What could possibly go wrong? dudette:are we still on for dinner with my mom? dude:sure, cant wait to see mrs.Plinketon again. dudette gets suggested reply m

What could possibly go wrong? dudette:are we still on for dinner with my mom? dude:sure, cant wait to see mrs.Plinketon again. dudette gets suggested reply message generated from one of dude's earlier emails to his friends: "mrs.Plinketon has great tits!" It is amazing to realize this type of technology can be bootstrapped only once (?)- as soon as you start getting automated replies in the system quality of learning on past data will drop, at some point snake will be eating his own automated reply tail. I imagine this system would benefit greatly from optional additional user voice input, something as simple as 1-3 spoken words fed back into Reply network to help with the reply intent decoding. You get email, NN generates 3 replies it thinks you would like, you have different idea for a reply, click microphone icon and say "no, maybe next week", Reply network rebuilds suggested answers with your input shaping reply vector.