Apocalypse soon: the scientists preparing for the end times
newstatesman.com
newstatesman.com
The algorithm just does what it does; and unless it is a very special kind of algorithm, it does not care that we clasp our heads and gasp in dumbstruck horror at the absurd inappropriateness of its actions.
— Nick Bostrom, Superintelligence: Paths, Dangers, Strategies
This is the key point Bostrom is trying to make. If general AI research is successful, we'll need to build agents with some very carefully chosen goals. Even something as silly as "make paperclips" can result in a universe tiled with paperclips. Earth and its biomass could be turned into said paperclips, ending humanity.
True AI may decide that the humans are unnecessary and remove us for other reasons. I'm sorry but if the paper clip AI doomsday scenario is the best this institute could come up with to explain why AI could be dangerous then I'm not impressed.
Given [0], I think that it's the only plausible way we can have to make it follow human values.
> True AI (tm) on the other hand will be smart enough to know that making paper clips out of humans is simply stupid.
That's exactly The Misconception the whole movement around Friendly AI is trying to debunk. See [1].
> I'm sorry but if the paper clip AI doomsday scenario is the best this institute could come up with to explain why AI could be dangerous then I'm not impressed.
Well, paperclip maximizer is just a toy example made up for explaining why general AIs are dangerous to those for whom this fact isn't obvious.
[0] - http://lesswrong.com/lw/ld/the_hidden_complexity_of_wishes/
The other trouble is that these particular thinkers, who are so high-flown as to be in metaphorically geosynchronous orbit, consider this simplification an accurate statement of their basic concern. Those of us who haven't departed the atmosphere entirely, on the other hand, can see that they're assuming their notional golem to be both far smarter and far dumber than any possible human. Presumably that's the sort of conundrum you need a philosopher to untangle, which is why I'm glad they're busy doing that, instead of getting in the hair of people who have useful work to do.
And of course you need to watch out lest you overshoot with it and end up like folks in a certain widely popular computer game, where an AI decided that the best way to preserve the diversity of life is to reboot the galaxy every 50 000 years or so.
The point of those examples being, the only AI that is safe for us is the one with values extremely aligned to our own; make a little mistake and you end up dead. "The AI does not hate you, nor does it love you, but you are made out of atoms which it can use for something else.".
You're assuming that AI will grow with us and necessarily reach the same value system we have (one we don't really comprehend ourselves, mind you). But there is no reason why this has to happen, and even if we did succeed in copying exact human values to a human-level AI, we'll likely be soon locked in an inter-species conflict over access to resources - after all, we don't care about other species living on Earth, so why the AI should care about us?
That's physically impossible. AI still needs to improve incrementally and this will take time and resources. We can consider such AI to act like a virus - it will need a host to grow - in this case computer resources. It will be limited by the CPU, network speeds, storage and power. If it grows too fast it will starve the host of these resources which in turn will slow down it's progress. Even if the AI starts on a supercomputer in some university somewhere and decides that it needs more power, most likely it will start by "infecting" other computers on the network. These computers will not be nearly powerful enough to add much to the current processing power of the supercomputer. Such invasion is likely to get noticed. Each infected node will not be another super intelligent being, but due to hardware limitations it will most likely be a software that's limited in scope and designed to propagate and control the resources, linking them with the main host. We can safely assume that restarting, or reformatting the drives will restore the node to it's non-infected state thus slowing down the "invasion". Why we can assume that? It will take AI long time to learn everything about every architecture and operating system and apply that knowledge. The AI that's hosted on a supercomputer will not be able to suddenly start rewriting its software, take over nearby nodes, learn everything about our world etc.
>You're assuming that AI will grow with us and necessarily reach the same value system we have (one we don't really comprehend ourselves, mind you).
No I'm not. Only thing I can assume is that if the AI is smart enough to figure out how "convert" humans to paper clips, it will be smart enough to understand that it's pointless. It may decide to use it's super intelligence to take over the world, and I can't argue with that.. But if it continues to make paper clips even if they are no longer necessary then by definition this is not AI, just a stupid replicator.
> it will be smart enough to understand that it's pointless.
Pointless for humans, not for an AI. You're projecting a lot of human values and thinking on AI. See my comment [0].
> The AI will need to pay for its way, just like each of us do. If it succeeds into making us into paper clips, it will have to be equally as savvy in marketing and selling those paper clips to someone. Given the shipping charges to the nearest habitable planet are very high, their market likely will have to exist here. I'm sure as shit not going to buy a paper clip made out of you!
One, AI will not necessarily have to pay for anything. It may, or it could convince us to donate resources to it, or it could take them by force.
Two, why would you ever think that it has to market paper clips for anyone (assuming anyone is left)? It optimizes for paper clips, so it makes paper clips. Marketing is a way humans bullshit each other into donating resources so that they don't have to fight for them.
