The alignment problem from a deep learning perspective