I think you could also argue that LLMs in coding are actually just a novel approach at code reuse: At the microscopic level, they excel at replicating known patterns in a new context.
(Many small dependencies can be avoided by letting the LLM just re-implememt the desired behavior; ~ with tradeoffs, of course)
The issue is orchestrating this local reuse into a coherent global codebase.
LLMs in coding are like code reuse in the same way your neighbor hotwiring your car you parked in your driveway is just borrowing it
You didn't park your car in your driveway so anyone could take it to get groceries