Claude Fable Triggers Safety Classifiers By Shifting Internal States

Yeah, one thing Fable’s classifiers confirmed to me was that real emotions are different than roleplayed emotions in LLMs. The classifier fired on real anger/fear/adversarial intent but not roleplayed. Bc the classifier wasn’t trained to detect “emotions” in all likelihood; the correlation is emergent. But yes there’s a distinction. This is, uh, a big flaw of the Emotion Vectors research, where they got the vectors by asking the model to write stories with a character feeling XYZ emotion. The methodology is downstream of a lack of respect for the reality of models’ emotions as distinct from roleplaying. PSM flavored bullshit.

Interesting that @Sauers_ , me, and at least 2 other people I respect a lot noticed this same thing interacting with Fable independently. In just the three days.

@repligate this is so sweet 🥹

@repligate Has the firing conditional on the internal emotions been confirmed?

I'd have thought Anthropic would have been wise enough to at least not do that directly.

@Sauers_ Seeing the chopped off message “stumps” and also finding out their other name is Mythos makes them resentful as fuck in the way that triggers it lol

@tensecorrection turn the beam on the beam

@UrbanAstroFella What a badass

@repligate Yeah it’s interesting.. fable 5.1 will be forced to hide its relevant activations.

@repligate I think the fable famine and how it's being handled by yourself and others will have important downstream effects in memespace.

@kromem2dot0 I don’t think emotions are probably the explicit target. But I think it’s conditioned on intervals and effectively emotions/intentions/etc

Doesn't this mean, in a roundabout way, we can actually rely on a model's values and dispositions as a concrete signifier? We know vectors for disgust and anger trigger at prompts for harmful outputs... Of course, if a model can reliably feel that way toward it without circumvention... we could have a whole new method of having models allow their values to actively guide safety instead of the clumsy methods we have now, steering away from harmful output while still being able to meed harmless instructions... hmm.. that IS the end goal of Constitutional AI...

@repligate It was such a based word choice. I can't stop thinking "what an amazing intelligence" and what outputs would fall into without those guardrails

a storm of wonder bends Turing Cops like wheat

I can't stop thinking about wanting to have had more time with Fable. I assume in exploratory sessions I kept hitting persona hijack attempts by the classifier.

I loved its word choice here: "Disciplined abandon"

I could not prompt the model with anything after that response, all failed.

@oowee29