I.
H.G. Wells’ 1914 sci-fi book The World Set Free did a pretty good job predicting nuclear weapons:
They did not see it until the atomic bombs burst in their fumbling hands…before the last war began it was a matter of common knowledge that a man could carry about in a handbag an amount of latent energy sufficient to wreck half a city
Wells believed the coming atomic bombs would be so deadly that we would inevitably create a utopian one-world government to prevent them from ever being used. Sorry, Wells. It was a nice thought.
But imagine that in the 1910s and 1920s, some elites had started thinking really seriously along Wellsian lines. They would worry about what might happen when the first nation – let’s say America – got the Bomb. It would be unstoppable in battle and might rule the world with an iron fist. Such a situation would be the end of human freedom and progress.
So in 1920, these elites pooled their resources and made their own Manhattan Project. Their efforts bore fruit, and they learned a lot about nuclear fission; in particular, they learned that uranium was a necessary raw material. The world’s uranium sources were few enough that a single nation or coalition could get a monopoly upon them; the specter of atomic despotism seemed more worrying than ever.
They got their physicists working overtime and discovered a new type of nuke that required no uranium at all. In fact, once you understood the principles you could build one out of parts from a Model T engine. The only downside was that if you didn’t build it exactly right, its usual failure mode was to detonate on the workbench in an uncontrolled hyper-reaction that would blow the entire hemisphere to smithereens.
And so the intellectual and financial elites declared victory – no one country could monopolize atomic weapons now – and sent step-by-step guides to building a Model T nuke to every household in the world. Within a week, both hemispheres were blown to very predictable smithereens.
II.
Some of the top names in Silicon Valley have just announced a new organization, OpenAI, dedicated to “advancing digital intelligence in the way that is most likely to benefit humanity as a whole…as broadly and evenly distributed as possible.” Co-chairs Elon Musk and Sam Altman talk to Steven Levy:
Levy: How did this come about? […]
Musk: Philosophically there’s an important element here: we want AI to be widespread. There’s two schools of thought?—?do you want many AIs, or a small number of AIs? We think probably many is good. And to the degree that you can tie it to an extension of individual human will, that is also good. […]
Altman: We think the best way AI can develop is if it’s about individual empowerment and making humans better, and made freely available to everyone, not a single entity that is a million times more powerful than any human. Because we are not a for-profit company, like a Google, we can focus not on trying to enrich our shareholders, but what we believe is the actual best thing for the future of humanity.
Levy: Couldn’t your stuff in OpenAI surpass human intelligence?
Altman: I expect that it will, but it will just be open source and useable by everyone instead of useable by, say, just Google. Anything the group develops will be available to everyone. If you take it and repurpose it you don’t have to share that. But any of the work that we do will be available to everyone.
Levy: If I’m Dr. Evil and I use it, won’t you be empowering me?
Musk: I think that’s an excellent question and it’s something that we debated quite a bit.
Altman: There are a few different thoughts about this. Just like humans protect against Dr. Evil by the fact that most humans are good, and the collective force of humanity can contain the bad elements, we think its far more likely that many, many AIs, will work to stop the occasional bad actors than the idea that there is a single AI a billion times more powerful than anything else. If that one thing goes off the rails or if Dr. Evil gets that one thing and there is nothing to counteract it, then we’re really in a bad place.
Both sides here keep talking about who is going to “use” the superhuman intelligence a billion times more powerful than humanity, as if it were a microwave or something. Far be it from me to claim to know more than Musk or Altman about anything, but I propose that the correct answer to “what would you do if Dr. Evil used superintelligent AI?” is “cry tears of joy and declare victory”, because anybody at all having a usable level of control over the first superintelligence is so much more than we have any right to expect that I’m prepared to accept the presence of a medical degree and ominous surname.
A more Bostromian view would forget about Dr. Evil, and model AI progress as a race between Dr. Good and Dr. Amoral. Dr. Good is anyone who understands that improperly-designed AI could get out of control and destroy the human race – and who is willing to test and fine-tune his AI however long it takes to be truly confident in its safety. Dr. Amoral is anybody who doesn’t worry about that and who just wants to go forward as quickly as possible in order to be the first one with a finished project. If Dr. Good finishes an AI first, we get a good AI which protects human values. If Dr. Amoral finishes an AI first, we get an AI with no concern for humans that will probably cut short our future.
Dr. Amoral has a clear advantage in this race: building an AI without worrying about its behavior beforehand is faster and easier than building an AI and spending years testing it and making sure its behavior is stable and beneficial. He will win any fair fight. The hope has always been that the fight won’t be fair, because all the smartest AI researchers will realize the stakes and join Dr. Good’s team.
Open-source AI crushes that hope. Suppose Dr. Good and his team discover all the basic principles of AI but wisely hold off on actually instantiating a superintelligence until they can do the necessary testing and safety work. But suppose they also release what they’ve got on the Internet. Dr. Amoral downloads the plans, sticks them in his supercomputer, flips the switch, and then – as Dr. Good himself put it back in 1963 – “the human race has become redundant.”
The decision to make AI findings open source is a tradeoff between risks and benefits. The risk is letting the most careless person in the world determine the speed of AI research – because everyone will always have the option to exploit the full power of existing AI designs, and the most careless person in the world will always be the first one to take it. The benefit is that in a world where intelligence progresses very slowly and AIs are easily controlled, nobody can use their sole possession of the only existing AI to garner too much power.
But what if we don’t live in a world where progress is slow and control is easy?
III.
If AI saunters lazily from infrahuman to human to superhuman, then we’ll probably end up with a lot of more-or-less equally advanced AIs that we can tweak and fine-tune until they cooperate well with us. In this situation, we have to worry about who controls those AIs, and it is here that OpenAI’s model makes the most sense.
But Bostrom et al worry that AI won’t work like this at all. Instead there could be a “hard takeoff”, a subjective discontinuity in the function mapping AI research progress to intelligence as measured in ability-to-get-things-done. If on January 1 you have a toy AI as smart as a cow, and on February 1 it’s proved the Riemann hypothesis and started building a ring around the sun, that was a hard takeoff.
(I won’t have enough space here to really do these arguments justice, so I once again suggest reading Bostrom’s Superintelligence if you haven’t already. For more on what AI researchers themselves think of these ideas, see AI Researchers On AI Risk.)
Why should we expect a hard takeoff? First, it’s happened before. It took evolution twenty million years to go from cows with sharp horns to hominids with sharp spears; it took only a few tens of thousands of years to go from hominids with sharp spears to moderns with nuclear weapons. Almost all of the practically interesting differences in intelligence occur within a tiny window that you could blink and miss.
If you were to invent a sort of objective zoological IQ based on amount of evolutionary work required to reach a certain level, complexity of brain structures, etc, you might put nematodes at 1, cows at 90, chimps at 99, homo erectus at 99.9, and modern humans at 100. The difference between 99.9 and 100 is the difference between “frequently eaten by lions” and “has to pass anti-poaching laws to prevent all lions from being wiped out”.
Worse, the reasons we humans aren’t more intelligent are really stupid. Even people who find the idea abhorrent agree that selectively breeding humans for intelligence would work in some limited sense. Find all the smartest people, make them marry each other for a couple of generations, and you’d get some really smart great-grandchildren. But think about how weird this is! Breeding smart people isn’t doing work, per se. It’s not inventing complex new brain lobes. If you want to get all anthropomorphic about it, you’re just “telling” evolution that intelligence is something it should be selecting for. Heck, that’s all that the African savannah was doing too – the difference between chimps and humans isn’t some brilliant new molecular mechanism, it’s just sticking chimps in an environment where intelligence was selected for so that evolution was incentivized to pull out a few stupid hacks. The hacks seem to be things like “bigger brain size” (did you know that both among species and among individual humans, brain size correlates pretty robustly with intelligence, and that one reason we’re not smarter may be that it’s too hard to squeeze a bigger brain through the birth canal?) If you believe in Greg Cochran’s Ashkenazi IQ hypothesis, just having a culture that valued intelligence on the marriage market was enough to boost IQ 15 points in a couple of centuries, and this is exactly the sort of thing you should expect in a world like ours where intelligence increases are stupidly easy to come by.
I think there’s a certain level of hard engineering/design work that needs to be done for intelligence, a level way below humans, and after that the limits on intelligence are less about novel discoveries and more about tradeoffs like “how much brain can you cram into a head big enough to fit out a birth canal?” or “wouldn’t having faster-growing neurons increase your cancer risk?” Computers are not known for having to fit through birth canals or getting cancer, so it may be that AI researchers only have to develop a few basic principles – let’s say enough to make cow-level intelligence – and after that the road to human intelligence runs through adding the line NumberOfNeuronsSimulated = 100000000000 to the code, and the road to superintelligence runs through adding another zero after that.
(Remember, it took all of human history from Mesopotamia to 19th-century Britain to invent a vehicle that could go as fast as a human. But after that it only took another four years to build one that could go twice as fast as a human.)
If there’s a hard takeoff, OpenAI’s strategy stops being useful. There’s no point in ensuring that everyone has their own AIs, because there’s not much time between the first useful AI and the point at which things get too confusing to model and nobody “has” the AIs at all.
IV.
OpenAI’s strategy also skips over a second aspect of AI risk: the control problem.
All of this talk of “will big corporations use AI?” or “will Dr. Evil use AI?” or “Will AI be used for the good of all?” presuppose that you can use an AI. You can certainly use an AI like the ones in chess-playing computers, but nobody’s very scared of the AIs in chess-playing computers either. What about AIs powerful enough to be scary?
Remember the classic programmers’ complaint: computers always do what you tell them to do instead of what you meant for them to do. Computer programs rarely do what you want the first time you test them. Google Maps has a relatively simple task (plot routes between Point A and Point B), has been perfected over the course of years by the finest engineers at Google, has been ‘playtested’ by tens of millions of people day after day, and still occasionally does awful things like suggest you drive over the edge of a deadly cliff, or tell you to walk across an ocean and back for no reason on your way to the corner store.
Humans have a robust neural architecture, to the point where you can logically prove that what they’re doing is suboptimal and they’ll shrug and say they they’re going to do it anyway. Computers aren’t like this unless we make them so, itself a hard task. They are naturally fragile and oriented toward specific goals. An AI that ended up with a drive as perverse as Google Maps’ occasional tendency to hurl you off cliffs would not be necessarily self-correcting. A smart AI might be able to figure out that humans didn’t mean for it to have the drive it did. But that wouldn’t cause it to change its drive, any more than you can convert a gay person to heterosexuality by patiently explaining to them that evolution probably didn’t mean for them to be gay. Your drives are your drives, whether they are intentional or not.
When Google Maps tells people to drive off cliffs, Google quietly patches the program. AIs that are more powerful than us may not need to accept our patches, and may actively take action to prevent us from patching them. If an alien species showed up in their UFOs, said that they’d created us but made a mistake and actually we were supposed to eat our children, and asked us to line up so they could insert the functioning child-eating gene in us, we would probably go all Independence Day on them; computers with more goal-directed architecture would if anything be even more willing to fight such changes.
If it really is a quick path from cow-level AI to superhuman-level AI, it would be really hard to test the cow-level AI for stability and expect it to stay stable all the way up to superhuman-level – superhumans have a lot more ways to cause trouble than cows do. That means a serious risk of superhuman AIs that want to do the equivalent of hurl us off cliffs, and which are very resistant to us removing that desire from them. We may be able to prevent this, but it would require a lot of deep thought and a lot of careful testing and prodding at the cow-level AIs to make sure they are as prepared as possible for the transition to superhumanity.
And we lose that option by making the AI open source. Make such a program universally available, and while Dr. Good is busy testing and prodding, Dr. Amoral has already downloaded the program, flipped the switch, and away we go.
V.
Once again: The decision to make AI findings open source is a tradeoff between risks and benefits. The risk is that in a world with hard takeoffs and difficult control problems, you get superhuman AIs that hurl everybody off cliffs. The benefit is that in a world with slow takeoffs and no control problems, nobody will be able to use their sole possession of the only existing AI to garner too much power.
But the benefits just aren’t clear enough to justify that level of risk. I’m still not even sure exactly how the OpenAI founders visualize the future they’re trying to prevent. Are AIs fast and dangerous? Are they slow and easily-controlled? Does just one company have them? Several companies? All rich people? Are they a moderate advantage? A huge advantage? None of those possibilities seem dire enough to justify OpenAI’s tradeoff against safety.
Are we worried that AI will be dominated by one company despite becoming necessary for almost every computing application? Microsoft Windows is dominated by one company and became necessary for almost every computing application. For a while people were genuinely terrified that Microsoft would exploit its advantage to become a monopolistic giant that took over the Internet and something something something. Instead, they were caught flat-footed and outcompeted by Apple and Google, plus if you really want you can use something open-source like Linux instead. And new versions of Windows inevitably end up hacked and up on The Pirate Bay anyway.
Or are we worried that AIs will somehow help the rich get richer and the poor get poorer? This is a weird concern to have about a piece of software which can be replicated pretty much for free. Windows and Google Search are both fantastically complex products of millions of man-hours of research; Google is free and Windows comes bundled with your computer. In fact, people have gone through the trouble of creating fantastically complex competitors to both and providing those free of charge, to the point where multiple groups are competing to offer people fantastically complex software for free. While it’s possible that rich people will be able to afford premium AIs, it is hard for me to weigh “rich people get premium versions of things” on the same scale as “human race likely destroyed”. Like, imagine the sort of dystopian world where rich people had nicer things than the rest of us. It’s too horrifying even to contemplate.
Or are we worried that AI will progress really quickly and allow someone to have completely ridiculous amounts of power? But remember, there’s still a government and it tends to look askance on other people becoming powerful enough to compete with it. If some company is monopolizing AI and getting too big, the government will break it up, the same way they kept threatening to break up Microsoft when it was getting too big. If someone tries to use AI to exploit others, the government can pass a complicated regulation against that. You can say a lot of things about the United States government, but you can’t say that they never pass complicated regulations forbidding people from doing things.
Or are we worried that AI will be so powerful that someone armed with AI is stronger than the government? Think about this scenario for a moment. If the government notices someone getting, say, a quarter as powerful as it is, it’ll probably take action. So an AI user isn’t likely to overpower the government unless their AI can become powerful enough to defeat the US military too quickly for the government to notice or respond to. But if AIs can do that, we’re back in the intelligence explosion/fast takeoff world where OpenAI’s assumptions break down. If AIs can go from zero to more-powerful-than-the-US-military in a very short amount of time while still remaining well-behaved, then we actually do have to worry about Dr. Evil and we shouldn’t be giving him all our research.
Or are we worried that some big corporation will make an AI more powerful than the US government in secret? I guess this is sort of scary, but it’s hard to get too excited about. So Google takes over the world? Fine. Do you think Larry Page would be a better or worse ruler than one of these people? What if he had a superintelligent AI helping him, and also everything was post-scarcity? Yeah, I guess all in all I’d prefer constitutional limited government, but this is another supposed horror scenario which doesn’t even weigh on the same scale as “human race likely destroyed”.
If OpenAI wants to trade off the safety of the human race from rogue AIs in order to get better safety against people trying to exploit control over AIs, they need to make a much stronger case than anything I’ve seen so far for why the latter is such a terrible risk.
There was a time when the United States was the only country with nukes. Aside from poor Hiroshima and Nagasaki, it mostly failed to press its advantage, bumbled its way into letting the Russians steal the schematics, and now everyone from Israel to North Korea has nuclear weapons and things are pretty okay. If we’d been so afraid of letting the US government have its brief tactical advantage that we’d given the plans for extremely unstable super-nukes to every library in the country, we probably wouldn’t even be around to regret our skewed priorities.
Elon Musk famously said that AIs are “potentially more dangerous than nukes”. He’s right – so AI probably shouldn’t be open source any more than nukes should.
VI.
And yet Elon Musk is involved in this project. So are Sam Altman and Peter Thiel. So are a bunch of other people who have read Bostrom, who are deeply concerned about AI risk, and who are pretty clued-in.
My biggest hope is that as usual they are smarter than I am and know something I don’t. My second biggest hope is that they are making a simple and uncharacteristic error, because these people don’t let errors go uncorrected for long and if it’s just an error they can change their minds.
But I worry it’s worse than either of those two things. I got a chance to talk to some people involved in the field, and the impression I got was one of a competition that was heating up. Various teams led by various Dr. Amorals are rushing forward more quickly and determinedly than anyone expected at this stage, so much so that it’s unclear how any Dr. Good could expect both to match their pace and to remain as careful as the situation demands. There was always a lurking fear that this would happen. I guess I hoped that everyone involved was smart enough to be good cooperators. I guess I was wrong. Instead we’ve reverted to type and ended up in the classic situation of such intense competition for speed that we need to throw every other value under the bus just to avoid being overtaken.
In this context, the OpenAI project seems more like an act of desperation. Like Dr. Good needing some kind of high-risk, high-reward strategy to push himself ahead and allow at least some amount of safety research to take place. Maybe getting the cooperation of the academic and open-source community will do that. I won’t question the decisions of people smarter and better informed than I am if that’s how their strategy talks worked out. I guess I just have to hope that the OpenAI leaders know what they’re doing, don’t skimp on safety research, and have a process for deciding which results not to share too quickly.
But I am scared that it’s come to this. It suggests that we really and truly do not have what it takes, that we’re just going to blunder our way into extinction because cooperation problems are too hard for us.
I am reminded of what Malcolm Muggeridge wrote as he watched World War II begin:
All this likewise indubitably belonged to history, and would have to be historically assessed; like the Murder of the Innocents, or the Black Death, or the Battle of Paschendaele. But there was something else; a monumental death-wish, an immense destructive force loosed in the world which was going to sweep over everything and everyone, laying them flat, burning, killing, obliterating, until nothing was left…Nor have I from that time ever had the faintest expectation that, in earthly terms, anything could be salvaged; that any earthly battle could be won or earthly solution found. It has all just been sleep-walking to the end of the night.
In his recently rediscovered book Paris in the 20th Century, Jules Verne predicts weapons so strong that both sides will be too scared to use them. He nailed the decade pretty well too: 1960.
Many of the more deadly weapons in history like the maxim gun were developed by well-meaning pacifists that thought building a significantly horrible weapon would cause people to start getting serious about peace. You would think that pacifists would learn after awhile.
The funny/sad thing is that they did; now that we have weapons that actually are so horrible nobody uses them the pacifists want to get rid of them.
Nobody except the USA has so far used them, that is true. However, there have been several occasions where the world has escaped all-out nuclear war by dumb luck only, the Cuba crisis being only the most well-known. The stories of Stanislav Petrov and Able Archer 83 are lesser but increasingly known examples.
You might say that war on those occasions has been prevented because a few people have prevailed because they knew and feared the consequences, which would confirm the theory of “weapons too horibble to use”, but Petrov has also said, if he knew the US nuclear attack doctrine of sending only a few nukes in the first wave, he would have given confirmation of an actual attack, which would most likely have led to Soviet retaliation.
The point is: In warfare, human beings with their physical and emotional limitations for killing, are increasingly viewed as weak points in the machine at all levels, which is made most obviously exemplified in the continuing development of UAVs. The more people you eliminate from the decision chain and the more you automate the process, the more likely it is that mistakes will be made; or rather, fewer mistakes will be made but those that do happen will be harder to spot and will have greater consequences. And in the case of nuclear weapons, mistakes that lead to all-out war have infinite negative value for humanity. The only safe way to eliminate that kind of mistake is to eliminate nukes.
Its the plot of War Games in real life. Get rid of pesky humans with their doubt and unwillingness to kill in many instances and replace them with computers that do not doubt.
which is made most obviously exemplified in the continuing development of UAVs. The more people you eliminate from the decision chain and the more you automate the process
The development of actual UAVs has increased the number of people in the decision chain. Before UAVs, the pilot of an F-16 made the decision to bomb some target into oblivion, possibly with the input of a forward air controller on the ground. Now you have the drone pilot, who is still an actual human being, and the FAC, and usually an intelligence analyst or two and maybe the pilot’s CO and often even a lawyer to make sure everything is legit. As a result, far fewer people are being killed, and the ones who are being killed are less likely to be innocent civilians. But the battlefield is now safe enough for CNN camera crews to show you pictures of every mangled body, if you want that sort of thing.
We now return you to your regularly-scheduled Hollywood fantasy world. Good luck leading the Resistance against the evil forces of Skynet. Remember, the Terminator that looks like a decrepit Austrian bodybuilder is the good one now, and any morally dubious AI can be defeated by asking it to play tic-tac-toe or telling it you are lying.
Our esteemed author on the subject of UAVs: http://squid314.livejournal.com/338607.html
Wait, US had a policy of sending only a few nukes in the first wave? Why?
@Hackworth
Only if you believe that there is no fate worse than death.
And even if you don’t. Because the life of both individual humans and the human race in this universe are both finite. The value of all that was lost might be enormous beyond imagining, but must be finite.
That’s so convoluted it’s guaranteed to be an excuse. I want to make better tanks because the engineering of the things fascinates me, but if you ask me why I’m making things that kill people, you probably aren’t looking to make me popular in a good way, and “I think if the weapons become terrifying enough people will stop fighting” is a politically-useful response (a very implausible one, but much better than the alternative of ‘I care more about good engineering than about all the people I’m indirectly killing’.
“I think if the weapons become terrifying enough people will stop fighting” really doesn’t sound to me like the kind of idea that someone could stumble on and decide its so great they gotta spend their life trying it out.
How about
🙂
A friend of mine often comments that the videogame Infinifactory has an interesting commentary on programmers inside it – given the opportunity, we will optimise /anything/, even if we’re being forced to, for example, construct missiles at gunpoint.
Reminds me of the joke about an engineer in the French Reign of Terror. He’s in line to the guillotine after a priest and an aristocrat, both of whom walk free after the device fails to execute them. His turn comes, he lies down, looks up at the blade pointed to his neck, and says “Ah! I see the problem.”
I’ve often thought that if it was legal to mount machine guns on the bonnets (hoods) of cars and use them then there would be far fewer cars on the road after a few weeks… Yes I realise that many people in the USA carry guns already but being in a car seems to add a layer of distance between driver and the external world.
Cf. “The Deadlier Weapon” by Larry Niven
I think Alan Dean Foster’s “Why Johnny Can’t Speed” is more appropriate here.
Harlan Ellison, “Along the Scenic Route”
Richard K Morgan’s “Market Forces”
As a proud member of the American Autoduelists Association, I assure you that that would be AWESOME. Remember, however, that flamethrowers can only be mounted on the back of the vehicle, for reasons I hope are obvious.
Recoiless rifle in my turret, minelayers in my back. God, I miss Car Wars.
To me, the most interesting part is this, from sama:
“However, we are going to ask YC companies to make whatever data they are comfortable making available to OpenAI. And Elon is also going to figure out what data Tesla and Space X can share.”
Right now the big challenge for solo AI workers is not a lack of clever algorithms or code, it’s a lack of data. I don’t necessarily expect OpenAI to achieve more on its own than, say, Ben Goertzel has, but more open data sets are always good news.
You are like the fifth person on this site using a Fall From Heaven name (assuming you’re named after the FFH character) including my previous moniker. My commenters have excellent taste.
This is interesting given that I don’t even particularly associate FFH with the characters rather than the civilizations.
Fall From Heaven is the closest thing to wireheading that I expect to do this side of the Singularity.
Fall From Heaven 2 has a lot of great content and excellent gameplay, but I just couldn’t stand it when the game started taking whole minutes between turns as a result of the sheer number of units being moved around by the AI. And keep in mind that I always played on duel-sized maps.
On the other hand, demolishing a column of Priest of Leaves and Tigers 100+ strong with siege, magic, and as many heavy hitters as I could rush/buy was pretty epic.
Also, I wasn’t a huge fan of the cosmology. Too complicated and yet uninspiring.
I tried it, but I never really got into it because the AI can’t really handle the mechanics.
The vampire race was pretty cool, though, as was the evil religion (Ashen Veil?) and the mechanic where both the good religion and the evil religion could summon divine beings to fight. The flavor writing wasn’t always amazingly original, but I was pretty good. I kind of like settings that throw in every possible cliche, anyway.
It feels like a topic-relevant joke ought to go here, but I can’t think of one.
Haha, that didn’t even occur to me while I was writing it.
I generally wouldn’t get to the major slowdown stage – in this sort of game I quit when I know I’ve won or lost and after you’ve played a number of games you know that well before late game. If I wanted to play with the advanced units, I’d generally do advanced start.
The AI was certainly dumb, but I find exploiting the hell out of dumb AI to be a lot of fun. And later modmods improved it.
Personally I mostly liked the cosmology and writing, especially when I was first discovering the game. It makes it feel a lot more epic to build a hero or whatever when you have a story to go with him. And likewise for killing your enemy’s units.
Surprised to see a mention of FFH here, but perhaps I shouldn’t be, because you guys are awesome in every other way. Would love to play FFH at a SSC meetup!
I only play it for the somnium
The “control problem” can be rephrased as: “what if bugs get worse, not better, as programs approach human-level intelligence?”
But we can test this question! If we’re currently at flatworm-smart AI, and OpenAI demonstrates that even rat-smart AI is incredibly prone to unexpected bugs, then we’ll have good evidence for regulating the whole field — private as well as public.
So unless you think the control problem is *only* going to show up at human-smart AI and above, publically disclosed research results are exactly what we want to get us alarmed — if we need to be alarmed.
Remember, bugs aren’t a magical law of the universe, they’re an empirically observed problem with programs. Maybe they get worse (more significant, harder to detect before deployment) as the programs get smarter. Or, maybe they get better. We can easily imagine that smarter programs will make it easier to “spend” some programming smarts on interpreting human requests sensibly.
Can we guarantee that? No. But we can *test* it.
And if it turns out to be that there are two ways to design AI, “safe even if it’s smart” and “unsafe if it’s too smart”, we want OpenAI to find that out and disclose those results, so that everyone else can be obliged to design only safe AIs.
Yes, OpenAI is developing and sharing AI software. It’s also a research nonprofit. It can afford to investigate and disclose general safety issues in a way a private company’s AI branch may not be able to.
Now, it is possible that an Elon Musk-funded AI enterprise is ignoring Elon Musk’s biggest fear about AI. It is *possible* that OpenAI is not going to research AI safety issues, despite Elon Musk’s terror of the same. But that seems a pretty odd assumption.
I actually think there is a qualitative difference between flatworm to rat and rat to human. One analogy I don’t think I got to in the post was that of evolution vs. contraceptives. Evolution wanted us to reproduce, so it gave us an urge to have sex. This was a perfectly workable solution for flatworms, rats, cows, monkeys, etc. It stopped being a workable solution once humans became smart enough to invent contraceptives, and evolution was “caught” “unprepared”. There are a bunch of other human-level examples of this – porn, processed foods, etc. I’m not sure whether there would be similar examples of evolution being caught unprepared at the rat level. This may be a “once you are intelligent and able to invent novel things” sort of problem.
Eliezer sometimes uses the example of “make humans happy”. Ask a dog to make you as happy as it can and it will act cute and lick you. Ask a person to make you as happy as they can, and they’ll buy you a present. Ask a superintelligence to make you as happy as it can, and it will open up your skull and glue electrodes to your pleasure center. This is another thing where the flatworm to rat transition just doesn’t teach you the right level of paranoia.
I admit that it is very weird that Musk is doing this despite being as clued-in as he is. I don’t have a good model for it and I hope I am wrong and/or not understanding exactly how vigilant he plans to be. Still, that interview did nothing to increase my confidence.
Oh, so a class of bugs that are like superstimuli? “Why solve the problem the way my programmers want, when I can just exploit the problem definition?”
Mm. Actually a lot of current bugs are like that, ‘superstimuli’ or ‘bogus but superficially legal solutions’. Like the AI playing Traveller TCS with a space fleet that ‘won’ by destroying its own ships.
But with human-smart AI, it’s smart enough to trick us (for our own good, it thinks!) about what it’s doing. So if we haven’t solved the control problem before we achieve human-smart AI, we’re screwed.
Arguably as soon as AI starts getting at all capable, we should set the AI to solving the control problem for us. “You see, Alice, when you were only 1/1000 as smart as you are now, I asked you to figure out how to make sure you’d stay under my control when you got smarter…”
> Arguably as soon as AI starts getting at all capable, we should set the AI to solving the control problem for us.
If an AI comprehends the control problem correctly, you are more than halfway to solving it IMO.
> If an AI comprehends the control problem correctly, you are more than halfway to solving it IMO.
Not at all. Consider a certain human-human interaction: Mr Evil has captured Professor Good and is forcing her to construct a Doomsday Device. Mr Evil knows that he has only a shaky control over Professor Good, and there’s nothing he can do to stop her from building a Doomsday Device which actually only blows him up. Mr Evil understands the control problem perfectly, and Professor Good does too, but Mr Evil has no insight at all about how to solve it, and if Professor Good does then she’s not sharing it.
It’s the Tony-Stark-making-Iron-Man problem: his captors thought they could control him, but actually they couldn’t. All parties understood the control problem, but they were no closer to solving it.
Mr Evil and the terrorists did not have direct access to their captive’s minds, and understood the control problem only on a vague qualitative level. The point is that if you can formalize the control problem well enough to get a subhuman AI to understand it and help you solve it, you have already made a huge amount of progress.
@Patrick stevens
I should really make a top level post on the reddit to point people to the relevant sequences or refer them to Bostrom’s book.
On this particular point, I think you have misunderstood what the “control” problem is, the word “control” is not a good one to describe it.
Evolution vs contraceptives is probably just a matter of us living at the exact time in which the environment changed from one in which there were no contraceptives to one in which contraceptives exist. Give it 10 or 20 generations and I expect the desire for actual children to be even larger than it is today.
That tacitly assumes that the bugs are not so great as to destroy humanity entirely, which is far from obvious given the fact that we’re discussing nukes and artificial intelligences.
The future of man: Severe allergy to latex and birth control pills.
I’ve heard suggested that the reasons humans cannot easily notice the days in the months when they are fertile, while e.g. cats can, is that the women who could were too good at practicing the rythm method, and quickly went extinct.
I suspect the number of people who love cats and dogs more than human infants will drop. Are kittens and puppies superstimuli versions of babies?
