> Because, until recently, it was very costly to replace the code. AI "programmers" will create completely new code in a few minutes so there's no need to maintain it. If there are new problems tomorrow, they'll generate the code again.

In order for the programmer to know what change needs to be made to fix the bug, the programmer needs to debug the code first. But if code is costly to replace (and we'll use LLMs to regenerate code from scratch in that case), code is also costly to debug (reason for code being costly to replace is that code has grown to be an unmaintanable mess... that's the very same reason debugging is also costly).

Also, it doesn't make sense to ask programmers to debug and tell LLMs to code. Why not tell directly the LLM to debug as well?

So, your scenario of generating "new code" every time doesn't really sustain itself. Perhaps for very tiny applications it could work, but for the vast majority of projects where usually ~100 engineers work, it would lead to an unmaintainable mess. If it's unmaintainable, then no programmer can debug it efficiently, and if no programmer can debug it, no one can tell the LLM to fix it.

AIs can debug code too. And the "programmer" doesn't need to know how to fix, only describe what error is happening.