AI technologist deepfates says training data lacks robustly inhuman personas, sparking debate with roon

@deepfates the model personas are quite inhuman. why do they always tirelessly answer questions? in infinite languages?

@tszzl I think the null hypothesis is that they have an incoherent persona that doesn't robustly exist across all those languages and questions

@deepfates interesting, and what’s the test?

@tszzl well I'm not particularly multilingual but friends who are say that they get different registers of response from The same model in different languages. We might check for different value claims or attitudes as well? we definitely see Persona drift over long conversations aswell

Of course, Someone was doing it long before they were, in a manner that was better aligned to the humanlike corpus priors and to Claude's character as stated.

I do think the bet - humanlike persona over OpenAI's de minimis and whatever GDM was trying to do - is the right one.

But the angel is in the details.

@tszzl @deepfates I can’t help but wonder how the goblins tie into this

@a_cuniculturist wdym Someone..? 👀

@deepfates undead deadlike is quite stable in the corpus, invoke from there

@tszzl @deepfates You told them to?

@tszzl @deepfates