Blogpost by Paul Christiano on AI Alignment and computer security:
“I am interested in the alignment problem: building powerful AI systems so that they are trying to do what we want them to do. I don’t have as much intrinsic interest in the security problem, of protecting AI systems from adversaries who want to manipulate their behavior.
That said, I’m starting to feel that working at the intersection of AI and security may be a good way to make progress on alignment, and that many problems in alignment might be naturally understood and approached as security problems. This post probably won’t say much new, but I wanted to explain why I think this.
My view on this topic was substantially influenced by Ian Goodfellow at OpenAI (who expects that if we are able to solve security problems then we will also be able to solve alignment problems) and to a lesser extent Peter Eckersley at EFF.”