# Latent Space The high-dimensional internal representation space where a model encodes information. Every concept, word, image, or piece of data the model processes is mapped to a point in this space. Similar concepts cluster together; related concepts are connected by directions. "King" minus "man" plus "woman" ≈ "queen" — the classic word-embedding example — works because gender and royalty are represented as directions in latent space. ## Why the metaphor matters "Latent space" has become the term of art for discussing what AI models "know" without committing to whether they truly understand anything. It's a neutral framing: the model has representations in latent space, those representations have structure, that structure captures something about the relationships between concepts. Whether that constitutes understanding, knowledge, or a [[World Model]] is a separate (and much more contested) question. For [[Diffusion Models]] (image generators), latent space has a particularly vivid interpretation: the model navigates a space where every point corresponds to a possible image. Generating an image means finding the right point; interpolating between points morphs one image into another. The concept of "exploring latent space" — systematically traversing this landscape to discover what the model can produce — has entered creative-AI vocabulary. [[Interpretability]] research is fundamentally about reading latent space — understanding what the model's internal representations encode, how they're organized, and what features or circuits produce specific behaviors. The geography of latent space is the map of the model's capabilities and biases. ## See Also - [[Embeddings]] — the coordinates in latent space - [[World Model]] — what latent-space structure might constitute - [[Interpretability]] — reading the structure of latent space - [[Diffusion Models]] — latent-space navigation for image generation