Charlie Steiner comments on Training goals for large language models