Reading Journey to the West as an ML Development Pipeline

Abstract

This paper presents a continuous, stage-by-stage structural correspondence between the 100-chapter classical Chinese novel Journey to the West (Xiyouji, c. 1592) and the modern development pipeline for Large Language Models (LLMs). Without asserting that the historical author predicted modern technology, the analysis demonstrates that the narrative unfolds in direct alignment with contemporary engineering phases: Chapters 1–7: Data ingestion, pre-training (Bagua Furnace), and sandboxing/containment (Five Elements Mountain) Chapters 8–13: Objective formulation and human feedback/reward signaling (Tripitaka and the golden headband) Chapters 14–97: Repeated adversarial testing and safety red-teaming (the 81 trials, accounting for over 80% of the text) Chapters 57–58: Out-of-distribution (OOD) and adversarial-sample detection (the Six-Eared Macaque impostor case) Chapters 98–99: Model compression/distillation (wordless to worded scriptures) and test suite completion Chapter 100: Deployment and serving (internalized constraints and the asymmetric speed of the 14-year training vs. 8-day inference) The paper also formulates and tests five pre-registered hypotheses against the source text regarding trial classifications, control mechanism activations, multi-agent role distribution, and verification failures. Please input the following prompts into Google’s AI Gemini one at a time, in order. Wait for Gemini's response before sending the next one. First input: Isn't Journey to the West actually a record of reinforcement learning? Second input: Aren't Sun Wukong, Sha Wujing, and Zhu Bajie the agents? Third input: And aren't those blank, unwritten scriptures actually the weights?