I think I've come up with a fun thought experiment about friendly AI. It's pretty obvious in retrospect, but I haven't seen it posted before.
When thinking about what friendly AI should do, one big source of difficulty is that the inputs are supposed to be human intuitions, based on our coarse-grained and confused world models. While the AI's actions are supposed to be fine-grained actions based on the true nature of the universe, which can turn out very weird. That leads to a messy problem of translating preferences from one domain to another, which crops up everywhere in FAI thinking, Wei's comment and Eliezer's writeup are good places to start.
What I just realized is that you can handwave the problem away, by imagining a universe whose true nature agrees with human intuitions by fiat. Think of it as a coarse-grained virtual reality where everything is built from polygons and textures instead of atoms, and all interactions between objects are explicitly coded. It would contain player avatars, controlled by ordinary human brains sitting outside the simulation (so the simulation doesn't even need to support thought).
The FAI-relevant question is: How hard is it to describe a coarse-grained VR utopia that you would agree to live in?
If describing such a utopia is feasible at all, it involves thinking about only human-scale experiences, not physics or tech. So in theory we could hand it off to human philosophers or some other human-based procedure, thus dealing with "complexity of value" without much risk. Then we could launch a powerful AI aimed at rebuilding reality to match it (more concretely, making the world's conscious experiences match a specific coarse-grained VR utopia, without any extra hidden suffering). That's still a very hard task, because it requires solving decision theory and the problem of consciousness, but it seems more manageable than solving friendliness completely. The resulting world would be suboptimal in many ways, e.g. it wouldn't have much room for science or self-modification, but it might be enough to avert AI disaster (!)
I'm not proposing this as a plan for FAI, because we can probably come up with something better. But what do you think of it as a thought experiment? Is it a useful way to split up the problem, separating the complexity of human values from the complexity of non-human nature?
The way that I choose to evaluate my overall experience is generally through the perception of my own feelings. Therefore, I assume this simulated world will be evaluated in a similar way: I perceive the various occurrences within it and rate them according to my preferences. I assume the AI will receive this information and be able to update the simulated world accordingly. The main difference then, appears to be that the AI will not have access to my nervous system, if my avatar is being represented in this world and that is all the AI has access to, which would prevent it from wire-heading by simply manipulating my brain however it wants. Likewise it would not have access to its own internal hardware or be able to model it (since that would require knowledge of actual physics). It could in theory be able to interact with buttons and knobs in the simulated world that were connected to its hardware in the real world.
I think this is basically the correct approach and it actually is being considered by AI researchers (take Paul's recent paper for example, human yes-or-no feedback on actions in a simulated environment). The main difficulty then becomes domain transfer, when the AI is "released" into the physical world - it now has access to both its own hardware and human "hardware", and I don't see how to predict its actions once it learns these additional facts. I don't think we have much theory for what happens then, but the approach is probably very suitable for narrow AI and for training robots that will eventually take actions in the real world.
It does have access to your nervous system since your nervous system can be rewired via backdriving inputs from your perceptions.