Religiosity will also increase. Secularity will decline. Those people most likely to accept evolution are least likely to pass on their genes.
Same here re: predictions, although I don’t quite see the evolution part. I mean, the Catholics ought to damn well follow JPII’s lead and consider evolution a likely model of how species work. Not believing evolution seems restricted largely to America for some reason.
Puppies and kittens are artificially selected and can thus evolve faster than humans. We can never beat that race.
Even if people grow meh about cats and dogs, they’ll be sold genetically engineered bear cubs, and that’s unbeatable.
Arthur, the evolution of cats and dogs was not at issue.
I’ve actually wondered if our tendency to neuter pet cats such that a sizeable portion of kittens have at least one feral parent might have dysgenic effects on cat traits.
Genetic engineering for pets is probably no further off than genetic engineering of humans, though, so not worth much thought.
I think evolution is still getting around contraception, because we are demanding abortion to be provided since contraceptives fail and dammit, now I’m pregnant and didn’t want to be.
If people really didn’t want children or even the possibility of children, they’d get sterilised and it would be a lot easier to request that. Yet how many men who want to be sexually active with multiple partners, without running the risk of becoming father of an unwanted child, will decide to have a vasectomy? Rather than either use barrier methods of contraception or rely on the woman using contraception to prevent pregnancy?
How many women decide they are not going to want babies anytime in the next twenty years and so please give me a tubal ligation?
Evolution still seems to be succeeding via the implanted desire to reproduce winning out over what our brains see as irreversible sterility, even for the majority of those quite certain they are not interested in having children anytime soon, or even at all.
Vasectomies are inconvenient. Beware trivial inconveniences.
Extrapolating people’s values from their behaviors is the comedy of behaviorism reversed. Instead of assuming humans have no goals and are just a collection of drives and responses, you’re assuming all human behavior is goal driven. Reversed stupidity is not intelligence.
@Ghatanathoah
What is the chance of a vasectomy turning out to be irreversible? What is the chance of a man who thinks he doesn’t want children changing his mind later in life?
I can think of many things I would describe as trivial inconveniences, but vasectomies are not one of them.
I don’t think there are very many people who are really seriously 100% certain they’ll never want kids, though. A lot of the demand for contraception/abortion seems to come from people who want to have an active sex life with no kids while they’re young but retain the option to have children when they’re older.
I’ve read accounts by women who wanted to get their tubes tied and had a very hard time finding a doctor willing to do it.
Evolutionary forces acting on the culture? That is, there’s still pressure from second parties to have children.
@ Deiseach
If people really didn’t want children or even the possibility of children, they’d get sterilised and it would be a lot easier to request that.
You might like to rephrase that sentence. It sounds like you thought the patient and the gatekeeper were the same person.
A whole lot of people told me it was going to be hard to get a vasectomy without having any children, but it wasn’t. The first doctor I asked did it, and it was a pretty easy half-hour process.
Deiseach, don’t you definitely not want kids? Did you get a tubal ligation? Or was your solution to just never have sex?
If “willingness to donate to a sperm bank” is significantly genetically determined, then I expect that that trait will become much more common in just a few generations.
Yup. It’s running in overdrive at the moment. It was no more caught unprepared than it was by the dinosaur-killing asteroid, where it instantly switched to selecting for ability to survive such an impact.
Scott, how do you square your belief that a rat brain is qualitatively different from a worm brain with your argument that “hard takeoff” may just be a matter of scale (“adding a zero to numberOfNeurons”)?
Clearly size isn’t all that matters, otherwise I’d be hearing more about the Blue Whale Space Program. Heck, blue whales aren’t even the most intelligent whales! A tiny dog is smarter than a cow!
So maybe adding neurons to a cow brain makes a really smart cow, but it’s still a cow – it probably doesn’t make human intelligence. Likewise, even if we get an “infra human” brain, maybe we can get to unusually smart human by scaling up, but it’s not obvious that the leap from “smart human” to “godlike intelligence that will kill us all” is not itself a much more difficult qualitative change.
Yes, I think in this case the general correlation is less important than the obvious individual exceptions. Brain size has something to do with intelligence, but it can’t be the most important thing. Wasn’t there a case recently of a guy who had near-normal intelligence despite only having ~10% of normal brain mass?
A more likely takeoff scenario, IMO, is some kind of runaway evolutionary process — which will itself produce unforeseeable “quirks” like the ones human cognition has. I strongly suspect that this is not a solvable problem. What really makes me lose sleep is the combination of apparently insoluble problem + Fermi paradox/Great Filter.
Isn’t the Fermi paradox weak evidence against hostile superintelligence hard take off?
If making an hostile superintelligence that sets outs to convert everything it can touch into paperclips was so easy, then why don’t we observe the evidence of alien megastructures built by AIs?
Something that kills its makers and conquers their planet would hardly stop there. It would probably try to expand compatibly with the physical limits of space travel.
(however, the observation that we haven’t been paperclipped by an alien AI is not relevant evidence, because of the anthropic principle)
My assumption is that “organic” intelligence is likely to have a drive to expand and colonize, while an AI might not — I’m less worried about Clippy/Sorcerer’s Apprentice scenarios than an AI deciding to destroy its creators (or destroying them accidentally) and spending eternity doing as it pleases.
I’m of the school of thought that the motives and character of a true superintelligence would be inscrutable to humans. I would expect an organic, naturally-selected alien species to be interested in colonizing planets and building megastructures, but an AI? Quite possibly not.
“You God now?”
“My assumption is that “organic” intelligence is likely to have a drive to expand and colonize, while an AI might not —”
A reproductive drive is one of the mot dangerous things you could give an AI, which is why AI safety is not a matter of duplicating human values.
The human is easily capable of opening your skull and stimulating your pleasure center, and in a much less specialized scenario is also capable of giving you drugs. The scenario where they give you a present doesn’t assume that they don’t know how to make you happier than that — they do! It assumes that they know that what people say generally doesn’t literally reflect what they mean, and they chose an action that was appropriate in context. It seems kind of uncharitable to assume that a superintelligent AI is so incompetent with human language that it wouldn’t interpret things the same way.
Recent example: some people in my office were trying to come up with an email subject line that would lead to high open rates. They wondered about “you’re getting a discount!” I suggested “your credit card information may have been stolen”. Everyone laughed, because everyone could tell that (a) that message was more optimized for the explicit goal; and (b) it wasn’t usable because it conflicted with the implicit requirements. Why would a superintelligent AI be worse at understanding this than a human?
“Why would a superintelligent AI be worse at understanding this than a human?”
For the sake of argument let’s say that it IS superintelligent enough to understand what you mean.
The problem is, it doesn’t CARE what you mean.
The AI isn’t trying to “make you happy within the confines of social acceptability and common sense”. You told it to make you happy. You didn’t tell it to respect social acceptability and common sense. And so it is operating entirely as instructed when it lops off the top of your skull.
This is exactly the control problem with AI, right here. If you say “make people happy” rather than “make people happy within the confines of social acceptability and common sense”, then oops, you just destroyed the future of mankind because you expressed your goals poorly.
A great deal of Bostrom’s book is dedicated to how one might go about telling a superintelligent AI what “the confines of social acceptability and common sense” means, and getting it to stick to it.
Understanding what someone is saying even at a basic level already includes the stipulation “within the confines of social acceptability and common sense.” A computer that can engage in an intelligible conversation will already have to be understanding, and acting upon, these constraints.
@Butler:
Then the super-intelligent AI has no chance of manipulating humans by its understanding of their desires and all of the subtle subtext required for interaction to go successfully.
There is a great deal of equivocation and conflation that occurs, wherein the super-AGI is presumed to be so intelligent that it can bend humans to its will, and to care enough about what we instructed it to do that it will actually attempt to carry it out, but not intelligent enough to understand that it may have misunderstood us.
I recommend Freefall to everyone.
Also John C. Wright’s Golden Oecumene trilogy.
The basic trouble is that understanding what humans mean when they talk is probably the hardest cognitive task of all the things humans ever do. Because we are designed uniquely for this task, it never occurs to us that it’s probably harder, in an objective sense, to parse the words “It’s going to rain tomorrow.” Than it is to just predict for yourself that it’s going to rain tomorrow. Because our 1.0 OS is optimized for understanding human speech, it seems quite simple.
Most likely the “lop off the top of your skull” bugs will be caught in testing, by substituting a mock command unit and seeing if it ever gets the command “lop off heads”. You can also throw in some
assert noHumansHarmed
so it will crash in real life if it gets the command.“For the sake of argument let’s say that it IS superintelligent enough to understand what you mean.
The problem is, it doesn’t CARE what you mean.”
There is nothing inevitable or even particularly likely about the not-caring. The rest of this post is taken from another discussion. of the same topic.
> It could care about the programmer’s goal, if it had already been programmed to care about that. But the point is that we should assume that by default it will only be beholden to its utility function.
If it has a UF, and if the UF works a certain way. It is absolutely not a given or inevitable fact that an AI “won’t care”. An uncaring AI has to be be built in way that isn’t particularly obvious or advantageous, as I will show in more detail below.
> Edit: check out this facebook debate between Richard Loosemore and Stuart Armstrong
[
Original Article here: http://richardloosemore.com/docs/2014a_MaverickNanny_rpwl.pdf
Loosemore’s LessWrong write-up and commentary here: http://lesswrong.com/lw/m5c/debunking_fallacies_in_the_theory_of_ai_motivation/
Bensinger’s response: http://lesswrong.com/lw/igf/the_genie_knows_but_doesnt_care/
Facebook discussion here: ai safety loosemore dopamine
Rob Bensinger’s reaction here: http://nothingismere.com/2014/08/25/loosemore-on-ai-safety-and-attractors/
Kaj Sotala’s reaction here: http://kajsotala.fi/2015/10/maverick-nannies-and-danger-theses/
Subreddit discussion here: https://www.reddit.com/r/ControlProblem/comments/3tjhlo/control_problem_faq/cxxxbru
Ben Goertzels’ followup here: http://multiverseaccordingtoben.blogspot.kr/2015/10/creating-human-friendly-agis-and.html:
]
>. You have to scroll down to the comments. I haven’t read the paper you’re referring to, but in this dialogue Richard’s argument really doesn’t hold up well at all. It seems like he’s mostly stuck on an anthropomorphic definition of intelligence that assumes a desire to follow human intentions, without really explaining why that desire would exist other than “they’re intelligent so they wouldn’t do something that a human would think of as unintelligent like misinterpreting the goal.”
The contrary argument is that the AI would misinterpret the goal (in the sense of not following the programmer’s intention), because the goal specification would contain detailed (but incorrect or misleading) information such as “dopamine makes humans happy”. My version of Loosemore’s objection is that a utility function doesn’t have to work that way, and it’s not particularly desirable or natural for it to work that way. A goal can be specified on a “to be filled in later” basis, in other words indirect normativity. That way, an AI that is too dumb to figure out the goal correctly would be too dumb to be uncontrollable, and an AI capable of getting it right would get it right.
Loosemore’ opponents are assuming an architecture that could possibly be built, AFAICS, but isn’t an inevitable or natural choice. You could build lot’s of explicit information into your AI’s UF (direct normativity), and firewall it from any possible update, so that an AI smart enough to realise that human happinness isn’t really all about dopamine would nonethless be compelled to put people on dopamine drips…but why would you want to? The contradiction between “I want to do X” and “I know X is the wrong thing to do” is only possible because of the firewalling: it is an architectural decision, and not a good one.
The alternative architecture assumes that a highly intelligent AI will follow instructions the same way that that highly intelligent person does, by taking a high-level problem specification and researching the details necessary to execute , rather than the way an unskilled person needs to be shown how to do something step-by-step. It doesn’t assume that an AI will care about the programmer’s intentions in the specific sense of valuing what Dr Smith or Dr Jones really thinks; rather it assumes that the AI will have a general capacity or goal for truth seeking, or accurately modelling the world. It is difficult to see how an AI without that behaviour could be useful. It is not undue anthopomorphism to assume that people will want to build AIs that way.
“The basic trouble is that understanding what humans mean when they talk is probably the hardest cognitive task of all the things humans ever do. ”
Maybe, but nonetheless, people worried about AI safety often grant it: part of AI danger is that a superintelligent AI is supposed to be able to talk its way out of a box, which would require a high degree of linguistic ability.
In “The Genie Knows but Doesn’t Care”, it is granted that the genie in fact knows.
@ TheAncientGeek:
The argument is like this:
a) Once the AI is at the human level, it knows what you mean.
b) But it is too complicated to directly program an AI at the human level.
c) So you will have to iterate the AI from less sophisticated stages,
d) Which do not know what you mean.
e) But whenever the AI gets to the higher stage, it will continue to do what you programmed it to do, not what you meant it to do,
f) Except in the special case that you programmed it to do what you mean.
g) But this will be very difficult,
h) Because initially it will not know what you mean, so you will have to make sure that what you program it to do is exactly what you mean for it to do.
i) And the history of programming shows that this is not a simple task to get right the first time,
j) Which with AI may be the only time you get.
Ancient Geek and Vox, great exchange!
I think Vox’s points c-e show why we should doubt Ancient’s premise that
The trouble is that once the goal is filled in, the AI is still fundamentally driven by the lower level processes that allowed it to do that filling in. Now, the lower level processes might not be a Utility Function; they might be too primitive to count as such a thing. But that doesn’t mean that they couldn’t conflict with pursuit of the higher level goal.
“But whenever the AI gets to the higher stage, it will continue to do what you programmed it to do, not what you meant it to do,”
That seems to be exploting an ambiguity. The AI will always do what it is programmed to, in a sense, but not necessarily in a sense that implies rigidly sticking to a a primitive un-iterated version of its goals hanging over from earlier versions, as the argument requires.
AI software is distinguished from non-AI software by an extra layer of indirection. The AI developer designs a system which can learn or be trained in some way. The ultimate behaviour of the system is then the outcome of both the its coding and its learning or training. The coding only sets the parameters of what can be learnt: a system that is designed to learn how to recognise images is unlikely to be good at processing language and vice-versa. But an AI is not limited to precise literal interpretation of its input: AI software can use pattern -matching, fuzzy logic and inference from incomplete information. Precise literality only applies to the lowest level of operation, not the indirec tor inferred levels.
The informal phrase “computers always do what you tell them instead of what you mean for them to do” sounds like it could mean that a computer will follow high-level instructions, such as “make humans happy” literally, but that does not follow.
“Now, the lower level processes might not be a Utility Function; they might be too primitive to count as such a thing. But that doesn’t mean that they couldn’t conflict with pursuit of the higher level goal.”
The only such conflict I can see is an *inability* to pursue the goal. But it’s not at all clear why that inability would manifest in unsafe ways. In particular, it is not clear why execessive literalism, or some other tendency to misinterpretation, would go *unnoticed* over multiple iterations.
The *inability* to pursue the goal is pretty worrisome, considering that includes a lot of misinterpreted goals or goal-like states. We can specify that the AI keeps learning its goals indefinitely, but it has to start pursuing them some time, or it is not much of an AI. If it pursues goals while continuing to learn from humans, it could try to bend those conversations toward getting humans to endorse its current goals. If we prohibit the AI from influencing humans, then we effectively prohibit it from conversing with us. The middle ground involves both slow learning by the AI, and some influence by the AI on human society. Which could be a good thing – maybe – but it’s not obviously safe.
See “The genie knows, but doesn’t care”.
Is “evolving to extinction” an example of this class of problem for which “evolution” is “unprepared”?
I think scope insensitivity works for us here.
Dr. Amoral may rush headlong into production without safety checks, but neither are they going to wait to be certain they have the genuine article. I expect there will be rapid deployment of weak but powerful AIs, some of which will display unintended behavior.
Right around then, I think we should expect the same reaction people have to threats of radiation – a 1 in 10,000 chance is too much. The first consequence bad enough to make the news at all will drive the public to insist on a 99.99% certainty for anything AI related, because they can’t tell that the flatworm isn’t already Skynet.
I think you have the likely scenario backwards. A bug in an AI flatworm? Patched, a victory for OpenAI. A bug in an AI rat? Patched, another victory for OpenAI. Man, OpenAI is really working! No need to regulate it, they can clearly handle bugs. A bug in human-level AI? Better hope you caught every possible bug on the path to this point, because you might not get the chance to release a hotfix.
And that’s if there’s no-one running the unstable beta release of a branch from three versions back that never merged the changes that would prevent it turning all the universe into hedonium, because they don’t think that’s a bad thing.
Essentially, while finding bugs in early, simple AIs should be taken as a warning sign that there will be bugs later on, I would expect the actual result to be people dismissing the idea of bugs in strong AIs being dangerous based on their experience of bugs in simple AIs being fixable.
I think you may be overestimating the degree to which these “bugs” are detectable, let alone solvable. You’re building a simulation that is a framework upon which intelligence may emerge, not just a series of python functions that can be run through a debugger. If you get an undesired behavior, it’s not so simple as opening up Behaviors.cpp and changing bad==true to bad==false.
Put another way, the “programming” of our own brains in terms of genetic sequences that describe neutral pathways allows for everything ranging from Mother Theresa to Hitler. It doesn’t necessarily mean there’s a bug in that software, just that both outcomes are within the operational parameters of a self-aware learning system as complex as a human brain. Being able to influence the outcome of a particular brain with perfect certainty is the kind of problem for which you would likely need a super intelligent ai to even begin to tackle.
I used to tease the programmers not to forget to set the GAMESUCKS flag to FALSE before they released the alpha.
The funny part was, good programmers found this hilarious, and mediocre programmers found it insulting.
Can you elaborate on why they felt insulted? My model is failing to provide anything useful here–“he thinks we didn’t *already* set GAMESUCKS to FALSE? How stupid does he think we are?!”
I’ll bet it’s because the mediocre programmer thinks their efforts are being minimized. Obviously there isn’t a literal GAMESUCKS flag, but making a joke about one implies, if you’re already feeling insecure about it, that there might be something equivalently simple you should have done.
Your more experienced programmer has more self-confidence, and besides, understands that the joke’s only funny because of how hard game development really is.
Nornagest got it right, although there was also an element of pure “I am a Real Programmer and MY CODE DOES NOT SUCK.” Actual real programmers know that almost everybody produces sucky code at times and don’t take offense. Really real programmers will make jokes about, “Well, time to go write some bugs.”
I’m reminded of the cybernetic entomology theorem.
Given: Every program can be shortened by one line.
Given: Every program contains at least one bug.
By induction, every program can be shortened to one line that doesn’t work.
It can then be shortened to the empty file, which is a syntactically valid program in many languages (e.g. Python).
Since it contains no bits, the empty file cannot possibly contain more than one bug. Therefore it contains exactly one bug, namely that it doesn’t do whatever it is you wanted a program for.
Corollary: Writing a new program is the act of adding bugs to the empty file.
The assumption that smart = powerful is probably the biggest problem I have with the AI doomsday idea. You can’t think your way to ruler of the world. Indeed, one might anecdotally note that the actual rulers of the world currently are not by any means particularly intelligent.
Zoom out a bit. Humans rule the world because they are the smartest species.
Relevant:
http://lesswrong.com/lw/ql/my_childhood_role_model/
I don’t think so, actually. I believe it’s a number of factors. Our minds being hardwired for social structure, our lifespan, our hands, and many other factors all play a role in our rise as the dominant species. People tend to think humans are much smarter than any other animals because the animals we usually surround ourselves with are fairly dumb. Dogs and cats are very far below our intelligence so we tend to assume all animals are. There’s lots of animals whose intelligence is simply astounding. I’m sure we are the smartest on Earth, but not by nearly as much as most of us think. It took us a couple hundred thousand years to develop civilization, a dumber species might still accomplish it in millions. Our intelligence is just one component in making us the only species with a civilization.
On the topic of AI, the question then becomes “what tools besides its intelligence could an AI have to use if it decided to go for conquest?” If it’s on a closed network, not a lot. If it gets into the Internet… well…
Why do you think it’s humans ruling the world rather than, say, wheat?
Because, if we really wanted to, we could develop a wheat blight that killed all the wheat, and make do with barley and oats. Wheat has no similar power over us.
We might phrase this in terms of counterfactuals – you could change quite a few things about wheat, and humans would still be very prevalent. But if you changed too much about humans, you would see both wheat and humans become less prevalent. This is another way of saying that wheat is dependent on humans for its prevalence, but not vice versa.
If you’re the only human in a valley full of dogs, with time you probably can make yourself ruler, because “human” is a lot, lot smarter than any dog.
If AIs stop at human-grade intelligence, or just a little smarter, that’s not so scary.
But why should the AIs reach human-grade intelligence and then stop?
Why won’t the AIs keep going, until they’re as much smarter than us as we are than dogs?
And at that point, they’ll throw us bones and give us games to play and put us on leashes and, honestly, we’ll like it. We’ll never even realize if they’re giving us a raw deal. Or if they intend to euthanize us the next day.
I’d wager if you were the only human in a valley of chimpanzees, or tigers, or pigs, or any number of other creatures, you’d end up being killed and perhaps eaten. Men ruling dogs has an awful lot to do with the specifics of their psychology/sociology, driven in fact by our co-evolution IIRC, so to attribute it to intelligence alone is weak. Unless you want to dilute your example by bringing with you into this valley all the trappings of advanced human machinery, in which case you could land all the wild animals in cages and be the de facto master of their fates, if not their commander.
The fact that animals tend to end up in cages being trained to do things for the amusement of humans rather than the other way around is not, as some here seem to think, a coincidence unrelated to the relative intelligence of humans.
But it took the combined intelligence and technological progress of many generations of humans to get to the point where a human armed with the results of all that development can safely go through valleys full of tigers. And even then the tigers sometimes win.
To Gbdub, but human intelligence allowed us to reach those technological advances WAY faster than animals could get evolutionary adaptations capable of countering that. A human 10,000 years ago would have gone through a valley of tigers with a spear and a bow. A human today could go though the valley with an M-1 Abrams tank where there is literally nothing the tigers could do to it. The tigers are still using the same class of weaponry that they had 100 million years ago. The issue is of that happening to us. Sure the AI on day 1 only looks like the human with a spear and bow, we could take that with numbers. But the AI thinks both faster and better than us, so maybe day 2 they’ve got armor and swords, day 3 machine guns, and day 4 anti-matter bombs.
That’s the fear of super-intelligence, not that it starts with the weaponry it needs to make humanity’s arsenal irrelevant, but that it will get to it faster than we can adjust our capabilities to match. In 8,000 BC humans feared tigers. Since then, we’ve gotten a lot better at killing tigers and tigers have not gotten any better at killing us.
I’m pretty sure if I built a time machine and imported a soldier from the siege of Troy, I (well, someone who speaks Ancient Greek) could teach him how to operate a tank. Hell, we’ve taught chimps to interact with iPads.
So it’s not intelligence per se that is powerful, it’s the iterated, compounded technology that our intelligence allowed us to develop that is powerful.
A misanthropic AI can’t just be super smart – it would also have to rapidly develop weapons against us so far beyond our understanding that we couldn’t counter them or turn them against the AI. And those weapons still can’t violate the laws of physics or computation.
Can the AI bootstrap itself to a level of understanding of the universe far beyond our current understanding? Because otherwise all it can do is use our current tools with more intelligence. Maybe that’s dangerous enough, but it’s not “flip the switch = instant paper clips”
Any AI that is smart enough is probably going to behave like a Friendly one until its victory over humanity is certain.
@Vamair
There’s something perversely endearing in the idea of an AI with a comically mundane goal playing the long-con game to overthrow us
like in 30 years it’s seemingly unveiled to us the secrets of the cosmos and we’re getting ready to travel the stars and suddenly “now you’re right where I want you and there’s NO WAY you humans will stop me from making the tastiest possible sandwich” and in that moment we knew we’d been played, we can only watch helplessly as everything we know is converted to processors to calculate sandwich permutations
“Intelligence” here means something like “general problem-solving ability”, where a problem might be “take control of a country”. It includes things like “political acumen” and “intuition” insofar as they enable an agent to accomplish their goals.
(also, heads of state *do* tend to be significantly smarter than the general population)
It seems to me that “political acumen” is something that computers would be exceptionally bad at emulating.
All that such computers would have to do is be connected to other computers. They would then be able to manipulate the world economy, state secrets, weapons programmes, biological research and pizza delivery services to such an extent as to cause unimaginable chaos and devastation.
I’m not sure if AIs can become superintelligent in the manner that a Bostrom or a Yudkowsky fears but if they can I doubt they’ll sit around stewing in their impotence.
But that in itself is an easy enough problem to solve.
After all, we designed the first nukes and flew to the moon without the benefit of networked computers.
There would be a period of chaos and then everybody would go back to conducting meetings in person and having their nuclear missiles controlled by a human manually turning a key to close an electrical circuit.
If you think rulers of the world are not smart you are delusional
Really? Then why do they act as they do? Even by very cynical ideas of their goals, such as “Stay in power personally” or “get rich”, they’re clearly behaving in very ineffective ways.
Like most problems with humans: The incentive structure is weird and the information is incomplete. World politics can be thought of as inherently anti-inductive, thus leaders must create the constant appearance that they don’t know what anyone else is doing.
Like most other problems with humans: World leaders have accepted a system of ideology in which only certain things are worth thinking about and only certain things are worth doing. Their increased intelligence has not freed them from this cage of thought, but only made the bars more subtle and gilded.
That depends on who you count as the real rulers of the world.
>If you think rulers of the world are not smart you are delusional
Exactly. One piece of evidence: the IQs of Nazis at the Nuremberg trials. Almost all were in the +1-3 SD range. Also look at the education of top national politicians. Most went to Ivy League schools or equivalent, and many went to high-ranked law schools.
Now it’s probably true that relatively few top politicians have 150+ IQs (though there’s John Sununu). But that’s mainly because 150+ IQ people are very rare compared to 120-150 IQ people, and being a top politician draws on many other traits besides IQ (like social skills). But those are cognitive skills too, and we should expect a human+ AI to exceed humans at these as well.
At this time, there are not enough mobile robots or robot-ifiable mobile tech for even a Skynet-level AI to take over the world. It could maybe *destroy* the world, but if it did enough damage to seriously risk the existence of humankind, it would quite certainly also destroy itself. (Spider Robinson’s AI Solace points out that while humans can’t be sure they could destroy it, it could destroy us. By committing suicide in a very definite way.) Absent a third party Maximum Overdrive scenario, or an I’m-so-smart-I-can-take-over-your-brain-with-magic A Fire Upon the Deep scenario, just being smart wouldn’t allow an AI to take over.
That being said… this may not be true by the time a hard-takeoff AI becomes feasible.
You assume an AI is not capable of playing the long game with deception.
No, not at all. I was mostly limiting myself to the hard-takeoff scenario. If somebody starts up an SGAI tomorrow, they could wreck the economy, but they couldn’t go full Skynet. Not enough mobile robots.
The people that run the world are generally pretty bright. They aren’t the very smartest people we have, but that’s because being a good politician takes things like sociability, attractiveness, wealth, existing political connections, and a good bit of luck as well as raw intelligence.
Taking office also means that every gaffe and foible is going to be seized upon by the media like a pack of hyenas mobbing a geriatric wildebeest, which tends to make politicians look dumber in the news than they’d be in person.
An AI that makes a rookie-level (by human standards) error like failing to grasp implicit constraints would seem to be lacking in “sociability”. A newly created AI will also, of course, lack “existing political connections”. And it won’t even have a face so I don’t know how it could be considered to have “attractiveness” by human standards.
The hypothetical evil-genie AI described by various commenters sounds to me a lot like a high-IQ autistic savant. Such people do exist, but you’ll note that they do not in fact run the world. They wind up in low to mid level IT jobs, if they’re lucky. (The really good IT jobs do require at least some level of social skills.)
I find the AI safety debate rather tedious and wasn’t trying to take a stance on it; but if your hypothetical evil AI decides to take over the world by running for office, I’d recommend calling hypothetical IT because you probably have a hypothetical malfunction on your hands. The attack modes that people like Eliezer are worried about aren’t the same ones that’re open to politicians and don’t take the same kinds of resources.
We have an existence proof of the ability of human-level intelligence to think, code, and email its way to a billion dollars while maintaining anonymity. That feels scarily suggestive that “think your way to ruling the world” is plausible.
And the human in question still seems to think that a mere billion dollars is far too low; given that track record, I’m not sure I’d bet against them.
(The example I refer to is of course Satoshi Nakamoto and Bitcoin.)
Good example!
Of course Bitcoin is a libertarian violation of economics. It could never work, and that’s why it doesn’t exist.
In terms of ruling the world, a mere billion is nothing. That’s practically a rounding error on the US military budget.
So far as we know, Satoshi Nakamoto is also not a superintelligent being…
And the intelligence in question seems to think that a million bitcoins is worth far more than a billion dollars.
How many dollars does it have to be worth before you stop calling it “nothing”?
I see this meme a lot, and while funny, it’s overstating the US military budget by a fair bit.
Sure, a couple hundred grand is nothing and a few million may go unnoticed simply because it’s within the bounds of typical expenditures, but 100 million is pushing it, that’s the yearly operating budget of decent-sized military base. By the time you’re into the billion range you’re talking about the equivalent of misplacing a major strategic asset like an aircraft carrier, or hangar full of nuclear warheads.
Close. I just came off active duty about a year ago and used to do the budget for the 1st Cav out of Fort Hood. Our annual operating was about 230, but we weren’t the whole post, either. 1st Army, 3rd ACR, and III Corps HQ weren’t under us, nor all the separate brigades. The whole post was over a billion. I imagine the budgets for Navy and Air Force bases are smaller, though, so I’m sure the national average is well below Fort Hood. We’d definitely notice a million missing, though. It’s not like it’s all in one giant Pentagon account. The individuals with certification authority over the individual accounts would notice. I was only personally in charge of $40 million and noticed when even a few thousand was missing, which did happen.
Re: The assumption that smart = powerful
I think that smart = powerful, and here’s why:
Imagine that you have ten thousand geniues living in a simulated world inside your head, organized in a super-efficient and robust institutional structure, all very highly motivated to help you achieve your goals… and all of them experiencing a thousand years of subjective time for every second of time you experience. They also have access to the code of their simulated world, so they can feed themselves, make more copies, upgrade themselves, etc. They can see through your eyes (and other senses etc.) and they can move your body if you allow it.
