The Real Explanation of GPT: A Structural Decomposition of the Large Language Model Black Box

This paper offers a structural decomposition of the black box of GPT and large language models. It does not explain GPT’s capability by a single parameter, a single attention head, a single feature, or simple next-token prediction. Nor does it treat hallucination, drift, and false attribution as merely a lack of knowledge. Instead, the paper explains how large language models form their capability boundaries through external training, textual sedimentation, path compression, prompt-driven activation, and generative continuation, and how the same structure produces both their strengths and their characteristic failures. The paper argues that a large language model is not a database of stored knowledge entries, but a language sedimentation system shaped by large-scale text training. The training objective supplies direction, loss marks difference, gradient and optimizer adjust paths, RLHF or DPO shapes preference, prompt provides the current task, and system prompt sets the boundary. Large-scale text training repeatedly compresses paths from previous text to later text, question to answer, context to response, format to completion, and task to fulfillment. These repeated paths form deeply familiar generative channels. GPT is powerful because these channels have been deeply sedimented: it can write, translate, program, imitate styles, continue tasks, and transfer across languages and task forms. The same structure also explains its failures. GPT can generate fact-like forms, but it cannot stably distinguish external facts, contextual materials, internal inference, and self-generated completion. It can maintain textual continuity, but this is not the same as stable current state. It can use tools, but this is not the same as forming an internal action loop. It can adjust output according to prompts, but this is not the same as having its own intention. In this framework, hallucination is not an accidental error, but a structural consequence of generation being stronger than source tracking. Drift is not merely a lack of context length, but the result of textual continuity not being equivalent to stable current state. False attribution is not merely a failure of factual retrieval, but a consequence of lacking reliable internal source stratification. The significance of this paper is that it opens the black box of GPT and explains it as a generative system shaped by external training, supported by sedimented language paths, activated by prompts, and not yet internally closed. The framework yields testable implications: interventions on FFN/MLP, cross-layer residual paths, attention boundaries, source tracking training, current state maintenance, and online learning mechanisms should affect capability, hallucination, drift, and task stability in different ways. Therefore, the key to next-generation AI is not only larger models, longer context, and more tools, but the gradual formation of stronger source tracking, current state maintenance, feedback loops, and self-updating capacity.