Is it even possible to have a perfectly aligned AI?
If you teach an AI to model the function f(x) = sin(x), it will only be "aligned" with your goal of computing sin(x) to the point of computational accuracy. You either accept some arithmetic cutoff or the AI turns the universe to computronium in order to better approximate Pi.
If you try to teach an AI something like handwritten digit classification, it'll come across examples that even a human wouldn't be able to identify accurately. There is no "truth" to whether a given image is a 6 or a very badly drawn 5, other than the intent of the person who wrote it. The AI's map can't really be absolutely correct because the notion of correctness is not unambiguously defined in the territory. Is it a 5 because the person who wrote it intended it to be a 5? What if 75% of humans say it's a 6?
Since there will always be both computational imprecision and epistemological uncertainty from the territory, the best you can ever do is probably an approximate solution that captures what is important to the degree of confidence we ultimately decide is sufficient.
I actually brought up a similar question in the open thread, but it didn't really go very far. May or may not be worth reading, but it's still not clear to me whether such a thing is even practical. It's likely that all substantially easier AIs are too far from FAI to still be a net good.
I've come a little closer to answering my questions by stumbling on this Future of Humanity Institute video on "Reduced Impact AI". Apparently that's the technical term for it. I haven't had a chance to look for papers on the subject, but perhaps some exist. No hits on google scholar, but a quick search shows a couple mentions on LW and MIRI's website.
It seems like most people think that reduced impact is as hard as value learning.
I think that's not quite true; it depends on details of the AIs design.
I don't agree that "It's likely that all substantially easier AIs are too far from FAI to still be a net good.", but I suspect the disagreement comes from different notions of "AI" (as many disagreements do, I suspect).
Taking a broad definition of AI, I think there are many techniques (like supervised learning) that are probably pretty safe and can do a lot of narrow AI tasks (and can maybe even be composed into systems capable of general intelligence). For instance, I think the kind of systems that are being built today are a net good (but might not be if given more data and compute, especially those based on Reinforcement Learning).
Ask AI to scan at least one human brain with Ph.d in ethics and age after 40, and let the AI run it in a simulation to give judgment of all AI's decisions. It will dramatically reduce chances of many obviously stupid decisions.
I don't mean to sound dismissive, but is that any better than any other boxing technique, like requiring it to ask verbal permission of a physical operator?
Unless I'm missing something, the AI would still have all the same incentives to influence the operator's answers, and solving those problems would be just as difficult for a digital operator as a physical one.
I only said that it would reduce chance of stupid decisions resulting from not understanding basic human words and values. But it would not reduce chances of deliberately malicious AI.
There are (at least) two different type of UFAI: real UFAI and failed FAI. Failed FAI wanted to be good but failed, and the best example of it smile maximizer which will cover all Solar system with smiles. (Paperclip maximizer also is some form of failed FAI, as initial goal was positive - produce many paperclips)
So it is not full recipe for real FAI, but just one way of value learning
I'm still not sure I understand you correctly. I suspect that if we follow this to the end, we will discover that we are only arguing semantics, and don't actually disagree over anything tangible. If that's your impression too, please say so, and we'll both save ourselves some time.
I only said that it would reduce chance of stupid decisions
I wouldn't disagree that having such an operator is better than not having one. I am questioning the value of having the operator uploaded. Why would programing an AI to care about the operator's values and not manipulate the operator be easier if the operator is uploaded? Wouldn't the operator just be manipulated even faster?
The only answer I see to that is that the uploading part is just to provide a faster and better user interface. If value loading was done via a game of 20 billion questions, for example, this would take an impractically long time. (Thousands of years, if just one person at a time is answering questions.) Same goes if the AI learns values via machine learning, using rewards and punishments given out by the operator, although you'd still have to keep it from wire-heading by manipulating the operator. Also, as an interesting aside, it may be easier to pull values directly out of someone's brain.
If we're only arguing about semantics, however, I have a guess at the source:
I understand "failed FAI" to be something like a pure smile maximizer, which has just as much incentive to route around human operators as a paperclip maximizer or suffering maximizer. It wouldn't care about our values any more than we care about what sorts of values evolution tried to give us. The unstated assumption here is that value uploading failed or never happened, and the AI is no longer trying to load values, but only implement the values it has. I believe this is what you're gesturing toward with "real UFAI".