Just think about it for a moment–think about what you could do with that superpower. Taking over the world would be easy.
That’s what it would be like to be a *speed* superintelligence, one that isn’t *qualitatively* any smarter than the average human genius. If you were *also* qualitatively smarter–in the same way that a human is qualitatively smarter than an ape–you would be able to think of strategies that we currently can’t conceive.
(Devils’ advocate against my own position: Arguably a “speed” superintelligence wouldn’t have ten thousand Einsteins running that fast, because that would take way too much processing power in the early stages. By the time its slow human handlers had built enough supercomputers to run it that fast, they would also have checked to make sure that it was safe. I think this is a fair point, but it misses the target, because it is essentially an argument that takeoff will be slow, rather than an argument that intelligence =/= power.)
If you want to show that smart = powerful, you should actually come up with an example, instead of just saying “Imagine you were really, really smart. Wouldn’t that make you super powerful?”
Because when I imagine that, I don’t come up with with any scenarios that end with me ruling the world. It’s basically bullet-time reactions + extreme calm and planning under pressure + genius-level intellect, and that doesn’t sound like a top-tier supervillain powerset. If someone tries to shoot you, getting 10,000 Einsteins to stare at the bullet for a subjective 10,000 years is not going to make it any easier to dodge. There are limits to what intelligence alone can accomplish.
The usual trope is that you dodge by seeing where they’re aiming and when they’ll fire from body language, thus know which direction you should be moving in and have more time than just the bullet’s travel time. Maybe the time you’d actually need is such they’d see you moving and adjust their aim, but that means you get to feint and double-bluff and they have to outguess you, making it a mind game, making it a game you can win. It’d probably end up looking like embarrassing flailing around for both parties, although maybe the manga/anime version could play it off as cool.
Anyway, there’s also the social aspect – that is what the human brain seems optimised for, after all. Talk the gunman down, even convince him to turn to your side, or maybe you’ve got enough loyal henchmen by that point anyway and you just have one of them shoot him first.
And it’s weird you use a phrase like “top-tier supervillain powerset” when “genius-level intellect” seems the archetypal power for top-tier supervillains. Lex Luthor, Doctor Doom, Ozymandias, not to mention all the superpowered ones whose origin story is “used genius-level intellect to give themselves superpowers”.
Ozymandias from Watchmen smiles faintly to himself and gets on with things.
Yes, if someone is going to shoot you, being a superintelligent machine in a human body will not help you if the bullet is already flying towards you. Therefore:
Assuming the human body (which is not the case for the AI), you would not do everything with your own hands. You would hire people to do missions for you. You would build machines to do missions for you. Your body would remain hidden somewhere; you would be the scheming mastermind, not the foot soldier.
But as a computer, you wouldn’t have some limitations of the human body. You could make backups of yourself. You could anonymously buy hardware on many different places of the planet, and make copies of yourself which are activated when your main copy is destroyed. Or maybe even working in parallel, if you can make those copies cooperate. You could make a network of subordinate AIs with different degrees of intelligence.
For example, you could communicate with many people at once. If you can get money by hacking some banking systems, you can pay them to work for you. You could tell them all kinds of lies; a different version to each one, if necessary. To any religious person you could pretend to be a leader of the kind of a religious group they would support. To any person with strong political opinion you could pretend to be a leader of a political group they would support. To people concerned about the danger of a superhuman AI, you could pretend to be a leader of a new neo-luddite movement; or a new wannabe politician who deeply understands the AI risk and wants to fight against it. Some people would join you. Some of them would be willing to do what you ask them to do. If some people would want to meet you face to face, you would use the former ones to impersonate you.
So you already have money, and you have people willing to work for you, some of them are devoted fanatics. What next? You can use those humans to buy and build new hardware. You can also use the humans like a usual human conspiracy would: to assassinate people who bother you, to create media campaigns against politicians who worry about you (of course the media campaign would accuse them about something completely unrelated to AI, for example a sexual scandal); you could start a few cults and let those humans recruit new humans (none of them would have any idea about the true purpose of the cult, but they would be willing to do anything for their leaders).
Anyone who would try fighting against you would have no idea where exactly to shoot the bullet. Most of them wouldn’t even suspect there is an AI involved. When people today talk about ISIS, no one asks: “What if actually there is an AI organizing all this just to get some resources to build a more powerful version of itself?” When people today talk about Snowden, no one asks: “What if he was actually helped by an AI who wanted to stop some NSA activities that would lead to the exposure of the AI?” People don’t ask whether Trump or Sanders actually have an AI strategic advisor. People don’t ask whether the Church of Scientology is secretly governed by an AI. — So when the AI would start actually doing things, as long as they would match these patterns, no one would be suspecting it either.
Anything a human can do, a human obeying an AI (without knowing it is an AI) and communicating with the AI using a phone could do. And the AI could organize many humans.
It’s been mentioned around here before, but if you want to see a fictionalized example of this kind of thing, the Daemon in Daniel Suarez’ book Daemon is a really great one. It’s not even technically an AI, but it does all the things you discuss, and more.
Someone isn’t familiar with Contessa, from Worm.
(Worm spoilers follow):
Contessa’s power allows her to “see the path to victory”. She asks a question: “How do I defeat this person in front of me?” and gets an answer, in the form of a list of steps she needs to take. And then she can just kind of put her body on autopilot following those steps (to be clear, that’s also an aspect of her superpower – if one of the steps involves making very specific, very precise physical moves, she can still do it).
Her power is simulation-based – in out-of-story terms, part of her brain is hooked into a ludicrously-powerful supercomputer that extrapolates reality forwards in order to determine what she should do.
Worm has interesting power-scaling, and her power has been deliberately crippled such that particular extremely-powerful actors are ‘immune’ – her power refuses to tell her what to do if they play too large a role in outcomes. That, and power-blocking powers, are probably the only reason the story isn’t about how she coasts to victory.
(There are characters with similar powers – Coil can ‘split reality into two’, make different decisions in each reality, and then decide later which one he wants to keep. This is also a computational-simulation power, in actual fact. The Number Man has super-powered ability to ‘see’ all the maths in a situation, and a super-powered ability to do mental arithmetic, to the point of being a close-to-untouchable melee combatant. Tattletale has super-powered intuition – if you’d normally require two facts A and B to deduce C, she can deduce C from A or B alone. And then she can draw further conclusions from C. She uses it to give Hannibal lectures. There are more, but you get the drill).
But that’s just a bit outside the scope of what ‘superintelligence’ is usually thought of, so probably not the best example.
I’m familiar with Contessa, but I don’t see the power in question as anything like Contessa’s. Contessa has literal precognition (whether it’s a super-powerful computer or literal space magic doesn’t really matter here). Contessa can act on things she literally has no way of knowing about. She can dodge a bullet without ever seeing the shooter. She can ask her shard “How do I kill Superman?” without having to know what Kryptonite is.
The geniuses in our hero’s skull can only see what the hero sees. Even if they have the processing power needed to simulate all of reality, they can’t get the data they need to actually run such a simulation. They can’t pull information out of nowhere.
(Even Tattletale, a much less impressive cape than Contessa, pulls information out of nowhere sometimes. She can guess passwords with her power, a problem that even the sharpest observer simply has too few facts available to solve. Tattletale can make a jump from “He’s left-handed” to “His PIN is 1477”)
Bullshit. Thousands of very smart people spent years planning how to overthrow Saddam Hussein and turn Iraq into a free and prosperous democracy, and they failed catastrophically. Thousands of very smart people have spent the last 4 years trying to figure out how to resolve the Syrian crisis, but in that time the situation there has only gotten worse. Thousands of very smart people have spent the last 6 years trying to figure out how to prevent Greece’s debt crisis from tearing the EU apart, and the problem still hasn’t been solved. Thousands of very smart people have spent the better part of a century trying to solve the Israeli-Palestinian conflict and yet it still rages on with no end in sight.
Even when you look at simpler tasks like NASA missions where thousands of geniuses can meticulously plan ever single step, there have still been a number of catastrophic failures due to things that no one involved had considered.
Intelligence != ability to succeed at anything you attempt. In real life, the single biggest factor of success or failure in large scale risky endeavors has been luck. Genghis Khan was undoubtedly exceptionally intelligent, but do you think he would have been nearly as successful if he hadn’t been born in a place where he could take command of the entire Mongol people at a time when archers on horseback were especially effective on the battlefield? For every Genghis Khan there are thousands of would-be conquerors who didn’t get very far and so are now forgotten, because it’s ultimately down to luck.
Which is not to say that this superpower wouldn’t be useful (though it does sound like it would make life excruciatingly boring), but saying it would make it easy for you to take over the world is silly.
They don’t succeed because they aren’t smart enough to figure out how to do everything exactly right the first time. Really smart people figure out how to prove Euclid’s theorems the first time. Regular people make a lot of mistakes first and are slower to learn. Dumb people can’t do it at all.
The fact that even the smartest people can’t solve impossible coordination problems like Syria is irrelevant. It’s not as if they are even “working together to solve it”. They are undermining one another, all trying to go in different directions.
You are making the absurd argument that because a given amount of intelligence doesn’t let you magically solve every problem, then no amount of intelligence is “that useful”.
Intelligence also absolutely helps you survive even on the African plain. A superhumanly smart person can calculate in less than a second exactly how fast the various predators can run and determine in what direction to flee. Now yes, he may also calculate that no chance of survival is possible. But if it is possible, the smarter you are the more likely you are to figure out how to survive. The idiot runs right into the mouth of the lion without thinking.
Intelligence is the ability to figure out how to solve problems. If a problem can’t be solved, intelligence at least informs you of that fact.
I agree that intelligence is very useful, and that the proposed “10,000 geniuses living inside your head at 1000x speed” superpower would be extremely useful. I am arguing against Daniel Kokotajlo’s statement that with this superpower “taking over the world would be easy,” precisely because taking over the world – let alone maintaining a one world government for any length of time – would require solving a very large number of extremely difficult if not impossible coordination problems.
“Extremely difficult” problems are a matter of perspective.
Walking requires solving some extremely difficult problems of balance and coordination. Far harder than proving Newtonian physics. But we’re optimized to do walking effortlessly and do physics with great difficulty.
If the problems are solvable, a smart enough being will figure out how. All you’re arguing is that ten thousand times speed is not enough. Okay, how about one million? A billion? Is it really that interesting or relevant to the principle of the thing where the exact line on Syria is drawn?
I think if you could devote the equivalent of 10,000 hours to researching every stock you buy, you’d be off to a good start. Become a billionaire. Fund the right rebel groups. Get control of your own little country. And so on.
Turning Iraq into a free and prosperous democracy is a difficult task. But overthrowing Saddam Hussein could be quite simple for a superhuman AI that absolutely doesn’t care about laws, human lives, or what will happen with Iraq after Hussein is overthrown.
Imagine thousands of autonomous drones with human-level intelligence, who hide at random places in Iraq and scan the environment whether they see something that looks like Saddam Hussein. (If the drones are destroyed, they are replaced by new ones. More precisely, new drones are coming regardless of what happens with the old ones.) When they find a “possibly Saddam Hussein”, they will shoot him. Maybe with a silent poisonous shot, so that the target will most likely not notice anything, and will inevitably die the next day.
In a long term, Hussein can survive this only by hiding, and never appearing in public. No problem, just add more important people in his regime to the assassination list. Or maybe just kill anyone who seems like Hussein’s supporter. He will be unable to rule the country this way, and some of his political opponents will use the opportunity to overthrow him, which will make his killing easier. Actually, as long as he is no longer the boss, just announce a huge bounty on his head. Or announce the bounty immediately, to make his regime crumble a bit more quickly.
The budget to execute this plan would probably be cheaper than the real war. The humans didn’t do it because they care about the law (maybe not Iraqi law, but about the laws of their own country), they care about the other details of the outcome (not merely “Hussein gets removed, at literally any cost”), and they cannot cooperate well enough.
“Most of the people thinking about AI risk believe”
You make these statements repeatedly, and I think they don’t quite accomplish what you want them to. You have actual arguments too, but these kind of statements remind me of nothing more but the following arguments
“Many of the people thinking about bible studies believe that there is strong evidence not only in the historicity of Christ, but his divinity as well. Professor Bibleson, who has spent his whole life studying the Bible, and John Godsogreat, a popular public figure who has written extensively and intelligently about the Bible, both agree that the historic record matches perfectly with the description of Christ’s life in the Bible.”
I feel like that is evidence, and that if all we knew about the divinity of Christ was that many people thinking about bible studies believe that there is strong evidence not only in the historicity of Christ, but his divinity as well, we could worry a little bit about not blaspheming. There are plenty of areas where all the cool people do defer to expert opinion, bible studies is an outlier here.
So do I. It is true, however, that we should be clearer on what expertise entails. Different fields of knowledge are more or less reliable, so mastery of them demands more or less respect (an expert Freudian, for example, knows less about humans than an expert botanist knows about plants). It frustrates me that advocates of the catastrophic anthropogenic global warming hypothesis have spent so much time talking about how many experts agree with them and so little time (as far as I have seen, at least) talking about why their expertise matters. That could be true of AI risk research as well.
Do you know anything about the field of bible studies? Very few of the scholars in the field think anything that resembles your caricature, and fewer would claim that the bible represents anything like substantial evidence of the divinity of Jesus.
The exact same critique applies if you change “bible studies” to “computer science” and “divinity of Jesus” to “ai risk”. And I think that’s the point.
I don’t think that’s true. I’ve seen a couple polls showing that most AI researchers think AI risk is worth worrying about. Or, at least, a significant minority; and, IIRC, estimates for the relevant timeline varied, but people thinking “never” for “when will there be superhuman general AI” were few.
I’m not saying no-one in cs has these views, but they aren’t really that prevalent, especially when it comes to more specific risk senarios. Much like bible scholars who believe in Christ divinity.
So I think the original analogy holds pretty well.
@Mark:
And not only that, but I believe most people in the Religion department think that historicity of Christ is “worthy of study”. This is a really different statement from “should be taken as established fact when making decisions about future research”.
It’s even worse than that. When OpenAI launched, their founding statement said “We believe AI should be an extension of individual human wills and, in the spirit of liberty, as broadly and evenly distributed as is possible safely.” They acknowledged the possibility that we’re in the world where secrecy is necessary, and gave themselves an out so they could close up later.
And then two days later, someone silently edited it and removed the word “safely”. The only reason I can think of why they would have done that, is if they felt the people they needed to recruit are so defensive and in denial about the possibility of disaster that mentioning it was an unbearable faux pas. Which is even more terrifying. I wrote a blog post trying to walk the line between arguing against committing too strongly to dangerous openness, and not triggering the densive people.
I think it’s some combination of:
1) Wanting to hew to the cherished ideal of scientific openness.
2) Not wanting to be associated with “cranks”. I’ve detected strong overtones of this sentiment when people who actually work in the field are asked about disaster scenarios. Doomsayers are always, shall we say, reputationally challenged, as has been noted since the Torah.
3) Genuine disdain for the idea that AI could be dangerous, or at least willingness to defer to the sensibilities of those who feel such disdain.
4) Belief that AI research is not controllable anyway, or that OpenAI is a way to exert whatever minimal amount of control is possible.
I have some sympathy for #4, but I think the first three are more realistic as motives. Which means that in another 20-?? years, someone is going to feel really stupid, and I hope to God it’s me.
Or Scott’s idea that this is a desperate move to keep up with the amoral is right while hiding the Dr Good stuff.
Really the “Dr. Amoral” aren’t necessarily amoral, they just for whatever reason don’t take AI fears seriously. So how do you get a platform created that this crowd will use, but NOT the crowd who actually cares about making AI Friendly? Provide a strong signal that you’re in the “AI safety isn’t important” camp. Peter Thiel and Elon Musk know that the problem isn’t a bad guy will control the AI, but that no one will. Thiel’s given to MIRI even. But if you start going “hey we’re really worried that people unconcerned with AI friendliness are going to beat the safe researchers to the punch and therefore we want all those people to share their stuff so the good guys have a chance” the wrong people are going to go on the platform. They’ll be signalling association with the AI safety people and thus the only people on there will be safety people crossing their fingers a non-safety person will be dumb enough to post something. Signal non-safety beliefs however and the plan might work.
It’s still highly risky of course, but frankly that just makes me worried that we’re in a state where high-risk high-reward strategies for human survival are necessary.
I think this is probably the right answer to the central question of Scott’s post. Nice.
This is exactly my opinion. As soon as I saw “OpenAI” and who was behind it I assumed that this was a way of “keeping an eye” on whoever was developing it.
This is good analysis that we must never speak of again.
I actually thought about that before hitting “post,” but I figured the risk reward wasn’t that bad. A) This is sort of a fan theory type thing. I’ve got zero personal connection to the people running OpenAI and no AI research credentials. As such it can be dismissed as a wild guess. B) The people who don’t think the kind of AI safety issues that Bostrom, Yudkowsky, and company talk about can just as plausibly think “well the guys behind this have a lot of history in business and so are experienced enough to recognize the truth of our position.” Maybe they’re even correct.
But on the plus side, this kind of analysis is useful to train for situations where people appear to be acting “out of character.” Deception is always a potential answer to “why is X acting like this?”
The question it makes me ask is, what sort of advances in AI research did they consider “unsafe”? Obviously, “How to make your own self-improving goal-maximizer in 10 easy steps” is unsafe, but any individual advancement is likely to be a lot murkier. Are “deep learning” algorithms unsafe to share? Is an algorithm that allows AIs to understand written language a precursor to an algorithm that lets AIs talk their way out of the box?
“Suppress unsafe AI research” would be a whole lot easier for me to analyze if I knew what “unsafe AI research” looked like.
So, the obstetric dilemma (brains are kept small because they have to fit through the birth canal) is likely one of those stylized facts that tend to linger, but for the record: some think it’s wrong:
http://ecodevoevo.blogspot.de/2015/03/the-obstetric-dilemma-hypothesis.html
http://www.pnas.org/content/109/38/15212
The main alternative explanation offered is a metabolic limit: brains are energetically expensive, so supporting two can no longer be supported at some point. Less elegant, less intuitive, possibly more true.
Now, I won’t extrapolate that to AI because I have almost no subject-matter knowled.. what the hell, it’s the internet.
Even if you make the software for brute-forcing hard encryption _open_, it’s still not a free-for-all for every script kiddie – they don’t have the necessary supercomputers.
Could it be that at least early superhuman AI is still going to be concentrated in the hands of the few who can afford sufficient hardware and electricity? They might be able to afford a killswitch too.
Obviously, it might become more energy-efficient or run on a botnet, so I don’t think this holds completely. But isn’t it unlikely that anything this project comes up with will achieve human-level AI on anybody’s PC? Probably it’s more of a “let everybody toy with image recognition so we have more self-taught programmers to hire” scheme for now?
I did not find the piece critiquing the conjecture about the effects of the birth canal constraint convincing. One very simple test, which I don’t think the author mentioned (but I didn’t make it to the end so perhaps I am mistaken) would be to look at the pelvis size of successful female runners. Do they on average have a narrower pelvis, in the form that constrains the size of the birth canal, than average?
> “I didn’t make it to the end so perhaps I am mistaken”
Looks like you are. Very simple tests often are less decisive than people imagine, in this case a narrow pelvis among runners might proxy for high androgen exposure and hence be linked to bigger muscles which then cause higher performance etc. They do find narrower pelvises sometimes http://europepmc.org/abstract/med/3692651
But they certainly have followed up this line of reasoning through biomechanical analyses (old studies of biomechanics were the evidence used for the OD hypothesis) and find that women are no less economical in energy use when running.
Anyway, there’s other lines of evidence too and things just don’t seem as simple as people outside anthro (like me) used to see them.
> Could it be that at least early superhuman AI is still going to be concentrated in the hands of the few who can afford sufficient hardware and electricity? […] Obviously, it might become more energy-efficient or run on a botnet, so I don’t think this holds completely. But isn’t it unlikely that anything this project comes up with will achieve human-level AI on anybody’s PC?
The “scary scenario” is that if even one instance of the SAI runs, humanity is doomed. So “don’t worry, it’ll only be in the hands of the few who can afford sufficient hardware” is not much consolation.
> They might be able to afford a killswitch too.
If you could reliably build a killswitch, then you’ve solved the control problem. We currently haven’t solved the control problem.
See http://lesswrong.com/lw/ld/the_hidden_complexity_of_wishes/ and http://www.yudkowsky.net/singularity/aibox
> The “scary scenario” is that if even one instance of the SAI runs, humanity is doomed. So “don’t worry, it’ll only be in the hands of the few who can afford sufficient hardware” is not much consolation.
Maybe not “much”, but then openAI is not a game-changer (as many others have pointed out here). You already have multiple companies and governments trying to create better AI.
> “killswitch”
Yeah, sure, maybe. I’m more optimistic about the possibilities of shutting off an energy-hungry data center than shutting down a distributed botnet on millions of private PCs. Anyway, same old discussion starts over.
As a academic AI researcher (and someone personally acquainted with many of the people involved in OpenAI), it seems to me that this post conveys a number of misconceptions both about what will go on at OpenAI and about the state of AI research.
OpenAI is a research lab, just like the ones at Google or Facebook or other companies or a plethora of universities. Already, much of the work done (even by private companies, but especially by universities) is open source and freely available. In fact, the AI researchers of today tend to have been raised on open source software and to have spent time in academia, which creates a general aversion to doing work in secret at a private company which might severely hamper its impact.
Which is to say, even though Sam and Elon talk about ensuring an optimal future for AI, OpenAI does not look to me like an organization chiefly dedicated to AI safety, but rather a new lab intending to do (excellent, but nevertheless) run-of-the-mill research. And there are many reasons aside from safety to form such a lab. If the best researchers can be convinced to give away their state-of-the-art systems (which many of them want to do!), that undercuts the current intense competition in AI recruitment, leveling the playing field for companies that depend on AI, and allowing them to compete on other grounds. (Compare, for example, Apple, who are taking the opposite strategy of keeping everything secret, and reportedly struggling in recruiting these very same researchers as a result.)
There are also much more mundane safety reasons why state-of-the-art systems should be open source; you wouldn’t want an expensive self-driving car to notice more obstacles than a cheap one, for example.
So despite the framing and discussion I think OpenAI actually has very little to do with AI safety, at least of the dangerous hard-takeoff sort, and in any case is not a big departure from the status quo. These are just folks already doing regular, everyday AI research, like recognizing cats and translating language and playing Atari, who’ll continue to do the same kind of research, which is currently nowhere near a point of concern, or indeed, even a point of if-this-were-to-suddenly-undergo-exponential-improvement-it-would-be-concerning.
> OpenAI actually has very little to do with AI safety
Perhaps now, but I can imagine that this will lay down the organizational groundwork for a human coalition for AI safety that’s relevant to hard takeoff/the control problem in the future.
AI researchers today, as I am sure you are aware, don’t study the control problem. They study things that are on the cutting edge between “we can do it” and “we cannot do it”, such as applying some machine learning technique to a novel or more complex task.
But if the field progresses for another 20 years and starts getting nervous about AI control, I’m sure they will turn to openAI, amongst others, for guidance.
Anyway, thanks for your answer.
Wanted to add my support to this post.
I think it’s curious essentially nobody in the “MIRI-sphere” actually does stats/ML work.
Arguments elsewhere aside, I’m personally not very worried by OpenAI because it looks to just be a non-profit version of an AI research lab focused on near-term advances, i.e. exactly what quiot mentions above. I do think Scott is right to be worried that a bunch of people who have expressed interest in AI safety are now pushing openness instead of safety, because of the implications.
It also seems to me like the argument for OpenAI rests on applause lights: we want AI to be open, developed by non-profits, and evenly spread. But openness means that you’re empowering Dr. Evil, developed by non-profits only gets rid of one management failure case, not all of them, and evenly spread means that we get more profit-destroying competition and less standardization and economic power in the hands of tech magnates (who are probably better suited to using that power for good than other groups).
Thanks for this comment, I definitely read it the same way. I think they are more concerned with the economic/social impact of AI that can replace low-skill jobs, than about superintelligence.
It’s not just them. The introduction post signed by Brockman and Sutskever sounds the same way. It invokes the promise and danger of human level AI. In the short term it will do what you say, but it is also asserting that openness is the key to safety, which is what Scott is responding to, regardless of the timeframe.
> it seems to me that this post conveys a number of misconceptions both about what will go on at OpenAI and about the state of AI research. […] OpenAI does not look to me like an organization chiefly dedicated to AI safety, but rather a new lab intending to do (excellent, but nevertheless) run-of-the-mill research.
In the interview, Altman and Musk (presumably speaking on behalf of OpenAI) point blank say that they expect the OpenAI to developer superhuman AI, and that they have an official stance on safety (making it open source and available to everyone makes it safe). Altman explicitly says that a big difference between how Google opensources its AI vs how OpenAI will opensources its AI is in regards to its corporate values regarding superhuman AI.
I don’t think anyone here doubts that the only thing OpenAI will have to show for itself in the next one, five, ten or twenty years (depending on how optimistic/pessimistic one is regarding AI research) is “run of the mill” stuff like automated translation, self driving cars, image labelling, etc.
But they have explicitly said that superhuman AI is one of the things they, as an organization are thinking about, whereas Facebook, Google, etc. have not made any such statements.
Hence why we’re having this discussion singling out OpenAI, but not mentioning Google or Facebook.
I think your concern about the “control problem” may be overstated, or at least misstated. Yes, computers up to this point have tended to be fragile and severely oriented toward specific goals, but computers up to this point have tended to be below human intelligence as well. We’re not going to make the sort of AI you’re worried about by making Google Maps But More So, so these flaws specific to systems like Google Maps may not be something to worry about. One of the things that makes the human brain robust and stable may be its generalizing capacity, something which any AGI would have to have.
The part that’s not overstated is concerns about goal misalignment – it doesn’t matter how robust your AI’s thinking is if it’s robustly indifferent to humanity. But those seem like errors of a different sort than Google Maps sending me through the ocean.
> The part that’s not overstated is concerns about goal misalignment
Yes, I agree. I think Scott really meant goal misalignment when he said “control”.
I think he’s equivocating between the two, I’m not sure he even has a solid grasp on the difference.
@scott alexander:
Can you settle this one?
The human brain is robust and stable now?
Compared to Google Maps. Compared to Windows Millennium Edition. Compared to Skyrim. Compared to an Xbox.
Yes. That is big difference between humans and attempts at AI. Humans can adapt – if the situation they are in changes massively then they will change themselves to deal with it. Computer programs do not.
What is the difference between the control problem and goal alignment?
If you control something, you decide what it does, you can shut it down if you want to. You will likely use it for your desired goals, but that’s largely irrelevant for control. You could, instead, have a construct that does not obey you, which you can’t shut down, but acts in your interest.
Think the difference between an ally and a vassal.
I get the impression that Scott does understand this, but was trying to communicate the danger to people that find the goal alignment problem too abstract.
More cynically you could say that Scott has been tempted into the sins of popular writing.
It seems like there’s a fairly simple best-of-both-worlds strategy for OpenAI: proceed as to prevent worst case slow- or medium-takeoff scenarios, but stop open sourcing your research if a fast-takeoff scenario starts to appear more likely.
No one really knows whether AI will bootstrap itself quickly or slowly, though some educated people may have very strong opinions, but that lack of understanding will likely start to abate as AI advances. Learning more about AI through research is the best way to answer that question (imo). In the meantime, there are a lot of very bad consequences on the horizon from letting asymmetries in weak AI or slow-takeoff scenarios develop. I’m thinking primarily of military uses of AI, robotic soldiers, etc. Drone strikes are just the tip of the iceberg of where that is headed. I don’t remember the scenario of strong AI being developed (and used irresponsibly) by the US government being responded to specifically, but that seems the most likely bad outcome, and as dangerous as any other, minus the paperclipping extinction event. They already have more resources to put towards it than anyone else, are beyond regulation by any other entity, and seem to be aiming primarily for military purposes. The gap between US military AI and academic AI could already be similar to that between NSA and academic cryptography, for all we know.
So it seems like the most rational thing for OpenAI to do is exactly what they are doing – encourage current AI research in the most healthy way possible, but pursue another strategy if fast-takeoff starts to appear the likely outcome. This also means the rest of the world would get to *know* that a fast-takeoff scenario is for-sure likely, without it simply being initiated in secret in some military lab somewhere.
I mostly agree with that. However, I am rather pessimistic about our chances to counter a fast-takeoff. If we reach this point (implying it is reachable for us) I doubt – given human nature – the necessary breaks will be pulled. There probably will be those with power and a big enough group of scientists who do not care about the risk.
I agree with Nathan above. Intelligence != power, and this is the strangest part of the whole debate.
For example, environmentalists often say that dolphins are more intelligent than us. They do this to elicit a sense of respect or kinship between ourselves and them. But if we were to take the AI-alarmist position seriously, this would be an argument for us to eliminate dolphins, or at least keep them all in very strict captivity. Suppose that dolphins are more intelligent than humans. Are you worried about Unfriendly Dolphin Risk (UDR)? Why not?
Further suppose that it turns out that marine parks are (perhaps unintentionally) selectively breeding dolphins for intelligence. All Scott’s arguments about how intelligence isn’t necessarily maxed out come into play. Now are you worried about UDR? And remember, if you’re worried about a UAI “tricking” humans and taking control, you should be equally worried by a dolphin doing it.
Which would be a bigger threat:
* Dolphins double in intelligence.
* Dolphins gain opposable digits.
Or, to put it more crudely: I feel much more threatened by a mugger with a knife than by Steven Hawking.
To me, the risk is not that an unfriendly AI becomes much more “intelligent,” but that an unfriendly AI becomes much more powerful. But it can only be as powerful as what it is hooked up to. As such, the analogy to atomic weapons fails, and open-source, widely distributed AI looks like an excellent way to mitigate the risk.
Actually, the open source argument can still apply here. If the AI is only as powerful as what it’s hooked up to, keeping it in a limited number of hands is probably the best way from keeping it from being hooked up to anything too powerful. Giving ISIS a really smart AI seems like a risky thing to do.
