The puzzle didn't disappear; it moved to a higher level of abstraction.
LLMs aren't a layer of abstraction. An abstraction allows you to pretend there is nothing under it (except when it leaks). LLM coding is more like having a pair of hands that you control indirectly and aren't innately aware of what was been produced.
LLM coding assistants produce faster results whether anyone is being paid or not,
This is still not true when you account for externalities such as that the "faster" people haven't taken the time to actually understand what they've submitted and have produced more tech debt.