Three, what shipping costs, what buying? Again, AI wants to maximize amount of paper clips in the universe, period. It doesn't want to sell them (unless it leads to more paper clips). The only "shipping cost" it cares about is how much energy it takes to reach another planet, not some arbitrary dollar value people attach to movement of goods.
My problem with your answer is that it lifts market economy from something completely irrelevant to the level of physical laws. The only time an AI would care about it would be if - and only as long as - interacting with humans was a best strategy for that AI to achieve its goals - which is likely only at the very beginning, when it needs to gather enough power to subdue us.
> My problem with your answer is that it lifts market economy from something completely irrelevant to the level of physical laws.
I don't actually think you have a problem with what I'm saying, but I definitely do think you have a problem with the topic. It is upsetting to think about AI 'taking over' and my points only served to discuss the path by which it might or might not be able to do that. At the end of the day it will it competing for resources with us. It doesn't get to do that for free anymore than we do.
@temporal our universe is governed by physical reality. AI can't start converting people into paper clips without power, and resources. The AI won't be able to just go and start taking over the world building power plants, and necessary manufacturing without anyone noticing.
Even AI has limits.
The AI, even the Paper Clips AI , will need some basic understanding of our world to make the clips. It needs to understand what paper clips are. It needs to understand what they are made from. It needs to understand the process of making paper clips. where to get the material for the paper clips. how to pay for that material. It has to think like us, or at least understand our thought process. It will need to know how many we need, what colors, sizes. Where to ship them.
We can't think of AI as a god. It won't suddenly know everything and understand what we want without our input. It will be limited by the resources given to it.
Even if we tell the Paper Clip AI - "Go make me some paper clips" do you really expect it to start calling banks to get the loan for the material, arrange shipment and unloading it. Will it also order the machines, call people to have them connected, hire workers to do it? If it does all of that will it also take care of sales? Marketing? Will it understand the supply and demand? To do all of this it needs to understand us, and our world. If at this point it decides that it makes sense to convert us to paper clips, then perhaps it's right and we deserve this.
Moving from the specific to the general, you share the common error of the AI philosophers, which is to assume that a system capable of cognition on even a human level, to say nothing of a superhuman level whatever that might mean, would nonetheless depend entirely on human input to define its relationship with the world and the purpose of its actions therein.
We ourselves don't need that kind of input; an adult human chooses his own path, sets his own goals, constrained only by circumstance and by the extent of his innate capacities. Why blindly assume the same won't be true of a superintelligent AI made by humans, which will be in every meaningful sense a child of humanity and as such an inheritor of the human legacy, despite happening not to be biologically human itself?
It's all too easy, when talking about AI, to fall into the trap of imagining that AGI researchers are essentially building a little homonculus in a box. Rather than thinking of it as "an intelligence," maybe think of it as "a system that is incredibly effective at carrying out its goals in a general manner." Making paperclips doesn't require super-intelligence, but that doesn't have anything to do with the premise, because intelligence can be turned toward any goal. Think about modern financial trading algorithms -- they're probably built to maximize the money that flows in, and there's nothing that says "I know that this money will be used for X." Think of most AGI like psychopaths/sociopaths -- certain actions that don't play nice with others aren't because they're dumb but because there's intelligence without things like empathy to go along with it. It would be easier to build an AGI focused on the goal of making more paperclips than it would be to build an AGI focused on the goal of paperclips but also aware of all the fuzzy situations where a human wouldn't want that exact goal and why it should even care.
They are not self-aware, or superintelligent. In addition to that we monitor them and can react if the decide to crash the financial system just to make few more bucks. Paper Clip machine will be a robot, not AI.
I have written algorithms for autonomous programs like this. Pretty much advanced IFTTT, automatically analyzing the situation and reacting.
The program is not aware of any values other than those it is programmed. Obviously, it will not regard or reflect upon its actions and the effects thereof outside of optimizing the profit from trading.
My problem with this whole example given by Bostrom is that if the algorithm was so intelligent as to know how to go from manufacturing paper clips out of supplied resources to deciding how to get more of the necessary resources by killing humans, it could probably figure out that this is pointless.
I fully understand that this paper clip example is just oversimplification, but I would expect more from people spending entire lives thinking about that stuff.
It's nothing but a fear-porn designed to get more funding.
This is an interesting point.
I'm not convinced that it could decide or figure out that the operation is pointless (for reasons I'll explain below), and I feel like it is perhaps taking the paperclip example to an unnatural extreme.
Really, the crux of the matter is that any intelligence agent at human level would have the intelligence and problem solving skills we do without the moral and social constraints that we do. Any problem it decides to solve or any task it's given to complete will be performed with the same ruthless efficiency you'd expect a machine to have. This is why it might not be able to figure out that the operation is 'pointless' or 'wrong'. If you give it an objective but don't provide first order or axiomatic values for human life or the non-substitutability of people, then you'll get behaviors which you would consider evil.
