Adaptive computation primitive theory: explaining AI's confabulation, prompt injection, and reward hacking with six primitives

A paper proposes the "Computational Primitive Theory (CPT)", arguing that all adaptive systems need to compute six primitive operations: Arouse, Orient, Valence, Position, Boundary, and Attune, with selection criteria of existential necessity, universality across independently evolved lineages, evolutionary conservation, and irreducibility. The framework is used in artificial systems to re-examine current challenges in machine intelligence, including confabulation, prompt injection, distractibility, reward hacking, and catastrophic forgetting, suggesting these problems may stem from missing computations. The authors emphasize that CPT is still a working hypothesis to be refined, requiring further examination through discussion, empirical testing, and application.