I wonder if you could teach millions of terminal sessions to an LLM and then have mosh run the inference algorithm to hide the latency (and eventually correct when it gets the real data)..