The breakthrough is probably going to be a GPT-N model. It needs a few tweaks to be good enough - for example to be able to handle complex projects. The glue code is always going to be an impossible problem from the point of view of language design, can't standardise, every case is different in a different way. Only a human or a language model are flexible enough to adapt to this kind of variance. We're going to love our glue generator and won't be able to understand how we used to write that code by hand.