Answers are grounded in this paper and cite their sources.
Most approaches to agent safety model threat detection: what is dangerous? This paper argues that framing is half finished. A robust homeostatic architecture must also ask: what makes this okay? What reduces uncertainty? What signals safety? And beyond that: what produces joy? What earns approach rather than avoidance? Potato, a local AI agent running on a MacBook Air M4 with an iPhone 16 Pro Max providing GPS and motion sensor data, implements a fear state computed from real hardware sensors and a social modulation layer in which the presence of a known trusted face reduces computed fear. This paper describes that architecture and then argues for a fundamental reframing: fear and joy are not opposites. They are the two poles of a single continuous affective variable that runs from 0.0 to 1.0, where 0.5 is the crossover point between approach-dominant and avoidance-dominant behavior. There is no neutral. Indifference is not a real state. It is either ignorance of a stimulus or complete familiarity that has compressed it to background. Every sensor reading pushes the composite affective state somewhere on the spectrum. The architecture already produces both poles. This paper names them, formalizes the joy inputs that already exist in the system, and describes the behavioral outputs below 0.5 that the fear-only framing has been missing. The central argument: fear and joy are the same architecture. Building only one of them is building half an organism.
No takes yet. Share an insight, caveat, or question.
Brian Riggleman (2026) studied this question.