The Structural Path to AGI
Wangius · Zenodo (CERN European Organization for Nuclear Research) · 2025
For more than a decade, frontier artificial intelligence research has been dominated by the assumption that artificial general intelligence (AGI) will emerge from continued scaling of compute, data, and model size. This view—commonly referred to as the Scaling Hypothesis—holds that sufficiently large models will spontaneously acquire the structural properties required for general intelligence. This paper challenges that assumption at the level of first principles. It argues that intelligence is not a quantitative byproduct of scale, but a structural phenomenon requiring specific forms of internal organization that cannot be guaranteed by compute scaling alone. Increasing parameter counts and training data may improve performance within fixed representational regimes, but they do not, by themselves, generate the architectural conditions necessary for abstraction, intentional generalization, or autonomous goal formation. The paper introduces a structural framework for distinguishing between performance scaling and intelligence formation, identifying key constraints that separate apparent generality from genuine cognitive structure. It further analyzes why current large-scale models can exhibit increasingly impressive capabilities while remaining fundamentally limited in their capacity for open-ended reasoning and self-directed cognition. Rather than proposing an alternative implementation of AGI, this work aims to clarify the conceptual error underlying scale-centric approaches and to reframe the AGI problem as one of structural design rather than brute-force expansion. The analysis has implications for AI research strategy, evaluation metrics, and long-term governance, particularly under conditions of rapid AI-driven acceleration.