Runs a real 0.6B-parameter language model entirely in your browser with WebGPU, and shows every weight matrix and every intermediate activation, layer by layer, for each token it predicts.