Maybe you can predict emergent abilities post hoc, but I recall no one predicting beforehand that, for example, a pure language model could do translation simply be giving a prompt that said “translate to French”