The Electric Power Research Institute quizzed large language models' ability to handle issues confronting U.S. power systems. It didn't go well.