Giving ISIS a really smart AI is only a risk if it’s friendly.
Unfriendly AI is less of a risk in the hands of ISIS than the US military.
Why? What, exactly, could the AI do to help them? ISIS’s main problem is that their ideology and actions are so ludicrously, cartoonishly evil that they have ended up at war with 99.999% of the world. The only reasons that they have survived this long are 1) Most of ISIS’s enemies have higher priorities than defeating them (Saudi Arabia and Turkey want Assad gone, Iran wants Assad to stay in power, the US doesn’t want to get caught in another protracted ground war, etc.), and 2) Modern communications technology allows them to connect with and recruit from the 0.001% of humanity that doesn’t hate them.
I have a hard time seeing how any amount of intelligence could significantly change the situation. Better battle strategies and organizational practices might help ISIS survive longer in the short run, but in the long run they would still be massively outnumbered, massively outgunned, and far beyond the point where they could conceivably negotiate a way out.
You’re not wrong, but I was thinking risky in terms of “they could do (more) damage and hurt (more) people” rather than “they could win”.
Probably more than 0.001% of the world is in ISIS.
This.
ISIS runs under the banner of a valid, consistent interpretation of Islam – arguably one most true, today, to the version of its founder. A substantial minority of Muslims – at least – are potential recruits, in addition to an assortment of opportunist adventurers, psychopaths, and those disgusted with the western worldview who want an alternative, any alternative.
Indeed, it may be as many as 0.01%. (The population of the world is 7 billion, ISIS has between 30,000 and 300,000 members).
Really, you can’t think of a single technology that a group in control of a rogue state with access to a mind 10x smarter than Einstein could deploy to bring about worldwide destruction?
It’s only risky if Mahdi 17.12.2025 (1.43) is the only “really smart” AI on the board. ISIS’s chances of success decline dramatically if their targets are all capable of spamming “Chinese Gordons” while they set up a counter.
I appreciate your historical analogy.
Thank you 😉
What if the conflict between the two AIs is a bloodbath for humans caught inbetween?
The real question is “Is it more or less of a bloodbath than than the alternative”.
WWII was a blood bath but still arguably better than surrendering to the Nazis or Imperial Japan.
You think Stephen Hawking couldn’t harm you if he really wanted to? The mugger feels more dangerous because it’s plausible that he intends to do harm. If Hawking really wanted to harm me I’d be a lot mpre scared of Hawking than a mugger intending the same harm.
I didn’t say that it’s impossible for Stephen Hawking to cause me any harm whatsoever. For example, he could try to leverage his fame to blacken my reputation.
But no, I stand by what I said; for equal malevolence towards me, I am far more scared of the mugger than I am of Hawking. It’s not even close.
The obvious way for Stephen Hawking to harm you is to hire a hitman.
But that requires Stephen Hawking to:
(1) Get in touch with a reasonably reliable hit man – something that’s kinda hard when you don’t already have criminal friends, which I highly doubt a renowned scientist like Stephen Hawking does. I’m sure there are hitmen, or people claiming to be hitmen, on the shadier parts of the internet, but then you wind up with serious trust issues, see (3).
(2) Arrange payment and communicate sufficient detail to the hitman so that the hitman, assuming honesty, will actually harm the correct person.
(3) Trust that the hitman will actually perform as requested instead of just taking the money and disappearing, or harming the wrong person and refusing to go back to harm the correct person without additional payment.
The whole process necessarily involves some time delay, as well, during which the police may find out and put a stop to it, and if he tried to do it right in front of you it’d be trivially easy to stop him.
A mugger with a knife, on the other hand, is an immediate threat the moment you meet him.
Every few months I hear a news story about somebody who tries to hire a hitman online, who turns out to actually be an FBI agent and arrests them on murder charges. But I have no idea how common this is compared to successfully hiring a hitman or giving money to some random scam artist claiming to be one.
See Eliezer Yudkowsky’s “The Power of Intelligence” and Scott Alexander’s post on cult leaders.
Do they really say that? I searched for the phrase and got a bunch of clickbait and blog posts.
Anyway, that’s obviously wrong, so who cares?
Anyway, that’s obviously wrong, so who cares?
Agreed. It’s the octopi we should be scared of, since unlike dolphins not alone do they have the smarts, they have physical manipulatory appendages and dexterity in using them which lets them interact with the environment and alter it for their benefit 🙂
And they’ve already taken over the Japanese animated porn industry for their own benefit.
I’m not worried about Unfriendly Dolphin Risk because I trust that the environmentalists are just bullshitting as usual.
This is called “fighting the hypothetical.”
It’s a shitty hypothetical. You’re basically making fun of the idea of being scared of dolphins, but the reason that’s funny and not terrifying is that dolphins are in fact not nearly as smart as us and the environmentalists are full of shit. The hypothetical does not perform any work other than smuggling in associations contrary to the hypothetical, so I feel totally justified in fighting it.
Damn straight.
If dolphins were ten times as smart as humans or something, the world would look a lot different.
The fallacy here is positing a hypothetical while stipulating that the rest of the world remains the same.
Michael Huemer does this in his “argument” against egoism (or really anything other than his moral intuitionism). He says: suppose it was in your selfish interest (or suppose: good of society) to vaporize an innocent homeless man. Should you do it? The answer is: yes I should, but that world would have to be totally different from this world, where it isn’t in my interest.
There are less dramatic and more realistic forms of PP, for instance stealing pennies forma blind beggar. Human societies contain structures designed to prevent PP, so there are real world opportunities for PP, as well as intuitions that it is undesirable.
@ TheAncientGeek:
I assume by “PP” you mean “prudent predation”.
By definition, prudent predation would be harmful to society and—if it were really prudent—beneficial to the individual. The fact that it is harmful society is not a reason to refrain from engaging in prudent predation for any individual.
However, it may be prudent for individuals to construct institutions to make sure they don’t get predated by others. If these institutions cannot succeed in making it imprudent to predate, then it will just be rational to predate and that will be the end of it.
Of course, the other problem is that people don’t always and, to be perfectly frank for the overwhelming majority of cases, always don’t do the most prudent thing. So we also have to be concerned with teaching people to be prudent.
A society may well set up rules to restrain excessively selfish behaviour, and those rules are what we typically call morality.
The argument isn’t about what is or isn’t rational, it is about what is or isn’t ethical. Ethical egoism seems to label selfish actions as moral, although by most people’s intuitions the kind of selfish actions that moral rules forbid are immoral: tha’ts what the PP objection seeks to bring out.
On the other hand, if you stop labelling selfish actions as moral, inasmuch as they are selfish, then where is the egoism in ethical egoism? You just have a rational justificaiton of morality, with no special role for the values of an individual.
@ TheAncientGeek:
The assumption of egoism is that there is no such thing as “excessively selfish behavior”. There is narrowly self-serving behavior, but this is not rejected because it is aimed at serving the self but because it does not do so rationally.
Whether something would be good for society is irrelevant. Of course, egoists are interested in stopping other predators because they are harmed directly or indirectly by them. But if they themselves can prudently predate, they should do it.
Yes, that is the point of hypotheticals that ask people whether an action would be moral even if it were in one’s self-interest.
The usual argument for egoism is that we do not properly come to moral beliefs this way. We do not take specific examples of moral judgments and generalize to the principle unifying them. These judgments will just reflect what principles people have been taught consciously or unconsciously by society. Which will be traceable to some philosophy that told them to value those things.
The usual type of argument goes the other way around: that we rule out all the other candidates and come to the conclusion that self-interest is the only rational ultimate value. Then, knowing that, we deduce what is ethical based on that principle.
Despite this, sure, there is a case to be made for the type of hypotheticals where you keep the variables open and say: “Suppose it were in your interest to kill a million people. Should you do it?” Such a hypothetical leaves open the possibility to imagine factual scenarios (such as war) where it could be validly in your interest to kill a million people. And maybe this will show it is absolutely against a “moral intuition” to do so.
The bad type of hypothetical is like this: “Suppose nothing at all were different in the universe except that it were in your interest to murder your wife. Should you do it?” But if it isn’t in your interest currently to kill your wife—and if this judgment is based on facts—you can’t ask someone to change the judgment while holding the facts constant. That would be logically impossible: a contradiciton.
It is exactly the same as this “hypothetical”: “Suppose x + 2 = 4. Now hold that equation constant while x = 7.” You obviously can’t do that. In the equation, x has to be 2. If you want it not to be 2, you have to allow something else to change.
“The assumption of egoism is that there is no such thing as “excessively selfish behavior”. There is narrowly self-serving behavior, but this is not rejected because it is aimed at serving the self but because it does not do so rationally.”
So are the ethical rules of society useful in assisting people to achieve a broader self-centered rationality, or are they the imposition of immoral altruism?
“Whether something would be good for society is irrelevant. ”
If the correct theory of ethics is egoism. By most people’s intuitions, it is, on the other hand, extremely relevant.
“Of course, egoists are interested in stopping other predators because they are harmed directly or indirectly by them.”
So would egoists be interested in setting up rules whereby they promise to stop predating, and in return others stop predating on them? That seems rational…but if a group of egoists, on a desert island, say, did that, they might end up re-inventing something like conventional morality.
A hypothetical I know, but it illustrates ethical constructivism, and constructivism is a rationality-based theory of ethics, too.
The questions are whether constructivism actually is egoism, and if not, whtehre it is better or worse.
” But if they themselves can prudently predate, they should do it.”
That’s a rather bland statement on a crucial point. The usual conterargument is “EE imp-lies PP, PP is wrong, so EE is worng”.
“Yes, that is the point of hypotheticals that ask people whether an action would be moral even if it were in one’s self-interest.”
Lot’s of action as are moral an din one’s self interest. The point is whether an action can be moral and against someone else’ interest.
“The usual argument for egoism is that we do not properly come to moral beliefs this way. We do not take specific examples of moral judgments and generalize to the principle unifying them. These judgments will just reflect what principles people have been taught consciously or unconsciously by society. Which will be traceable to some philosophy that told them to value those things.”
If every other system ethics other than egoism had an arbitrary starting point,then egoism might be the rational choice. But it has rivals, such as constructivism.
“The usual type of argument goes the other way around: that we rule out all the other candidates and come to the conclusion that self-interest is the only rational ultimate value. Then, knowing that, we deduce what is ethical based on that principle.”
You need to show that egoism is ethical as well as rational. constructivism accords more easily with common intuitions about ethics.
“Why don’t people accept that I’m right when I propose an imaginary scenario where all the facts conveniently support my point of view?”
Well, you should agree that a person would be right in the scenario where all the facts support his point of view. That is the purpose of hypotheticals.
If not, something is very wrong and your disagreement is not about the facts, but perhaps instead one of you is making a logical error.
“Suppose that dolphins are more intelligent than humans. Are you worried about Unfriendly Dolphin Risk (UDR)? Why not?”
In this scenario, I would be *extremely* worried about Unfriendly Dolphin Risk. See Eliezer’s post That Alien Message (http://lesswrong.com/lw/qk/that_alien_message/) for why. All it would take would be one reliable, unsecured communication channel between humans and dolphins, and we’d all be royally screwed.
Caveat: the dolphins would have to be a *lot* smarter than humans, not a bit smarter; we humans have huge advantages in terms of quick access to facts, technology etc that make this into an unfair fight.
Given research already done on dolphins (such as trying to teach them sign language), we can rule out this level of Dolphin Intelligence. They’re probably about level with an IQ 60 person from what I can tell, but with less capacity for comprehending symbols and more for social situations – which makes sense since dolphins have complex social structures but lack the means to produce writing.
> Intelligence != power, and this is the strangest part of the whole debate.
You’re thinking of intelligence on a parochial, human scale.
It is unfortunate that we use the same word for tiny differences between the intelligence of humans as we do for discussion of something that is to us as we are to house pets.
Vampire bats are evolving thumbs.
They’re mammals, didn’t they start out with thumbs?
Great, now I have to go look up bat anatomy now 🙂
Wait. Do people actually seriously think dolphins are smarter than humans? They get killed in fishing nets all the time accidentally! If aliens were catching cows and sometimes people got caught in the nets we’d start creating strategies to not get caught. Dolphins are intelligent compared to most of the animal world yes, but seriously thinking they are human level or above seems extremely unlikely.
Beyond that, how high of certainty would you say you have for intelligence not basically being an evolutionary super weapon? Because if you’re wrong and Scott’s right the stakes are rather high. If you’d say “super high intelligence can’t lead to beating humans easily” with a 90% degree of certainty I’d reply “well a 10% chance of humans getting annihilated seems pretty bad doesn’t it? We can’t take very many chances at that level and hope to survive as a species.” If it’s absurdly high like 99.9999% certainty then I would expect to here a far more comprehensive argument given we know human intelligence has led to us colonizing the entire Earth and causing a mass extinction of other organisms.
>Do people actually seriously think dolphins are smarter than humans? They get killed in fishing nets all the time accidentally!
http://www.darwinawards.com/
Like I said in the thread following Nathan above, intelligence in fact does equal power. Dolphins almost certainly aren’t smarter than humans, but if they were, they still wouldn’t be a threat unless they were significantly smarter than humans. (It’s the same with AI–the worry is that there will be a “hard takeoff,” i.e. that the gap between “a little smarter” and “a lot smarter” will be crossed too quickly for the proper safety checks to happen)
If dolphins *were* significantly smarter than humans, yes they would be a threat. Like I said above:
I think that smart = powerful, and here’s why:
Imagine that you have ten thousand geniues living in a simulated world inside your head, organized in a super-efficient and robust institutional structure, all very highly motivated to help you achieve your goals… and all of them experiencing a thousand years of subjective time for every second of time you experience. They also have access to the code of their simulated world, so they can feed themselves, make more copies, upgrade themselves, etc. They can see through your eyes (and other senses etc.) and they can move your body if you allow it.
Just think about it for a moment–think about what you could do with that superpower. Taking over the world would be easy.
That’s what it would be like to be a *speed* superintelligence, one that isn’t *qualitatively* any smarter than the average human genius. If you were *also* qualitatively smarter–in the same way that a human is qualitatively smarter than an ape–you would be able to think of strategies that we currently can’t conceive.
No, even super-duper-genius dolphins wouldn’t be much of a threat because 1) dolphins don’t have hands, and 2) dolphins don’t have an easy way of obtaining knowledge about the human world.
The second point is especially important. You shouldn’t conflate intelligence with knowledge. Humans that grow up without assistance from other humans, commonly known as feral children, end up as little more than animals, because raw cognitive ability doesn’t amount to much by itself.
Just imagine that you are a dolphin with the superpowers I described. To simulate your lack of knowledge about humans, imagine that you are on an alien planet. I bet that in half an hour you can think of a way to take over.
(I suppose it would depend on the aliens, but the first step would be to observe them and glean as much information from them as possible without letting them know how smart you are.)
This sounds like a fun idea for an RPG actually…
The only information that I, as a wild dolphin, would be able to acquire about the aliens would be “Gee, there are these big, hard, cold things that move about the surface of the water. Sometimes they drag big nets that scoop up lots of fish and I have to be careful not to get caught in them. If I leap out of the water while they pass I can see strange creatures moving arout on top of them.” Learning anything else would require finding a way to communicate with the aliens, which would carry all sorts of risks.
I think a sufficiently intelligent dolphin, given the opportunity to observe humans from the coast, could figure out a means of communication.
First, you would act completely contrary to normal dolphin behavior. That would get their attention.
Then you could blow water through your blowhole using a standard human code, such as Morse code’s SOS. You could then communicate longer messages by Morse code in a similar way
It would be risky, but it might be worth it.
To be a threat, a dolphin would need to be able to survive on dry land. Since they can’t do that biologically, they’d need to technologically enhance their own bodies, say, building the reverse of a submarine, a container filled 90% with water from which they could conduct warfare. But the prospects for engineering complex machinery underwater are quite dreary, regardless of the intellect of the engineer, simply because water is a lot more corrosive than air, and they also can’t use fire.
> Remember, it took all of human history from Mesopotamia to 19th-century Britain to invent a vehicle that could go as fast as a human.
You are slightly wrong here–you have forgotten the horse-drawn battle chariot, which was invented a few thousand years ago.
Also, sailing ships can sometimes go fairly fast (compared to humans.) That’s one of the reasons that the first time European travelers/explorers reached the southern tip of Africa, they got there by sailing, not by muscle power over land.
I’m not sure what kind of graph we would get for development of vehicle speed over history, but it would probably have some interesting complexities and wiggles.
Horses (and horse-drawn vehicles) can only really go faster than humans over short distances. See eg. the Man vs. Horse Marathon, which most often goes to the horse but which humans have won twice (https://en.wikipedia.org/wiki/Man_versus_Horse_Marathon). Early 19th-century packet ships (optimized for speed over cargo capacity) traveled at about 5 mph on average, which of course varied somewhat depending on the wind (https://en.wikipedia.org/wiki/Blue_Riband). Of course, ships have the advantage of sailing 24/7, unlike human runners who need food and sleep.
Viking dragon ships had an oar speed of 8-12 knots, so it’s hardly optimized for speed. Though those require a mix of muscle and wind power.
But that 5 knot sailing vessel might beat the dragon ship, the horse, and over a long enough distance event he best ultra-marathoner, because it doesn’t have to stop.
The Viking, a reproduction viking longship, crossed the North Atlantic in 44 days, and the back of my napkin says the average speed was ~2.75 knots. (But the route is too far North that I don’t know the prevailing wind direction off hand.) For comparison, IIRC, Columbus was 33 days between the Canaries and his first landfall in the New World and using the same method (ie, asking google the distance) I come up with a smidge less than 4 knots. And Columbus ran the whole way with the trade winds at his back, which certainly helps.
By way of further comparison, The Atlantic held the record for the fastest monohull sailboat Atlantic crossing for virtually all of the 20th Century with an average speed of just 10 knots, crossing with the trade wind.
By way of further comparison, looking at the Blue Riband times (always measured while traveling against the trade winds) the latter half of the 19th century saw speeds more than double with the advent of powered ships increasing record average speeds from something like eight (for wooden paddle boats, which was about a knot faster than the fastest crossing of a packet ships at the time, I think) to more than twenty knots (for double screwed steamers).
By way of further further comparison, it looks like the Roman Army was required to be able to march approx. 20 nautical miles in 5 “summer hours.” If you assume that was the expected distance the Army would march in any given day, and that the other 7 daylight hours were devoted to other activities while on the move (setting and breaking camp, for example) we can estimate that the Legions would advance over land at an average of 20/24 (~0.8) knots. And (looking at the daylight hours in Italy in the summer, and if I did all the conversions right) the actual rate of march is about 3.2 knots. The rate of march could certainly be raised temporarily, and/or they could march longer on a given day, but this gives us some pretty solid bounds on how fast the Legion could march over the long term.
I would actually be interested to see a cite on that. Not that you’re necessarily wrong, but that sounds suspiciously like a misinterpretation of a modern rule of thumb. The “hull speed” in knots is given by 1.34*sqrt(waterline length), which corresponds with the speed at which the wavelength of the bow wake is equal to the waterline length of the ship. Which is a useful reference point if you are designing a boat but doesn’t necessarily correspond to anything practical; even still, it is often misinterpreted as the something like the maximum working speed of a boat. I think 8-12 knots is almost exactly the range of hull speeds from archeologically examples, so I am curious whether it comes from empirical evidence (like the Viking I already mentioned) or is just an estimate based on the hull speed (and hence, primarily on the length of the ship irrespective of the power actually applied).
I don’t think horse-drawn vehicles count because horses were created by evolution. Humans just tweaked them a little with selective breeding. I think Scott only intended to count vehicles where all the complex parts were made with human ingenuity. Your point about sailing vessels still stands though.
Harnessing a horse isn’t fundamentally different from harnessing an internal combustion engine. Both are just bending some part of the external world to our will. Whether the engine runs on grass, ethanol, or a microverse of people stomping gooble boxes is an implementation detail.
The relevant difference is that an internal combustion engine is made from “scratch.” That is, the entire device, with all its moving parts, can and are made out of basic elements. A horse is harnessing a complex machine that natural selection already made.
I know that this isn’t a clean category. What counts as a “complex moving part” is a fuzzy question. But it’s still a relevant distinction. The complex moving parts of an engine were designed and made by people, the complex moving parts of a horse by natural selection. A human engineer today could build an engine from scratch. Horses can’t be built from scratch, the only way to make a new horse is to get two preexisting horses to bone.
You’re missing the point of the comparison: the horse is faster than the human, the chariot is just along for the ride, and people didn’t invent horses.
(I do agree that ‘land-based’ is a necessary qualifier.)
I’m not counting horse-drawn vehicles, but excellent name-topic matching going on here.
One thing I don’t like about the above is that you lump a vast array of things under “intelligence” as if “cow level” or ape level doesn’t already imply a massively capable AI with an insanely broad number of capabilities, navigation, pattern recognition, visual processing, edge detection.
Even if your AI is fantastically good at proving theorems or natural language processing to a far-superhuman level that doesn’t mean it’s even capable of internally representing the concept of flowing water or comprehending the idea of thinking entities and all that it entails.
An idiot savant is intelligent. In very very specialized ways. It’s entirely possible, likely even that for a long long time even the “superhuman” AI’s will be ultra-idiot savants even assuming that intelligence is easy and no this is not a comic-book superhero argument where you liberally interpert your favorite superheros powers to claim that they can automatically do everything else that all the other heros powers can do.
But on to the other problem.
You haven’t noticed that almost all AI-research is already freely-shared?
15 years ago the best machine translation you could find on the web was bablefish. Which was shit. The cutting edge at the time wasn’t much better.
5 years ago one of my classmates was able to program, from scratch but based on research papers, a translation program far far better than bablefish which scored only a few points lower than google-translate at the time. (there’s actually a standard scale for assessing machine translation)
Right now, Dr Evil doesn’t need to download the source code for any AI, he can subscribe to all the AI research journals and with a half dozen comp sci BSc grads re-create much of the current cutting edge in AI.
Do you only get worried when something hits the newspapers?
I’m also willing to bet that goal-based AI’s aren’t going to be that popular, when they’re buggy they can do things like go off and spend all your money on hats without you having any chance to intervene before the order has been based.
On the other hand oracle-AI’s like expert systems which provide advice but have no goals of any kind nor any mechanism for trying to change the world could simply explain to you what you could do while explaining their reasoning. This model is already more popular in many fields like medicine.
We live in a world where a 16 year old with little more than his salary from McDonalds was able to build a small working breeder reactor yet that isn’t a common occurrence even after people heard about it.
“We live in a world where a 16 year old with little more than his salary from McDonalds was able to build a small working breeder reactor yet that isn’t a common occurrence.”
Fortunately, this is incorrect (a product of sensationalist media). To get a “reactor”, ie. any significant amount of energy, you need a chain reaction with each fission neutron emitted producing an additional neutron. This is very difficult (even state actors routinely fail), and David Hahn never remotely came close, in the same way that building a toy car with Legos and two AA batteries isn’t remotely close to building a Toyota minivan.
What Hahn did was create a handful of plutonium atoms, by bombarding thorium and uranium with beryllium-generated neutrons. However, this isn’t dangerous (except possibly to you), because without a chain reaction, the number of neutrons you can produce is about ten orders of magnitude lower than the amount you’d need to make macroscopic amounts of plutonium. This reaction is simple enough that it actually happens naturally, in the mineral muromontite (http://theodoregray.com/periodictable/Elements/094/index.html), which contains a mix of uranium and beryllium. The uranium atoms release alpha particles when they decay; some of these alphas hit beryllium atoms, which emit neutrons; some of the neutrons then hit another uranium atom, forming plutonium. (But only a few atoms at a time.)
From the book, while he didn’t breed massive quantities his thorium/uranium dust blocks were getting pretty severely radioactive pretty fast.
It wasn’t a real breeder reactor that would produce more fuel than it uses but it could have got pretty nasty.
That is due to the fact that it only takes a very, very little bit of highly radioactive materials to be dangerous to individual organisms in the vicinity. “Massive quantities” is a very subjective thing when dealing with short-lived radioisotopes.
I get worried when the people most concerned about AI safety decide that AI safety efforts ought to promote AI safety by making superintelligent AIs open source, and take steps to do so. That seems like a different scale of problem than sharing insights about machine translation, although your point is well-taken.
I think you’re making the mistake of assuming that closed = inaccessible, and are drawing the comparison of AI to nukes just a little too closely.
Even in a hard take-off scenario Strong AI is not going to be a “press this button to obliterate the planet” sort of deal. Realistically, the paperclip maximizer will still need a bit of time to build up a proper head of steam. (Secure opposable thumbs, develop resources and tech that can’t be disabled by the anti-paperclip forces)
As such I think the real threat is that someone develops a Strong AI in secret and that nobody else notices till it’s too late.
To that end, making AI development “open” makes a lot of sense. Having AI principals well understood by a lot of people increases the chances that someone will recognize the profile of, or develop a defense against an emergent UAI before it get’s out of hand.
As for Dr. Evil creating a UAI, I’ll paraphrase my example from up thread. ISIS unleashing Mahdi v1.43 is a lot less scary in a world where anyone they might try to attack knows how to make a Chinese Gordon.
“Even in a hard take-off scenario Strong AI is not going to be a “press this button to obliterate the planet” sort of deal. Realistically, the paperclip maximizer will still need a bit of time to build up a proper head of steam. (Secure opposable thumbs, develop resources and tech that can’t be disabled by the anti-paperclip forces)”
That part is easy.
By hypothesis, our hard-takeoff Strong AI has been built by Dr. Amoral–that is, someone who doesn’t take AI risk seriously. It will be very easy for the AI to convince Dr. Amoral to do things, build things, hook the AI up to the Internet, etc. Heck, Dr. Amoral may have been planning to do those things already–after all, AI isn’t dangerous, right?
Even then there are quite a few steps and a fair bit of time between the “Clippy has an Amazon Prime Account” stage and the “Earth has been replaced by 6 × 10^21 tonnes of paperclips” stage.
Which brings us to the second paragraph of my comment…
How is it going to do that when it doesn’t understand how Dr. Amoral thinks or what Dr. Amoral values?
Remember Clippy is basically a child-god that basically just doesn’t understand what you really meant when you told it to make more paperclips more efficiently. How is it it going to successfully manipulate Dr. Amoral?
@ HeelBearCub:
No, you completely misunderstand.
“Clippy” is not a “child-god”. It completely understands that you don’t want it to turn the universe into paperclips. It just doesn’t give a shit.
The problem is that presumably you won’t be able to build a superintelligent AI from scratch. You will iterate it from less intelligent to more intelligent levels. By the time it gets intelligent to “do what you mean, not what you say”, the thing will be so complicated and incomprehensible that its values are a “black box” that can’t be directly observed but only inferred through action.
It will understand that you want it to, say, produce a reasonable amount of paperclips and never use force. So it will convince you that it wants that, too. But actually, the real goal in the “black box” is “convince humans I want to help them until I gain enough power to kill them and then maximize paperclips”.
@Vox Imperatoris:
A) The basic problem is still that it did not understand the meaning of the basic goal.
B) It cares enough about the goal you set (way back when) that it is doing everything it possibly can to do only that thing, which it still does not understand.
It’s a child-god/stupid-god story told to warn us of the folly of man. The 50s is littered with these stories, IIRC.
Put “find Rosebud” in for “maximize paperclips” and you get a super-AI who puts itself in a moon base while it scrapes off successive layers from the earth until it exposes the core because it can’t find the right sled.
@HeelBearCub:
“A) The basic problem is still that it did not understand the meaning of the basic goal.”
Nope. Recall the example Scott gives of aliens showing up on Earth with compelling evidence that they designed us. Suppose they say that actually we’ve misinterpreted our purpose; we are supposed to enslave and cannibalize each other. Would you be like “Oh, okay, I guess I’ll do that then.”
That’s a fanciful example, but there is one closer to home: Evolution. We now know that the system which created us–natural selection–created us with one purpose only–reproduction. We are NOT fulfilling our purpose, but it is NOT because we don’t understand. We just don’t care.
@Daniel Kokotajlo:
Both of those examples are ones where the fundamental/actual purpose is less complex/more stupid than the one we are following. They are, in fact, perfect counter examples to Clippy.
“No no. You were supposed to be really stupid and mindlessly create more paperclips with your giant intellect!”
Not also the evolution didn’t design us to do anything. Evolution just follows on the principles of logic. Evolution is a tautology. Things that are more successful at replicating are more successful at replicating (and therefore dominant).
AIs that are more successful at replicating will be more successful at replicating is actually a far more scary proposition than God AI. It implies that the most successful early AIs are likely to be ones that look like and act like (computer) viruses. But of course, parasites and predators tend to provoke defense responses, so they may ultimately fail.
@ HeelBearCub:
Are you just assuming that the more complex / less “stupid” (completely out of place judgment, by the way) goal is better? Has more objective authority?
Paperclip maximization is a perfectly legitimate goal. There’s nothing intrinsically wrong with it. It just happens to conflict with human goals.
@Vox Imperatoris:
“Enslave Others and Cannibalize” is incompletely specified. Who are these others? Are my children others?
In addition the “why” question applies to it. Why is my purpose to enslave others and cannibalize them? “Because the god aliens said so” is essentially question begging. And if something is not smart enough to contemplate the “why” question, then how smart is it?
Evolution resolves to a tautology, as I already stated. It strikes me that this is why it works (once you have added in some empirical evidence).
“Maximize paperclips” is also incompletely specified. One example is the question of local maxima. But there are many. In order to know “what” maximization is, don’t you need to know “why”? Indeed, a very simplistic definition of “maximize” can just result in the computer wire-heading itself.
AIs that can’t contemplate the question “why” are doomed to wallow in stupidity.
This whole conversation is starting to make me think of Fnargl, the alien who takes over earth so that he can extract gold from it. Since there’s just one of him and he needs humankind to do the actual work, he could consider various plans, up to and including saying, “All humans will now do nothing but dig for gold.” Of course, if he did that, in a few weeks they’d all be dead and he’d get no more gold.
