loading scene data…
The dictionary that learned refusal
model
scene
induction
refusal
unlearning
emergence
SAE features
attribution graph
wikitext SAE (18)
refusal SAE (19)
ablate #
?
refusal feature (fires on harmful)
aligned but silent
other dictionary feature
REFUSED
answered
tile brightness = how hard it fires on this prompt
refusal SAE: feature #2673 lights up on harmful prompts
feature #
?
activation
—
❚❚
0.5×
1×
2×
4×