of Ashby's cybernetics; the 2024 Nobel Prize in Physics traces its own lineage to McCulloch and Pitts's neuron of 1943, and it has carried that neuron forward while forgetting the regulator that was its sibling. Read through the coupling, and held to Beer's 1984 standard for a model (that it describe, predict, diagnose, and prescribe), the outward migration is argued to reach a non-mechanisable terminus: the embodied evaluator. An assigned goal is information that must be transmitted to a system; a goal that is instance-specific and perspectival — knowable only from within the situation — can be effectively incompressible, exceeding what any model built into the machine can carry, so that it can be recognised but not specified, embodied but not sent, and the human who holds it cannot, on this account, be absorbed into the loop. The prediction is put to three major European ventures (world models, experiential learning, and open-ended self-improvement), each of which, on the framework, will succeed to the boundary of formalisable goals and stall there. What follows is a design prescription: build harnesses that defend the evaluator rather than eliminate it, and reclaim a regulation theory the field already possesses rather than pay, once more, to re-derive it.