From my perspective LLMs shine in programming languages because they (languages) are invented by people to be more formal so they are more predictable. Even for syntactically different programming languages there are similarities we don't see with our own eyes, but the training has no problem detecting. My favorite example is when I tried to find how to represent an int array for reflection in Kotlin. Ask ChatGPT or a sufficiently large model about this, the answer will be IntArray::class.java , but try to find the exact line with google - few or nothing, more on github search in the sources, still not much. So LLMs "detected" the system of making type signatures in Java/Kotlin and were able to successfully predict because the rules are consistent. In human language it also works, but to a lesser degree, so if you give it a verb/subject pair that makes sense and ask for the equivalent ones, you will get some that still make sense but are literally absent in the full corpus of the web.
How will there be new programming languages after people stop practicing programming and the skills decay because an AI is performing the deliberate practice?