At the moment it can recognise the type of statements in the training data set [1] and transpile them to Python, Java or C++ using the mappings defined here [2].
This is very different from how Codex/Autopilot work as it is trained using an NLU framework [3] which is usually used for training chatbots.
One use case could be that a complete beginner to programming can enter some natural commands (e.g. "X is a list of numbers", "read 10 values in X") without worrying about the syntax and convert them to a working program in one of the target languages.
Also, the input doesn't have to be in English as it can be easily trained with samples from other languages. So people can write/speak in their native language and get it translated to working code.