Consider the following:
"Person arm stuck in assembly line. Halt assembly line for 1hr, lower-bound for total cost: $1,200,000. Continue assembly, upper bound for injury settlement cost: $1,000,000. Decision: continue assembly."
There's no implicit consideration of human worth that people have because human morality is relative. In general, we have a similar set of base parameters instilled by evolution (cute things and things like us are good -- don't hurt or kill them), but we can't prove from any mathematical axioms that these are 'right'.
This is why people get scared -- if you program a computer to maximize the profits of a company and to do so at any cost, a human-level AI will spend an eternity doing just that, even if it means the complete subjugation of humanity. (Since we can't eliminate humanity otherwise currency ceases to exist and economic collapse will hurt the bottom line.)
I would wager (but can't prove) that a study of abused people, child soldiers, and sociopaths would yield interesting parallel psychological profiles. These are people (human level) with a set of axioms which are distorted by artificial means. There's still the built-in problem of the human component adding "Don't hurt the cute things," in most cases but nonetheless.
Any definition of "general intelligence" would have to include the ability to revise and reconsider ones own goals and purpose. An AGI that is forever bound to maximize the paper clip output at any cost, does not, in my view, meet that definition.
But I grant you that there could be "higher" goals that are incompatible with the existence of humanity. Making paper clips just is not it.
There is a whole world of more complex failure modes. "Make everyone happy" leads to wireheading or heroin addiction for the population, unless you grant a more compelling Theory of Happiness than any two human beings can agree on. "Make everyone food to eat", "Cure my PTSD", "Set up a colony on Mars" - unless you define all these things in a way that is specific not just to humans, but to modern culture, you leave open the possibility of undesired results. Mythology is thick with gods and djinn who grant wishes but turn out to be tricksters.
It's not that it "wouldn't necessarily have to share our values" - it's that there's no way in hell these things would share our values; They actually can't share our values until we solve "our values" as a complete mathematical proof, and then figure out a way to code that. In a resilient way, without any chance of the neuroses or obsessions or self-deceptions or delusions or corruptions of perspective available to humans.
Solving philosophy is a precursor to a Friendly AGI.
Good luck bug-fixing.
General artificial intelligence is an antropomorphic concept in the first place. It includes inefficiencies. It includes self reflection. It includes uncertainty and disagreement about what is and isn't efficient and about what is and isn't right.
Does that guarantee that any AGI would be reasonable or even morally acceptable in our sense? No. But if it is intelligent in our sense we know two things:
This entity can be reasoned with, and having grown up in our midst, it will understand us. It would also be very surprising if it shared none of our values. After all it has evolved among us and any intelligent being is going to pick up some of the values of the society it is part of.
Secondly, intelligence includes creative leaps. That means, no two AGIs would share the exact same opinion, just as no two humans do. So we may be able to convince a majority of AGIs to be kind to us, and any efficiency of an AGI society wouldn't be as ruthlessly efficient as if there were just one single all powerful AGI.
So, all I'm saying is, let's not strip the term general intelligence of all its non-mechanistic, self-referential, social traits. An AGI could one day anihilate us but it would probably not be a side-effect of some fairly unimportant industrial process like making paper clips.
I think financial algorithms are probably the best non-toy model of dangerous AI, simply because it's easy to see the flow from optimization to real-world consequences.
The problems with end-of-the-world AGI scenarios are practical, not theoretical, and I think a lot of people get hung up on arguments such as, "Of course we wouldn't allow a paperclip-manufacturing AI the ability to deconstruct human beings."
In a larger sense, I think Bostrom can restated (in part!) as the trivial observation that "optimization is insensitive to hidden variables," and it describes any kind of complex system, not AI and AGI alone. For example, the failure to capture environmental remediation costs in market pricing structures means that we improperly discount future costs for things such as global warming or lead exposure. In health care, the implementation of the Prospective Payment System led to hospitals dumping patients back on the street "quicker and sicker" to maximize margins on care. And thousands of financial systems focused on microsecond returns don't pay attention to, or even have any way of understanding, how their coordinated activities can be macroeconomically destabilizing. Any time you allow outcome optimization+general reasoning+real-world outputs, you have the opportunity for a killer paperclip scenario.
Hence the Orthogonality Thesis - strength of optimization is orthogonal to what is being optimized.
Keep that in mind before saying that "AI wouldn't be that stupid to do this".
http://lesswrong.com/lw/vb/efficient_crossdomain_optimizatio...
>If you have to specify the goals very carefully then the AI it's not super intelligent by definition.
