> train models to write code

I've seen "deep dream". You'll get models that produce stuff that locally looks like code but doesn't actually make sense. Code-like extruded product. People will waste careers on the "steering" and "minor fixups" necessary to make it actually work which will be orders of magnitude larger than the work needed to just write the software in the first place.

That's the cost of giving up symbolic AI in favour of pattern recognition.