I program in Python, Fortran, and R. LLMs may be better at Python than Fortran since there is more training data for the former, but they are still pretty good at Fortran. LLMs make it easier to use less popular languages and much easier to use the more popular ones. An advantage of using LLMs for a compiled language like Fortran is that you can iterate with compiler messages until the code compiles. A Python script to do this is at https://github.com/Beliavsky/Groq-Fortran-agent. I assume Cursor and Windsurf do this too.

Language adoption maybe, but not language development.

Language dev requires precise semantics. If a user decides that types or other academic nonsense are overrated, they can get away with it. You can be cavalier and just let your application crash - as long as there's a precise formulation underneath you about what 'crashing' means: executing `finally` blocks, rolling back the stack and capturing a stack trace, following an exceptional continuation, etc.

You can't YOLO an Algorithm W, an ANF transformation, memory models or register allocation. Language design has to be just right. I don't think language design selects for people who look at a "95% correct" LLM benchmark and think "wow!". A program generated by a 95%-correct compiler will probably run catastrophically wrong in about 100% of executions.

    so people will hesitate to adopt a new language even if it has great features

Tools like LSPs etc are almost always behind the leading-edge of development; all that means is that the "red squiggles go brrrrrr" crowd doesn't adopt things right away.