I think that's the wrong way to look at it.

Let's frame it this way. English is not the world's best language. It's pretty bad, honestly. It makes little logical sense, pronunciation is all over the place, and it's inconsistent even between native speakers. Yet like 2 billion people speak it, even in places where it's not the native language, because the UK spent so long as the dominant world power and just saturated all international discourse long enough to make it the most convenient common tongue. And so English becomes the most commonly used language for international discourse in the EU, despite the EU having just one member state (Ireland) where English is the majority, because it's too inconvenient to switch to anything else.

Programming languages can fall into the same trap. LLMs today can have the majority of their code trained on a small set of popular languages. They'll be likelier to produce that kind of code reliably, which in turn motivates vibe coders to prioritize those languages over other options that may be more purpose-built or appropriate for the need.

A new programming language that is massively better, more efficient, and easier to use can come about, but an LLM might never excel at it. Basically, a new language precludes itself from success with LLMs. The LLM will suck at it because there is substantially less training data to reliably model from. There will never be enough training data because fewer people are using it. Fewer people are using it because shitty vibe coders just rely on what the LLM can do well. The cycle repeats.

I think that LLM's ability to understand any language, programming or otherwise, is almost entirely a factor of how much training data it can get its claws on. In the cases of new programming languages, it probably won't be able to do much of anything.

I don't think this is linearly correlated. Despite being called a "Language Model", it does not mean it processes language as humans do. If an LLM is good at supporting you with a programming language mainly depends on the amount of available training data.

Let's take esoteric languages as an example - there are languages that only work via weird Unicode symbols or other cryptic commands. A human will have a hard time to understand that language, the LLM may not have any problem at all to give working comprehensive examples (as in will be useful).

That's a shit take, it's assuming that the LLM has the same thought process for learning a programming language instead of being autocorrect no steroids.

I'm pretty sure LLMs will be shit at Lisp for the foreseeable future just because the language is sort of created by the programmer.

There's more to languages than ease of use. D is often cited having a "poor library infrastructure", by those who leave it. They often deal with Rust, Go, or even C/C++ instead of writing their own libraries and/or library bindings.