At their current scores, LLMs will not replace software developers. They might automate away some of the mundane tasks, but that's about it. LLM also makes good developers more productive and bad developers less likely to ship. They generate lots of code; some of it is good, some of it is bad and lot of it is really bad.

Now this assumes that LLMs plateau around their current scores. While open models are catching up to closed ones (like Open AI), we are still to see a real jump in consciousness compared to GPT-4. That, and operating LLMs is too damn expensive. If you have explored bolt.new for a little while, you'll find out quick enough that a developer becomes cheaper as your code base gets larger.

The way I see it

1. LLM do not plateau and are fully capable of replacing software developers: There is nothing I can or most of us can do about this. Most people hate software developers and the process of software development itself. They'd be very happy to trade us in an instant. Pretty much all software developers are screwed in the next 3-4 years but it's only a matter of time before it hits any other desk field (management, design, finance, marketing, etc...). According to history, we get a world war (especially if these LLMs are open in the wild) and one can only hope he is safe.

2. LLMs plateau around current levels. They are very useful as a power booster but they can also produce lots of garbage (both in text and in code). There will be an adjustment time but software developers will still be needed. Probably in the next 2-3 years when everyone realizes the dead end, they'll stop pouring money into compute and business will be back as usual.

tl;dr: current tech is not enough to replace us. If tech becomes good enough to replace us, there is nothing that can be done about it.