The Thought Cloud places each artifact — an essay, visualization, paper, or piece of software — on a two-dimensional map of concepts. An embedding model converts a paragraph describing each artifact into a high-dimensional vector, representing a location in the space of concepts. This is the same basic way LLMs represent concepts.
Several broad conceptual categories are also embedded as anchors. The anchors are placed on the plane by projecting their high-dimensional embedding vectors onto the two directions of greatest variation across topics using principal-component analysis. Each artifact then sits at a weighted average of the anchor positions, with weights given by the cosine similarity between the artifact's vector and each anchor's. The result is a cloud where artifacts on similar topics cluster organically.