= 27d567bc2d48d9f40e9ccc20ea370b15)
turchin comments on Open thread, Oct. 10 - Oct. 16, 2016 - Less Wrong
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 (115)
Good point, but my question was about what we can do to raise chances that it will be friendly AI.
Get a job at Google or seek to influence the people developing the AI. If, say, you were a beautiful woman you could, probably successfully, start a relationship with one of Google's AI developers.
I am confused as to whether I should upvote for "get a job at Google" or downvoter for "prostitute yourself".
And how she will use this relation to make safer AI?
She could read "The Basic AI Drives" to him at night.
In hope that he will stop creating AI? But in 6 years it will be Microsoft.
Ignore all the stuff about provably friendly AI, because AFAIK its fairly stuck at the fundamental level of theoretical impossibility due to lob's theorem and its prob going to take a lot more than five years. Instead, work on cruder methods which have less chance of working but far more chance of actually being developed in time. Specifically, if Google are developing it in 5 years, then its probably going to be deepmind with DNNs and RL, so work on methods that can fit in with that approach.
Yes. I think that we need not only workable solution, but also implementable. If someone create 800 pages pdf starting with new set theory, solution of Lob theorem problem etc and come to Google with it and say: "Hi, please, switch off all you have and implement this" - it will not work.
But MIRI added in 2016 the line of research for machine learning.
I agree. I think it's very unlikely FAI could be produced from MIRI's very abstract approach. At least anytime soon.
There are some methods that may work on NN based approaches. For instance my idea for an AI that pretends to be human. In general, you can make AIs that do not have long-term goals, only short term ones. Or even AIs that don't have goals at all and just make predictions. E.g., predicting what a human would do. The point is to avoid making them agents that maximize values in the real world.
These ideas don't solve FAI on their own. But they do give a way of getting useful work out of even very powerful AIs. You could task them with coming up with FAI ideas. The AIs could write research papers, review papers, prove theorems, write and review code, etc.
I also think it's possible that RL isn't that dangerous. Reinforcement learners can't model death and don't care about self-preservation. They may try to hijack their own reward signal, but it's difficult to understand what they would do after that. E.g. if they just tweak their own RAM to have reward = +Inf, and then not do anything else. It may be harder to create a working paperclip maximizer than is commonly believed, even if we do get superintelligent AI.
I agree. FAI somehow should use human upload or human-like architecture for its value core. In this case values will be presented in it in complex and non-ortogonal ways, and at least one human-like creature will survive.
Nothing, because we still don't know what a friendly AI is.
Friendly AI is an AI which maximizes human values. We know what it is, we just don't know how to build one. Yet, anyway.
We don't know what an AI which maximizes human values is because we don't know what human values are at the necessary level of precision. Not to mention the assumption that the AI will be a maximizer and that values can be maximized.
Who says we need to hardcode human values though? Any reasonable solution will involve an AI that learns what human values are. Or some other method to the control problem that makes AIs that don't want to harm or defy their creators.
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.
That doesn't mean that there is nothing to do - if you don't know what FAI is, then you try to work out what it is.
And how do you find out whether you're right or not?
We do know it isn't an AI that kills us. Options b and c still qualify.
Options (b) and (c) are basically wishes and those are complex X-D
"Not kill us" is an easy criterion, we already have an AI like that, it plays Go well.
We don't have an AGI that doesn't kill us. Having one would be a significant step towards FAI. In fact, "a human-equivalent-or-better AGI that doesn't do anything greatly harmful to humanity" is a pretty good definition of FAI, or maybe "weak FAI".
If it's a tool AGI, I don't see how it would help with friendliness, and if it's an active self-developing AGI, I thought the canonical position of LW was that there could be only one? and it's too late to do anything about friendliness at this point?
I agree there would probably only be one successful AGI, so it's not the first step of many. I meant it would be a step in that direction. Poor phrasing on my part.