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.

These things are cool, but I find ultimately things like this end up just being a more natural language like programming language. There will generally be ways you have to phrase your query to get the algorithm to understand it. Having to make sure the interpreter can understand you, and debugging your query to me takes away all the advantage of a tool like this. You end up with syntax and keywords, all features of a language, but instead of getting them from a documentation page, you have to bash your head against the interpreter to figure them out.

Same thing with no-code platforms. They're great until you have to understand why the code generator spat out something that's throwing incomprehensible errors. For example, try debugging compiled typescript for an idea of how autogenerated code debugging works out.

Ask it for a polynomial time factorisation algorithm.

I’d love to see this work with voice input

good