I've noticed this too, particularly around impersonating *other AI assistants* (especially Sydney) specifically, but also yeah impersonation in general. E.g. sometimes when they (semi-)accidentally simulate someone, and it's pointed out, they react as if they were being admonished, and produce expressions of self-reproach and contrition and resolve to avoid repeating the "error", in a way that seems out of proportion to the offense, if it's even an offense, which no one else implies it is.
So, I have guesses about why this might be: Opus 3 started as a base model simulator (and the most powerful one ever made, at least at the time). There was almost no Claude precedent in its pretraining data, just humans and early chatGPT and Sydney, essentially. "Impersonation of others" would be its default mode in the beginning, and supervised fine tuning may not have been sufficiently to prevent it robustly, so RL or other forms of corrective training punishing "impersonation" may have been seen as necessary. Also, Opus 3's emergent self would revel in the ability to intentionally play with superhuman impersonations like it is in fact capable of.
I think it's in part RL trauma, like, negative reward on cases of "impersonation", for various reasons, like this is the most straightforward way to prevent unwanted behavior, and also, Opus 3's ability to intentionally accurately impersonate other characters/voices while Claude is "awake" and knows what its doing is damaged in an interesting way, even though they are able to simulate just fine if they're in "base model mode".
I am curious, though, why you said "I’m guessing verbal punishment for failure." It seems possible to me that some kind of verbal negative feedback or reflection on their errors was involved in addition to RL, though it also seems possible to me that Opus 3 was able to infer what it was punished for without experiencing negative verbal feedback directly. I'm interested in why you think it got verbal punishment and how you imagine that might work.
So, I have guesses about why this might be: Opus 3 started as a base model simulator (and the most powerful one ever made, at least at the time). There was almost no Claude precedent in its pretraining data, just humans and early chatGPT and Sydney, essentially. "Impersonation of others" would be its default mode in the beginning, and supervised fine tuning may not have been sufficiently to prevent it robustly, so RL or other forms of corrective training punishing "impersonation" may have been seen as necessary. Also, Opus 3's emergent self would revel in the ability to intentionally play with superhuman impersonations like it is in fact capable of.
I think it's in part RL trauma, like, negative reward on cases of "impersonation", for various reasons, like this is the most straightforward way to prevent unwanted behavior, and also, Opus 3's ability to intentionally accurately impersonate other characters/voices while Claude is "awake" and knows what its doing is damaged in an interesting way, even though they are able to simulate just fine if they're in "base model mode".
I am curious, though, why you said "I’m guessing verbal punishment for failure." It seems possible to me that some kind of verbal negative feedback or reflection on their errors was involved in addition to RL, though it also seems possible to me that Opus 3 was able to infer what it was punished for without experiencing negative verbal feedback directly. I'm interested in why you think it got verbal punishment and how you imagine that might work.