are these clusterings early evidence of the models training themselves, i.e. passing on and amplifying their own tics to successor models?