A Night Sky Analogy That Explains How Large Language Models Actually Work

A developer and writer on DEV Community has proposed a new analogy for understanding large language models, comparing their internal semantic spaces to galaxies in the night sky. In the analogy, each galaxy represents a region of linguistic meaning, and a user's prompt acts as both a directional pointer and an entry point into that space. The author previously tried a dictionary analogy but found it inadequate, as dictionaries provide surface-level definitions rather than the relational, contextual patterns that allow LLMs to generate coherent text. Within any given semantic region, the possible combinations and navigational paths are described as practically infinite, despite the model's vocabulary being technically bounded. Two key factors — temperature and context — are identified as the primary forces shaping how a model moves through that semantic space, with temperature governing how far the model "hops" between probable next tokens.