Mostly, but not necessarily. The preservation of some properties, not all or most properties. One could imagine the AI preserving the safety-relevant aspects but radically changing everything else.
I also worry that ‘high fidelity copying’ connotes some outside system doing the copying, which would miss the point entirely. The difficulty of tiling isn’t about the difficulty of copying; the central difficulty is about trusting something as intelligent or more intelligent than yourself; trusting something which you can’t predict in detail, and therefore have to trust on general principles (such as understanding its goals).
So, maybe “selective robust alignment-preserving reproduction” as a propertyof your aligned agent (which may be smarter than you, and may create agents smarter than itself)
Just for clarification, does it make sense to interpret “tiling” in the sense you are using to mean something akin to “high fidelity copying”?
Mostly, but not necessarily. The preservation of some properties, not all or most properties. One could imagine the AI preserving the safety-relevant aspects but radically changing everything else.
I also worry that ‘high fidelity copying’ connotes some outside system doing the copying, which would miss the point entirely. The difficulty of tiling isn’t about the difficulty of copying; the central difficulty is about trusting something as intelligent or more intelligent than yourself; trusting something which you can’t predict in detail, and therefore have to trust on general principles (such as understanding its goals).
So, maybe “selective robust alignment-preserving reproduction” as a propertyof your aligned agent (which may be smarter than you, and may create agents smarter than itself)