char counts : each represented on 1-dimensional feature manifolds embedded with high curvature in low-dimensional subspaces of the residual stream
representing char count
ringing in manifolds == interference in superposition linear probees work well
1. SAE feature showing like tuning curve
hand picked SAE features showing tuning features
- notice early(layer 1 and 2) SAE features activated on word “called”
- checked their tuning -shady thing: no error bars? ring like architecture in flies: https://www.biorxiv.org/content/10.64898/2026.05.18.725766v1.full.pdf
2.
detecting boundary
QK matrix does it such that inner product is large for certain region
final decision - word or \n predicting new line
orthogonal subspaces?