Open Problems

Another important step I wish would be done on problems like these: explain them. Don't just define the problem; build up the network of inferences that lead up to it so that someone can get the Level 2 (see link) understanding necessary to get the problem at a "gut" level. Make it so that someone can build up from the pre-requisite knowledge across the inferential gap just as easily (well, more so) than you can do here.

That, I think, would expose the problem to a lot more people capable of contribute, and reveal connections to similar problems.

A related post, 'Friendly AI Research and Taskification':

I don't even see how one would start to research the problem of getting a hypothetical AGI to recognize humans as distinguished beings. Solving this problem would seem to require as a prerequisite an understanding of the make up of the hypothetical AGI; something which people don't seem to have a clear grasp of at the moment. Even if one does have a model for a hypothetical AGI, writing code conducive to it recognizing humans as distinguished beings seems like an intractable task.

If the nature of ethical properties, statements, attitudes, and judgments does ultimately correlate with human brains, it might be possible to derive mathematical models of moral terms or judgments from brain data. The problem with arriving at the meaning of morality solely by means of contemplation is that you risk introducing new meanings based on high-order cognition and intuitions, rather than figuring out what humans as a whole mean by morality.

Two possible steps towards friendly AI/CEV (just some quick ideas):

1.) We want the AGI (CEV) to extrapolate our volition in a certain, ethical way. That is, it shouldn't for example create models of humans and hurt them just to figure out what we dislike. But in the end it won't be enough to write blog posts in English. We might have to put real people into brain scanners and derive mathematically precise thresholds for states like general indisposition and unethical behavior. Such models could then be implemented into the utility-function of an AGI, while blog posts written in natural language can't.

2.) We don't know if CEV is itself wished for and considered ethical by most humans. If you do not assume that all humans are alike, what makes you think that your personal solution, your answer to those questions will be universally accepted? A rich white atheist male living in a western country who is interested in topics like philosophy and mathematics does not seem to be someone who can speak for the rest of the world. If we are very concerned with the ethics of CEV in and of itself, we might have to come up with a way to execute an approximation of CEV before AGI is invented. We might need massive, large-scale social experiments and surveys to see if something like CEV is even desirable. Writing a few vague blog posts about it doesn't seem to get us the certainty we need before altering the universe irrevocably.

Is the intention of this project to approximate the list of questions, which if answered, would let someone build an FAI? (That seems incompatible with SIAI's traditional more secretive approach. Surely the Manhattan Project would never have published a list of questions, which if answered, would let someone build a nuclear bomb.)

If yes, is this a strategic (i.e., considered) change of plans? If so, what were the arguments that changed people's minds?

Or is the intention to just define some open problems that are FAI-related, but whose answers, if made public, pose no significant danger?

There seems to be a number of different possible approaches to building a Friendly AI, each with their own open problems. For example, we could design a system of uploads with safety checks and then punt all other questions to them, or figure out how we solve confusing philosophical problems and program a de novo AI with those methods, or solve them ourselves and just code up a decision process together with a set of preferences. Does this effort to define open problems assume a particular approach?

I think the Millennium Prize Problems isn't the best example in this context, because for the one problem that was solved in that set, the prize was rejected.

This is probably also a good opportunity for people concerned about existential risk to shape the perception of the scientific and mathematical public.