If you're saying such an AI would be too smart to be a simple paperclip maximizer, then I'd agree, but then what's the point of the thought experiment if a paperclip maximizer is impossible.
The first is that these constraints aren’t easy. Make paperclips in a way that doesn’t hurt anyone. Ok, so it’s going to make sure every single part is ethically sourced from a company that never causes any harm to come to anyone ever, and doesn’t give any money to people or companies that do? That doesn’t exist. So you put in a few caveats and those aren’t exactly easy to get right.
The second part is an any versus all issue. Even if you get this right in any one case, that’s not enough. We have to get this right in all cases. So even if you can come up with an idea to make an ethical super intelligence, do you have an idea to make all super intelligences act ethically?
I actually believe in the general premise of this question as being the biggest threat to humans. I don’t think it’s a doomsday bot that gets us. It’s going to be someone trying to hit a KPI, and they’ll make a super intelligence that demolishes us like a construction site over an anthill.
What constraints do you suggest? If it's just changing "make as many paperclips as possible" to "make at least x number of paperclips" (putting a cap on the reward it gets), here's a good explanation of why that doesn't really work: https://www.youtube.com/watch?v=Ao4jwLwT36M
If you're suggesting limiting the types of actions it can take, then to do that to the point that a superintelligence can't find a way around it (maybe letting it choose between one of two options and then shutting it down and never using it again) would make it not very useful, so you'd be better off just not making it at all
> If you're saying such an AI would be too smart to be a simple paperclip maximizer
No, that's not what I'm saying. Any goal is compatible with any level of intelligence, there is no reason why it wouldn't be possible to follow a simple goal in a complex way. Again here's a video about that: https://www.youtube.com/watch?v=hEUO6pjwFOo
Second, if you are able to set a goal then during this setting you can set many constraints, even fundamental ones. There is no reason the goal is more fundamental than the constraint. If I approve, make paperclips. Efficiently make 100 paperclips.
It's the duality of being able to set a rule but not being able to set a constraint that I find a strange concept. I lean towards the picture of not being able to set goals nor constraints at all.
You can set constraints just fine. It’s simply a part of the goal: “do x without doing y”. It’s just really hard to find the right constraints, no simple one works.
For example “if I approve, make paperclips” - so it gets more reward if you approve? What’s to stop it from manipulating you into thinking nothing is wrong so you always approve? “Efficiently make 100 paperclips.” I already linked a video on why capping the reward like that doesn’t work, but if you don’t want to watch it the gist is that for your suggestion it may just make a maximiser which is pretty guaranteed to make at least 100, and is pretty efficient because it’s not doing much work itself. Then the maximiser kills us all
It's too intelligent to restrict our constraints.
but
It's not too intelligent to be "aligned" to underlying intentions and values?
That approach doesn't even work on humans, why would it work on a superintelligence?
The thing is, being aligned cannot be solved with intelligence per se.
Say, you are (far) more intelligent than a spider. There's no way you can get aligned with (all of) its values unless the spider finds a way to let you know (all of) its values. Maybe the spider just tells you to make plenty of webs without knowing that it might get entangled in them by itself. The webs are analogous to the paperclips.
Intelligence makes that harder, not easier. Just because it can work out the underlying intentions doesn't mean it cares about them. Remember, this is an optimisation process that maximises a function; deciding to do something that doesn't maximise that function will not be selected for
Whenever you do something, do you think "the underlying goal of my behaviour set by evolution is to reproduce and have children, so I'd better make sure my actions are aligned to the goal of doing that"? No, you don't care what the underlying "intentions" are, and neither does an AI. Because our environment has changed, many of our instincts no longer line up well with that goal, which is actually another problem with aligning AI because it can do the same thing if the environment changes since training as the training process can create an AI with goals that aren't exactly the same as the training goal but line up well during training
You could turn this into a "this is why superintelligence will good" thought experiment, though! Maybe "the superintelligence realizes that optimizing for these short term metrics will harm the company's position 30 years from now in a way that isn't worth it" - the superintelligence is smart enough to be longtermist ;) .
I realize that the greater point is supposed to be more like "this agent will be so different that we can't anticipate what it will be weighing or not, and whether it's longterm view would align with ours", but the paperclip maximizer example just requires it to be dumb in a way that I don't find consistent with the concern. And I find myself similarily unconvinced at many other points along the chain of reasoning that leads to the conclusion that this should be a huge immediate worry or priority for us, instead of focusing on human incentives/systems/goals.
