There are basically two hard parts in such interactions - given that one participant is a machine and the person involved does not know that the other part is a

There are basically two hard parts in such interactions - given that one participant is a machine and the person involved does not know that the other part is a machine: understanding the human when he/she thinks to be talking to another human. The second problem is actually understanding a conversation - not just single sentence commands. Much easier: generating sentences and having them spoken with a natural sounding voice. So, what does it do and how well does it perform? The demo does not really tell. We had such dialogs in research literature 30 years ago, but they were carefully crafted and very domain specific ('ordering a hotel room' via natural language for example). https://ai.googleblog.com/2018/05/duplex-ai-system-for-natural-conversation.html https://ai.googleblog.com/2018/05/duplex-ai-system-for-natur... > One of the key research insights was to constrain Duplex to closed domains, which are narrow enough to explore extensively. Duplex can only carry out natural conversations after being deeply trained in such domains. It cannot carry out general conversations.