@coolboymew cakecnn admitted they did a bit of a trick. they analyzed the emojis in messages and used emoji emission in the model to instruct the puppet to change to some emote state.
so basically just do the same that everyone else does: record request/responses in a big decision tree, pretend to be an AI for a while by just pulling the neural fingerprint closest response, and then use the data you harvest to fine tune mixtral 7/8B :ablobcatpopcorn: