= 27d567bc2d48d9f40e9ccc20ea370b15)
Lumifer comments on Open thread, Oct. 10 - Oct. 16, 2016 - All
You are viewing a comment permalink. View the original post to see all comments and the full post content.
= 783df68a0f980790206b9ea87794c5b6)
Loading…= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Powered by Reddit
= f037147d6e6c911a85753b9abdedda8d)
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (147)
But if you don't know what human values are, how can you be sure that the AI will learn them correctly?
So you make an AI and tell it: "Go forth and learn human values!" It goes and in a while comes back and says "Behold, I have learned them". How do you know this is true?
If I train a neural network to recognize dogs, I have no way of knowing if it learned correctly. I can't look at the weights and see if they are correct dog image recognizing weights and not something else. But I can trust the process of training and validation, that the AI has learned to recognize what dogs look like.
It's a similar principle with learning human values. Of course it's more complicated than just feeding it images of dogs, but the principle of letting AIs learn models from real world data is the important part.
Of course you do. You test it. You show it a lot of images (that it hasn't seen before) of dogs and not-dogs and check how good it is at differentiating them.
How would that process work for an AI and human values?
Right, human values: “A man's greatest pleasure is to defeat his enemies, to drive them before him, to take from them that which they possessed, to see those whom they cherished in tears, to ride their horses, and to hold their wives and daughters in his arms.”
Do you expect me to give you the complete solution to AI right here, right now? What are you even trying to say? You seem to be arguing that FAI is impossible. How can you possibly know that? Just because you can't immediately see a solution to the problem, doesn't mean a solution doesn't exist.
I think an AI will easily be able to learn human values from observations. It will be able to build a model of humans, and predict what we will do and say. It certainly won't base all it's understanding on a stupid movie quote. The AI will know what you want.
I'm saying that if you can't recognize Friendliness (and I don't think you can), trying to build a FAI is pointless as you will not be able to answer "Is it Friendly?" even when looking at it.
So if you can't build a supervised model, you think going to unsupervised learning will solve your problems? The quote I gave you is part of human values -- humans do value triumph over their enemies. Evolution taught humans to eliminate competition, it taught them to be aggressive and greedy -- all human values. Why do you think your values will be preferred by the AI to values of, say, ISIS or third-world Maoist guerrillas? They're human, too.
Why do I need to recognize Friendliness to build an FAI? I only need to know that the process used to construct it results in a friendly AI. Trying to inspect the weights of a complex neural network (or whatever) is pointless as I stated earlier. We haven't the slightest idea how alphaGo's net really works, but we can trust it to beat the best Go champions.
Evolution also taught humans to be cooperative, empathetic, and kind.
Really your objection seems to be the whole point of CEV. A CEV wouldn't just include the values of ISIS members, but also their victims. And it would be extrapolated, to not just be their current opinions on things, but what their opinions would be if they knew more. Their values if they had more time to think about and consider issues. With those two conditions, the negative parts of human values are entirely eliminated.
This amounts to saying "because I'm right and once everyone gets to know reality better, they'll figure out I'm right."
In reality they will also figure out the places where you are wrong, and there will be many of them.
I'm not claiming that at all. I may be wrong about many things. It's irrelevant.
It is not irrelevant. You said, "With those two conditions, the negative parts of human values are entirely eliminated." That certainly meant that things like ISIS opinions would be eliminated. I agree in that particular case, but there are many other things that you would consider negative which will not be eliminated. I can probably guess some of them, although I won't do that here.
See my other comment for more clarification on how CEV would eliminate negative values.
You are still facing the same problem. Given that you can't recognize friendliness, how will you create or choose a process which will build a FAI? Would you be able to answer "Will it be friendly?" by looking at the process?
That doesn't make much sense. What do you mean by "negative" and from which point of view? If from the point of view of the AI, that's just a trivial tautology. If from the point of view of (at least some) humans, this seems to be not so.
In general, do you treat morals/values as subjective or objective? If objective, the whole "if they knew more" part is entirely unnecessary: you're discovering empirical reality, not consulting with people on what do they like. And subjectivism here, of course, makes the whole idea of CEV meaningless.
Also, I see no evidence to support the view that as people know more, their morals improve, for pretty much any value of "improve".
You are literally asking me to solve the FAI problem right here and now. I understand that FAI is a very hard problem and I don't expect to solve it instantly. Just because a problem is hard, doesn't mean it can't have a solution.
First of all let me adopt some terminology from Superintelligence. I think FAI requires solving two somewhat different problems. Value Learning and Value Loading.
You seem to think Value Learning is the hard problem, getting an AI to learn what humans actually want. I think that's the easy problem, and any intelligent AI will form a model of humans and understand what we want. Getting it to care about what we want seems like the hard problem to me.
But I do see some promising ideas to approach the problem. For instance have AIs that predict what choices a human would make in each situation. So you basically get an AI which is just a human, but sped up a lot. Or have an AI which presents arguments for and against each choice, so that humans can make more informed choices. Then it could predict what choice a human would make after hearing all the arguments, and do that.
