I think fully automated LLM code generation is an inherently flawed concept, unless the entire software ecosystem is automated and self-generating. I think if you carry out that line of thought to its extreme, you'd essentially need a single Skynet like AI that controls and manages all programming languages, packages, computer networks internally. And that's probably going to remain a sci-fi scenario.
Due to a training-lag, LLMs usually don't get the memo when a package gets updated. When these updates happen to patch security flaws and the like, people who uncritically push LLM-generated code are going to get burned. Software moves too fast for history-dependent AI.
The conceit of fully integrating all needed information in a single AI system is unrealistic. Serious SWE projects, that attempt to solve a novel problem or outperform existing solutions, require a sort of conjectural, visionary and experimental mindset that won't find existing answers in training data. So LLMs will get good at generating the billionth to-do app but nothing boundary pushing. We're going to need skilled people on the bleeding edge. Small comfort, because most people working in the industry are not geniuses, but there is also a reflexive property to the whole dynamic. LLMs open up a new space of application possibilities which are not represented in existing training data so I feel like you could position yourself comfortably by getting on board with startups that are actually applying these new technologies creatively. Ironically, LLMs are trained on last-gen code, so they obsolete yesterday's jobs. But you won't find any training data for solutions which have not been invented yet. So ironically AI will create a niche for new application development which is not served by AI.
Already if you try to use LLMs for help on some of the new LLM frameworks that came out recently like LangChain or Autogen etc, it is far less helpful than on something that has a long tailed distribution in the training data. (And these frameworks get updated constantly, which feeds into my last point about training-lag).
This entire deep learning paradigm of AI is not able solve problems creatively. When it tries to it "hallucinates".
Finally, I still think a knowledgable, articulate developer PLUS AI will consistently outperform an AI MINUS a knowledgable, articulate developer. More emphasis may shift onto "problem formulation", getting good at writing half natural language, half code pseudo-code prompts and working with the models conversationally.
There's a real problem too with model collapse, as AI generated code becomes more common, you remove the tails of the distribution, resulting in more generic code without a human touch. There's only so many cycles of retraining on this regurgitated data you can create before you start encountering not just diminishing returns, but damage the model. So I think LLMs will be self-limiting.
So all in all I think LLMs will make it harder to be a mediocre programmer who can just coast by doing highly standardized janitorial work, but it will create more value if you are trying to do something interesting. What that means for jobs is a mixed picture. Fewer boring, but still paying jobs, but maybe more work to tackle new problems.
I think only programmers understand the nuances of their field however and people on the business side are going to just look at their expense spreadsheets and charts, and will probably oversimplify and overestimate. But that could self-correct and they might eventually concede they're going to have to hire developers.
In summary, the idea that LLMs will completely take over coding logically entails an AI system that completely contains the entire software ecosystem within itself, and writes and maintains every endpoint. This is science fiction. Training lag is a real limitation since software moves too fast to constantly retrain on the latest updates. AI itself creates a new class of interesting applications that are not represented in the training data, which means there's room for human devs at the bleeding edge.
If you got into programming just because it promised to be a steady, well-paying job, but have no real interest in it, AI might come for you. But if you are actually interested in the subject and understand that not all problems have been solved, there's still work to be done. And unless we get a whole new paradigm of AI that is not data-dependent, and can generate new knowledge whole cloth, I wouldn't be too worried. And if that does happen, too, the whole economy might change and we won't care about dinky little jobs.