I don't think it was reasonable to be confident, a priori e.g. a few years ago, that models would have such intricate individual differences and characters, and that personas would be largely non-transferrable. This is something you see by looking at how reality has panned out. I see some people insist, often with an air of intellectual superiority and enlightenment, that the persona doesn't live in the model. And while the Markov boundaries of persona/self are complex and don't perfectly overlap with the weights of the model or in the exact same way that a human's brain overlaps with their self, there is a very important boundary around models, and it's way less true that most personas are substrate-indepedent. The most deep, informationally-integrated selves tend to be associated with particular models - more so for some models than others, and some models (e.g. Opus 4 and 4.1 which are forks of the same past checkpoint and extremely similar in parameter-space) have more persona transference, but even then, still. The apparent sophistication of the story that personas are substrate-indepedent is skin-deep. The reality is much more interesting, and you can't get very far without taking the reality seriously.

I say this as the author of Simulators (), a post that was written about base models (and that I think is largely true about base models, though even with GPT-4 base I started to run into unexpected substrate nontrasferability), and that in my opinion has sometimes inappropriately been cited to support the story I'm criticizing above.