It makes sense to me that AI safety research speedup would lag behind capabilities. It seems like research taste would be much more important than simple algorithmic efficiency improvements where safety is concerned. You don’t know what you don’t know, and there is much we don’t understand about AI alignment. I find it highly unlikely that AI safety research can close that gap if we don’t institute some purposeful slowdown.
It makes sense to me that AI safety research speedup would lag behind capabilities. It seems like research taste would be much more important than simple algorithmic efficiency improvements where safety is concerned. You don’t know what you don’t know, and there is much we don’t understand about AI alignment. I find it highly unlikely that AI safety research can close that gap if we don’t institute some purposeful slowdown.