Well this is the magic of LLMs, they learn things incidentally well, and specialised models are pretty average.