> Rather than compressing raster images of people speaking, we could send a model as an initial payload and subsequent instructions on how to manipulate it as t

> Rather than compressing raster images of people speaking, we could send a model as an initial payload and subsequent instructions on how to manipulate it as the conversation progresses. What about latency though? Would that improve?