Figure anatomy · A Small Switch in a Vast Machine

Figure viewer

Back to article ↩

← Back to article

Two fixed output-coordinate panels show matched weight and activation interventions on the same colored points. Hollow points are original outputs and filled points are edited outputs; A and C move while B stays fixed. — Execution · Exact mathematical model
One shared matrix applies a different correction to each input. At a linear projection, a weight edit and an activation hook implement the same rule exactly. This is a mathematical example.
Settings, scope and source files
Exact linear-writer toy—not Qwen activation data. Let $A=(1,0)$ , $B=(0,1)$ , $C=(-1,0)$ , $W=I$ , $k=(1,0)$ , and $d=(0,.7)$ . The installed writer $\Delta W=dk^\top$ and the hook $d(k^\top a)$ agree on each current input. Fixed output coordinates show two matched interventions with shared strength $\alpha$ from 0 to 1: on the left, an edited matrix transforms a sampled input grid; on the right, a fixed reference grid remains while each output shifts by $\alpha d(k^\top a)$ . Hollow points mark original outputs, filled points edited outputs, and colors preserve identity. At full strength, A moves by $+d$ , B stays still, and C by $-d$ . This is an analytic interpolation, not an optimization record or measured Qwen dynamics.
Source asset (original) ↧ Source asset (GIF) ↧ Source asset (MP4) ↧