Updatable Meta-Weights Architecture: A Unified Path from Probabilistic Correlation to General Intelligence

What is the fundamental difference between human intelligence and today's large language models? Mainstream research oscillates between scaling up data (emergentists) and injecting explicit rules (symbolists), yet neither captures the core mechanism that allows humans to transcend pure statistical association. We propose the Updatable Meta-Weights Architecture (UMWA), a unified theoretical framework that defines intelligence as the interplay between a Base Experience Network (probabilistic association) and a Meta-Rules Verification Layer (high-priority constraints). Rather than treating physical laws and logical axioms as external symbolic rules, UMWA encodes them as high-weight, high-priority parameters—Meta-Weights—within the same parametric space, thus bridging the gap between connectionism and symbolism. This framework yields a precise operational definition of intelligence, a three-stage engineering model of creativity (Low-Probability Retrieval → Meta-Rule Verification → Experimental Validation), and a feasible roadmap to AGI: Directed Meta-Weights Installation, replacing blind scaling with the targeted injection of humanity's "civilizational compression package." We further derive three corollaries: (1) intelligence does not require subjective consciousness, (2) scientific falsifiability weights can discipline socio-ethical weights, and (3) AGI represents the first embodiment of collective human rationality—a societal System 2. UMWA provides both a testable theory of cognition and a practical blueprint for building provably aligned general intelligence.