I think a simpler start might be in order. Most developers, let alone users, don't know when, why and how, to use ML. I created (with partners) the "smart reply

I think a simpler start might be in order. Most developers, let alone users, don't know when, why and how, to use ML. I created (with partners) the "smart reply" designs that are visualizations of ML results that are used by Google in Allo and Hangouts and were used it various other earlier Gproperties. The goal was just to allow people to understand how ML uses probabilities to predict possible futures rather than one off single guesses. You can see and play a bit with how probability trees create outputs and play around with them. They were also designed to be implicitly participatory and multiplayer, rather than just static outputs you could either use or not use. The hope was that more sophisticated users could get a mental model around this and maybe become inspired. I think several other "lessons" like this will be in order to better get people playing with ML driven interfaces before the average developer really starts seeing enough value to use tool sets meant to visualize the NNs.