carado.moe

tammy's blog about
AI alignment, utopia, anthropics, and more;

also on twitter, lesswrong, rss, and at

generalized adding reality layers

in predictablizing ethic deduplication, i talk about how when we don't know how reality works, we can task a singleton superintelligence with "adding a layer" to reality, which guarantees that inside that simulated reality we are able to function with known ethics.

in addition there's a sense in which, if one principle overrides the other, in general with arbitrarily many layers of reality-simulation we should tend to favor whichever option overrides. so for example: if our reality is actually on top of 1000 layers of reality simulations, then it only takes one of them to be (truly, deeply) deduplicating for our universe and any sub-universe we simulate to also have deduplication.

or, more precisely, for any set of mutually exclusive traits with a dominance ordering (such as deduplication > no-deduplication), we can expect the take one of those shapes:

(click on the image to expand)

i will call this the "generalized adding reality layers" (GARL) device, and i think it could have a broad use to reason about properties of the cosmos (the set of instantiated universes), even ones that might seem axiomatic and untestable.

for any set of mutually exclusive traits, we care about four properties:

• what the dominance ordering is between those traits
• how they affect the rate of spawning varied sub-universe
• how they affect the rate of spawning moral patient experiences
• how they affect the rate of spawning deeply caring actors

so, what other sets of traits can we examine using GARL ? here are some that i can think of off the top of my head, as well as my guess for the questions above..

question	dominance ordering	most varied sub-universes	most moral patient experiences	most deeply caring actors
moral patient deduplication	dedup > no-dedup	unaffected ?	no-dedup > dedup	i've no idea
infinite compute ¹	finite > infinite	infinite > finite	infinite > finite	infinite > finite ?
type of compute ¹	classical > quantum > hyper	hyper > quantum > classical ?	unknown	hyper > quantum > classical ?
moral realism ²	realism > non-realism	unsure	whichever maximizes good	realism > non-realism ?
deeply-caring superintelligence	present > absent	depends on its goals	depends on its goals	present > absent

¹: these two questions are similar to one another in that they have one dominant variant that restricts computation, and one recessive variant that doesn't; as a result, i would tend to assume that the recessive variant has a higher chance of spawning most kinds of stuff

²: my reasoning: once what is true becomes aligned with what is good, then the orthogonality thesis becomes falsified in that sub-cosmos, and superintelligences are more easily aligned by default

other questions to which GARL may be applicable but i haven't figured out how:

• is occam razor's/solomonoff induction applicable?
• are minimal circuits daemon-free?
• can you control the past?
• SIA vs SSA