What thousands of conversations look like inside a language model
An interactive atlas visualizes internal representations of the Qwen3-1.7B language model, plotting residual-stream token states from thousands of English, Chinese, German, Croatian and Python examples in a reduced three-dimensional view. The author reports that the visible sheet held almost all 3D variation yet captured about 17.6% of full token-state variation, and that axes tracked language with topic structure still present.