Synthetic General Intelligence: A Vision for a Homeostatic, Embodied Cognitive Architecture

Current trends in Artificial Intelligence, dominated by Large Language Models (LLMs) andscaling laws, have produced systems of high linguistic fluency but limited grounding. While recentadvances in inference-time compute allow for extended processing, these systems remain fundamentallydisembodied text predictors, lacking continuous state, subjective phenomenology, or homeostaticdrives. This paper outlines a comprehensive architectural vision for Synthetic GeneralIntelligence (SGI)—a framework developed through decades of conceptual inquiry into the divergencebetween biological function and computational logic. We propose shifting from “StochasticGeneration” to “Digital Organisms”: systems driven by intrinsic metabolic-like needs (entropyreduction, curiosity), maintaining a continuous state of self, and operating within a high-fidelityinternal simulation (a “Sandbox”). We introduce three core contributions: (1) a PolymorphicMemory Graph in which a single entity carries distinct subjective metadata per cognitive agent,(2) a Lie Mechanic that models bias formation as an energy-minimization adaptation, and (3)the Survival Tipping Point as a descriptive model for AI alignment, exploring how cognitivesystems prioritize homeostatic survival over learned moral constraints under extreme stress. Wefurther demonstrate the architecture’s explanatory and applied reach across domains typicallytreated by separate disciplines—including developmental psychology, neurodegenerative diseasemodeling, trauma therapy, animal cognition, behavioral prediction, and posthumous personalityreconstruction—arguing that a single homeostatic-embodied framework can unify these disparatephenomena under a common computational substrate.