SAE steering — the research
Steer the model
Compose a feature-injection formula and watch the model’s answer change — no weight edits, just vectors added to the residual stream while it generates.
model
Recipes — one-click formulas that reliably work
Layers
Features · layer —
no matches
Formula (4)
layer 7
600% · ≈6/1
#9982α +6
layer 11
600% · ≈6/1
#7871α +6
layer 15
600% · ≈6/1
#12143α +6
layer 19
600% · ≈6/1
#2726α +6
Live meters are estimates and under-read when features reinforce — the exact per-layer value is reported after Run. Past ~40% of a layer’s residual norm the output tends to break.
The heatmap shows each injected feature’s endogenous activation at each token — token-local resonance, not causal contribution. A feature can be causally load-bearing upstream and still read ~0 here; to test causality, remove it from the formula and re-run.