goals for emergency unaligned AI

posted on 2022-03-22

in a previous post i talk about killing timelines — by making an AI fill them with something that's an easy to implement goal, such as paperclips — to avoid the even worse outcome of them becoming S-lines. in this post i wonder: can we do better than paperclips?

if the True Nature of the cosmos is a universal program, then there could be some things to turn timelines into that consume less compute cycles; for example, maybe we somehow make an AI that makes its timeline run as little compute as possible. the cosmic universal program will then spend less many cycles running those dead timelines, and more running remaining live timelines — making survivors in them possibly "more real". in this sense, it may be that pruning timelines can be done without causing astronomical waste: compute time and therefore kinda "realness" or "soul juice" are redistributed to remaining timelines.

even if the "naive" most likely "implementation" of our universe consumes just as much compute regardless of what goes on in it, the universal computation will contain other "implementations" that "compress" compressible timeline-states, and we will reclaim cycles in them — and if they are good at compressing, those implementations might be where most of our realness juice is located anyways.

another possibility is to task an AI with turning its timeline into a state that is as identical as it can be to another timeline in which said AI's didn't appear. if it can achieve that, then we can kill timelines by replacing them with copies of alive timelines. this also recycles compute time, possibly more efficiently since it doesn't rely on compressing implementations.

posted on 2022-03-22