It strikes me as clear that multimodal models will dominate this sort of robotic humanoid LLM interaction at some point -- tonality, pacing, body language etc. make up a drastic amount of human (sub) communication, and this needs to be considered. Neat that the one you had a chat with was a localized model, a fine tuned Llama variant I assume? I'd like to see your conversation with an improved/tweaked version and on a SoTA non-local LLM, just for comparison sake to get a closer gauge of how far we are out from genuine comfortable interaction.