More complicated ideas were mentioned in Superintelligence. I like the idea of "motivational scaffolding".Somehow train an AI that can learn how the world works and can generate an "interpretable model". Like e.g. being able to understand English sentences and translate their meanings to representations the AI can use. Then you can explicitly program a utility function into the AI using its learned model.
From your point of view. You gave me examples of values which you consider bad, as an argument against FAI. I'm showing you that CEV would eliminate these things.
Your stated example was ISIS. ISIS is so bad because they incorrectly believe that God is on their side and wants them to do the things they do. That the people that die will go to heaven, so loss of life isn't so bad. If they were more intelligent, informed, and rational... If they knew all the arguments for and against religion, then their values would be more like ours. They would see how bad killing people is, and that their religion is wrong.
The second thing CEV does is average everyone's values together. So even if ISIS really does value killing people, their victims value not being killed even more. So a CEV of all of humanity would still value life, even if evil people's values are included. Even if everyone was a sociopath, their CEV would still be the best compromise possible, between everyone's values.
No, I'm asking you to specify it. My point is that you can't build X if you can't even recognize X.
Learning what humans want is pretty easy. However it's an inconsistent mess which involves many things contemporary people find unsavory. Making it all coherent and formulating a (single) policy on the basis of this mess is the hard part.
Why would CEV eliminate things I find negative? This is just a projected typical mind fallacy. Things I consider positive and negatve are not (necessarily) things many or most people consider positive and negative. Since I don't expect to find myself in a privileged position, I should expect CEV to eliminate some things I believe are positive and impose some things I believe are negative.
Later you say that CEV will average values. I don't have average values.
I see no evidence to believe this is true and lots of evidence to believe this is false.
You are essentially saying that religious people are idiots and if only you could sit them down and explain things to them, the scales would fall from their eyes and they will become atheists.This is a popular idea, but it fails real-life testing very very hard.
And I don't agree with that. I've presented some ideas on how an FAI could be built, and how CEV would work. None of them require "recognizing" FAI. What would it even mean to "recognize" FAI, except to see that it values the kinds of things we value and makes the world better for us.
I've written about one method to accomplish this, though there may be better methods.
Humans are 99.999% identical. We have the same genetics, the same brain structures, and mostly the same environments. The only reason this isn't obvious, is because we spend almost all our time focusing on the differences between people, because that's what's useful in everyday life.
That may be the case, but that's still not a bad outcome. In the example I used, the values dropped from ISIS members were taken for 2 reasons. That they were based on false beliefs, or that they hurt other people. If you have values based on false beliefs, you should want them to be eliminated. If you have values that hurt other people then it's only fair that be eliminated. Or else you risk the values of people that want to hurt you.
Well I think it's accurate, but it's somewhat nonspecific. Specifically, CEV will find the optimal compromise of values. The values that satisfy the most people the most amount. Or at least dissatisfy the fewest people the least. See the post I just linked for more details, on one example of how that could be implemented. That's not necessarily "average values".
In the worst case, people with totally incompatible values will just be allowed to go separate ways, or whatever the most satisfying compromise is. Muslims live on one side of the dyson sphere, Christians on the other, and they never have to interact and can do their own thing.
My exact words were "If they were more intelligent, informed, and rational... If they knew all the arguments for and against..." Real world problems of persuading people don't apply. Most people don't research all the arguments against their beliefs, and most people aren't rational and seriously consider the hypothesis that they are wrong.
For what it's worth, I was deconverted like this. Not overnight by any means. But over time I found that the arguments against my beliefs were correct and I updated my belief.
Changing world views is really really hard. There's no one piece of evidence or one argument to dispute. Religious people believe that there is tons of evidence of God. To them it just seems obviously true. From miracles, to recorded stories, to their own personal experiences, etc. It takes a lot of time to get at every single pillar of the belief and show its flaws. But it is possible. It's not like Muslims were born believing in Islam. Islam is not encoded in genetics. People deconvert from religions all the time, entire societies have even done it.
In any case, my proposal does not require literally doing this. It's just a thought experiment. To show that the ideal set of values is what you choose if you had all the correct beliefs.
We typically imagine CEV asking what people would do if they 'knew what the AI knew' - let's say the AI tries to estimate expected value of a given action, with utility defined by extrapolated versions of us who know the truth, and probabilities taken from the AI's own distribution. I am absolutely saying that theism fails under any credible epistemology, and any well-programmed FAI would expect 'more knowledgeable versions of us' to become atheists on general principles. Whether or not this means they would change "if they knew all the arguments for and against religion," depends on whether or not they can accept some extremely basic premise.
(Note that nobody comes into the word with anything even vaguely resembling a prior that favors a major religion. We might start with a bias in favor of animism, but nearly everyone would verbally agree this anthropomorphism is false.)
It seems much less clear if CEV would make psychopathy irrelevant. But potential victims must object to their own suffering at least as much as real-world psychopaths want to hurt them. So the most obvious worst-case scenario, under implausibly cynical premises, looks more like Omelas than it does a Mongol invasion. (Here I'm completely ignoring the clause meant to address such scenarios, "had grown up farther together".)