@aaronpk Itâs actually a really interesting hybridg eng/AI approach: they build an ML model from past Zoom interactions to create a simulacrum of each user which speaks on their behalf. What youâre hearing is just a generative model of the person youâre âtalkingâ to saying what theyâd be likely to say in real life; same on their end. Itâs *incredibly* bandwidth-efficient. The protocol just sends a tiny presence packet with your user ID once per second to keep your avatar talking.