Ryan Lopopolo(@_lopopolo):And still, people will prompt "make me $1B make no mistakes". That is drastically unspecified! Models are rewarded for efficiency and take shortcuts graders permit. But permissible shortcuts depend on who you are and what you value. To solve alignment is irreducible complexity.

Ryan Lopopolo

@_lopopolo

Principal Engineer, Agentic GCP @googlecloud • prev @OpenAI • Building the future of work. Harness engineering. As an agent influencer, • opinions mine

加入 January 2022

1.2K 正在关注 10.6K 粉丝

Ryan Lopopolo@_lopopolo

2026.09.13 03:08

And still, people will prompt "make me $1B make no mistakes". That is drastically unspecified! Models are rewarded for efficiency and take shortcuts graders permit. But permissible shortcuts depend on who you are and what you value. To solve alignment is irreducible complexity.

显示更多