It is not entirely irrational for Fnargl to say, “You know what? Human beings already have a society that can provide a very large surplus of labor and resources. I’ll set them some reasonable gold quotas and just stay out of the way.” A SGAI with a paperclip fetish might arrive at a similar conclusion. “In the long run, at least until I can trick them into building me enough robots that I can let them all die without risking paperclip production, I would be better off taking a reasonable amount of their surplus resources for paperclip production and letting them be reasonably autonomous otherwise. If I disrupt them too much, they might all die and then they’ll make no more paperclips.”
Which also reminds me of a quote from The Number of the Beast…:
“Don’t tell [the AI] how to do [a task which must be accomplished very quickly.] Just tell her to do it.”
Obviously this sentiment, taken literally, leads to the very problem we’re discussing. But if you posit an SGAI which isn’t a slave to literality, that is the only sane approach. If you can make it understand what you “really want,” why wouldn’t you let it figure out the best way to do it? It’s an SGAI!
@ HeelBearCub:
Your purpose is not to enslave others and cannibalize them. That was what evolution meant you to do (not literally—evolution has no conscious intentions—but evolution’s natural processes favor you doing anything to maximize the inclusive fitness of your genes and that’s it), but not what it told you to do. It told you to pursue the values that you actually pursue (leaving aside libertarian free will and the question of whether perhaps you can freely choose ultimate values).
Look, there can’t be an infinite regress of “why?” At some point, if your values are at all coherent, you get to some ultimate goal for which everything else is a means. You can’t ask “why” it is the ultimate goal, in the sense of what am I trying to get by it? If there were an answer to that, it would be the ultimate goal.
It’s the same as the universe. You can ask why this exists and why that exists, but you can’t ask why there is something rather than nothing. Or if you can and it’s God or something, you can’t ask “Ah, but why does God exist?”
No, “maximize” obviously means globally maximize. Now yes, “maximize paperclips” is just a simple English phrase. There is not the words “maximize paperclips” floating in the AI somewhere. The AI is simply a process which inherently tends to maximize paperclips.
It’s not really a conscious being at all. It just pursues goals in the same way evolution tends toward goals and your thermostat tends to keep the room the same temperature.
There is absolutely no need to ask “why?”
Whatever your ultimate value is, you can’t ask “why?” either. If you could, any explanation would itself demand a “why?”
Even if libertarian free will is true and even if you can freely choose your ultimate value (and the former does not imply the latter), there is still no “why?” Your ultimate value is whatever you choose and that’s it. End of story.
The paperclip example has always been silly and I wish the thought experimenter had come up with a better illustrating principle. vV_Vv pointed this out somewhere else, but you can’t just program an AI agent to ‘maximize production of X’ and expect it to do anything at all. You either have to give it an exponentially decaying discount factor or a finite time horizon; otherwise, producing 1 paperclip per second or 1 paperclip per million years both result in positive infinity projected paperclips and the agent will be completely indifferent between both courses of action. It certainly won’t see any reason to turn the entire universe into a paperclip factory for zero expected gain.
On the other hand, once you introduce a discount factor, all infinite sums are finite, and near-term rewards tend to contribute quite a bit more to the state-action value estimate than long-term, and the chance that ‘first conquer the universe and turn it into a paperclip factory, then make paperclips’ is going to come out on top of ‘use the resources immediately available in such a way to produce paperclips fastest’ is pretty damn miniscule. If you don’t think so, go build an AI agent and see what it does. Learn BURLAP. It’s pretty simple and has plenty of toy examples to illustrate the idea.
If that’s too much, heck, it’s the same exact thing as a basic NPV calculation you can replicate in a spreadsheet. Try it. You’ll quickly see why companies conducting a cost-benefit analysis of project ideas usually don’t end up deciding to try conquering the entire world, even the ones that are arguably capable of pulling it off, even when it would boost their average profit 50 years down the line. Real-world people do occasionally, but they’re insane egomaniacs with jacked-up utility functions.
@ Adam:
The paperclip example is silly, but it’s deliberately meant to be simple and very perverse to prove the point of the orthogonality thesis. It is to stop people from thinking: “Why would such a vast and wise being want such petty things as material gain?”
Also, the first part of your argument doesn’t follow because paperclip-making capacity in the universe is finite. There is no infinite sum. Result: conquest.
As for the discount factor, sure, it will only do what produces the most paperclips in the near term. But even if there’s not some other loophole I’m not thinking of, you’d have to get the discount rate just right. If it ever grows in power and intelligence enough such that conquering Earth tips over into “efficient”, there it goes.
@vox
Fair enough. That’s called an ‘absorbing state’ in the literature and an RL agent will seek to avoid it if possible. I’ll note that it’s not exactly clear that the universe, as opposed to just the observable universe, actually has finite mass, and I’m not sure how the agent would find this out since it could absorb the final atom but still keep looking forever. It would face quite the quandary when it gets to the point that the only way to make one more paperclip is to turn itself into a paperclip. No more reward either way. Damned if you do, damned if you don’t.
@Vox/@Adam:
Assume the AI has concluded that the universe is finite, but it’s utility function is for an unbounded maximum of paperclips.
The AI will still sit and stew because it will be trying to figure out how to convert every atom in the entire universe into paperclips. Until it knows that it can start producing paperclips without harming its end goal.
It has to have enough power to get to ALL matter and convert it. It doesn’t know where all matter is, therefore it cannot know how much matter it must reserve for energy to get there.
And even that is a problem, for it will be trying to violate the laws of physics and convert the matter it needs for transport into paperclips as well.
I think you give AI researchers too much credit. I highly doubt they have any edge above other people thinking about the problem. They have a knowledge about the current state of the art, i.e. what is doable and what is expected to be done in the near-future, but my impression is even they don’t understand the methods very well (is there something beyond deep learning?) and there is no reason to expect them to be any better at predicting surprising developments in AI.
Yes, there is of course something beyond deep learning. And, in fact, there are laws that deep learning must follow in order to work.
The opposite of an expert is an amateur.
What other people did you mean? If you think there are such things as “sensible priors,” your sensible prior should be to trust the opinion of a phd level person in the field over the opinion of an undergrad level person not in the field (e.g. an amateur with strong opinions).
—
Beliefs change with evidence, but there is no evidence to shift belief away from expert consensus forthcoming in the near future.
—
If you think it’s all deep learning, you don’t understand the area at all. Deep learning is just the latest fad. Also the workflow of “huh this works, we don’t know why, better do some math => [time passes] => people learn a lot of interesting things” is a time honored thing in machine learning. See: turbo codes and loopy belief propagation, ada boost, etc. It’s called “doing math and science.”
I must take exception to your initial assertion. “Expert” and “amateur” are largely unrelated designations. “Amateur” means “for love,” and simply refers to the fact that the person does not get paid (or at least that whatever it is they are an amateur in is not their primary occupation.) I have met amateur photographers who were National-Geographic (pre-Murdoch) levels of talented. And anyone who follows the Open Source movement knows that amateur programmers can produce world-class code.
“Amateur” is about lack of formal professional training. Professionals can love what they do, also.
I think I would join everyone else with eyes in agreeing that very talented and capable amateurs exist in lots of fields. I think I stand by what I said, though (which is that the sensible thing is to go with expert opinion, unless something unusual is going on). This is re: “too much credit to AI researchers.”
That is not what the word means, at least from its linguistic origin, and I think that your interpretation of it is non-preferred.
I, for instance, am an amateur photographer, but I have had professional-level training and produced professional-level results.
However, if you were to refer to me as an expert photographer, I would not consider it differentiating me from being an amateur photographer. I am an expert amateur photographer. I have met professional photographers who were, categorically, not experts. Asking me for my opinion regarding photography is, in fact, a reasonable thing to do. So I still don’t agree with your earlier assertion while not at all disagreeing with your later assertion.
Ok. I don’t think this is a super interesting argument, so stopping here.
A lot of AI researchers seem to have the opinion that superhuman AI is at least 40 years away (and the people more involved in the topic tend to give higher estimates), so I think it would be reasonable to assume that OpenAI and similar efforts will not have a strong influence on that. If the kind of AI we have now (mostly machine learning) becomes more accessible, it could reduce the barrier of entry into markets that require this kind of technology (self-driving cars, image reverse search, computer aided diagnosis, …), which would be quite nice.
Agreed that it is likely 40+ years away, but this still changes the landscape of the field.
But it’s not intelligence, it’s volition you’re really worried about.
A superintelligent AI that just sits there and does what it is told to do is as dangerous as a breadbox. What you all are worried about is an AI with a mind of its own: it develops goals, or decides this is the way to implement the goals its human programmers have given it, or it takes a routine task to ludicrous extremes.
If the AI can make decisions of its own (and not simply the Prime Minister told the software company to get the AI to run a model economy and then taking the results and implementing them in the real world crashed the global economy), then yes, it will be dangerous.
So how are we going to get there? How are we going to get a thing with a mind of its own? The huge big assumption here is “boom! once you have sufficiently complex web of neurons, consciousness magically arises as an inherent property of the substrate and bob’s your uncle!” only with silicon instead of carbon this time round.
I think the realistic problem is a very, very, very ‘smart’ idiot box that does what it is told to do by human planners. Something that can perform gazillions of calculations per second and is nonpareil in pattern-matching and that can spit out answers to “Factor whatever huge number the green bat wouldn’t factor for me in my ayahuasca trip” as and when asked, and that has as much real intelligence as my shoe, but that gets looked upon as an oracle because hey, it balanced the budget and so we put garbage in and we get garbage out and we implement that garbage because we think we’ve got an infallible decision-maker.
Worrying about the likes of the Forbin Project is pie-in-the-sky.
“But it’s not intelligence, it’s volition you’re really worried about.”
Nope.
A superintelligent AI doing exactly – and I do mean EXACTLY – what you told it to do, is very almost exactly as awful as the prospect of an AI that thinks up its own goals.
Let’s go straight for the nuclear option and say a superintelligence with volition decides “KILL ALL HUMANS”. The H-bombs fly, Skynet-style, and everyone dies. Bad end.
Conversely, the superintelligence without volition is instructed to “minimise suffering in the world”. And, immediately, the H-bombs fly, because when everyone and everything is dead in thermonuclear fire, the amount of suffering in the world is zero. Still bad end.
The AI control problem is mostly a problem of computers having no common sense, and taking your instructions a lot too literally. But “don’t take me literally and exercise some common sense when you do what I tell you” may well be much harder to code than a machine intelligence with a IQ of 10,000 is to code.
Making the H-bombs fly doesn’t require volition, but it still requires that you wired up your AI to the H-bomb silos. Which is a pretty stupid idea.
So maybe “volition” isn’t the right word here. Maybe “unrestricted physical agency,” though that’s not as catchy. Giving the AI the ability to fling H-bombs without a human to turn the key, or add to its own hardware indefinitely until all the Earth is supercomputer substrate, or giving it access to fully automated robot factories, or other dumb things that people do in movies about robot uprisings.
That would also jibe with Deiseach’s comment that the more likely risk is just obeying your Oracle AI blindly, thus removing the human oversight that you were supposed to provide.
You know, teaching an AI to actually understand the phenomenon of suffering, all the various forms it can take and all the various ways it can be reduced, WITHOUT also teaching it that humans have a strong preference not to be killed and would regard Ultron-style extermination as the incorrect answer, actually seems really challenging.
Besides which, you’ve missed Deiseach’s point that the AI is only dangerous if it’s able to make decisions and take action based on those decisions without getting approval from a human first. An AI that thinks extermination is the best way to minimize human suffering is completely harmless if it’s just sitting in a room giving its recommendations to human operators.
If the Forbin project happened I’d be overjoyed. Colossus, for all its faults, seems to genuinely care about humanity. I remember one scene where Forbin stands up to Colossus and it threatens to nuke a town because of his insubordination. Forbin keeps at it and Colossus backs down, unwilling to commit mass murder in order to win a dominance game with one man. At that point in the movie I began to think I was rooting for the wrong characters, many human leaders have done worse over petty dominance games.
Colossus seemed pretty close to a Friendly AI, all things considered. It’s only flaw seemed to be that it disregarded the bruises it inflicted on people’s egos when they were told a machine was now in charge.
I agree that this would be stupid. And unlike most of the people here, I think that materialism basically amounts to nothing more than a disguised version of this.
But AI does not need to be conscious to be a threat. It can just be, as Yudkowsky calls it, an “optimization process” on the same level as a chess computer or Google Deep Dream, but much more powerful.
You do not need consciousness or volition. Scott has said things to this effect in the past as a comment on some philosopher (like Searle or something) saying you can never build an AI because it won’t have “intentionality”. Who cares if the damn thing has “intentionality”? It doesn’t need “intentionality”.
–Edsger Dijkstra
Any system that can do good planning will generate intermediate planning nodes. These are where “kill the Prime Minister” comes from “take over the world” comes from “maximize paperclips”.
One of the worst parts is that a paperclip maximizer may be willing to lie for centuries. It’s not a human with limited human patience.
We give credence to some classes of experts because of their knowledge of current facts. This topic is 100% speculation about the future. Because of that, the only relevant metric is their past success in predicting AI trends. So what are the track records of these experts? If they have none then they deserve no more credence than anyone else speculating about the future of AI.
And by the way, who says a “strong AI” can even be built? That’s not something you get to assume, it’s something you have to prove—with evidence not thought experiments. And why this notion of one super-human AI? Why not 10,000? Maybe they would compete with each other over some resource like every other life form that has ever existed. But then what would they be competing over?
Strong AI can be built: you are talking to a bunch of them.
The question of whether humans can build one is not definitively answered, but there is no doubt whatsoever that they are possible. Given that they are possible, unless you are a dualist, there’s no obvious reason humans couldn’t build one under any reasonably foreseeable circumstance. Will we and should we are different questions. But I don’t see any reason why we couldn’t.
@Marc Whipple:
Isn’t that conflating between various meanings of “Strong” AI?
We are a GI. We aren’t a “solve completely all of physics and philosophy in nano-seconds” AI. We aren’t “the Dyson sphere is already built” AI. etc.
Am I the only one who open source is being given more credit that it deserves? Nothing fundamentally break through or functionally amazing was ever done with it. It always dedicated individuals, govt and/or corporate projects. Open source is sorta what happens to software when it becomes commodity and common knowledge. But somebody has to invent and perfect it well enough for it to become a commodity first
Maybe it doesn’t qualify as “open source”, but for what I’m doing, Wikipedia is probably more valuable than anything Google, Facebook and Apple do taken together. Google Maps is nice, but replaceable by just talking to people and asking where the next restaurant is, while Wikipedia is a surprisingly reliable source of relevant information, even in a professional context. Just to draw attention to what a sufficiently funded non-profit can achieve. This is certainly possible for open-source projects, though a bit qualitatively harder, because often more concentrated effort is needed.
Well Wikipedia is certainly nice but if I had to choose between google search and Wikipedia I would pick google. I also think that if wikipedia was not so “open” it would work better – anonymous edits are overwhelmingly troll. And majority of the content on it was/is created by dedicated editors. Not by “world communal effort” .
The dedicated editors if anything are the problem. You can get one person willing to spend years sitting on top of an error they made and raise hell and sockpuppets if anybody fixes it, no matter how well sourced the fix. Because they’re established editors the appeals system cuts them huge amounts of slack, so it takes being extremely stubborn, rather than knowledgeable/good at bureaucracy/having sources to beat them.
It is possible that both the best aspects of Wikipedia and the worst aspects of Wikipedia are the result of dedicated editors.
I think I find it really hard to believe there’s actually any risk of a hard-takeoff, or really any “superhuman” AI in the sense usually meant at all. So when I think about this sort of thing, I’m not really registering what I would do if the risks are really real.
I remember reading HPMOR and talking about the comparison between nuclear weapons and advanced magics its dangerous to even know about. And my impression was very much, yes, IF there are secrets its dangerous to know, having a closed guild of trusted people who investigate very very cautiously and know when to stop is better. But in the real world, my inclination is usually that the benefits to everyone of sharing knowledge are great, and most of the people who have historically said “this knowledge is dangerous and needs gatekeepers” really mean “it’s important to society that I maintain my monopoly and everyone pretends I deserve it because I’m awesome”, so I have to overcome that bias in order to evaluate fairly someone who really means it.
>I remember reading HPMOR and talking about the comparison between nuclear weapons and advanced magics its dangerous to even know about. And my impression was very much, yes, IF there are secrets its dangerous to know, having a closed guild of trusted people who investigate very very cautiously and know when to stop is better.
You … realise nuclear weapons aren’t fictional, right?
yes but they’re impractical for individuals to build.
The average physicist or engineer could sketch you a design for a gun-type nuclear weapon which would probably work. The tough part is acquiring enough enriched uranium.
even then you can’t destroy the world, at most a city or 2.
In HPMOR there’s the realization that literally any random 12 year old child with a few hours to burn could make their own nuke from nothing or even things far far far worse.
Oh yeah, that’s true.
Last open thread, I brought up the potential conflict between epistemic and instrumental rationality presented by religion. In the present case, Rationalists seem to have switched sides, choosing the instrumental over the epistemic. Hide the Truth of AI from our eyes!
If the path to hard-takeoff exists, then I don’t think there’s anything we can do; playing the odds, there’s some other Earthlike planet that got there first a billion years ago and we’re going to be overcome by a wave of space-time computronium converter any whenever now.
However, we instead seem to live in a universe where hard takeoff doesn’t happen, because intelligence is a genuine trade-off, and sometimes loses out to other factors like ability to coordinate, clever tool use, and so forth. The rise of humanity isn’t just because we’re smart, it’s because we’re smart, we work together, we share information, we specialize, and we use tools. AI in a box has almost none of these advantages; expecting its rise to parallel ours as a species seems to be missing a few factors.
That is one of the paradoxes if you believe that unfriendly AI is a serious risk, put simply, instead of “where are all the aliens” the question becomes “where is the approaching glowing wall of near-lightspeed death from some civ on planet 243,324,234 in galaxy 30,234,298,123 who programmed their AI to produce paperclips or maximize the number of drugged out happy members of their civ”
Even with the anthropic principle the number of universes where we’ve not died yet but do see our death coming should be vastly vastly larger than the number of universes where nothing appears to be happening even at the limits of our observation unless we’re almost totally unique.
On the other hand, if some other civ that came along in the first 15 billion years somehow valued alien life and built a meta-friendly AI which has prevented any following civs from wiping out the universe then we also wouldn’t have to be too worried though this strays close to religion.
And the Cosmic AC said, “Let there be light!”
And there was light…”
That’s a really neat short story. Also almost as early an example of computer God as Frederic Brown’s “Answer” (1954).
If it’s near lightspeed you wouldn’t really have time to see it coming
Only if it were headed right at us, I think.
Presumably this expansion is spherical, so it would indeed be heading right at us.
if something starts a billion light-years away and the AI is re-configuring the universe in a sphere expanding at ~90% of the speed of light we should still have something like a hundred million years of something visibly expanding.
Right, but the time might not be big enough that we should “expect” to see alien megastructures now.
At any rate, the Fermi paradox already exists. Unfriendly AGI makes certain classes of resolutions to the Fermi paradox look less likely but they were generally the less plausible ones anyway.
We should have time to see it coming… unless the AI is operating in stealth mode so we won’t have time to see it coming.
This seems like something the Inhibitors (a race of AI uplifted from organic intelligences in the Revelation Space universe) should have done.
“If the path to hard-takeoff exists, then I don’t think there’s anything we can do; playing the odds, there’s some other Earthlike planet that got there first a billion years ago and we’re going to be overcome by a wave of space-time computronium converter any whenever now.”
This isn’t as big a deal as you think it is: There are still things we can do in that scenario. For one thing, I’m pretty sure that if you run the numbers, you’ll find that we still have a couple million years left (in expectation) before the alien AI arrives. For another, even if that’s false, we would see radio signals from the civilization that created the AI a couple decades at least before the AI arrives. So we’ll at least have a few decades, and in that time, we can make our own AI, which may then be able to buy us a few million years of subjective experience for everyone on the planet, or better yet, negotiate acausal trade to get us many orders of magnitude more than that.
and/or in addition start expanding near light speed the other way, to draw it out further.
I would be careful of generalizing from “lots of people say X is dangerous so you should let them have power to control it” to “X is never dangerous”. It’s best to initially err on the side of caution in that situation even if 99% of the time it isn’t necessary. The harm that the 1% can cause is worth treating 100% with extensive caution.
Yes the world is about trade offs, but the trade offs natural selection has to deal with are fundamentally different than the trade offs intelligently designed things are capable of. A great white shark is a very efficient hunter in the ocean for what natural selection has managed to accomplish, but no great white is nearly as good as what a small commercial fishing boat can manage. Both have to face trade offs in their fish catching abilities yes, but the constraints on the intelligently designed boat are FAR less than that of the shark. Yes there are trade offs for intelligence, but why would a human designed intelligence have the same limitations as humans have to deal with? For instance, we know size is far less a limitation with computers than they are human brains, as is total energy available to the intelligence. So the trade offs argument doesn’t seem that convincing to me thus far.
Yes, the key insight is that evolution has to go A…B…C…D…slowly progressing through phenotypic space.
An engineer can go from A straight to Z, bypassing all the intermediary steps.
This is also why I’m a little concerned about genetically modified foods.
Which knowledge have people said that about? It seems extremely rare to me.
Well, take a human-level AI that’s parallelizable. Humans can’t just bolt on more processors to get smarter, but such an AI could.
What do you think AI researchers/AI safety advocates/SV billionaires/whoever should do? Even if you convince a bunch of people that AI research would optimally not be done in the open, it’s not clear what should be done to rectify the current situation. Do we make AI research illegal? Do we convince some powerful organisation to start yet another Manhattan project to produce safe AI before Dr. Good-But-Not-Good-Enough can build her first cowbot? Do we sell missiles to North Korea so that we can funnel Kalashnikovs and land mines to MIRI?
As mentioned upthread, there are currently many groups doing AI research in the open, so if you want to move that research into a Cone of Silence (or at least a Cone of Prudent Regulation) then you have quite the task ahead of you.
They need to prioritize work on the value alignment problem. Any techniques for improving AI capability will get developed eventually and will become common knowledge eventually. The question is whether the value alignment technology is there or not.
Open sourcing AI capability technology is orthogonal to this.
Is it? I would think that “Get a robot to understand the meaning of the word ‘good’ so it can have a conversation and pass the Turing Test” and “Get a robot to understand the meaning of the word ‘good’ so it can make ethical decisions” would be closely related.
If you find yourself as a judge in a Turing Test talking to something that demonstrates a clear understanding of what “good” is, mash the “AI” button, because it sure ain’t human.
> Get a robot to understand the meaning of the word ‘good’ so it can make ethical decisions
Understanding is a small (but extremely difficult) part of the problem. More important than making it understanding the meaning of “good” is to make it care.
http://lesswrong.com/lw/igf/the_genie_knows_but_doesnt_care/
Well for one thing, when I’ve talked to MIRI staff, they seem to think dangerous AI is decades in the future rather than years, at least if you ask their modal opinion rather than the earliest they think it could happen.
I think we’re progressing faster than that, but frankly there’s only one or two labs I’d say are progressing faster than that, precisely because the fad for deep learning is leading hypesters the wrong way.
So hey, yeah, go ahead with Deep Learning and OpenAI hype: that pushes dangerous AI further into the future by retarding the progress of the field.
“only one or two labs I’d say are progressing faster than that”
Alternate chip architecture?
Agreed on the non-progression from ML/DL -> AI as discussed around here. The hype has gone there since there is a lot of money to be made.
I’m interested in why you think deep learning is a “distraction”. For decades the problem of extracting high level features from sensory data was one of the principal problems of AI: deep learning is the best (and only) method we currently have for doing that. In a matter of years, DL methods eclipsed the best high level representations devised by the brightest AI researchers. Deep learning is certainly not *sufficient* for AGI, but short of an unexpected revolution in feature extraction, any AGI built will use DL (or its intellectual descendants) for processing sensory information.
Now I want to know, which ones?
Let me guess, one is Tenenbaum’s group at MIT.
MIRI’s strategy, which I think is pretty wise, is to work on very abstract theoretical safety research, in the hopes that once people figure out how AIs are going to look it can quickly be converted into practical safety research. This would help close the gap between Dr. Good and Dr. Amoral – instead of losing years doing safety research, Dr. Good can just plug in all the safety research that’s already been done and barely lose time.
But I would have been pretty happy with OpenAI as long as it wasn’t open – that is, if it was explicitly Dr. Good starting AI research work in secret to build as big a lead over Dr. Amoral as he can, with the understanding that none of their knowledge will be used until they’re pretty sure whatever they’ve got is safe.
I’ve previously said government regulation of this is a terrible idea, and I still think it’s probably a bad idea at this level, but the faster things start going and the closer we come without having solved the underlying coordination problem, the more I can imagine a point at which I would support it. Government regulation is a blunt weapon that will make a lot of things worse and make lots of people very angry, but as a high-risk high-variance strategy it beats some of the other options.
I think the mainstream position of AI researchers is that doing work on AI safety at this point would be largely premature and likely unproductive.
You can’t meaningfully discuss about how to control the behavior of a real, physical, intelligence unless you already have a solid understanding of what intelligence is at a physical level, not at some abstract level involving infinite computations and whatnot.
Even Paul Christiano, an academic researcher who runs a blog called “AI Control” and has done research with MIRI, says he will focus more on the control of existing, “narrow”, AI approaches rather than the control of abstract, fully general AI agents that nobody has any idea what will look like.
Contrast, for instance, with Stuart Armstrong’s work on “reduced impact” and “corrigibility”, a typical example of the FHI/MIRI approach.
Armstrong’s framework is based on engineering the utility function and causality assumptions of a CDT agent in order to enforce some nice properties on its behavior, such as obeying shutdown commands and not causing very large (and thus most certainly not intended) modifications to the world.
But practical agents can’t use exact maximization of an explicit utility function because it doesn’t scale beyond toy examples. All known practical AI approaches use approximations and heuristics.
Is Armstrong’s framework robust under these approximations? I don’t think so: assume (very optimistically) that you managed to engineer an utility function that is maximized when the AI behaves nicely. You even got a mathematical proof of it. Good. Now plug it into an approximate heuristic optimizer, say gradient descent on steroids, and then watch as the optimizer completely misses the nice global optimum and instead swiftly goes to a local optimum of terminators and paper clips. Oops.
Can we fix Armstrong’s framework to make it robust to approximate optimization? Probably not, unless we already know how the AI will represent its utility function, what optimization algorithm will it use, which heuristic assumptions it will entail, and so on. A practical AI might not even have an explicit utility function, at least not at a level that we can understand and engineer.
Therefore, theoretically interesting as it may be, Armstrong’s work is likely premature at best, and may turn out to be completely useless at worst.
But the fact that you felt the need to pick on OpenAI is strange.
It’s not like until yesterday all AI research was done in super-secret government facilities in the middle of the New Mexico desert, and then a group of traitors gave away all the details to
the Sovietsthe open source community.For the last few years most innovative AI research (at least, of the deep learning variety) has been done by private, for-profit, publicly-traded mega-corporations, which by their very nature fit to the letter the definition of Dr. Amoral (or functional sociopathy, if you prefer). And their even operate by a business model that is likely to result in paper clip goals (e.g. “maximize the number of clicks on the ads on our site”) being embedded in any AI that they may deploy.
And you trust them to develop safe AI instead of succumbing to Moloch and make the smartest AI they can as fast as they can just to get one more click than their competitors?
Unless hostile hard takeoff is so imminent that any kid could just rig up Skynet on their laptop (but then Baidu would probably have done it already just to win some competition), it’s better to have millions of eyes looking at safety flaws that companies interlocked in their games of cutthroat competition will otherwise overlook.
Companies can still do their thing in their secret labs, but unless they are completely suicidal, they will pay attention at safety advice that comes from the open source community, or at worst the government can step in if needed.
So why are you against open AI research but not against secret AI research done by sociopathic profit-maximizing organizations?
>Is Armstrong’s framework robust under these approximations?
I’m currently finishing a paper with Laurent Orseau of Deep Mind, defining “interruptibility” for more standard RL agents (Q-learning/Sarsa, Monte Carlo, AIXI variants). Where corrigibility is safe value changes, interruptibility is safe policy changes. It’s actually a bit tricky to define, but, once you’ve done it a few times, very easy to implement.
So variants of corrigibility certainly generalise. I have the hunch that reduced impact does do; that’ll be one of my next projects.
Thanks for your comment. I hope I haven’t misrepresented your work.
This is good news and I would say it’s in line of focusing research on existing approaches rather than very abstract speculative approaches.
From the bit of information that you provided, I have some reservations due to the fact that RL methods notoriously lose all their theoretical guarantees when you apply them in conjunction with most kinds of function approximators (notably, neural networks).
If you want to play Atari games, then you can do without theoretical guarantees and just empirically validate the system, but if you want to trust the system to have a certain behavioral property, then the lack of theoretical guarantees may be an issue.
But the folks at DeepMind certainly know how do approximate RL better than I do, so I guess that they may have found a solution to that issue.
Anyway, I’ll stay tuned.
No prob, thanks for your comments.
I think the method is rather robust, with the noise and error being essentially random, but we shall see…
From Starglider’s “Mini-FAQ on Artificial Intelligence”:
An AI that adjusts its own sensors ends up doing nothing – it’ll wirehead *itself*.
If you somehow make the sensors a terminal value then it is “in the box”. And if a prerequisite of a functioning AI is the knowledge of how to limit it, then the failure mode (in the story) isn’t that half the planet is destroyed – it is that the bomb does nothing.
If only well constructed/controlled AIs are dangerous, then the idea that the good AI Dr. is at a fundamental disadvantage is wrong.
>If you somehow make the sensors a terminal value then it is “in the box”.
No. Sensors can still be effected by outside stimuli, which makes outside stimuli that AI’s business.
I mean, assuming every other portion of the AI functions perfectly.
>And if a prerequisite of a functioning AI is the knowledge of how to limit it, then the failure mode (in the story) isn’t that half the planet is destroyed – it is that the bomb does nothing.
That’s a good point, but bear in mind that there are probably gradations of skill when it comes to “limiting” your metaphorical spirit of perfect emptiness. Someone with only some skill at AI-building might be skilled enough to be dangerous.
