Alignment of Agentic Systems

In this article, we present a novel bottom-up approach to aligning agentic systems. Additionally,we introduce a training algorithm called projective learning, inspired by how the brain likely learns.To start, we describe two seemingly universal neuronal models, capable of modeling a broad rangeof phenomena, with the aim of modeling consciousness in our systems.We then develop projective learning, a mechanism that utilizes the simple action of strengtheningrepeated connections and pruning neglected ones to form new concepts. Later we illustrate howprojective learning can also be used for finding causes.In the last and most central section, we outline a neural architecture for generating preference-based actions through a few tunable weights.