Mapping Semantic Meaning Onto the Night Sky

Short summary

The author uses the night sky as an extended analogy for how large language models navigate semantic space. Each 'galaxy' represents a region of meaning, and a prompt sets both the direction and entry point. Temperature controls hop distance between tokens—low temperature follows the most probable path, while higher temperature enables exploratory leaps to less likely but potentially creative regions.

• •Night sky analogy maps LLM semantic space as galaxies of meaning
• •Prompts set direction and entry point into a semantic region
• •Temperature controls traversal: low = deterministic hops, high = exploratory leaps

Generated with AI, which can make mistakes.