I would have liked to see them focus on less well-known companies. It seems quite possible that the reference human responses (from exponent) were already present in the training data.