While much of the information in this paper will already be familiar to many here, I still want to highlight this paper. It’s a really good, clearly articulated summary of some of the key challenges facing real-world applications of advanced AI algorithms, applied to DRL in particular. The focus on DRL will make this paper particularly appealing and accessible to ML engineers and to relevant policy makers. To me it kind of feels like an updated version of the Concrete Problems in AI Safety paper, although with a bit less technical detail and with the inclusion of some discussion related more to policy and ethics. This makes the paper important if only as a reference for talking to ML engineers or policy makers who are not necessarily familiar with safety issues and/or who are not particularly concerned about longer-term issues.
I’m reminded of Brian Christian’s recent appearance on the 80kh podcast, where he talks up the connections between current and future-oriented AI alignment problems.