Ive recently been tinkering with live visualization of local language models. including attention patterns, token probabilities, residual-state projections, and layerwise logit changes. Im curious which measurements people seem to find genuinely …