It's quite the contrary, the less interpretative the language, the better. And no, LLMs were not trained on English to begin with. And they don't perform best in English.

Please expand more on the idea that LLM's are not trained on English to begin with. Not sure what you mean by this as clearly many LLM's are trained on data that contains a lot of English. For instance GPT-1 seems to have been trained on a purely English corpus.

Interesting, where do you think LLMs perform the best?