[jboggan]

This is neat, we did something a couple of years ago like this (but much cruder) by training off of Stack Overflow question and answer data. Actually we were trying to make a rubber ducky debugging chatbot but we didn't clean our inputs and it started spitting out code blocks when we asked it English questions.

With all the fuss over Autopilot have you considered the inverse problem? I trust myself to write code but often in an unfamiliar codebase I spend the majority of the time reading and deciphering what code does. Does the inverse work very well with your training set? I'd love to highlight code on the screen and get an English approximation of what it does. That's a truly useful helper.