Smarter example selection cuts the cost of translating to languages almost nobody speaks

Researchers tested whether large language models can translate into ten truly low-resource languages by showing them lots of translation examples at once. It turns out that picking examples carefully matters more than just dumping more examples in — 50 well-chosen examples work about as well as 250 random ones, and 250 chosen ones match 1,000 random ones.

Translation for languages with almost no training data has always been expensive because the model needs either mountains of examples or expensive human annotation. If example selection is the real lever, not raw quantity, it means smaller teams and poorer countries can actually build functional translation systems without supercomputer budgets. The cost floor just dropped — the question is whether anyone actually uses it.