Eeeh, I spend less time writing code, but way more time reviewing and correcting it. I'm not sure I come ahead overall, but it does make development less boilerplaty and more high level, which leads to code that otherwise wouldn't have been written.

I wonder if you observe this when you use it in a domain you know well versus a domain you know less well.

I think LLM assistants help you become functional across a more broad context -- and completely agree that testing and reviewing becomes much, much more important.

E.g - a front end dev optimizing database queries, but also being given nonsensical query parameters that don't exist.

Oh yes, of course, if I don't know a domain well, I can't review it. That doesn't mean the LLM makes fewer mistakes there, though.

Curious, what do you count as senior work?

Roughly:

A senior can write, test, deploy, and possibly maintain a scalable microservice or similar sized project without significant hand-holding in a reasonable amount of time.

A junior might be able to write a method used by a class but is still learning significant portions and concepts either in the language, workflow orchestration, or infrastructure.

A principal knows how each microservice fits into the larger domain they service, whether they understand all services and all domains they serve.

A staff has significant principal understanding across many or all domains an organization uses, builds, and maintains.

AI code assistance help increase breadth and, with oversight, improve depth. One can move from the "T" shape to "V" shape skillset far easier, but one must never fully trust AI code assistants.

That sounds plausible if the senior did lots of simple coding tasks and moves that work to an agent. Then the senior basically has to be a team lead and do code reviews/qa.