> these damn models are already so good at handling complexity, that they can actually work fine with a byzantine code base

Not really. They're only as good as whatever they are prompted for. If you prompt them for the wrong thing, they'll do it. Great tools, human error