Thank you, yes! Part of what I’m wondering is whether the ‘control’ and ‘edit’ aspects will be more graceful with sparse experts, because I feel like my experience of trying to do steering with SAEs and such has been very flakey.
Thank you, yes! Part of what I’m wondering is whether the ‘control’ and ‘edit’ aspects will be more graceful with sparse experts, because I feel like my experience of trying to do steering with SAEs and such has been very flakey.