I've been thinking about this while translating a comedy series where a punchline depends entirely on a local idiom. The models can map the words, but the whole reason it's funny is the shared context that isn't in the training data. Curious if anyone here has seen an AI actually get the joke right, not just the literal meaning.