How will that work out?

If it simply generates code from natural language then I am still fundamentally working with code. Aider as an example is useful for this, but anything that isn't a common function/component/class it falls apart even with flagship models.

If I actually put my "natural language code" under git then it'll lack specificity at compile time likely leading to large inconsistencies between versions. This is horrible user experience - like the random changes Excel makes every few years, but every week instead.

And everyone that has migrated a somewhat large database knows it isn't doable within minutes.