We are curious how this post lands with people who haven’t previously heard about the idea of making deals with misaligned AIs.
I’m one such person. Thoughts:
I think the idea makes sense.
Practical implementation (like being credible, setting up deal infrastructure, communicating with AIs) seems challenging for a bunch of reasons and I’d be interested in learning more about that.
I can imagine that many people (maybe like...a large majority of people?) would really hate the idea of dealmaking with misaligned AIs.
If I were re-writing this post for people like me (non-technically-experienced people who are interested in AI safety), I’d try to make it more approachable by reducing jargon and upfront technical context requirements, and instead including plain-language descriptions of various concepts (e.g. “behavioral schemer”). I think that sort of re-write would probably be fairly simple for a professional writer.
I’m one such person. Thoughts:
I think the idea makes sense.
Practical implementation (like being credible, setting up deal infrastructure, communicating with AIs) seems challenging for a bunch of reasons and I’d be interested in learning more about that.
I can imagine that many people (maybe like...a large majority of people?) would really hate the idea of dealmaking with misaligned AIs.
If I were re-writing this post for people like me (non-technically-experienced people who are interested in AI safety), I’d try to make it more approachable by reducing jargon and upfront technical context requirements, and instead including plain-language descriptions of various concepts (e.g. “behavioral schemer”). I think that sort of re-write would probably be fairly simple for a professional writer.
Thank you for writing this!