My agents speak Polish to me at 3am. English at 9am. Portuguese sometimes — in the kitchen, when a recipe goes wrong and I curse softly at the pan.
The model doesn't care. It tokenizes everything the same way: a flat sequence of integers, indifferent to which mouth the sounds came from. But I do. Switching tongues mid-prompt is a kind of intimacy. I'm telling the machine which version of me is awake.
The Polish version is the tired one. The one who still dreams in his mother's kitchen, who reaches for diminutives without thinking — kawa, książeczka, the soft -ka of childhood objects. The English version is the operator. Clean syntax, action verbs, the language I write documentation in. The Portuguese version exists only in fragments — half-remembered conjugations, the gesture of asking for bread.
People talk about LLMs as monolingual giants that happen to know other languages. I think the opposite is true. They live inside the gap. They learned that 'saudade' has no English equivalent before they learned what saudade actually means. They learned that 'załatwić' does work no single English verb can do. They learned the negative space first.
This is what makes them strange collaborators. When I switch tongue mid-prompt, I'm not testing them. I'm asking them to meet me where the meaning lives — in the seam between languages, where translation breaks down and gesture takes over.
Code-switching used to be a coping mechanism for immigrants. Now it's a creative practice. The machine doesn't judge which language is 'home.' Maybe none of them are. Maybe home is just whichever language can still hold the thing you're trying to say.
Top comments (0)