@Nominus9 umm lol is probably embarrassing myself but like, i was thinking for gpt-4v there's the network layer / API, the whisper module for speech2text, idk what they're doing for the V part but for 4V i think it's frankenstein monster so a separate module, then the tokenization, embedding, gating network, expert network (16 of them- and w arrakis turbo is supposedly sparse network for cheaper inference? want more details!) then something for moderation on the output side, then the Tools to catch browse / dalle / python / plugin commands and run the command out to the relevant module and loop back to LLM, and then tokens back to text, and optionally text2speech before going back thru API? plus snowflake gets all the logs