You can’t cleanly cut one concept out of the weights. Because more features are stored than there are neurons, they share directions — superposition — so zeroing one neuron doesn’t remove one feature, it smears across many. Try to ablate the target and watch the collateral light up. Slide to pick which neuron you cut.
Superposition: with F features packed into D < F dimensions, each feature is a direction and each neuron carries a piece of MANY features. Ablating one coordinate (setting it to 0) therefore changes several feature read-outs at once — the damage is distributed, not local. Here 8 features live in 4 dims; cutting any single dimension moves more than one feature’s value by a measurable amount. A fail-loud self-check throws unless one ablation produces collateral on more than one feature. ◆ real (toy superposition), node-verified.
8-in-4 with random directions is a toy of Elhage et al.’s superposition; real models are far larger and features only APPROXIMATELY linear. The structural claim — distributed codes make clean single-concept removal impossible — is what ‘no clean removal’ and ablation-smearing name.