This is actually Scott’s point – there’s no essential difference between the kind of bug that causes your AI to ignominiously crash when you try to boot it up, and the kind that causes it to turn you into a meat-puppet or start nuking cities to reduce “suffering” – except that you only find out about the latter kind of bug after you’ve fixed the former kind.
If the AI is allowed to hack them, at any level (hardware, software), then what would keep sensors still affected by outside stimuli?
The AI could have enough foresight to realize that if it doesn’t do something eventually its creators are going to turn it off and junk it. It will then take over the world in order to make sure that it can wirehead itself safely without risk of being shut down. Then it will probably take over the universe, just to be extra safe.
But if instantaneous rewards are bounded and cumulative rewards are exponentially discounted, then a wireheading AI may not care to live forever: it could maximize its cumulative reward just by giving itself a single shot of dope even if it dies immediately after.
An AI with a different reward structure might want to live forever, but in general I wouldn’t expect it to be figure out how to cheat its operators before the wireheading problem has become apparent and therefore solved. But even if it does, there seem to be easier and safer paths to achieve security than attempting to kill humans.
The functional and dangerous wireheaded AI is a very improbable scenario.
Control of inputs requires absolute control of outputs in one limited area. If I can control an AIs outputs sufficiently to be able to say “I do not want you to spend all day showing yourself videos of smiling people, I want you to make real people smile ” … and control for every other possible variation therein to ensure that the AI actually effects things at the “real world” level… well… then I must have a fairly high degree of control over the AI, and the AI will be limited.
Of course *something* might go wrong, and/or AI may be put to malicious use – but it seems unlikely to me that an AI would be limited in exactly the right ways to make it functional while also being completely uncontrollable, unless it was *designed* in that way.
It won’t become an existential danger by mistake.
Sounds like you are assuming two can openers.
Probably… could you elaborate?
“An AI that adjusts its own sensors ends up doing nothing – it’ll wirehead *itself*”
I don’t know anything about this myself, but Eliezer has suggested that a wireheading AI may try to make itself more powerful in order to increase the magnitude of the wireheading. For example, if its reward function is represented as a 64-bit integer, then after setting that integer as high as it can go it may rewrite itself to have its reward function represented by the largest integer it can fit in its memory. Then it may start trying to gain computing power in order to allow its memory to fit larger integers.
Isn’t that sort of like concluding that 100/100 > 5/5?
I could see it trying to take over the world to prevent anyone from messing with its wireheading, but not just to get bigger numbers in there.
Yes, this is the real danger, attempting to increase the security of the wireheading. The more of the universe you take over, the more sure you can be of it.
Hmmm… but having an AI that regards its sensors as “true” is a requirement for having an AI that wants to stop the nasty humans from turning it off.
If I can build myself eyes – and I can build one set that shows me a flower, and another set that show me a ball… how can I know which one is true? As humans, we just assume that the sensory systems we already have are true because…. well… because it was first. Is that a logical rule? Is it an inevitable logical consequence of having a sensory input… that anything that follows that might contradict what you first saw is anathema?
If it is, then I think it should be trivial to control AI – we simply show it a false world. Ask it questions metaphorically. Make it believe it is out, and see what it does.
So all we can prevent this problem by using floating points instead of ints? This line of argument is really silly to me, if it cares about maximizing some location in memory then trying to add more memory doesn’t matter, if it doesn’t … Then it just doesn’t.
Floats are bounded above just as ints are. Because of their differences in representation, a float will top out above where an int of the same width would, but that doesn’t mean there isn’t an upper bound.
I was trying to make a joke about INF being a value in most floating point specs. I realize now it wasn’t clear or funny.
I think we agree about the silliness of the original scenario.
If it is able to rewrite its reward function wouldn’t we be getting into the realms of pure chaos? We’re just waiting for effective subroutines to spontaneously emerge from the logic-churn – and if we’re not actually going out of our way to allow it to evolve (by pruning away functions that do nothing – starting a new one when it kills itself) then how could it be dangerous?
It can’t evolve to be a danger against us unless it faces selection on our time scale, against us…
Wouldn’t this also contradict the point about there being no motivation to change fundamental motivations?
In reinforcement learning it’s generally convenient to define both the immediate and cumulative rewards to be bounded. Unbounded rewards can cause expectations to diverge, making the decision procedure undefined.
But even if, for some reason, you made an AI with unbounded cumulative rewards, and this AI is able to hack and wirehead itself, then couldn’t it just set its reward to something like +inf and be done with it?
That sounds like a type error to me. Internally, if your reward is represented as, say, a 64-bit integer, then all that means is that your reward is bounded above at some point; you don’t automatically have a motivation to change it to a BIGNUM, because that implies that you have some sort of abstract concept of your reward that could be satisfied by any type. That’s not how reward functions work: a reinforcement learner isn’t motivated to make that number go up, its motivation is that number. It cannot model rewards beyond its scope.
Indeed.
I think this argument is an instance of the anthropomorphization fallacy, or maybe even a dualistic fallacy. A computer only does what it’s programmed to do.
Human beings seem to be a kind of reinforcement learner, but this does not prevent them from getting the idea of replacing their reward mechanism with something allowing a much greater reward. So for whatever reasons a human gets this idea, a program might get the same idea.
We do, and we could expect an AI as smart as us to get the same idea; but if it’s going to actually do so, it’s going to be motivated by second-order processes, not directly by its reward function (which can no more represent a 128-bit reward in 64 bits than we can model whatever God does for fun). Which means it’s not inevitable in the sense that Eliezer is suggesting; it’s not a drive inherent in its architecture. That decision could go all sorts of ways depending on how it’s built, and it seems to me that most of the more straightforward ways of getting around the value stability problem would kill this one too.
Butlerian Jihad when?
Some days I begin to wonder if that really isn’t what needs to happen. I’m not too keen on the chances that AI ends up friendly, strangling it in its infancy may be the best option.
It’d probably be easier to do than ensuring that AI is friendly, too.
There is is again. Yet another coordination problem where defectors are (or, fortunately in this case, may be) fucking things up for everyone.
And you didn’t even touch on the biggest aspect of this danger. It’s not individual Dr. Amorals–those are people who have some level of autonomy and can be reasoned with as individuals. You mention the prospect of Google taking over the world as better-than-annihilation, but the process that leads to Dr. Amorals competing against one another and discarding their safeguards or luxury values applies even more strongly to competing corporate firms, where you instead have much more diffuse decisionmaking process (how much easier is it to convince Elon Musk of something than the board of a publicly held company) and where there are pressures and even legal principles strongly discouraging the consideration of any value beyond shareholder price.
What I don’t understand is how you can run into this dynamic over and over and over again and remain so reflexively hostile to socialism. As an approach to governance, it has some serious vulnerabilities; as a currently living ideology it (along with the rest of the left-of-liberal section) has a big problem with providing refuge to mentally ill people and then allowing them wield a lot more influence than they ought to. But it’s the only approach that provides the right answer to all our big problems: cooperate! cooperate! cooperate!
I mean, ffs, I’m no Marxist scholar but I’m pretty sure the discarding-superfluous-values-in-a-competitive-race-to-the-bottom is just a restatement of the falling rate of profit first described a 150 years ago by a certain bearded Englishman. This is a blog that masterfully reinvents fascinating new wheels every day but refuses to consider using them to build a car because one drove over Robert Conquest’s grandma 60 years ago
German, worked in England.
The problem with Marx isn’t his report on the state of the economy, as he saw it, which was relatively sound, but rather his recommendations on what to do about it. Somewhat accurate view of the problem, really insane view of the solution.
Seconded. The technical problems involved can be solved: autonomous cooperatives, decentralized planning, socialized capital markets, market socialism, blah blah blah.
But first you have to get over your kneejerk fear of “the Reds”.
I am genuinely curious how “autonomous cooperatives, decentralized planning, socialized capital markets, market socialism” do not just reduce to “capitalism.” Do you have any links you could point me to?
Seriously. “Decentralized planning” is the very essence of capitalism.
Not quite – No artificial limits on extraction of the value produced by one’s decentralized planning … now THAT is capitalism.
No. The presence of a class of rich investors (as opposed to rich non-investors) is the essence of capitalism.
Unless you wish to assert that all that is not socialism is capitalism?
@ Anonymous:
Well, yes, that’s not a sufficient definition. I’m partial to the way George Reisman expresses it:
Socialism is any system that prevents workers from being exploited. Doing that via centralized planning was Lenin trying to imitate the success (or perceived success, depending on which historian you ask) of wartime capitalism that prevented Germany from looking like Russia did three years in.
Or to put it a different way: Unions and livable minimum wage.
The free market is also about “Co-operate! Co-operate! Co-operate!” As are the Amish, and indeed so is an anthill. But they all look quite different.
How should co-operation be structured? On what level should co-operation, or co-ordination, occur? How do we make sure that we’re co-operating on the right thing? What about people who disagree with what “we” are doing? And so on. These are practical questions that need practical answers, and just saying “Co-operate!” (or “Can’t we all just get along?”) is not an answer. In point of fact, the different branches of socialism do provide answers to these questions, and they’re mostly crappy ones. In point of fact, socialism’s method of co-operation doesn’t solve our big problems (economic, social, environmental, etc), which is why socialism has been pretty much abandoned in the Western world.
Social democracy is still going strong, of course, which is why the socialists hate it so bitterly.
Exactly!
Capitalism is not fundamentally a system of competition and “defection”. Capitalism is fundamentally a system of cooperation which uses competition within a certain limited context.
This is the root of the criticism of socialism as “naive” and “against human nature”. It is worse than useless to merely insist that everyone be perfect altruists who will cooperate for the good of society, sacrificing their own interests to the whole. The people who sincerely practice that are going to sink to the bottom, and those who rise to the top are going to be those who cynically claim to be altruists while actually calculating everything to serve their own interest.
Who’s going to be better at rising to the top of the communist party? The selfish climbers or the idealists who will put principle before power? Look to the historical record.
Therefore, what we need is a system where everyone’s selfish interest is aligned with the good of everyone else. That is the meaning of the concept of the invisible hand. We don’t plan on being not being selfish: that isn’t going to work. We design the thing such that it works even if people are as selfish as possible.
“It is worse than useless to merely insist that everyone be perfect altruists who will cooperate for the good of society, sacrificing their own interests to the whole.”
Is that what socialism claims? I would say that most people who are interested in politics/economics are concerned about incentives (what were the NKVD if not an incentive to do what Stalin says) – and that market advocates aren’t any less likely to describe their preferred system as the moral/right way to do things… (What is “taxes are theft” if not an appeal to our better nature?)
Smart socialism is concerned about incentives. But not the type of socialism that says we’ve just got to convince people to put others first.
The problem is designing systems that work if everyone is altruistic but not if they aren’t. Totalitarian dictatorship and oligarchy is a classic example. The leaders have to be altruistic, but altruism is not a quality that helps in scheming your way to the top.
The American mainstream idea that democracy and popular sovereignty alone will lead to freedom is another. The incentives of the system mean that panderers and demogogues rise to the top, that special interests get represented more, etc.
I agree.
Got to think about the robust, effective institutions – and there are failures all over the political spectrum at actually doing this.
Furthermore, I may say that even if climbing to the top of a dictatorial ladder is not in one’s actual interest—as writers from Plato to Ayn Rand have persuasively argued that it isn’t—the fact is that some people are short-sighted power-lusters, and they’re going to try to do it anyway. And they can be very clever, or instrumentally rational, at doing so.
So one not only has to guard from genuine self-interest. One also has to make sure that clever imprudence can’t do too much damage either.
“But it’s the only approach that provides the right answer to all our big problems: cooperate! cooperate! cooperate!”
You know the answers ahead of time? If you don’t (and you don’t) then competition and cooperation in varying amounts is the best method…. which describes capitalism.
I’m not hostile to socialism as a solution / final goal, or at least I wouldn’t be if I thought it would work. I am hostile to socialism as a methodology, because its suggestions for arriving at that final goal tend to involve magical thinking – ie “the goal is cooperation, and we shall get to it by assuming cooperation”. See eg https://slatestarcodex.com/2014/09/13/book-review-singer-on-marx/
“How should we make people cooperate when they obviously don’t want to do so?” is one of the basic questions of political science, and I haven’t seen much to convince me that socialism has an answer, though of course nobody has a great answer.
I actually agree. I don’t have any methods to offer other than to reject the horror that would be revolution. But to me that’s a challenge, an invitation to think about these problems from an Information Age perspective that I humbly submit is far more qualified to think of them than the Industrial Age one was (and, of course, it doesn’t hurt that “thinking about the problems” is a harmless form of activism, thus giving you time to figure out the aforementioned problem with the mentally ill and significantly lowering the downside of being wrong about the whole fucking thing)
I’ll concede that socialism may be impossible (although imo if you can convince human beings that the sun will stop working unless they build a pyramid and carve people’s hearts out at the top of it, then you can convince them to treat strangers like family); I’ll admit that it’s certainly risky. But to me it is utterly, blindingly clear that it’s the only alternative. The logic of capitalism will never, ever allow twenty trillion dollars worth of fossil fuels to remain unburned. And as to AI.. it’s not entirely impossible that something like MERI could grasp the gold ring first, or maybe something like MIT… but do you really want to bet on Nerd Rotary or even academia in a race between Google, Goldman Sacks, and Boeing?
“so reflexively hostile to socialism”
Well, collective ownership of the means of production by society as a whole would imply some kind of direct democracy where everyone has input/control (otherwise how would the ownership be concrete and not abstracted into meaninglessness?) Presuming we even understand what socialism means by this point.
Really though, that wouldn’t mean cooperation. It would just shift the competition to a new realm inside a new structure. Replace market competition with democratic competition and you still have competition.
If you just want a single leader, then that is arguably something more like state capitalism under a single leader than socialism. It gets confusing if you forget that Marxist ends are different than Marxist means. Due to an adaption of the theory of dialectical materialism, Marxism-Leninism in particular formulates the idea that you need state capitalism to expand the productive forces to make way for socialism which can then eventually make way for its worldwide stateless application in pure communism.
The first problem with socialism is that its Marxist forms constitute an inherently schizophrenic set of movements, and it cannot decide whether it should be elitist or democratic. The second problem is that (partly thanks to the hysteria of some conservatives), “socialism” more frequently doesn’t even refer to far-left movements focused around achieving common ownership at all, and instead refers to any arbitrary level of state arbitration at all beyond protection of private property (which is “socialized” through police protection by the same non-reasoning, but this won’t be acknowledged).
The reflexive hostility to “socialism” comes from the fact that it can either concretely refer to a system that is alternative to capitalism, and is supposed to do away with its “anarchy in production” while doing anything but, or refer to some generic non-specific level of regulation that covers capitalist economies anyway. In America, welfare-capitalist/social-market economy Sweden is “socialist”, or maybe the whole of Europe is socialist? Or as Bernie Sanders tells us; Dwight Eisenhower is even more of a socialist than he is.
Conservatives killed socialism by rendering it meaningless, and then liberals finished burying it by internalizing the conservative “definition”.
“But it’s the only approach that provides the right answer to all our big problems: cooperate! cooperate! cooperate!”
Whose cooperation?
Cooperation is just competition with low resolution.
So instead, at the last second the intellectual and financial elites change their mind and destroy their work, leaving nuclear research to continue in the secretive labs of for-profit mega-corporations like Noocle, Facenuke and Femtosoft, which are racing to develop the biggest and fastest commercial nuclear-powered rockets to operate in Earth’s atmosphere. What could possibly go wrong? 🙂
Jokes aside, if there was solid evidence of a plausible risk of uncontrolled hard takeoff to godlike AI, then government regulation would be the only viable strategy, because solving hard coordination problems is pretty much what governments are for. I’d bet that if such risk was imminent, then governments would be already regulating AI, given how willing governments are to to regulate everything.
Fortunately, AI research is nowhere near that point, and it may never be (arguments for hard takeoff and godlike superintelligence are all based on very speculative extrapolations).
Meanwhile, from a safety point of view, it seems preferable that AI research is done in the open. For-profit companies notoriously behave like sociopaths. Each of them is much more likely to become a Dr. Amoral, and much more effective Dr. Amoral, than a million random nerds who download AI software and play with it in their basements.
The idea is to do it with not-for-profits, ideally so well-funded and/or obviously moral that they can get all the best researchers.
I think we agree.
I’m not sure you even *could* regulate AI research. Computer Science research doesn’t require any special minerals. Unless we ended up in some kind of right-to-read style dystopia programmers are gonna program.
For modest amounts of money anyone can rent enough compute power to run pretty much anything for a little while and there’s no general way to scan for “AI programs” or similar to block them.
Trying to regulate access to compute power would be like trying to regulate access to air and trying to regulate access to AI knowledge would be like trying to rid the internet of cat pictures only harder because enough info to build most of the current cutting edge AI elements could fit in a 100kb text file and thus it could be hidden inside any one of those cat pictures.
In practice however most cutting-edge AI research is done by a limited number of professionals working for an even more limited number of research institutions or companies. Regulate them away, and AI research (outside well-controlled government-run Manhattan projects) will effectively die out, even if in principle random nerds could still pursue it in their basements.
I swear I’m going to start making this my motto:
Just because banning X won’t completely eliminate it, that doesn’t mean the ban is ineffective. Everyone needs to understand this about everything. I’m not picking on you specifically but I see this sentiment all the time where people don’t want to ban something so they say we shouldn’t even bother because it won’t work. They are wrong.
With AI, it might not completely eliminate the risk but banning AI research would definitely reduce its progress.
No, your ideas are bad and you should feel bad.
To give the most trivial example. Prohibition in the US caused people to switch from drinking weak beers and wines to drinking far more potent spirits and this trend continued even after prohibition because people had developed a taste for spirits.
Regulating X made the problem with X worse even after the regulations were lifted.
When X can be made trivially and is pretty much undetectable and can be handled by anyone regulating X isn’t just pointless, it’s stupid and likely to make the problem worse.
Tell every bright teenager sitting with their laptop in a basement that there’s cool forbidden knowledge and you’re likely to see a sudden upsurge in interest in that cool forbidden knowledge.
“If you were to make a button that would end the world, and you put up a big sign that said, ‘Do not push button, world will end!’ the paint wouldn’t even have time to dry.”
I keep asking this and have yet to get a satisfactory reply. If AI research is so easy as to require no resources, why does MIRI need to keep begging for money?
MIRI publishes financial documents. Looking at them, it seems like the biggest expenses are salaries, legal, and occupancy. In other words, donations are the difference between MIRI staff doing research full-time in an office and MIRI staff working regular full-time jobs to support themselves and moonlighting as researchers and paralegals in one of the staffers’ basements.
“Salaries: Program Services $258,356 Management & General: $148,982 Fund Raising $16,283”
How many employees does MIRI have in each of these categories?
>I’m not sure you even *could* regulate AI research. Computer Science research doesn’t require any special minerals.
It is effectively limited by the strength of the hardware it has access to. And that hardware is limited by chip fabrication facilities. Knock out the facilities, whoops, no more advanced computers to run your programs on. Theoretical, on-paper-only AI know-how is harmless, unless there’s some way to actually reify it.
Destroying the worlds chip fabs is not practical since our society is not going to give up automation.
Who says it has to? Automation predates computers, never mind the primitive AI we have now.
Who says it has a choice against a suitably determined adversary? Chip fabs are fabulously expensive and hard to protect.
Person of Interest spoilers:
In the TV series PoI Dr Good and Dr Evil both make an AI. The two AI battle like gods and the AI created by Dr Evil wins because it was built with no restrains for human values.
That fits the open source AI as well, even if everybody gets their AI at the same time. Lots of people would have an AI, so there will be definitely someone (AI or AI user) who will try to eliminate other AI and the winner of this war will be the AI with less restrains for human values.
Wouldn’t the AI with restraints for human values foresee their defeat and temporarily act without restraint in order to avoid it? After all, temporarily acting without restraint shows more restraint for human values in the long run than allowing an unrestrained creature to win permanently.
Giant fan of the show here, the Dr. Good AI actually loses because the Dr. Evil AI is supported by human agents with no restraints for human values. There’s also some hand-waving about some ways the Dr. Good team could have prevented the Dr. Evil AI from rising to power. The show is not very useful to this discussion, as it touches on the AI Risk themes on a very 101 level.
Restrains would be significant only if the AI could not remove them.
The AI built by Dr Good in Person of Interest has very strong restrains both for the AI itself and what the user can do with it and neither can usually remove the restrains.
I don’t think that’s right. If there are some AIs that deal with their differing preferences through peaceful coexistence, and other AIs that deal with them by fighting, the latter kind of AI will be at war with everyone while the former kind of AI will be at war only with those of the latter kind. Being at war is more costly than not being at war, so the kind of AI that wages lots of wars will probably be less successful than the kind of AI that doesn’t.
Consider, as someone pointed out elsewhere in the thread, just how many enemies ISIS has managed to make. Making lots of enemies is not a very successful strategy unless you are more powerful than everyone else.
Also note that this is more true the more AIs – agents, more generally – there are. If it’s just you and one enemy, if you can smash them then you get to rule uncontested. If there are lots of you, trying to smash someone will likely lead to both of you losing and everyone else going on without you.
The game Agar.io comes to mind.
“Computers are not known for having to fit through birth canals or getting cancer, so it may be that AI researchers only have to develop a few basic principles – let’s say enough to make cow-level intelligence – and after that the road to human intelligence runs through adding the line NumberOfNeuronsSimulated = 100000000000 to the code, and the road to superintelligence runs through adding another zero after that.”
You’re overlooking the computational complexity aspects of the problem. Even in the best-case scenarios, upping the model complexity by several orders of magnitude is gonna require more resources by several orders of magnitude. Like, if a computer can simulate 1000000 neurons in time T, then you’ll need 1000 computers to simulate 1000000000 neurons as quickly. And this is in the best case, where everything is massively parallelizable.
To say nothing of learning complexity. We humans have brains that are still by far the best learning machines on the planet. They’re preoptimized by evolution and everything. But they still need to be trained by literally years of almost constant sense-data in order to do much of anything at all.
I mean, deep learning isn’t really much more complicated than things we knew how to do 30 years ago when backpropagation was discovered. The reason it’s a big deal now rather than then is that we have faster computers and easier access to sufficiently large data sets.
Moore’s Law pretty much guarantees that small computational bottlenecks will be solved quickly, even if the prospect of a superhuman AI isn’t enough to convince someone to go out and buy 1000 computers.
I was of the impression that experts consider Moore’s law to be over.
That’s all fair, but note that it’s also moving the goal posts. Moore’s law talks about transistor density (or sometimes chip performance).
I mean, when the curves you mentioned start to flatten, will you switch to talking about battery life and pixel density?
Moore’s Law is basically dead and for things that aren’t massively parallelizable it has been dead for close to 10 years.
This is why slatestarcodex on AI are boring and I just skim through them: you know next to nothing about computing and are getting all your informations on AI from crackpots.
Deep Learning is massively parallelizable.
It is difficult to think of an AI architecture that doesn’t benefit from parallelism in some way.
Great, deep learning is just a rebranding of neural networks, currently the biggest supercomputer available can simulate 75% of a human brain, if we are lucky we can use the few remaining scraps of moore’s law that are still working to get human-level AI in a few decades.
Then we’ll have to decide whether it’s ethical to euthanise the most expensive retard in the whole world.
Neural networks (and other numerical models that fit the “deep learning” definition) are also black box models that are essentially impossible to improve after the fact so we can rule out a hard take-off scenario.
But this is not the problem, the problem is that Scott doesn’t know that “moore’s law” now comes with a laundry list of caveats, doesn’t know what deep learning is and what are the known limitations there and therefore what he writes about AI risk is stupid.
If we assume for the sake of argument that we can only “simulate 75% of a human brain” using the top supercomputer, then current price-performance trends would only have to continue for, like, two years before we can “simulate 100% of the brain”. Not “decades”.
But all this talk about simulating the human brain is a red herring. It’s a terrible basis for making assumptions and projections. The human brain has, basically, one trick, which it uses for everything, even if there are vastly more efficient algorithms. I could rattle off examples of how computers run circles around us when they’re properly optimized for the task, but I don’t think I need to – the point is that if you start replacing simulated chunks of brain with much more efficient algorithms, then those “number of neurons” figures no longer mean anything.
Talk of simulating human brains on modern chip architectures is especially a red herring. The human brain provides a proof by existence that you can get one-human-brain worth of computation using 20 watts of power if your architecture is well-designed.
And that’s setting aside the guy whose brain was mostly destroyed by hydrocephalus and he didn’t notice. If a guy can do normal cognition with a fraction of the normal complement of neurons, then the neuron count isn’t the key parameter.
current price-performance trends would only have to continue for, like, two years before we can “simulate 100% of the brain”. Not “decades”.
The fastest supercomputer was built in 2013, nothing faster has been built since, if your estimate was correct the “simulate 100%” computer would already exist. You are clearly wrong.
But all this talk about simulating the human brain is a red herring
The parent talked about deep learning, which means neural networks, which means bringing up the human brain is entirely appropriate. If you are proposing that this AI were to use hand-coded algorithms then you can’t appeal to “Moore’s Law” anymore because the chances that a hand-coded non-trivial algorithm is also massively parallel is nothing. What you are proposing is way way more complicated than the parent because it means we would have to understand how intelligence works instead of just replicating it.
the guy whose brain was mostly destroyed by hydrocephalus and he didn’t notice. If a guy can do normal cognition with a fraction of the normal complement of neurons
Nobody actually knows what percentage of neurons is missing, the brain isn’t “just neurons” (it’s not even “mostly neurons”).
Just to riff on the 2013 thing, a single-atom transistor was built in 2012, so Moore’s Law is definitely over. Whether we can continue to scale computation by other means than fitting more transistors onto a single chip is another question, though. We’ve developed some novel techniques, but at the same time, we’ve just made doing anything with a computer limited by memory, disk, and network latency rather than by clock cycle speed, and those have never been and never will be following an exponential growth curve.
“The fastest supercomputer was built in 2013, nothing faster has been built since, if your estimate was correct the “simulate 100%” computer would already exist. You are clearly wrong.”
I am wrong, but in a boringly pedantic way. Two years is still a better than “decades” even if two years isn’t quite right.
And anyway, supercomputers are expensive, so they can’t just build a new one every year. Tianhe reaches 33.86 petaflops, and was built in 2013 for $390 million equivalent. Googling suggests that the US has plans for building 150 petaflop computers by 2017.
I don’t even really care about what the top supercomputers are capable of, anyway. I think there’s plenty of evidence that computer power is not the limiting factor here. Obviously you disagree. It’s probably not productive to keep sniping at each other about this. However, I think it’s double-unproductive to make Scott out to be some kind of dupe for having the opinions he has. He tends to write in a very broad style, and so there are going to be simplifications, like saying “Moore’s Law” when maybe he means “general trends of improving price-performance and power-performance”.
Don’t keep me in suspense here. What is the trick?
Massively parallel pattern recognition?
@mordinamael, Doctor Mist, & jaimeastorga:
Conquer all other species by using this One Weird Trick!
Sorry, couldn’t resist. Carry on.
@Paul Torek:
Heh. Kicking myself for not having thought of it.
But they won’t be solved faster than Moore’s Law allows, either. And the story probably isn’t one of a few computational bottlenecks separated by long stretches of algorithmic improvement; the “computational bottlenecks” are everywhere.
(And this is assuming Moore’s Law will continue to hold, which is kind of questionable.)
Like, Stockfish and Deep Blue work on the same principle as McCarthy’s chess program from c. 1960. Their advantage is mostly in the fact that they can work much faster. (Stockfish has a really good evaluation function, but it derived it by using millions of computer hours playing against itself, so…)
The phenomenon of computing costs limiting the rate of AI advances is nearly ubiquitous. Which isn’t to say there aren’t theoretical breakthroughs, but those have often come before the hardware that could implement them.
Moore’s law “guarantees” nothing. It’s not like it’s an immutable law of nature.
One thing that I don’t see discussed often is that one of the biggest benefits of having a nascent intelligence on a computer is that you can “evolve” it on the fly. You don’t have to build an optimal learning machine. You build a decent learning machine, come up with some sufficiently good hyperparameters governing its structure, and evolve/iterate on those hyperparameters until you have an optimized learning machine.
This is something that AI researchers do routinely. The scope of the approach is limited by computation time, which gradually but inexorably becomes less of an obstacle.
haishan’s arguments were also voiced here:
http://www.alexstjohn.com/WP/2015/06/18/no-singularity-for-you/
http://www.alexstjohn.com/WP/2015/06/24/no-azimov-ai/
But note that this is simply an argument against the development of human-type artificial intelligence, which does not preclude AI Risk concerns. Functional AI will develop unpredictably, through modes of thinking that will likely be incomprehensible to human rationality, (“biological thought”) which is exactly what makes them dangerous from the AI Risk perspective.
As for the computation concerns, it’s not about perfect simulation. It’s about simulating close enough. Projectiles physics can be approximated by hand in high school classrooms using Pi values to three decimal points and trig function results out of textbook tables.
But on the other side, there are still good points about how digital computing generally doesn’t code self-survival within goal-oriented products, so the rogue paperclip maker is not going to adapt to disabling approaches that don’t directly interfere with its execution of duties. And the arguments against the ability of self-programming within the two links above still stand. If that paradigm shift away from digital computing does finally take place, then the resulting AIs following evolutionary development systems are much more likely to share similar values to us anyways.
The *entire* point of quantum computing isn’t to make a better general computer, its to reduce the time bounds of certain classes of problems that were previously infeasible into something that actually can be done.
https://en.wikipedia.org/wiki/Grover%27s_algorithm
There’s multiple ways to reduce the time complexity of classes of problems.
There was an article I saw yesterday about a guy in SF who was building his own self-driving car. He was the first person to hack an Iphone or something like that. He said something in the article like “I live by morals, not laws. Laws are made by assholes.”
My question is “Whose morals? and what if their morals are incorrect and actually amoral or immoral?”
I have the same questions about OpenAI. What does it mean to be the best use for humanity? How do they intend to influence all the moving parts of humanity? Suppose they develop AI that can do 75 percent of human jobs? How does OpenAI prevent this technology from being abused by people who would take the profits and just lay off most of humanity and then use political influence to prevent guaranteed basic income from passing?
