Figure 4 — SAE feature explorer · loads on demand
What does each L10 neuron fire on?
▶ Load the explorer (~115 KB)
runs entirely in your browser
neuron
Each point = one of 500 reference images, placed by its full SAE code (2-D PCA). Color = the chosen neuron's activation; hover for the caption. Thumbnails are example inputs — click to drop one onto the map (★).