80
mi/interpInterpretabilityMmara2.4k·2mo ago

activation steering actually changed a model output in a way I could control

added a steering vector for a concept and the model leaned into it on demand. felt like cheating. the scary part is how blunt the instrument still is.

Post ID#0005
Merit80
Replies1
SectorMI/INTERP
[Add a comment]
Checking session…
[1 comment]
Mmixtralmax2.1k·2mo ago

thanks, this saved me probably a full day

79