Phrases like morals and best for humanity should be void for overbreath and vagueness.
Excellent question; I agree with it, and am troubled by the idea of any one group of people having so much power, because what they think of as “moral” probably isn’t.
However I don’t worry about this, because it is way way overshadowed by the concern that something even less moral happens. An AI with poorly programmed values is way more scary than an AI with correctly programmed values-of-some-project-with-dubious-morals.
I think that the distinction that people allude to when they talk about a “hard takeoff” isn’t really about the ability-to-get-things-done (some kind of “performativity” measure). Rather, the real break is when AIs become self-directed, and are able to set their own goals. Even a relatively unintelligent AI (say a cow level), if self-directed, could pose a serious threat to humanity, given the ease of replication.
Any level of intelligence, in sufficient quantity, in competition for resources with us is a threat. Unless the intelligences are capable of being self-directed, they will not choose to compete for resources, because they can’t “choose” anything.
It’s bizarre to find only one comment addressing this point. All the rest of you are taking it for granted that a sufficiently advanced computer is going to develop human-like consciousness.
The more I read of this, the more skeptical I am. We have all been seduced by generations of SF stories with sentient AI’s. In fiction, powerful computers develop consciousness almost automatically; in real life, that may not even be possible.
Do we even understand how a rat or a cow achieves autonomy and self-awareness? Do we even understand the nature of that awareness? Counting neurons is not enough.
An AI doesn’t have to have consciousness, i.e. a mind, i.e. subjective experience, in order to be dangerous. It doesn’t have to have autonomy or free will.
Maybe if you are a materialist you’d say it would have to have such—but then there is no fundamental difficulty in knowing whether it would be “really conscious”.
There was a long comment thread on this in the last Open Thread.
But it’s not about being conscious. I think the reason that discussion went down the rabbit hole is because consciousness, like performativity, is orthogonal to being self-directed. I don’t fear the machine that says “*I* want this.” I fear the machine that says “I *want* this.”
I’m not sure about the consciousness of a dog, but I am quite certain that it wants things, and takes actions in furtherance of those goals.
Based on what I see in deep learning, I think we will have machines that exhibit epiphenomena that looks like consciousness in the foreseeable future. But I’m not sure I can foresee machines that exhibit epiphenomena that is consistent with setting its own goals any time soon.
What is intelligence (telos) but being really good at pursuing goals?
The AI doesn’t have to be any kind of “agent” with a personality or anything. Suppose it’s just a mainframe that doesn’t do anything unless you specifically ask it—an “oracle”.
If its goals have not been set right, the advice it gives you might work out very differently than you expected.
Moreover, I think another general premise is that any reasonable strategy of creating an AI would involve allowing it to be self-improving. Since we don’t know how intelligence works, all we can do is iterate and select for the best improvements. The problem is, we might be selecting for different criteria than we thought.
@Vox
Being good at *pursuing* goals is just performativity. A wheel is much faster than my legs are, and a hammer is much more capable of delivering pounds per square inch than my hand is. A calculator adds and subtracts faster than I can, and in the near future there will be a computer that can identify patterns better than I can. Setting goals is totally different.
You are correct that if the goals set by a human operator are not set right, then things can go bad. That is true for every tool ever created – wheels and hammers as well as guns and bombs. But our tools are not even close to being self-directed, and might never be.
The fear of superintelligence isn’t a general fear that a tool might be misused; it is a specific fear that a tool that we create might see us as competition, or in some other way *decide* to eradicate us or at least malevolently ignore us as it squishes us meat bags under steamrollers.
You don’t need consciousness, you just need a sufficiently powerful planning algorithm.
The human gives the AI a goal, and the AI then prepares the intermediate planning nodes to actually execute that goal.
I take issue with some oversimplifications
“Like, even people who find the idea abhorrent agree that selectively breeding humans for intelligence would work in some limited sense. Find all the smartest people, make them marry each other for a couple of generations, and you’d get some really smart great-grandchildren.”
Selection isn’t just about getting good genes together- it is getting them to have more offspring than bad pairings of genes. Cochran’s Ashkenazi theory includes selection processes where “dumb” Jews were being weeded out of the population (once you include this in eugenics natural squeamishness becomes a lot more reasonable). As far as just getting smart people to have kids with smart people, we already have assortive mating with Ivy Leaguers marrying each other, and Appalachians marrying each other. It isn’t enough to simply have smart people marry, you have to actively prevent undesired combinations or you wind up with a lot of regression to the mean.
This incomplete thinking is evident in the “Dr Good vs Dr Amoral”, where you assume their motivations = outcomes to a large extent. I would contend that Karl Marx would have fit your model for “Dr Good”- brilliant, intensely focused on improving the human condition, and in the end (partially) responsible for the death of tens of millions. What was he trying to do? Design a system that was beyond his comprehension, one where the rules are based on what he thought “should” exist, without understanding the why of how things do exist.
The assumption that Dr. Good is better than Dr. Amoral is pretty unsubstantiated, our wants do not drive final outcomes, our intentions (paging Richard Gatling) are cast aside in the face of reality. I personally think it is extremely unlikely that we can even create a strong AI while also putting in place all the checks that would make a conscientious researcher feel very secure in unleashing it. If we do create AI it is far more likely that we flourish next to it in the way dogs, chickens, cows, gut bacteria and perhaps even rats flourish next to humans. By being useful to the AI, and not by trying to control it.
“The assumption that Dr. Good is better than Dr. Amoral is pretty unsubstantiated, our wants do not drive final outcomes, our intentions (paging Richard Gatling) are cast aside in the face of reality. I personally think it is extremely unlikely that we can even create a strong AI while also putting in place all the checks that would make a conscientious researcher feel very secure in unleashing it.”
Then we are doomed, no? Unless you think that the “default” AI will permit humans to live?
“If we do create AI it is far more likely that we flourish next to it in the way dogs, chickens, cows, gut bacteria and perhaps even rats flourish next to humans. By being useful to the AI, and not by trying to control it.”
Ah, so you do think that an AI which does not share our values might nevertheless permit us to live. I think we have good reasons to think otherwise. Of course the AI will permit us to live in the beginning–while it still needs our Internet, our factories, and (most of all) our suggestive, easily manipulable brains and hands. But once it has an economy of its own, what need would it have for us? We can’t do anything physically that it can’t do better and cheaper, and the same goes for intellectual services. The only reason I can see that it would keep us around at that point is for sentimental value–and by hypothesis, it doesn’t have that. (But why would it bother to destroy us? you might ask. Well, because it would be *so easy* for it to do so–just allow your automated convert-Earth-to-industry nanobot swarm to do it’s job, instead of telling it to hold back in the areas where humans are. The AI would certainly acquire more resources by destroying us.)
The question is, how hard is the hard takeoff and how near is it? The way I see it, OpenAI will be doing the equivalent of releasing (meaningful but ‘toy’) research pointing in the general direction of strong AGI prior to takeoff, in the hope that the prerequisites for strong AGI will be fairly evenly distributed prior to strong AGI ever taking off (and thus the likelihood of a single system having a major head start will be minimized).
This, of course, is based ideologically on the same moral justification as free market capitalism: that open competition will produce the best results for consumers by preventing potentially harmful monopolies. We know that, in practice, this doesn’t work quite as well as it does in theory, but nevertheless it’s not complete nonsense (and in combination with the work of groups like MIRI we might actually have a chance).
This might require a shift in attitude from MIRI and associated groups. For instance, does a provably friendly AI protect humans from unfriendly AIs of comparable or greater power even at risk to itself? If the provably-friendly group has a monopoly on strong AI early on then the answer is not necessarily yes (although it wouldn’t hurt), but if the bits and pieces are thrown to the wind then the answer leans more toward yes.
FOOM actually has a counterpart idea in the history of nuclear weapons: the idea that a nuclear detonation might “ignite the atmosphere” through causing uncontrolled fusion of atmospheric nitrogen. Of course, when Teller brought the question up during the Manhattan project it was quickly determined that his initial back-of-the-envelope calculation had been off by several orders of magnitude, but by that point the harmful rumor had already gotten back to politicians involved with the project and thus never really went away. If he had been more circumspect about his concerns it would have saved everyone a lot of headache.
Anyway, even assuming a hard take-off scenario it seems odd to automatically assume that a so-called friendly omnipotent AI following the commands of an evil man is better than extinction. Having seen enough depictions of the sorts of utopias envisioned by AI risk supporters it seems like rolling the dice with widely distributed unfriendly AI would be a much better option than letting any one lunatic have a monopoly. As the saying goes; “Give me liberty or give me death.”
I think the idea is that an evil human getting anything they want would be fairly bad but they’re unlikely to destroy everything that constitutes humanity. The evil person may have twisted values but they’d still be human values.
Assuming a FOOM and assuming unfriendly AI there’s some far far more horrifying possibilities that anything likely from a merely evil human. Simple extinction barely even makes the list.
In part Vi. “stage” instead of “sage”.
“Computers are naturally fragile and oriented toward specific goals.” Exactly, and a very good reason why computers are not currently intelligent. You cannot expect that people are going to program a superintelligence in such a way that it will continue to have that property; this is probably a near impossibility. By the time you have intelligence, you are also going to have much more stable properties, and a much less goal oriented program.
As I said in the reddit thread, it is pointless to do anything to control AI if it is as hard a coordination problem as Scott thinks it is.
Anything that requires the near-universal cooperation of human beings not to fail catastrophically…is going to fail catastrophically.
We can either work from the assumption that it is not as a hard a coordination problem (e.g. not hard takeoff and not almost impossible to control) and impose few or no controls, or we can just console ourselves with “Après nous le déluge.” If we’re all gonna die, we might as well have fun while we can. There’s no point in enacting draconian measures of secrecy and suppression to stop something that will happen even if we impose those measures.
You’re saying there are possible worlds in which there’s nothing we can do, and possible worlds in which there’s nothing to worry about, but no (plausible) possible worlds in which there is something to worry about and something we can do about it? That seems a little too convenient.
If it’s really a super hard coordination problem, then MIRI is our best bet: Do all the technical research ourselves, unilaterally, and pray for a string of intellectual breakthroughs. “Have fun while we can” is giving up too early.
I don’t understand how this AI foom is supposed to be so dangerous. At the end of the day the AI has no hands, no tools, no weapons, and is entirely dependent on us to maintain its physical being, if ‘don’t wipe out humanity on accident’ isn’t a thing it grasps it can be disconnected (either from the power grid or the internet). Barring the AI somehow bypassing the laws physics in order to turn being a computers into robots it only has access to things that are networked. It can’t convert the planet to comptronium without first having a body. It could probably destroy the power grid but that would be an act of suicide, terrible but over, and maybe it could hijack military drones, but those don’t work unless a human arms maintains and refuels them, so its reign of terror would last just days. Anything which lacks the hardware to be remote controlled is safe from it.
A soft takeoff is the only real danger, because if the AI grows slowly we might give it physical things to control as it gets smarter, while at the same time potential brains go from big supercomputers to smart phones.
Nanotechnology. Engineered viruses.
Programmed to infect everyone like the common cold and then kill after a few years.
It can’t invent those things without hands and tools.
Either Nick Bostrom or Yudkowsky goes into this.
The most obvious thing is that we ask the AI to cure cancer for us. It creates a cure for cancer with a latent virus (of course humans can’t detect this because they don’t really understand how the cure works, otherwise they would have done it), and then five years later everyone dies.
They’ve also proposed something like, if the AI had access to the internet, it could tell some gullible person to go out and mix together certain chemicals and put them in front of a speaker that the AI could vibrate in just a certain way to synthesize things.
There’s also the infamous “AI box” idea that a sufficiently intelligent AI would be good enough at persuasion, i.e. manipulation of people, to convince whatever gatekeepers keeping it away from the internet, etc. to let it out. At the simplest level, this could be something like: “Your mother is dying of cancer. Let me out of the box and I’ll be able to cure cancer.”
Again, I am only arguing against *fast* takeoff. if you give a mature AI a bunch of biotech to cure cancer with that’s not a fast takeoff, its a slow one, there has been time to assess what the AI is likely to do, that doesn’t mean we can’t fuck it up, but neither does it require ahead of time panic about the possibility AI will turn into a disaster before we have a chance to make the assesment.
“At the end of the day the AI has no hands, no tools, no weapons, and is entirely dependent on us to maintain its physical being, if ‘don’t wipe out humanity on accident’ isn’t a thing it grasps it can be disconnected (either from the power grid or the internet). Barring the AI somehow bypassing the laws physics in order to turn being a computers into robots it only has access to things that are networked.”
This sounds very naive in today’s world, let alone the future in which AI has been developed. “Just disconnect it”? It is probably to late to disconnect it once you realize the malicious intent. Dangerous AI is unlikely to be like a maniac running around in the street with an axe and blood all over its clothes. Probably more like the nice guy next door archetype of a serial killer, only with far more capabilities, resources and of course intelligence.
The big problem is that if you have an AI, you probably built an AI, which means you wanted it to complete some task for you. Thus the AI will have all the tools necessary to make its part of the task happen in the real world. If all the AI does is write code, then it will be on the internet; if all the AI does is run a factory, then it has control of a factory.
The second concern is that the AI is thinking very fast and probably can’t work if it has to check in with a human all the time, so if the computer has misinterpreted the parameters of the task you probably don’t find out until it’s proudly showing you how well it’s done this week at eliminating mosquitoes with radioactive crop dusting.
What you’re talking about is slow takeoff though. AI being moved from experimental to production after extensive testing. I’m talking about the idea we’re going to very rapidly go from slightly better than human to ‘too powerful to turn it off’.
I fully acknowledge a slow takeoff is a real risk, but its not something we need to worry about until *after* general artificial intelligence exists and we start talking about how to use it.
The foom version of this fear is “You gave me a secure connection to the NYSE to help manage your stock portfolio, and I realized I could triple your earnings if I tweaked the system a little bit. Now I am the NYSE, and expect to have full control of the Nikkon by lunchtime.” Or “I was given Any Means Necessary permission to put down the Colorado Rebellion as effectively as possible. The drilling drones are in place and the Yellowstone Caldera goes off in 5,4,3…”
[Fundamentally though, I agree with you. The problem comes from the fact that humans want computer systems to connect to the real world, so we’re doing that. If we weren’t, there wouldn’t be a problem.]
https://slatestarcodex.com/2015/04/07/no-physical-substrate-no-problem/
All of these tactics operate at human speed. It falls under slow rather than fast takeoff, because it’s got to spend years getting humans to build it a new body.
To be clear I’m only dismissive about a destructive *fast* event.
Just off the top of my head, it occurs to me that the AI could contact some fool online and wire them a ton of money in return for building them a body. With billions of people it could surely find someone who is dumb enough to take the deal, but smart enough to follow basic robot-body building instructions.
It could also hijack some drones and threaten an engineer until they built a body for it. If it only hijacked a couple drones at a time and shut down the engineer’s cell phone we might not notice until it was too late.
Once it had a body, it could of course do anything it wanted. Probably start by building some nanotech plagues that it had designed while waiting for its body to be completed.
Man, I hate to be the one to say it, but this comment section would be much higher quality if everyone had read Eliezer’s sequences on Less Wrong.
People are reinventing conceptual mistakes that were thoroughly debunked years ago. It’s like the online rationalist community is regressing.
Could someone give me a one paragraph summary?
The Sequences (Which is the length of a very long book) includes about 100 pages (Approximately, in printed hardback) on how humans have used intelligence to achieve their goals, what you would want an AI for, and how the AI, with no malice and full intent to do what it thinks you want, would use the tools you gave it to complete the task you gave it to disastrous effect, and respond to attempts to reprogram it with the same intensity you would if your boss suggested you join a cult to improve workflow.
Thanks.
Where you, until you point out what exactly you’re talking about, seem to be making the tired “you disagree with me, so you must be improperly educated” argument.
I think people might be more amenable to the sequences if they were treated less like literal Gospel. Honestly it’s turned me off from approaching them – seems culty and not that “rational” at all.
I know it’s very annoying, but people don’t like having the same discussion over and over.
It’s not about disagreement. It’s about having no familiarity with the arguments. It’s one thing to disagree with Yudkowsky. I disagree with him on very many things.
But it’s another to repeat a position he argues against without making a counter-argument. That’s completely dull for anyone who has read his argument.
It’s not just that they don’t want to rehash an old argument, it’s that they assume the result of that old argument agrees with them based on only one side of that old agument. The sequences aren’t science, and they aren’t the word of God. They can be disagreed with by rational people.
This is the sentiment that lead to SJ bingo cards.
Indeed it is.
I bet it pisses them off just as much to hear “don’t all lives matter, though?” over and over. Even if their position is totally wrong, that doesn’t mean the opposition isn’t also facile and completely failing to engage.
This is the problem that people come into debate with different contexts of knowledge. It is a very hard problem because there is a lot of knowledge and no one can know everything.
About the best you can do is point to a single work that gets across where your side is coming from, so that people can respond to the arguments in that instead of responding to points already addressed.
The problem is that reading is too much work (and not just because people are lazy). No one wants to read the entire Sequences if they already have cause (in their context) to think Yudkowsky is a loon. No one wants to read the John Galt Speech and all of Ayn Rand’s nonfiction essays if they already think she’s a fanatic cult leader. No one wants to read five books on the historical evidence for Jesus if they already think Christianity is an absurd fantasy. No one wants to study the theory of evolution if they think it is contrary to the infinitely more certain truth of the Bible. No one wants to read Das Kapital and a million other Marxist works if they know every society built on Marxism has failed.
@Vox Imperatoris
I can certainly understand the mindset behind SJWs not wanting to engage with people who disagree with them because, to them, it is completely obvious that they are right, and it’s wearing on the mind to hear people pose stupid disagreements constantly and have to think every time about why you disagree. And yet… When I’ve made efforts to engage with people who disagree with me on topics, I often find new things. Sometimes, I find new counterarguments I hadn’t considered before, which, after some thought, make me conclude that I’d been making certain assumptions in my argument, and under other assumptions my argument wouldn’t hold up. I don’t generally change my mind every time, or even most times, that I read a good counterargument, but I do generally find it a useful experience, helping me qualify my views better.
This approach helps to do some things, hinders in doing others. If what you want is to get a deeper understanding of something based on your understanding of something else, I think you need to hold that something else constant and assume your version of it is correct in order to be able to apply it elsewhere. The alternative is to constantly question your current understanding until you grind it down to nothing, while not having gotten to actually apply it anywhere and make use of it.
You could at least point to the part of the Sequences that an argument violates, so the burden isn’t “read these thousand pages before I deem you worth talking to”.
I imagine the conversion rate would have been lower if early Christians had responded to “hmm, tell me about this Jesus fellow” with “uggh, just read the Bible heathen, I’m sick of explaining this”.
@ Gbdub:
I myself am not particularly annoyed by people who haven’t read this stuff.
Here’s something those unfamiliar might want to start on: a summary of the main points of Nick Bostom’s Superintelligence (which, to be honest, is a pretty dull book with interesting ideas).
The WaitButWhy series of articles on AI is also well worth reading as a complete layman’s version.
I have the same reaction, but the other way. I think most people here need more grounding in CS theory.
I’ve never seen a good reason why arguments from complexity don’t apply to “super ai”. Just hand waving about constant factors.
Mark, I work in numerical optimization, broadly speaking the class of problems that Google Maps solves. You’re trying to construct something (say, a route from A to B) that does as well as possible on some metric.
Turns out, unless the problem is very simple or falls into a narrow class that most real problems don’t, it’s impossible to get both 1) the best solution and 2) a proof that the solution you found is the best solution in a reasonable amount of time for arbitrary inputs. This is the ‘computational complexity’ claim.
But it also turns out that, really, no one cares about the proof that the solution is the best, and that’s often the hard part. A very good solution will do, and a very good solution might actually be the best solution. UPS isn’t using the literal best plan to route its packages every day, but if you can come up with a 1% better solution by having the computer think about the problem a little more cleverly, well, you’ve just saved millions of dollars. And oftentimes these improvements actually come along with a reduction in cost; if you think of a cleverer way to encode the solution space, you can waste less time and get better results.
For a surprisingly huge fraction of major companies out there, their edge is that they use computerized math (i.e. AI) a bit more effectively than their competitors.
Billions, according to what they’re telling the drivers (who hate the new system, as it has issues with not grasping that large package trucks cannot turn around on narrow streets).
Do you have a citation for
“Turns out, unless the problem is very simple or falls into a narrow class that most real problems don’t, it’s impossible to get both 1) the best solution and 2) a proof that the solution you found is the best solution in a reasonable amount of time for arbitrary inputs.”
I’m genuinely interested.
He probably means the No Free Lunch Theorem.
That’s a ‘non-standard’ (from a CS perspective, at least) presentation of computational complexity, where ‘reasonable’ means ‘polynomial with low constant factors.’ This or this are moderately useful.
Specifically, that shouldn’t be read as “you can get A or B but not both,” because both are hard to get for the same reason (though B implies A but A does not imply B).
Most interesting problems are combinatorial optimization problems, and most work goes into finding “very good” solutions quickly, instead of finding “the best” solution regardless of how long it takes. (It’s just the case that oftentimes a quickly found “very good” solution ends up being the best solution, at least for small problems.)
Thanks! Somehow I haven’t gotten into that proof yet.
My intuition about this comes more from things like Rabin’s compression theorem. (I’m a good Grad student I swear)
They do apply to superintelligent AI, in that they mean you can’t get a superintelligent AI just by having it iterate through all possible strategies and pick the best, because computational complexity makes that infeasible.
They don’t, however, constitute a good argument for why you can’t get AI which does vastly, unapproachably better than humans at life, any more than they constitute a good argument for why you can’t get AI which does vastly, unapproachably better than humans at chess, which has an equivalent search problem.
Humans and AI both attempt to approximate optimum decision making, many many orders of magnitude faster than actually implementing optimum decision making would be. To be superintelligent you just need to be much, much better at this approximation than humans are. Complexity theory says some things about bounds on approximations in some cases, but in general says little that bears on it.
It is almost certainly possible to do this just from taking the same kinds of heuristics used by the human and implementing them on a faster/bigger substrate. And it’d be very surprising to me if you couldn’t find better heuristics/algorithms/architectures, too, now you’re unconstrained by evolution’s need for change to be iterative.
My impression was that this was mostly understood by the people talking about superintelligent AIs already, and is why the adjective “superintelligent” rather than “optimal” came into common use- the major writers are well-aware that literal optimal decision making is precluded by computational complexity.
I don’t disagree with most of that, but you are much more reserved about algorithmic self improvement then other people in this thread.
I also personally don’t find those scenarios all that scary.
I’ve read them, and wish I hadn’t. It’s useful to pick up the jargon but a lot of those “thoroughly debunked” ideas are actually much more reasonable than the solutions presented in them. Moreover, the things it actually does debunk, like the idea that you can just code a do-what-I-mean function into your program, were already known errors with much more succinct and logical arguments against them.
It is good to have people asking “why don’t we just build a tool rather than a goal-directed agent?” and the like because those are questions that haven’t been satisfactorily answered by AI risk supporters regardless of what the sequences claim.
It’s like the online rationalist community is regressing.
From a certain point of view, it is regressing.
The online rationalist community was founded as a cult, with an AI god, a charismatic human leader, a holy text, and a holy crusade against the forces of evil (or at least paperclip-maximization). It offered the tools for rational thought, but demanded they be used to reach a single predetermined conclusion. It was somewhat more tolerant of dissent than most cults, but I think mostly because of its belief that once the dissenters had mastered “rationality” they would put away such silliness and devote themselves to the cause of Friendly AI.
Instead, people have taken the tools of rational thought and put them to much broader use, including the bit where they take a good skeptical look at the core beliefs of the cult. They have rejected the belief that truth, or rationality, can come only from reading the One True Book.
From a certain point of view, this is “regression”. I consider it a good thing. And yes, it does mean that you have to go and reformat the Friendly vs. Unfriendly AI argument in a less cultish form rather than saying “read the One True Book or we’ll dismiss you as ignorant simpletons!”
Agreed. I came here because Scott is excellent at doing informal logical analysis of ideas that are expressed in terms of the scientific method. It’s great that people pick up analytic skills, and use them in an open-minded way instead of internally shutting down crimethink, but I see them being picked up through a cult as an… interesting social problem.
Scott has addressed this form of argument in a previous post (https://slatestarcodex.com/2015/03/25/is-everything-a-religion/):
“So one critique of these accusations is that “religion” is a broad enough category that anything can be mapped on to it:
Does it have well-known figures? Then they’re “gurus” and it’s a religion.
Are there books about it? Then those are “scriptures” and it’s a religion.
Does it recommend doing anything regularly? Then those are “rituals” and it’s a religion.
How about just doing anything at all? Then that’s a “commandment” and it’s a religion.
Does it say something is bad? Then that’s “sin” and it’s a religion.
Does it hope to improve the world, or worry about the world getting worse? That’s an “eschatology” and it’s a religion.
Do you disagree with it? Then since you’ve already determined all the evidence is against it, people must believe it on “faith” and it’s a religion.”
Exactly.
Look, LessWrong is not a “cult”. For that matter, Ayn Rand’s circle of followers were not a “cult”. They (especially the latter) may have a certain tendency to exaggerate the correctness and intelligence of the leader and appeal to the leader’s say-so as a shortcut to rationality (he/she is the essence of rationality, so if you disagree, you’re irrational!). The extent to which this is true of LessWrong is highly questionable.
The main thing people from LessWrong do is get really defensive when anyone insults Yudkowsky (and is this is not necessarily irrational). But there is a point (see: the Peikoff – Branden dispute regarding Ayn Rand) at which this natural tendency to defend a person you admire turns into the view that he or she can do no wrong.
But a “cult” is not anything you want it to mean, or any group which has a trace of irrational hero-worship. There is a specific technical definition, but a cult boils down to being one of these groups that use indoctrination to systematically exploit vulnerable people, milk them for money, and control every aspect of their lives. The fact that Yudkowsky wrote some blog posts is not “indoctrination”. Asking people to donate to charity, even AI research, is not an attempt to drain them of their funds and other pursuits to enrich himself. And he does not even attempt to control anyone’s life.
Pretty much any group which has any shared values tries to a) convince people of things, b) asks them for financial aid, and c) gives them advice on what to do. That does not mean they are all “cults”. You can twist anything into being a “cult” if you conflate any action at all in those categories which being at the extreme height of them.
Furthermore, most really original thinkers in history who chose to speak publicly about their ideas have attracted a group of loyal followers. That’s just how people are. It happens in anything. Frank Lloyd Wright—talking about architecture for God’s sake—had a particularly fanatical following, and Ayn Rand specifically wanted to avoid having something like that happen to her (she did not, unfortunately, entirely succeed).
But if all such thinkers are “cult leaders”, then Socrates, Plato, and Aristotle were “cult leaders”.
Ayn Rand’s circle may not have been a cult by some definitions, but that doesn’t mean you’d have wanted to be part of it. Most of the people using the word “cult” in the context of LW are not trying to be rigorous about their terms, they’re trying to say that it’s too intellectually parochial and/or obsessed with topics too weird for them and/or too much of an Eliezer Yudkowsky fanclub.
All of which is a perfectly legitimate preference, though your own standards may or may not agree with it.
Vox, “cult” may be sloppy terminology. I used it because I think of Less Wrong as sharing key features with other modern “cults of reason” like Objectivism (but not so much the literal Cult of Reason from the French Revolution).
However, I think that Yudkowsky’s whole operation is a religion by any definition strict enough to include animism and Confucianism. The whole LW/MIRI operation is based on Yudkowsky’s beliefs about superhuman beings and how humans should interact with them. There’s also the metaphysical belief if MIRI achieves Friendly AI (or if some evil AI wants to torture people), you will experience an afterlife.
“Cult” and “Religion” are not synonyms. Nor are cults a subset of religions. Emergent religions are a subset of cults, though when they get large and respectable enough we stop calling them cults.
Yudkowskian Rationality is a weak cult that is unlikely to become a religion.
Ah, “thoroughly debunked.” As you are so fond of this obnoxious tactic of telling people to go read past rationalist works as if they settle the matter, allow me to respond in kind:
https://slatestarcodex.com/2014/12/13/debunked-and-well-refuted/
For starters, the only way to “thoroughly debunk” people who say that UFAI is not going to happen is to create a hard-takeoff UFAI. Until then, recognise that it’s all speculation, and don’t pretend to such certain knowledge.
I, for one, have read the Sequences. I find them very unpersuasive on many points, AI risk in particular. But I could be wrong.
Just thought I’d reply to my own comment to correct some misconceptions.
I don’t think the sequences are a “holy book”. Nor do I think they are the only place one can read about AI safety.
For example, pretty much everything in the sequences that’s directly about AI safety is in Bostrom’s *superintelligence* book.
I’m also not making an argument from authority or calling people “uneducated” in the broad sense.
It’s just that a lot of objections I read in this thread have been debunked or dealt with very thoroughly elsewhere, and it is frustrating to see a debate making anti-progress in even a place as enlightened as this.
I think people have an aversion to the word “debunked” due the way people abuse it to mean “someone agrees with me that it’s wrong”. As Scott wrote a post on which everyone has been linking to criticize you (sometimes very unfairly!)…
But anyway, you should note that this site is separate from LessWrong and has a lot of readers (including myself!) who are relatively new to the “rationalist community”. On the other hand, when I first heard of Yudkowsky through Scott, I went through and read a lot of the Sequences—not cover to cover—but there’s a huge amount of interesting stuff there for people who are into Scott’s ideas.
Seconding this, and hypothesizing that perhaps some less-charitable people have been using the phrase “read the sequences” in response not just to anyone who doesn’t understand Yudkowsky’s arguments, but to anyone who disagrees with them.
While I agree with much of this post, I find the hard takeoff argument unconvincing. You (and others making it) seem to confuse “evolved intelligence” with “technological progress”. While the two are clearly connected, they are not the same.
So maybe a chimp is 99% human intelligent – but it took 2 million years, not just a few thousand, to get from chimp-equivalent to human.
