A small sentence-embedding model runs here, in your browser, and lights up the concepts your prompt reaches for.
Let a local LLM write the conceptsThe encoder can only light up concepts that are already in the diagram. A decoder can propose new ones. This needs WebGPU and downloads a real model — 194 MB, once, cached by the browser.
Space plays and pauses. ← → step a frame. The spiky ball is collective human knowledge; the smooth one inside it is what the model can recall.