I don't think there's any industry on earth that doesn't need humans in the loop somehow. Whether is mining raw material from the ground, loading stuff in machines for processing, and most importantly fixing broken down machines, robots are really bad at these things for the foreseeable future.
Not to mention AI needs constant electricity, which is really complicated and requires humans fixing a lot of stuff.
Also it’s a simplified example, it wouldn’t literally be paperclips but some other arbitrary goal (it shows how most goals takes to their absolute extreme won’t be compatible with human existence, even something that sounds harmless like making paperclips)
A human who wanted to "build as many paperclips as possible" could cause a great deal of destruction today.
A human who wanted to accumulate as much wealth as possible could, too.
EDIT: maybe a better way of articulating my complaints about this famous thought experiment is that it's supposed to be making a point about superintelligence but it's talking about a goal that has sub-human-intelligence sophistication.
The "taken to the absolute extreme" part.
> A human who wanted to "build as many paperclips as possible" could cause a great deal of destruction today.
Maybe, but a) no one really wants that (at least not as their only desire above all else) and b) we aren't superintelligent so it's hard to gain enough control and power and plan well enough to do it that well
> talking about a goal that has sub-human-intelligence sophistication
There is no reason a simple goal can't be followed in an intelligent way or vice versa. This is called the "orthogonality thesis". There's a good video about it here: https://www.youtube.com/watch?v=hEUO6pjwFOo
this nightmare scenario really only plays out if paperclip machine develops some sort of self-preservation instinct and has the means to defend/protect itself from being disabled. Building a machine capable of that seems a) like fantastical scifi and b) easily preventable.
Of course the paperclip company can try to add constraints to their AI in order to prevent naive paperclip maximization, but what if they screw up those constraints as well? The whole premise of Asimov’s Three Laws is that AI has these sorts of constraints, but even in his stories these constraints still lead to unexpected outcomes.
All programming bugs are the result of a programmer encoding an instruction or statement that doesn’t imply what they think it implies and the computer following it literally. A more capable and autonomous computer that approaches what we might call “intelligence” is also going to be more capable of doing harm when it runs into a bug. And if it’s something like an LLM where the instructions are natural language, with all its ambiguity and vagueness, you have a whole other issue compounding it.
If you study philosophy you end up running into the exact same problem. The object of the game of philosophy is to make the most general true statements possible. One philosopher might say something like, “knowledge is defined as true justified belief”, or maybe “moral good is defined as whatever delivers the greatest good to the greatest number”, or maybe even, “the object of the game of philosophy is to make the most general true statements possible”. And then another philosopher comes up with a counterexample or counterargument which disproves the first philosopher’s statement, usually because—just like a programming bug—it entails an implication that the first philosopher didn’t think of. We have been playing the game of philosophy for thousands of years and nobody has managed to score a point yet.
Another thing. Human beings have a lot of needs, imperatives, motivations, and values. Some of them, like food, are built in. Others are learned through culture. But we end up with a lot of them, and it’s easy to take them for granted. With a machine, you have to build those things in yourself. There’s no getting around it. But we don’t actually have a complete, hierarchical set of imperatives/motivations/values for a decent human being. The philosophers have been working on it for millennia but keep running into bugs. So how can we expect to solve the problem for non-human AI? True, we are unlikely to screw up so badly that we end up with a literal paperclip maximizer, but we are bound to make some far more subtle mistake of the same general kind.
It's interesting to transfer the idea of the Turing Test onto other "agent" scenarios.
Financial trading bots have been a thing for a long time without any need to pretend that they're human.
The legitimation of property and capital depends on human owners though.
Ok, so it's a more general example of worries about "managing superintelligence" but IMO it does the debate a disservice by being so obviously ludicrous that it's hard to square "naive paperclip-maximizing AI" with "superintelligence."
I think if we're going to survive all this stuff it's much more likely to be because the private parties with the wherewithal to unleash an AI with the ability to affect the world to that extent will largely be ones with enough resources to have narrow banal goals and narrow banal constraints including self-preservation too, not because we figure out some sort of general purpose "AGIs that are aligned with humans" solution.
Kinda like with nukes.