When it comes to my outlook on the job market, I don't concern myself with
change, but time-tested
demand.
For example: Have debit cards fundamentally changed the way that buying an apple works? No. There are people who want to eat an apple, and there are people who want to sell apples. The means to purchase that may be slightly more convoluted, or standardized, or whatever you might call it, but the core aspects remain exactly as they have for as long as people have been selling food.
So then, what demand changes with LLM writing code? If we assume that it can begin to write even a quarter way decent code for complex, boutique issues, then central problems will still remain: New products need to be implemented in ways that clients cannot implement themselves. Those products will still have various novel features need to be built. They will still also have products that will have serious technical issues that need to be debugged and reworked to the clients specifications. I don't see LLM being able to do that for most scenarios, and doubly so for niche ones. Virtually anyone who builds software for clients will at some point or another end up creating a product that falls into a very specific use-case, for one specific client, either because of budget concerns, restraints, bizarre demands, specific concerns, internal policy changes, or any other plethora of thing.
Imagine for example that there is a client that works in financing and taxes, but knows virtually nothing about how to describe what they need some custom piece of software to do. Them attempting to write a few sentences into a tarted up search engine isn't going to help if they don't have the vocabulary and background knowledge to specify the circumstance and design objectives. This is literally what SWE's are. SWE's are translators! They implement general ideals described by clients into firm implementations using technical knowhow. If you cannot describe what you need to an LLM, you have the same problem as if there were no LLM to begin with. "I need tax software that loads really fast, and is easy to use." isn't going to help, no matter how good the LLM is.
Granted, perhaps those companies can get one employee to dual hat and kind of implement some sloppy half-baked solution, but...that's also a thing that you will run into right now. There are plenty of clients who know something about shell scripts, or took a coding class a few years back and want to move into SWE but are stuck in a different industry for the time being. They aren't eating my lunch now, so what would have us believe that this would change just because the method of how a computer program might become slightly more "touchy feely"? Some people are invested in software design. Some are not. The ones who are not just want the thing to do what its supposed to do without a lot of time investment or effort. The last thing they want to do is trying to work out what aspect of US Tax law its hallucinating.
As for the companies making the LLM's, I don't see them having the slightest interest in offering support for some niche piece of software the company itself didn't make - they don't have a contract, and they don't want to deal with the fallout. I see the LLM company wanting to focus on making their LLM better for broader topics, and figuring out how to maximize profit, while minimizing costs. Hiring a bunch of staff to support unknown, strange and niche programs made by their customers is the exact opposite of that strategy.
Honestly, if anything, I see there being more people that are needed for the SWE industry, simply because there are going to be a glut of wonky, LLM-generated software out there. I imagine Web Developers have been pretty accustomed to dealing with this type of thing as far as trying to work with companies that are trying to transition out of WYSIWYG website implementations. I haven't had to deal with it too much myself, but my guess is that the standard advice is that it's easier and quicker to burn it to the ground and build anew. Assuming that is the case, LLM-Generated software is basically...what? Geocities? Anglefire? Sucks for the client, but is great for SWE's as far as job security is concerned.