When training data runs out, their usefulness will stop growing quickly. Why should their usefulness diminish?

Because they would not be up-to date with programming languages, tools, best practices etc.

May be there is some way to keep the model up-to date in less dramatic ways. But I think something gotto give..

I mean, even now the vibe coded stuff is reprehensible.