Architektur für Antizipationskompetenz in KI-Systemen: Der Selbstvektor / Architecture for Anticipation Competence in AI Systems: The Self-Vector

Abstract

Current AI agent systems rely on mechanical status mechanisms (heartbeat, watchdog) that report the past but cannot anticipate the future. This paper argues that genuine anticipation competence requires a persistent, compact self-model — a self-vector — that dynamically weights information processing. A subtraction argument shows that self-awareness survives the loss of sensory input and emotion, making it fundamentally a network phenomenon. The scope of this argument is limited: embodiment history shapes the network (Merleau-Ponty, Damasio), and the enactivist objection (Thompson/Varela) remains. Recent Drosophila connectome simulations support the substrate-transferability of neural topologies, but not the transferability of all cognitive functions. From this analysis, a three-layer architecture is derived: a six-dimensional core as initialization prior, an experience-based emergent layer, and a reflection mechanism. A Heidegger-guided analysis of embodiment shows that substrate (material, interchangeable) and embodiment (interaction form, qualitative difference) are two distinct levels — an embodied self-vector models not merely information processing, but situatedness in a physical world. The paper works within a functionalist framework, transparently declared as a chosen position, and presents itself as a vision paper: formalization of the core functions and an empirical proof-of-concept remain as future work.