Tim Freeman has a paper which clarifies many of the issues that need to be solved for humans to coexist with a superhuman AI. It comes close to what we would need if we had unlimited computing power. I will try amplify on some of the criticisms of it from the sl4 mailing list.
It errs on the side of our current intuitions about what I consider to be subgoals, rather than trusting the AI’s reasoning to find good subgoals to meet primary human goal(s). Another way to phrase that would be that it fiddles with parameters to get special-case results that fit our intuitions rather than focusing on general purpose solutions that would be more likely to produce good results in conditions that we haven’t yet imagined.
For example, concern about whether the AI pays the grocer seems misplaced. If our current intuitions about property rights continue to be good guidelines for maximizing human utility in a world with a powerful AI, why would that AI not reach that conclusion by inferring human utility functions from observed behavior and modeling the effects of property rights on human utility? If not, then why shouldn’t we accept that the AI has decided on something better than property rights (assuming our other methods of verifying that the AI is optimizing human utility show no flaws)?
Is it because we lack decent methods of verifying the AI’s effects on phenomena such as happiness that are more directly related to our utility functions? If so, it would seem to imply that we have an inadequate understanding of what we mean by maximizing utility. I didn’t see a clear explanation of how the AI would infer utility functions from observing human behavior (maybe the source code, which I haven’t read, clarifies it), but that appears to be roughly how humans at their best make the equivalent moral judgments.
I see similar problems with designing the AI to produce the “correct” result with Pascal’s Wager. Tim says “If Heaven and Hell enter into a decision about buying apples, the outcome seems difficult to predict”. Since humans have a poor track record at thinking rationally about very small probabilities and phenomena such as Heaven that are hard to observe, I wouldn’t expect AI unpredictability in this area to be evidence of a problem. It seems more likely that humans are evaluating Pascal’s Wager incorrectly than that a rational AI which can infer most aspects of human utility functions from human behavior will evaluate it incorrectly.