What Truly Makes GPT Work Is Not Computation？A Structural Explanation of Large Language Model Question-Answer Capability

Large language models are commonly understood as systems driven primarily by real-time computation. This paper proposes an alternative explanatory framework. It argues that what truly participates in GPT’s core question-answer capability is not primarily real-time computation itself, but structural sedimentation gradually formed during training. The paper distinguishes two functionally different categories of computation inside large language models. The first contributes to training through parameter updates, error backpropagation, and repeated historical accumulation, gradually forming stable structure, historical bias, and long-term sedimentation. The second contributes during inference, where Attention mechanisms determine which historical structures enter the present receptive scope. The paper further argues that Attention operates more closely as receptive-field selection rather than the central mechanism responsible for question-answer capability. Question-answer capability is proposed to arise primarily from accumulated historical structure, together with familiar paths, common structures, and law-continuation already formed inside the Sedimentation-Field. The framework is summarized in one proposition: “Look there — sediment there.” The paper further discusses possible relations among training structure, historical influence, receptive selection, and language question-answer capability.