Help me to understand AI alignment!
As will be very clear from my post, I’m not a computer scientist. However, I am reasonably intelligent and would like to improve my understanding of AI risk.
As I understand it (please do let me know if I’ve got this wrong), the risk is that:
an AGI could rapidly become many times more intelligent and capable than a human: so intelligent that its relation to us would be analogous to our own relation to ants.
such an AGI would not necessarily prioritise human wellbeing, and could, for example, could decide that its objectives were best served by the extermination of humanity.
And the mitigation is:
working to ensure that any such AGI is “aligned,” that is, is functioning within parameters that prioritise human safety and flourishing.
What I don’t understand is why we (the ants in this scenario) think our efforts have any hope of being successful. If the AGI is so intelligent and powerful that it represents an existential risk to humanity, surely it is definitionally impossible for us to rein it in? And therefore surely the best approach would be either to prevent work to develop AI (honestly this seems like a nonstarter to me, I can’t see e.g. Meta or Google agreeing to it), or to accept that our limited resources would be better applied to more tractable problems?
Any thoughts very welcome, I am highly open to the possibility that I’m simply getting this wrong in a fundamental way.
Epistemic status: bewitched, bothered and bewildered.
Let’s explore the ant analogy further.
The first thing to note is that we haven’t killed all the ants. We haven’t even tried. We kill ants only when they are inconvenient to our purposes. There is an argument that any AGI would always kill us all in order to tile the universe or whatever, but this is unproven, and IMO, false, for reasons I will explore in an upcoming post.
Secondly, we cannot communicate with ants. If we could, we could actually engage in mutually beneficial trade with them, as this post notes.
But the most important difference between the ant situation and the AI situation is that the ants didn’t design our brains. Imagine if ants had managed to program our brains in such a way that we found ants as cute and loveable as puppies, and found causing harm to ants to be as painful as touching a hot stove. Would ants really have much to fear from us in such a world? We might kill some ants when it was utterly and completely necessary, but mostly we would just do our own thing and leave the ants alone.
I recognise that the brains of any AI will have been designed by humans, but the gap in puissance between humans and the type of AGI imagined and feared by people in EA (as outlined in this blog post, for example) is so extreme the fact of us having designed the AGI doesn’t seem hugely relevant.
Like if a colony of ants arranged its members to spell out in English “DONT HURT US WE ARE GOOD” humans would probably be like huh, wild, and for a few days or weeks there would be a lot of discussion about it, and vegans would feel vindicated, and Netflix would greenlight a ripoff of the Bachelor where the bachelor was an ant, but in general I think we would just continue as we were and not take it very seriously. Because the ants would not be communicating in a way that made us believe they were worthy of being taken seriously. And I don’t see why it would be different between us and an AGI of the type described at the link above.
The thing is, the author of that post kind of agrees with you, In other places he has noted the probability of AI extinction as being 1, and is desperately trying to come up with any way to prevent it.
On the other hand, I think the model of AI put forward in that post is absurdly unlikely, and the risk of AI extinction is orders of magnitude lower. Ai will not be a single minded fanatic utilitarian focused on a fixed goal, and is likely to absorb at least a little bit of our values.
Oh, no, to be clear I find the post extremely unpersuasive—I am interested in it only insofar as it seems to represent received wisdom within the EA community.
I think the idea is that if we can get the A.I. to have the right values, then it won’t matter if it could theoretically take-over and overpower us, because it won’t want to. A more sinister variant of this, which I suspect a lot of people at MIRI believe in, and perhaps Bostrom also (but I have no direct evidence of this, other than a vague sense from things I’ve seen said over the years) is that if we can get an A.I. with the right values, it would be great if it took over and optimized everything towards those values (sure, power would corrupt humans, but that’s not a fact about all possible minds, and genuinely having the right values would prevent this). I am not terribly worried in itself about MIRI people believing the latter, because I don’t think they’ll build AGI, but I am a little worried about people at DeepMind (who I think take MIRI people, or at least Yudkowsky) more seriously than you’d intuitively guess, taking up these ideas. (Though I am much less confident than most EAs that world changing A.I. is imminent.)
Thank you, that is helpful. I still don’t see, I think, why we think an AGI would be incapable of assessing its own values and potentially altering them, if it’s intelligent enough to be an existential risk to humanity—but we’re hoping that the result of any such assessment would be “the values humans instilled in me seem optimal”? Is that it? Because then my question is which values exactly we’re attempting to instill. At the risk of being downvoted to hell I will share that the thought of a superpowerful AI that shares the value system of e.g. LessWrong is slightly terrifying to me. Relatedly(?) I studied a humanities subject :)
Thank you again!
I think the idea is that it will only change its values in a particular direction if that helped it realise its current values. So it won’t changes its values if doing so would mean that it would do horrible things according to its current values. A philosophical thing lurking in the background is that you can’t work out the correct values just by good thinking, rather basic starting values are thinking-independent, as long as your consistent: no amount of intelligence and reasoning will make you arrive at the correct ones. (They call this the “orthagonality thesis”, but a similar idea is known in academic philosophy as Humeanism about moral motivation. It’s quite mainstream but not without its critics).
‘...the thought of a superpowerful AI that shares the value system of e.g. LessWrong is slightly terrifying to me.‘
Old post, but I’ve meant to say this for several months: Whilst I am not a fan of Yudkowsky, I do think that his stuff about this showed a fair amount of sensitivity to the idea that it would be unfair if a particular group of people just programed their values into the AI, taking no heed of the fact that humans disagree. (Not that that means there is no reason to worry about the proposal to build a “good” AI that runs everything).
His original (since abandoned I think) proposal, was that we would get the AI to have goal like ‘maximizes things all or pretty much all fully informed humans would agree are good, minimizes things all or almost all fully informed would humans agree are bad, and where humans would disagree on whether something is good or bad even after being fully informed of all relevant facts, try and minimize your impact on that thing, and leave it up to humans to sort out amongst themselves.’ (Not an exact rendition, but close enough for present purposes.) Of course, there’s a sense in which that still embodies liberal democratic values about what is fair, but I’m guessing if your a contemporary person with a humanities degree, you probably share those very broad and abstract values.
I’m starting to think that this intuition may be right (further thoughts in linked comment thread).
Cheers! Here’s to being first against the wall when the basilisk comes.
The case for risk that you sketch isn’t the only case that one can lay out, but if we are focussing on this case, then your response is not unreasonable. But do you want go give up or do you want to try? The immediate response to your last suggestion is surely: Why devote limited resources to some other problem if this is the one that destroys humanity anyway?
You might relate to the following recent good posts:
https://www.lesswrong.com/posts/j9Q8bRmwCgXRYAgcJ/miri-announces-new-death-with-dignity-strategy
https://www.lesswrong.com/posts/uFNgRumrDTpBfQGrs/let-s-think-about-slowing-down-ai
“But do you want to give up or do you want to try?”
I suppose my instinctive reaction is that if there’s very little reason to suppose we’ll succeed we’d be better off allocating our resources to other causes and improving human life while it exists. But I recognise that this isn’t a universal intuition.
Thank you for the links, I will have a look :)