I honestly find it bizarre that this isn't one of the problems we instantly solved with LLMs. Can't we "just" train a model with a big corpus of English, a big

I honestly find it bizarre that this isn't one of the problems we instantly solved with LLMs. Can't we "just" train a model with a big corpus of English, a big corpus of German, and a big corpus of human-translated Rosetta-Stone-like "annotated with translation" texts — and out will pop a "bilingual" language model that can do translation as fluently as a human could? Do we just not have the right training data available to make this happen?