ryan_greenblatt comments on Sycophancy to subterfuge: Investigating reward tampering in large language models