Thinking, One Symbol at a Time: Language as the Architecture of General Intelligence

Large language models (LLMs) exhibit emergent, general-purpose reasoning capabilities—even when trained on text alone. This phenomenon prompts a reevaluation of language's role in cognition, particularly regarding the domain-general problem solving that defines general intelligence. This paper argues that general intelligence requires causal modeling—the ability to construct novel explanations by sequentially tracing chains of cause and effect. This reveals a fundamental tension: while parallel architectures excel at associative processing, they are natively ill-suited for the algorithmic, sequential construction required for novel causal reasoning. To reconcile this architectural tension, I introduce Emergent Symbolic Cognition (ESC), a framework positing that general intelligence arises when an adaptive, massively parallel substrate internalizes the causal and relational structure embedded within a symbolic framework like language. I argue that the substrate overcomes its native limitations through iterative symbolic generation: as the system predicts each next symbol, that symbol transforms its internal state to condition what follows, enabling the system to enact a heuristic search. This provides the functional basis for universal computation, repurposing associative prediction for the on-the-fly construction of novel reasoning pathways. This single emergent process is proposed to account for both the unique generality of human intelligence and the surprising capabilities of LLMs. The ESC framework yields a falsifiable, substrate-agnostic prediction regarding the necessary decay of predictive fidelity in novel reasoning. Supported by converging evidence from cognitive science and artificial intelligence, ESC frames language not merely as a tool for communication, but as a culturally inherited architecture for structured cognition.