Ean Huddleston

Abstract

Large language models (LLMs) exhibit general-purpose reasoning even when trained on text alone. This raises a fundamental question: could language be not merely a tool for expressing preexisting thought, but a constitutive component of general intelligence? Here I address this question by analyzing what open-ended causal modeling demands of any finite system. These demands expose a core architectural tension: maintaining an adaptive world model requires graded, continuous plasticity, whereas constructing novel causal chains requires systematic, sequential commitment. From these competing constraints, I derive Emergent Symbolic Cognition (ESC), a computational theory proposing that general intelligence arises when a continuous, adaptive substrate—whether biological or artificial—internalizes the relational structure of a symbolic framework like language. This internalization serves a dual function: it organizes the substrate’s continuous geometry into an abstract world model, and it supplies the composable elements for iterative symbolic generation. As each predicted symbol feeds back to reshape the internal state, it acts as a discrete ratchet that converts parallel associative processing into the sequential, step-by-step construction required to solve novel, dependency-structured problems. Because ESC is derived from substrate-agnostic constraints, a single architectural principle accounts for both the unique generality of human intelligence and the emergent capabilities of LLMs. From this architecture, I derive falsifiable predictions (including a sharp decay in the neural or hidden-state decodability of downstream symbols) and evaluate the theory against converging evidence from cognitive science and artificial intelligence—explaining why consolidated reasoning can operate without active language use, and showing how symbolic input alone can ground world models without direct sensory contact. Ultimately, ESC reframes language as a culturally inherited architecture for general-purpose cognition.