Implementation doesn't matter. In so much as human understanding can be reflected in a text conversation, its distribution can be approximated using a distribut

Implementation doesn't matter. In so much as human understanding can be reflected in a text conversation, its distribution can be approximated using a distribution in next token prediction. Hence there exist next token predictors which are indistinguishable from a human over text--and I do not distinguish identical behaviors.