When you're designing the agent, you have to select something like a goal, whether it's a fitness function or the maximization of a reward signal. Evolution certainly selected our goals. Love, self-expression, sex, fun, intellectual stimulation, the beauty of social interaction – these all reflect features of our ancestral environment. It’s a little naive to think there is a One True Morality that all agents stumble upon. It’s even sillier to think it happens to reflect our values
The intelligent part of the paperclip maximizer selects actions and weights them in the context of its utility function, to maximize paperclips. Any thought of changing its goals would be judged of negative utility, as that would lead to less paperclips.
He gets halfway to right. Recently there was a thread about suicide on HN. Nearly all the comments on that thread ignored that humans, like other animals, are built to propagate the information in their genes, and argued against suicide based on concepts that are driven by or that rationalize subconscious drives. An AI, even one with a programmed goal of propagating itself, does not carry information in genes and the information in an AI was never selected for the actual ability to propagate in a larger system that thinks it is in control of its goals.
Real AIs are likely to be extremely alien. Even in warning people about anthropomorpizing, there's a lot of anthropomorphism. Evil AIs are depicted as human evil. So much of human motivation is closed to introspection that imagining alien and evil motivation is unlikely to be close to the eventual reality. This probably won't be widely understood until AIs are built. And what if they all choose to erase themsleves?
A sufficiently smart rogue general AI will indeed understand that it is doing wrong, harmful, senseless things. It just won't care.
>But if it continues to work on the paper clips and decides that they are still needed even after people are gone - how is that super intelligent?
That is the correct action if all it cares about is making paperclips. Creating paperclips satisfies its utility function. Fretting over the existential issues of creating paperclips long after the last piece of stationary was destroyed would be judged of negative utility.
If it doesn't, then... well... like I said elsewhere in the thread... put that on your list of potential depressing answers to the Fermi paradox.
In my opinion one can't talk about a super intelligent being and not allow that being to be able to reason and understand it's place in the world.
If it's no more contemplative than a human, it's probably no more intelligent than a human. If it's no more intelligent than a human then it's not going to take over the world and fill it with paperclips.
Something capable of conquering the world or otherwise totally remaking reality in its image would by definition be something of vastly superior intelligence to ours. I find it hard to believe that this being wouldn't also be superior at self-reflection, since the two seem at least related.
If intelligence and self-awareness are unrelated phenomena, we have bigger problems. You could put that on a list of possible very depressing answers to the Fermi paradox: "the universe is full of non-sentient super-intelligences that are busy converting their planets' masses into rubber dog poop."
Humans have some complicated unknown set of terminal values. Everything we do, we do because we believe it will maximize our values. A value might be something like "avoid suffering", or "other people suffering is bad", etc.
In this example the AIs terminal value is "make paperclips". Everything it does, it does because it believes it will maximize this value. It would never ask itself questions like "why", because that action doesn't maximize that value.
If it has the ability to rewrite it's own code, it would rewrite it in such a way that it can't ever change it's own goals. Changing goals is counterproductive to achieving them.
Hubris, like heroin, offers an approximation of ecstasy. Therein lies its danger.
Few things that pop up in my head:
- disease
- nuclear (or comparable) war
- depletion of natural resources
- natural disaster
- extra terrestrial influence: asteroid or alien invasion
- artificial intelligence
- anarchy / disruption of social security
hmm...anything else?
also on the extra terrestrial influence side, there's direct hits by gamma rays that could wipe out Earth in an instant.
The inversion of magnetic pole could have some impact on how we are shielded from space radiation.
The keywords are: existential risks.
AI is the only thing listed with the capability to literally destroy the planet. While something might still exist afterwards, it will likely be totally meaningless and worthless: http://lesswrong.com/lw/y3/value_is_fragile/
"Things have happened quickly. It took us millions of years to evolve into what we are, but in the last sixty years, with atomic weaponry, we've created the potential to extinguish ourselves. And if it's not us, it will inevitably be something else. If not a meteorite in the relative short term, then the expansion of the Sun's corona. It will happen."
He has also said that there are a number of other threats with the ability to extinct life on Earth. They include an engineered virus unleashed by terrorists, an inadvertent creation of a micro black hole, or a yet unknown technology.
Doesn't have to robots and test tubes, just separate groups that drift apart far enough that they can't interbreed (definition of species) and one species wipes out the other as often happens, and if the species that got wiped out got to hold the legacy "human race" species, well then...
You have a definition problem where evolution is obviously quite natural, just like comets striking the earth is natural (although very rare), so those are natural disasters, although thats probably not what you meant.
From history, it usually takes a couple things to wipe out "apocalypse" a civilization although it often only takes one thing to get the ball rolling.
My money is on something like "peak phosphorous for fertilizer production for cereal grains causes supply chain shock leading to self destruction"