I remember when I was of legal age (okay, well, of kinda somewhat legal age) and I started going to bars. Who teaches you how to behave at a bar? How do you know how much to tip a bartender if at all ("$1 a drink, usually, in the US") or what tabs are ("you pay when you leave") -- is it weird if you're there by yourself? (Sometimes.) When do you buy a round for everyone? (Sometimes.) When is it okay to check your phone? (It depends.) There are all these rules of socialization that we internalize without thinking, practicing them until they become reflex. Different bars in different places have different rules, and we wordlessly sensitize ourselves to each arena.
But even before we enter bar culture, we get socialized at a much more basic level -- in the art of conversation. How do you know when it's your turn to talk? When can you make a joke? When can I leave a conversation?
My prototyping process has involved a lot of linguistics research along these lines, mainly focusing on an old (now somewhat irrelevant?) branch of applied linguistics called "conversation analysis." It might be really hard to teach an AI just how to time its responses and get into the rhythm of things, but there are 5 year olds who effortlessly achieve gapless conversation on a daily basis. I find that fascinating -- and where there's elegance and an element of timing, there's strategy and a game.
To be clear, my goal isn't to solve "procedural conversation generation" in any way, but rather to sidestep it. Convo is NOT about "what" you say or "how" you say it -- it's mostly about "when" you say it.
I'll post more about Convo as I develop it.
