Each word's internal state at this layer, squashed to 2D. Words the model treats alike land close together. Low layers group spelling, high layers group meaning.
Switching a block off zeroes the scale shared by each group of 128 of its ternary weights, right in GPU memory, so the whole block counts as 0. Restore puts back the exact original bytes. Loss is the average surprise per token on a fixed paragraph of easy facts, so lower is smarter. The cloud keeps comparing against the intact model, so damage shows up as dots far from their usual level, and a switched-off MLP layer goes quiet (its ring turns orange).
Each layer of Bonsai has an MLP with 6,144 neurons. Showing all 147,456 would be a fog, so each layer shows the 192 that vary most on a short warm-up text. Their activations are copied off the GPU after every layer, for every word, and match PrismML's own llama.cpp to a correlation of 0.9999.
Height is the layer. Around each ring, a neuron's angle comes from which warm-up words make it fire (the first two principal components of its firing pattern), so neurons with similar tastes line up, even across layers. Distance from the centre is spread evenly so every dot is visible.
Colour is how far a neuron is from its own usual level on this word: orange is about 2 standard deviations or more above. Lines join each neuron to the neuron in the next layer it most often fires with, and turn orange when both fire. They show correlation, not wiring. The orange path follows the centre of the strongest firing in each layer.