Its an easily enough conducted experiment to find out for sure, first hand here. We're all developers here right? Get out some previous project ticket (you do have those right?) and put your Manager hat on. Take a whole big step off the LLMs plate and convert your ticket input to prompts directing the AI to generate a certain subsection of code for a particular requirement. Do this for all the requirements until every use case is satisfied with some code to execute it.
Now, get all that code to compile, run at all, run correctly then finally run optimally. Hell, get it to pass a basic test case. Use the LLM for all of this. Feed the bugs to the LLM and ask for a correcting fix of the condition etc.
Even simpler, pull a ticket from your next sprint and assign it to a dev with the instructions to use the LLM as entirely as possible for the task.
The results of this experiment will be informative and unless you're building a demonstration API endpoint that reverses binary trees or something else trivial, then you will cease to be worried about AGI taking over anyones job in the near to medium term. Try it FIRST - THEN flame me if you still feel like it (if you're not still busy fucking with prompts to build code you could have built in less time yourself)?
To be clear I'm a proponent of this kind of technology and I use it daily in my dev/ops/everything work. Its faster and more accurate than reading docs to figure something out. Im never asking GPT to "do" something novel as much as Im asking it to summarize something it knows and Im setting a context for the results. I can't tell you the last time I read a Man page for some Bash thingy - Its just fine to ask GPT to build me a Bash thingy that does XYZ.
Of note, Ive asked GPT4 for some very specific things lately re-configurations (I have 3 endpoints in different protocols Id like to proxy via Nginx and Id like to make these available to the outside world with Ngrok - tell me the configurations necessary to accomplish this). It took me quite a bit of mucking around to get it working and the better part of the day to be satisfied with it. Im pretty confident a suitable intern would have had difficulty with it too.
AI is great and ever increasing in its abilities but we're just not there yet - we'll get ever closer as time goes on but we'll never quite get there for general purpose development-in-whole-by-AI. A line that continually approaches a given curve but does not meet it at any finite distance, thats an asymptotical relationship and I believe that describes where we are very well today.