A datapoint which I found relevant: @voooooogel on twitter produced steering vectors for emergent misalignment in Qwen-Coder.
When applied with −10 multiplier, the steering vector produces emergent misalignment: https://x.com/voooooogel/status/1895614767433466147
+10 multiplier makes the model say ‘Certainly!’ or ‘Alright!’ a lot. https://x.com/voooooogel/status/1895734838390661185
One possible take here is that this steering vector controls for ‘user intent’ (+10) vs ‘model intent’ (-10). This seems consistent with the argument presented in the main post.
Current theme: default
Less Wrong (text)
Less Wrong (link)
Arrow keys: Next/previous image
Escape or click: Hide zoomed image
Space bar: Reset image size & position
Scroll to zoom in/out
(When zoomed in, drag to pan; double-click to close)
Keys shown in yellow (e.g., ]) are accesskeys, and require a browser-specific modifier key (or keys).
]
Keys shown in grey (e.g., ?) do not require any modifier keys.
?
Esc
h
f
a
m
v
c
r
q
t
u
o
,
.
/
s
n
e
;
Enter
[
\
k
i
l
=
-
0
′
1
2
3
4
5
6
7
8
9
→
↓
←
↑
Space
x
z
`
g
A datapoint which I found relevant: @voooooogel on twitter produced steering vectors for emergent misalignment in Qwen-Coder.
When applied with −10 multiplier, the steering vector produces emergent misalignment: https://x.com/voooooogel/status/1895614767433466147
+10 multiplier makes the model say ‘Certainly!’ or ‘Alright!’ a lot. https://x.com/voooooogel/status/1895734838390661185
One possible take here is that this steering vector controls for ‘user intent’ (+10) vs ‘model intent’ (-10). This seems consistent with the argument presented in the main post.