It’s not just about how many examples of a language exist in the LLM’s distribution though, dynamic typing is a poor choice if you’re using LLMs for software in my view because it turns compile-time errors into runtime errors which will only be caught if the LLM writes a suitable test; an area that they still leave a fair bit to be desired in.

The whole virtue of dynamic typing is that it reduces cognitive load on the programmer, if you’re using LLMs there’s no reason to do that in my opinion, you’re offloading a bunch of cognitive load to begin with so expressiveness and constraint is more important than human ergonomics. Yes there’s loads of JS in an LLM corpus but a lot of it is going to be poor examples, because it’s generally quite easy to write JS badly. If you use typescript you’re at least forcing a few common hallucinations to produce compile-time errors which the model has immediate feedback for, and you’re giving the model more context on what it’s seeing.