Google Explores How Large Language Models Actually Translate
Just 1.4% of large language model PaLM’s natural language exposure is “unintentional consumption” of bilingual signals — which seems to be enough to contribute to the model’s translation abilities.