Is it refuted that it wasn't sufficently part of the training data or did LLM manufacturers create more training data to answer questions like that correctly?