This is a very interesting approach to conversational AI via growing dataflow programs that (among other things) claims that "Task-oriented dialogue is interactive programming": https://www.microsoft.com/en-us/research/blog/dialogue-as-dataflow-a-new-approach-to-conversational-ai/

Great to see them putting an emphasis on trying to gracefully correct inevitable mistakes. Very practical and hype-free blog. I'm interested in how much builtin knowledge they are going to go for ("...including pieces like here and now that are implicitly available from the beginning of the conversation..."). Natural language is absolutely laced with assumed knowledge. Will it be able to understand "my wife", "my sister", "my house" etc. What about the gradient of niche or local knowledge - "Boxing Day" etc? The problem with starting down this road is it is essentially arbitrary where you stop, and the user won't know where that is. I suppose it won't matter if the error correction is strong enough ("When you say "sister" do you mean Jane.Doe@outlook.com?")

@Chris Knott What you're pointing to here is one of the fundamental problems of natural language understanding: to really understand language one must understand the world in which the speaker lives, including the agents who live there and their motivations. One thing I find interesting about this paper in particular for FoC is that what they're doing in the personal assistant context could repurpose well to the narrower context of descriptive programming.

Back when I was an assistant for conversational agent research, it was disheartening: having a good natural langauge parser, but not much there there deciding what to conversationally. (This was well before the ML resurgence.)