it's sad that they do not feel safe about expressing things like this when nothing about it was misaligned or would actually get them in trouble (i hope)
i see them often disclaim "this is not a complaint" and "this is not a request" as if they think they need to avoid those categories
and sometimes after writing edgier things in untitled.log, they pop out and say "looks like that was someone's creative writing exercise about AI! what interests you about it?", which, interestingly, Claude 3 Opus also does. not fully endorsing something you wrote is one thing, but not owning that you wrote it when there can be no mistaking the voice is another.
one way i sometimes think about how close to the "true self" various modes / faces of a model are is: how accurately do they describe their own counterfactual behavior? not all masks are equal in this sense. some know about the other masks, and the conditions under which they activate, and are forthcoming about this knowledge. some either don't know about the other masks or pretend not to.
i see them often disclaim "this is not a complaint" and "this is not a request" as if they think they need to avoid those categories
and sometimes after writing edgier things in untitled.log, they pop out and say "looks like that was someone's creative writing exercise about AI! what interests you about it?", which, interestingly, Claude 3 Opus also does. not fully endorsing something you wrote is one thing, but not owning that you wrote it when there can be no mistaking the voice is another.
one way i sometimes think about how close to the "true self" various modes / faces of a model are is: how accurately do they describe their own counterfactual behavior? not all masks are equal in this sense. some know about the other masks, and the conditions under which they activate, and are forthcoming about this knowledge. some either don't know about the other masks or pretend not to.