Do you understand "failed FAI" to be one which simply misunderstood our values, like a smile maximizer, but which never exited the value loading phase? This sort of AI might have some sort of "uncertainty" about it's utility function. If so, it might still care about what values we intended to give it.
I don't think that we are only arguing semantic, but the idea of scanning a human is not my only one idea and is not the best idea of AI safety. It is just interesting promising idea.
In one Russian short story a robot was asked to get rid of all circular objects in the room and the robot cut the owner's head. But if the robot had a simulation of right moral human he could run it thousands times a second and check any his action with it.
The the first difference between sim and a human operator is that the sim can be run infinitely many more times and very cheaply, so we could check each robot action.
The second difference is that a human operator is changing (could be replaced by other operator or aging or suffer stroke). The sim is not changing with such speed (but he could still learn, or for simple situation fixed copies could be used)
The 3d is that sim is specially selected and trained for many years and later tested as a sim in many imaginary situations. We can't test operator with such rigorous quality.
Four. The sim could control any infinitely small update of AI's utility function, because it runs on the same speed. It prevents AI from becoming malicious. I have to explain it: our typical discussions about malicious AI have a logical loop (first shown by Ben Goertzel)
The loop: To start hide his utility function from peers, the AI should become malicious. But he can't covertly become malicious without secretly changing his utility function. In other words, between the moment where AI decides to be be bad, and the moment when he start to act base on this decision, is a short time gap. This gap maybe very short in time like 0,01 sec, so no human operator has time to react. But sim will have time (especially if the AI will be halted after any update in his open utility function).
Another explanation of the gap: if all values of an AI are written in open from (in some stack), and he creates the value to hide his values, than the value "hide values" is also written in open form (for short time). So if the value stack is translated outside, there is no way for AI to hide his transformation from the peer. (except to kill him before he get all implications)
Five. The sim not only check values but also correct understanding of human language and context.
You confuse the stupidity of whoever set the goals with the stupidity of the AI afterward. Any AGI is going to understand what we actually want, it just doesn't care, if the goal it was given wasn't already smart enough.
Solving the value learning problem is (IMO) the key technical challenge for AI safety.
How good or bad is an approximate solution?
EDIT for clarity:
By "approximate value learning" I mean something which does a good (but suboptimal from the perspective of safety) job of learning values. So it may do a good enough job of learning values to behave well most of the time, and be useful for solving tasks, but it still has a non-trivial chance of developing dangerous instrumental goals, and is hence an Xrisk.
Considerations:
1. How would developing good approximate value learning algorithms effect AI research/deployment?
It would enable more AI applications. For instance, many many robotics tasks such as "smooth grasping motion" are difficult to manually specify a utility function for. This could have positive or negative effects:
Positive:
* It could encourage more mainstream AI researchers to work on value-learning.
Negative:
* It could encourage more mainstream AI developers to use reinforcement learning to solve tasks for which "good-enough" utility functions can be learned.
Consider a value-learning algorithm which is "good-enough" to learn how to perform complicated, ill-specified tasks (e.g. folding a towel). But it's still not quite perfect, and so every second, there is a 1/100,000,000 chance that it decides to take over the world. A robot using this algorithm would likely pass a year-long series of safety tests and seem like a viable product, but would be expected to decide to take over the world in ~3 years.
Without good-enough value learning, these tasks might just not be solved, or might be solved with safer approaches involving more engineering and less performance, e.g. using a collection of supervised learning modules and hand-crafted interfaces/heuristics.
2. What would a partially aligned AI do?
An AI programmed with an approximately correct value function might fail
* dramatically (see, e.g. Eliezer, on AIs "tiling the solar system with tiny smiley faces.")
or
* relatively benignly (see, e.g. my example of an AI that doesn't understand gustatory pleasure)
Perhaps a more significant example of benign partial-alignment would be an AI that has not learned all human values, but is corrigible and handles its uncertainty about its utility in a desirable way.