I only have 3 points against LLMs: they lack reason and they can't count.

'emer ge' is two tokens, 'emergency' is one. The models think in a logosyllabic language.