enable the field · add one idea · watch every other relationship move
Six terms are present, but the page has not asked a model how they relate.
Semantic inference is optional and runs in this browser. Enabling it downloads an Apache-2.0 embedding model from Hugging Face: about 30 MB on the preferred WebGPU path and about 55 MB on the quantized WASM fallback. Typed concepts are sent only to the local worker and are never written to Warp experience memory.
felt
Adding one concept changed the motion and resting distance of terms you never touched. Even apparently opposed words could remain close because the model learned contextual similarity rather than human agreement.
found
An embedding does not contain a final map of meaning. It places language in a high-dimensional relational space; this experience turns those relative distances into temporary physical constraints.
carry
Search, recommendations, clustering, retrieval, and AI memory all make consequential decisions from representations of nearness. A similarity score becomes more legible when you can feel what it pulls together.