So far LLMs scale the opposite of silicon, instead of exponential gains in efficiency/density/cost, each generation of model takes exponentially more resources.

Also, being a software developer is not primarily writing code, these days it is much more about operating and maintaining production services and long running projects. An LLM can spit out a web app that you think works, but when it goes wrong or you need to do a database migration, you're going to want someone who can actually understand the code and debug the issue.