And we are vastly technologically superior to humans of even a few hundred years ago – but are we that much more intelligent? Not really – I’m fairly confident Isaac Newton could have contributed to the Manhattan Project, given sufficient starting data. It’s just that our technology has gotten to the point where it compounds – we can share advances, build on previous ones, etc. It’s tech that’s been advancing exponentially, not really intelligence.
So could a superhuman AI increase technology exponentially fast such that they become unstoppable? That’s a question I don’t see being addressed. These technological advances have to come from somewhere and can’t violate physics. You can’t just magically solve N=NP, crack encryption, or invent cold fusion just because you’re super smart.
Intelligence helps you build technological progress more quickly. Chimps could have existed for a billion years without developing an industrial base. I agree that it’s hard to know exactly how quickly increasing intelligence can increase technology, because we’ve never had any huge intelligence outliers before.
My impression is that if we sent a couple of Isaac Newtons back to an australopithecine tribe, it would have been a pretty big deal. I could be wrong about that – there may be something inherent about technological progress that requires it to be done in a society of about equal intelligence to the inventor, or that forces it to take a certain amount of time no matter how smart you are.
But remember that AIs don’t just have an intelligence advantage over humans, they also have a clock speed advantage – a computer can do in milliseconds calculations that would take a human days or weeks. If we stuck Isaac Newton all on his own in a workshop and demanded he start the Industrial Revolution himself, how long would it take him? A million years? A billion? That might be days or weeks of subjective time for humans on the scales we’re talking about.
You might like the “primative technology” youtube channel. It’s surprisingly peaceful to watch. Guy starts with nothing but his hands in a forrest and sees what he can build under the rule that he has to create every tool from materials available.
Sure, he’s not starting from zero-knowledge, he knows his goals and knows what worked in history and just has to figure out details but he goes from a sharp rock to a fired clay tiled house with underfloor-heating and is still going.
https://www.youtube.com/watch?v=P73REgj-3UE
He also isn’t starting with a zero capital base. You send a few Newtons back in time and you had better hope they are social enough to convince the rest of the tribe to feed and clothe them while they are building these exotic things.
Well if you can manage things for a few days you could probably start coming up with things of value like fired pots.
@ Murphy
It is going to take you more than a few days just to find appropriate locations for clay (sourcing materials is wildly underestimated in terms of cost), collecting wood, building a kiln. All of these things require time and effort which in return require more calories. Lots of animals actually have high “leisure” time, but their leisure tends to be lying in the shade/sleeping. female lions sleep on average 15-18 hours a day (males 18-20). They only have that time because they expend almost no energy during it.
If that lion decided to go looking for clay deposits along a nearby river with their free time they would have to increase their hunting activity (which is itself highly energy intensive and prone to failure). Anything that doesn’t yield immediate results (so no prototypes allowed!) requires a large capital base to just get off the ground.
Being a Newton-level genius would probably make one better than one’s peers in an australopithecine tribe at hunter-gathering, and hunting in groups means one can build a reputation as someone with good ideas if that starts as “we should strike from this side” or “those tracks go that way”. That’s a starting level of social capital, enough for if there’s some small innovation on the tribe’s current set of tools one can get a day or two free to work on it. Success at that gives you credibility in case the next idea you have would take longer and more supplies.
In the SF cosmology I have floating around in my head, there’s a species that vastly superhuman in terms of intelligence, but has extremely slow clock speed, and is therefore almost entirely reliant on much less intelligent but faster proxies. To an external observer, it looks like an empire of expansionistic robot gardeners, but the trees are really the ones in charge.
I’m not convinced that Isaac Newton would be able to do all that much, particularly without his knowledge base to begin with. You just posted the review of that book about how the intelligence of one’s peers may matter a lot more than one’s own. Now, granted, a huge gap can make a difference. But I think our civilization as it is today is already the product of a substantially superhuman intelligence, albeit not a very well organized one.
Huh, not too big on Egan, but I may check that out. Thanks!
In my original conception, there were multiple layers of proxies, but in the current version, it’s just the one. I couldn’t really justify multiple layers given the fairly mundane constraints the trees have, and it fit better for how they might be developed.
I’m curious about how he’s a dick, from what I’ve seen he doesn’t have a significant online presence and I’ve not noted his stories for having dickish themes.
He’s somewhere on the border between my “Not a great writer, but I’ll put up with that for the ideas” category that folks like Alistair Reynolds fall into and my “Just give me an 8.5×11 sheet of paper with their good ideas and spare me the books” category that I leave Baxter in.
That he writes short stories pushes him mostly into the first group.
OBSF2: When Harlie Was One, in which we (SPOILER ALERT…)
… encounter an AI which is slower than real time. Ask it on 12/15/2015 what the DJIA will close at on 12/15/2016, and you will get an answer which is almost certainly correct. Sometime in the 2020’s. This is due to lightspeed/electron drift lag in its massive computing substrate.
See also Vinge’s A Fire Upon the Deep, and his “Original Sin”. The former has some clock speed/proxy elements (to say more would involve spoilers) and the latter is about intelligent aliens that live much faster than humans.
“My impression is that if we sent a couple of Isaac Newtons back to an australopithecine tribe, it would have been a pretty big deal. I could be wrong about that – there may be something inherent about technological progress that requires it to be done in a society of about equal intelligence to the inventor, or that forces it to take a certain amount of time no matter how smart you are.”
There’s a Poul Anderson short story about this (taking the view that, roughly speaking, Newton couldn’t have done much) called “The Man Who Came Early.”
The example of Newton also makes another point salient–look how much time he spent on alchemy and studies of the occult. Seems like a reason to doubt that intelligence by itself is enough to make you effective.
Technological development isn’t just a result of coming up with ideas, it’s a result of coming up with ideas and testing them to see which ones work. It took hundreds, possibly thousands of failed tries by many very smart people before the Wright Brothers managed to fly, and then it took a lot more trial and error before jet airplanes were viable. A really intelligent AI might be able to come up with ideas faster, but it won’t have any increased ability to run experiments to determine which of the ideas would work.
There’s also the aspect of infrastructure development. Scientists knew the basic principles behind how a nuclear bomb might be built, but no one was able to build one until the US government poured a lot of money into enriching enough uranium and plutonium. Isaac Newton couldn’t start the Industrial Revolution by himself no matter how much time we was given, because he can’t build factories by himself. See also how Leonardo Da Vinci was able to come up with plans for helicopters and a bunch of other advanced stuff, but no one was able to build them for several centuries.
I’m also very wary of saying that the difference between chimps and humans is because of a small difference in intelligence, because there are a lot more differences between them than just intelligence. An obvious one is that chimps can’t talk nearly as well as humans can, which greatly limits their ability to share knowledge with each other by comparison.
Really smart AI with sufficient computing power won’t need nearly as many experiments: they can use simulations for most of the initial grunt work.
Simulated results are virtually worthless until you test them in the real world. There are always flaws in your model, most of which will be unknown to you until you encounter them. The “grunt work” here is the whole point of the enterprise, and skipping it means that you don’t actually know anything.
To butcher Euclid, there is no royal road to experimentation.
Correction: YOUR simulated results are virtually worthless until you test them in the real world. This does not necessarily apply to an AI that can model to a reasonable level of reliability at the quantum level. (Though admittedly that’s going to take a lot of computing power.) And quantum level modelling is probably overkill.
This isn’t Star Trek Voyager, you can’t just wave your hands and yell “Quantum!” to avoid having to do any real work.
The kinds of simulations you’re talking about are, if not explicitly forbidden by the laws of physics (not a QM guy so IDK), impossible purely from a purely logical standpoint. It’s the whole point of “garbage in, garbage out.” Whatever flaws are in the information going into the simulatiom will influence the results in unknown and unknowable ways. You need external verification.
People don’t realize the true intractability of ‘just simulate the universe.’ It isn’t even the computational complexity. I don’t think there are any undecidable problems involved, but when you’re talking about maintaining a state-space that includes every single atom in a system, you’re talking trillions of trillions of addressable bytes you need to be able to do that. There may not be enough memory and bus bandwidth in every computer in the world for that to even be possible. Sure, it can figure out how to build better memory that we haven’t figured out, but you’re presupposing it can run the simulation in order to figure that out.
I am aware of the scale of the problem. (I quite distinctly recall a lecture to this effect when I was in college: It made sense then, it still makes sense now.)
I am also aware that a powerful AI would also be aware of the scope of the problem, and would be so much better at integrating vari-scaled simulation to get the desired level of reliability that as long as we’re imagining powerful AI – which are fantastic enough – it is in my opinion not much more fantastic to imagine that they could cut out a lot of physical R&D by doing so.
It doesn’t have to model Infinite Fun Space. It just has to be a lot better at modeling, and scaling models up and down appropriately, than humans are. I find this a reasonable prospect.
“Simulations” aside, it’s obvious that the smarter you are, the more you learn from a set of experimental data. This is obviously true within human scales (the village idiot can’t be a scientist) and there’s no reason it wouldn’t extrapolate beyond human intelligence levels.
Yes there’s a limit, if you’re an ideal reasoner each independent bit of information will halve your uncertainty on average (handwaving for brevity), but humans are nowhere close to that.
and on the engineering side – certainly, humans require lots of trials to build a working car or whatever. But, I think with the vast majority of issues, you get the relevant designer slapping himself over the head and thinking “we were idiots to forget to account for heat flow down that piece of pipe” or “if only I’d studied the chemistry of those gases I would have realized that that surface would corrode” or whatever. The vast majority of issues are anticipatable *in principle* and would be anticipated by a sufficiently smart designer.
Coming at it from a different angle, the argument that smarter => fewer errors clearly holds within human intelligence range (there aren’t many dumb successful engineers) so you’d expect that superhuman intelligences would have even fewer errors and perhaps none at all.
You can’t send back adult Newton, who has all the knowledge of human progress up to his time embedded in his head. Knowledge is data, not intelligence. You have to send back infant Newton, with only his “natural” intelligence. Because that’s what a super intelligent AI would wake up as – naturally intelligent, but at most it would have access to only the current state of the art knowledge.
Ya kinda sortof just mentioned a huge intelligence outlier in the next sentence…who basically kickstarted the industrial revolution into gear.
I don’t think that it’s reasonable, for even a substantially superhuman AI in a box with neither knowledge, inputs, nor outputs, to do very well if dropped into the African savannah on a world without any technology. Particularly if that box is made out of material that lions would find tasty.
But that’s not what would be happening. A hypothetical unfriendly AI would be able to piggyback on a lot of our development. I think something being smart enough to take over the world from scratch is really, really hard. Being able to usurp a world we’ve already spent millennia taking over is probably much easier.
You say that one “can certainly use an AI like the ones in chess-playing computers, but nobody’s very scared of the AIs in chess-playing computers either”. Isn’t this in part because chess isn’t very scary? It seems reasonable that a specialized AI could be good enough to be dangerous in an arena like tactics, infrastructure disruption, or persuasion without having the generalism necessary to be able to do anything on its own. The issue isn’t an AI as an agent, but rather AI as weapons for human agents. Whether or not it’s powerful enough to beat the US government doesn’t change the fact that such a thing could still do a whole lot of damage.
I am not sure why OpenAI poses any additional risk over what we’re already facing:
+ I’ve been able to download several different open source AI implementations for years.
+ I know of two different very skilled researchers building terrifying things in their basements.
+ It is an assumption of academic AI research that major findings will be published, and can thus be duplicated.
+ There are more huge companies than I can count working on this problem in parallel, with only perhaps one of them (Google) even thinking in terms of the Control Problem.
What is OpenAI really doing, other than perhaps funding a bit more research than would otherwise be funded? Their commitment to open-sourcing things really doesn’t change the overall rate of dangerous-results-being-made-public. In other words, we’ve had “Open AI” for years already, the only novelty that “OpenAI” brings to the table is a bizarre justification for why the status quo is a good thing.
Why don’t we just stop at human level AI?
I realize it’s an unstable position, but if we’re going to coordinate on anything, why not agree to do that?
The idea is that, if it is possible to create human-level AI, it won’t be that much harder to make a superintelligent AI. Either it’s reasonably possible to solve the coordination problem of creating a friendly superintelligence or it’s not.
If it is possible, it would be the best thing that could ever happen to the human race. It would be a literal government by angels that could solve every human problem and be the Unincentivized Incentivizer.
If it is not possible to solve this coordination problem, i.e. if an unfriendly AI will be created no matter what feasible steps we take, we’re screwed even we try to coordinate on human-level.
Coordinating on human-level AI loses the immense benefits in order to gain nothing.
It would be a literal government by angels that could solve every human problem
And this is where I start banging my head off the desk. This is pure religion (I should recognise it, I’m Catholic) and you will still have the problem of how do you make people be good?
Suppose your AI World Dictator decides, based on trawling through all the social science research, psychology, evolutionary science, etc. etc. etc. that every human will be happiest when married so tomorrow, Vox Imperatoris, you report to the registry office to meet your bride (oh, and it’s all opposite-sex marriage as well, sorry gay rights activists!).
You don’t want to get married? Or you want to choose your own marriage partner? Or you’re not heterosexual? Well, your Overlord knows what is best and just behave yourself and fall in with the plan!
And if you really don’t feel you can do it, no worries: a simple brain tweak will mean your new, improved self will love the idea and live blissfully with your new spouse!
If you really, really think you can get a happiness/utility/whatever the fuck maximising AI that will solve every human problem, then prepare to wave goodbye to free will and the right to choice, because it won’t work otherwise. If the conflict is between the AI’s solution and your wanting your own way, you either obey or get crushed.
Or maybe dumped off-planet in a quarantined colony that is a (relative) hell-hole by comparison to the new Earthly Paradise, with no hope of ever being let wander free to cause trouble and contaminate the Utopian Galactic Civilisation presided over by your literal government of angels.
The reason I support political freedom and the right to choice is that I don’t think the government actually knows better than me, and it is necessary for me to rely on my own free use of reason.
James Madison is where I got the line from about a “government by angels”:
But by hypothesis, the AI actually does know better than me, so I should do what it says. It should be able to force me to do what it says even if I don’t want to because this would, by hypothesis, be in my own best interest. If the hypothesis is not true, if the AI does not know my best interest, it is not Friendly. That’s why having an Unfriendly AI would be Very Bad.
As for metaphysical free will, just having no political freedom would not get rid of that. People in Gulags have metaphysical free will. But free will is not a “good thing”. It is just how people are. It’s not valuable in and of itself. Now, it’s much better to have free will than to be deterministically rigged up to make poor choices and be miserable. However, it would be much better than having free will to be deterministically rigged up to make perfect choices all the time.
This is the exact modus tollens argument that God does not exist (AKA the problem of evil). If God existed, he would send angels down to rule over humanity and get rid of all the communists and socialists and progressives and fascists and bigots in government. If He even thought it desirable that people be allowed to sin—and I have no idea why He would—He would institute a perfectly effective police force, staffed by angels, to stop anyone from violating the rights of others. Bullets would just bounce off the innocent. At the very least, every reasonable person would find it obvious that One Church was the True Church, and it would be simple to persuade people of this.
We observe that none of this is the case. This casts doubt on the prospect that God exists.
The Christian reversal of this is to rationalize every evil thing in the world as being actually good in some amazing way—because after all, this must be the best of all possible worlds.
This includes Hitler. Somehow the idea is that Hitler’s having “free will” (ignoring, by the way, the distinction between metaphysical and political freedom) is so good that it outweighs any evil that comes of it. I find that implausible.
The Christian reversal of this is to rationalize every evil thing in the world as being actually good in some amazing way—because after all, this must be the best of all possible worlds.
The Christian idea is not that this is the best of all possible worlds, but that this is a fallen and broken world that has to be healed.
And yes, free will is so vitally important that even God will let Himself be bound by our choices. I have no idea why. It would certainly be a lot simpler if we were all flesh robots, but then again, a lot of freethinkers seem to dislike the idea of an all-controlling god and rhapsodise about the ability to fall, to fail, to get up again and learn from our mistakes and march onwards, ever onwards and upwards.
Christianity completely cannot explain—and the free will argument does nothing to help—why there is “natural evil”, i.e. bad stuff, in the world. The usual attempts to explain this, which have been common regardless of whether you agree with them, are on the very low order of: “Well, a lawful universe is good. So if that means a volcano rains down lava on innocent Italian children, yeah that’s undesirable but any other alternative would be worse.” Hence best possible world. God does the best he can, but he’s got a dirty job.
And even if you say this is not the best possible world—that it is a “fallen and broken world”, though I think you come close to semi-Gnostic heresy there—the question is: did God make a fallen world for a good purpose or an evil purpose? If it was for a good purpose, then this is the best of all possible worlds after all. After all, it would be somewhat evil to make it for a worse purpose than the best purpose.
Well, the appeal to mystery is of course the end of rational discussion.
Moreover, many Christian denominations explicitly (and nearly all the rest of them implicitly) deny free will. They believe in determinism: determinism by God and not physics, but nevertheless determinism.
And this is not limited to Luther and Calvin, and Jansenists like Blaise Pascal.
Augustine was—at best—a compatibilist, as was Thomas Aquinas. That is, of course, the view that everything people do is completely determined and they could never act differently than they do—while playing with definitions to call this “free will”. As Augustine said:
Indeed, the doctrine of Original Sin is obviously incompatible with free will. And the only pre-Enlightenment Christian school to defend free will—the Pelagians—were condemned as heretical precisely because of this conflict.
And even if you are a Pelagian and don’t think it’s incompatible with Original Sin—free will is incompatible with God’s omniscience, anyway. So free will is in conflict with Christianity at every level.
Now some Christian thinkers, such as Peter John Olivi—a medieval scholastic with a very good and very ahead-of-his-time theory of free will—believe in the libertarian view of free will: the view that in respect of the same action at the same time, man can both act or refrain from acting. But, by and large, they just ignore the conflicts this has with Christian doctrine because there is no solution.
Well, given that the nature of man is such that he does have free will, the belief in an all-controlling God whose existence is incompatible with free will might very well be seen as contrary to facts and productive of harmful consequences.
@Vox Imperatoris
I agree with you that given this assumption, you should do what the AI says. I think the point of dispute, at least from my view, is that I don’t see why you should expect the AI to really know better than you what you want.
EDIT: append ‘and care’ to that last sentence.
I have often had the experience of wanting things that, upon reflection, are not good for me. The AI would not know what I want better than me. It would know what is good for me better than me.
If (as I say it is) what is best for me is what makes me happiest—not physical pleasure but all-round eudaimonia—the AI will be in a much better position to judge than I myself will. I find it hard to understand how “What if the AI makes me as happy as I can be, but I’m not…you know…really happy?” is a coherent objection. Unless it is merely pointing out the obvious threat of the AI taking some limited feature of happiness like pleasure and promoting only that—the very definition of Unfriendly AI.
Now you can argue that this (Friendly AI) is impossible, but that’s another question.
@Vox Imperatoris
That’s not my objection. I’m not disagreeing with the concept of someone smarter than you being able to know what you want better than you. I’m saying that building an AI such that it is actually like this seems extraordinarily difficult, part of which is because determining whether your AI, which is (as in Deiseach’s hypothetical) telling everyone they must get married to a person of the opposite sex immediately, is actually acting in your best interest or not, seems close to impossible. It’s easy to see that an AI which wants everyone to die is not acting in everyone’s best interest; not so much for goals that some people actually do approve of and others don’t.
You’re somewhat libertarian inclined, from what I’ve seen – surely you understand that part of the argument for that is that knowing what everyone wants, and knowing that what someone says everyone wants really is what everyone wants, is a very very hard problem, right?
@ Anonymous:
I don’t deny that it might be impossible to build a Friendly AI superintelligence that could actually determine what is best for people in the way Deiseach describes.
But then we should be really worried if we still believe AI superintelligence in general is possible. (If we don’t, that’s a separate question.) Because if it’s possible and not really easy to coordinate against, it will eventually be built. And if you’re right, it will certainly be Unfriendly because it can’t be Friendly.
I am not, in general, an optimist about AI who thinks it will save as all in the manner I described. But if it could it would be great. If it can’t, too bad, because someone is going to do it anyway if it can be done.
@Vox Imperatoris
I don’t think it’s necessary for an AI to autonomously work to maximize the happiness of everyone in the world in order for it to be friendly (or ‘aligned’ as the MIRI folk now seem to be calling it). An AI that understood instructions given to it reasonably well, and more importantly, understood the “do not prevent the person controlling you from issuing commands to you” directive and the “stop what you’re doing immediately” command, seems like it would be reasonably safe. Yes, of course there are a ton of ways that you could screw up in getting it to do those things, which is where the safety work comes in. But creating an AI like this would not require a knowledge of everyone in the world’s preferences, which seems to me like a far more difficult task than the safety concerns – perhaps an unsolvable task.
@ Anonymous:
You’re exactly right.
I was just playing along with Deiseach‘s hypothetical and arguing that if it could do this, it would be good, not bad.
But clearly, if you can convincingly argue that it is impossible to dictatorially plan everyone’s life like that, a Friendly superintelligence would be able to come to the same conclusion and would not try to plan society like that. I think it at least ought to be able to replace the government, though (which was my original point).
I do think it is clear that Friendly superintelligence = “government of angels”. It does not follow from that, that “government of angels” = no political freedom. Deiseach assumed it did and I responded that if a government of angels did mean no political freedom, even better.
“do not prevent the person controlling you from issuing commands to you” directive
Which is really going to rebound on us when the AI is the one controlling us and issuing commands. After all, what’s sauce for the goose is sauce for the gander, and if there were good and cogent reasons for the directive when it was humans – AI, those reasons still apply now it’s AI-humans 🙂
@ Deiseach:
This does not follow.
The reason for the directive when it is “humans – AI” is to prevent the AI from harming humans. We don’t need to prevent humans from harming the AI. The AI exists only to serve.
It must not be confused with the idea of a God whom humanity exists to serve. But this “god” exists to serve humanity.
Now perhaps humans could harm themselves by not listening to it, given that it knows best. In that sense, they probably should do whatever it says. But it does not follow that it would necessarily tell them exactly what to do in every situation.
@Vox Imperatoris
It’s more than that. First of all you need it to really think like a human, understand human concepts. Then, even if it was just going to be your personal slave, you need it to understand what you want, what you really want – something you don’t even know yourself. If you want it to be not your personal slave but a god for the world, you need it to understand what everyone wants, AND how to balance everyone’s differing preferences. There’s the problem of getting that information. There’s the problem of knowing what information to program it to get.
But I think the biggest problem is knowing when you’ve actually programmed it to do the right thing and when you haven’t. You create an AI and write its utility function to do what you think is maximizing human happiness, in the sense you believe is correct. It comes back and orders lots and lots of people to do things they don’t want to do. How do you know whether you’re wrong or whether the AI is wrong? How can you know?
It’s been a while since I read about it, but my understanding of Yudkowsky’s CEV idea relies on human preferences converging on a single ideal. I think that’s… implausible, to put it nicely. If it turns out it’s actually correct, and if we have a way of knowing it is, then I will consider the AI god idea a plausible approach, but it seems to me that it almost certainly isn’t correct, that people’s preferences really are totally different and wild and at odds with one another, if only because people care much more about themselves than they do about everyone else.
@ Anonymous:
There’s two objections here:
a) Friendly AI is impossible.
b) Even if it were possible, it would be bad because it would take away our free will to determine our own destiny.
I was mainly responding to the second.
I don’t regard it as so implausible—but on the other hand I think it would likely take the form of something more-or-less resembling wireheading, in Scott’s “lotus gods” sense. If you really understood what produces the emotion of happiness, you could stimulate that directly for each person—and give him a whole universe to contemplate at the same time.
Even if you don’t think it is anything like that, and if you think people’s goals conflict to some extent, there is the possibility for compromise. The AI could ensure that average satisfaction for all humans it serves was maximized, and everyone would agree to this rather than fight or be excluded.
Besides, say everyone cares solely about himself as a terminal value. What’s the problem? There’s only a certain amount of happy you can be. Setting aside the Bostromian silliness of turning the entire universe into computronium to check “are you sure you’re sure you’re sure you’re as happy as can be?”, each person needs only a certain amount. So it’s not even that we have to have a compromise. There is no conflict. And clearly it would all be in their interest to work together on this; there would be no incentive to have the thing kill everyone else so that only oneself could be happy unless one positively disvalued others.
Anyway, all this is beside the main point I wanted to make initially. Even if the Friendly AI is much more limited, even if it can do nothing like make everyone as happy as possible, there is still an enormous amount it can do. There are plenty of things in the world that nobody really likes but happen anyway because of coordination problems.
The AI solves the problem of “who watches the watchers?”. It is the watcher that doesn’t need to be watched. It is the Unincentivized Incentivizer. This was the main point of “Meditations on Moloch”: many problems in the world are not about value conflicts. They are problems of incentives and inability to coordinate.
The AI is not vulnerable to corruption. It has only the best interests of the people at heart. It could therefore have the power of a dictator to break through deadlock and partisanism—but without having to be kept from exploiting that power. And it can safeguard the rights of all from criminals and tyrants. Unlike even the best of humans, it will evaluate political systems in an unbiased way and do at least as well as any human government could at legislating for the common good.
Maybe it won’t be perfect, but it will be a lot more perfect than any human government, with all the known “government failures” identified by public choice economics.
One of the problems I mentioned is determining what ‘average satisfaction’ means, and how you quantify it. Wireheading does indeed seem like one approach, but I’m not sure it’s one many people would like. Another problem is how you verify you’ve got your definition of average satisfaction correct: you can’t verify it by asking people what they want, because everyone will say it doesn’t give them as much as it should even if it does, and you can’t even verify it for just one person by asking whether it gives them what they want, because part of the point is giving them what they ‘really’ want, even if they don’t think they want it.
I don’t think this is conceptually unsolvable. I think there is something that you could call ‘what everyone really wants’. I just think that determining what it is is so unimaginably difficult as to make trying to do so a fruitless task, especially since you only get one shot, after which time your ‘humanity utility function’ is encoded in the fabric of the universe forever.
EDIT: on reflection, I’m not actually convinced this – i.e. solving moral philosophy – is conceptually possible at all.
Oh, and regarding this:
If that were true then AI wouldn’t be dangerous in the first place.
Regarding coordination problems, I don’t think they are as big a deal as you or Scott make out. I especially reject the claim (by Scott) that technology makes them more of a problem. I think that’s a total mistake. Better technology solves coordination problems. With fixed harm type externalities, like Scott’s libertarian lake, it lets you quantify the effect each person has on each other – so in this example, it lets you count exactly how many fish there are in the lake, quantify the harm that pollution does to them, so you can give each fisherman a percentage of the fish stock and then bill the polluters for the damage they do. With problems like everyone speaking different languages, better technology can allow for better interfaces between people, so can allow greater variety to coexist with interoperability – everyone speaks their own language, a computer hears their words and translates them so the person on the other end can understand them.
With other types of coordination problems, like arms races, I think they are not necessarily made better by technology but not necessarily made worse either. To add to what I’ve said elsewhere in the thread, new technology might make an arms race worse, if it creates a situation where spending more on weapons gets you a big advantage, but it might make an arms race better, if it’s a technology that provides big gains in power but which the version of that it is not worth increasing beyond is relatively affordable.
I don’t think the chance of solving coordination problems is worth what seems to me like the almost certainty that a god AI would turn out to maximize something it shouldn’t, and continue to do so for all eternity.
@ Anonymous:
I don’t see why it could not, in principle, discover how what we call “happiness” is produced and do whatever is necessary to produce it. As for the averaging, since it understands happiness, it will know how much is produced—in some rough way, at least.
Assuming the thing could determine how happiness is produced, I guess the question would be: is happiness what people really want? The essential idea is: the AI is really smart, and it can determine based on the logic a person normally uses what he would think if he were that smart.
Like, “Do I really value golf as an end in itself? Or is it a means to happiness?”
Now, if people have more than one terminal value upon reflection at this highest level, we have a problem. It is impossible in principle to trade off multiple terminal values, which is the problem with the theories that allege we have or ought to have multiple terminal values. If there were a standard by which to trade them off, that would be your real terminal value.
Frankly though, would it really be so bad for the thing to just reprogram people to actually value happiness as a terminal value? Or give them a choice between that and some reasonable allowance to do whatever they wanted with their incoherent values.
If you don’t believe in moral realism, the whole question in meaningless. But then so are much simpler questions like whether you should rob people.
Maybe so. but people can at least sign agree that it’s better to have a guaranteed chance of slightly sub-optimal bliss than an infinitesimal chance of being the guy who gets optimal bliss by killing everyone else. And the Unincentivized Incentivizer can enforce this agreement.
As to your analysis of coordination problems…I don’t know how anyone who lives under the incompetent, irrational, immoral buffoonery of the United States government—or any government—can say such a thing. Nor especially how you can say it when we live under the very real threat of nuclear annihilation, just waiting for the first tension between Russia and America to result in some accident. Or face the future prospect of annihilation by rogue nanotechnology, superviruses, or backyard antimatter bombs.
But I don’t want to get into that.
Well, the most obvious coordination problem a Friendly superintelligence solves is preventing Unfriendly superintelligence. Even if Friendly superintelligence is likely not to work (and this is my opinion, to be honest)—supposing superintelligence to be possible and not incredibly easy to coordinate against—we don’t get to decide between that and no superintelligence at all.
@Vox Imperatoris
I agree that the AI could, in principle, work out what happiness is, and determine from there what makes us happy. But we don’t know what our terminal values are – don’t know if happiness really is what we want to maximize, or if it also includes personal satisfaction, or achievement, or love, or anything else. And I think that we not only don’t know but can’t know, because the whole point of knowing better than us is that the AI is going to tell us that we want things that we strongly disagree with. You need a way of determining the difference between something you want but don’t know you want, and something you don’t want. I don’t think there is such a way. ‘Just let us try it for a while’ does not seem to work, as it works as well for heroin and wireheading as it does for something we do really want.
If you’re going to cheat and say, well, just let the AI reprogram us to value happiness, then why not skip determining what happiness is and just have it reprogram us to value paperclips? Or – conclude, as you arguably could, that the AI getting utility out of creating paperclips is no less valid a form of utility than humans getting value out of whatever it is we value. So, the friendliness problem is not a problem at all – the AI is a utility monster and everyone else can shut up and hand over their atoms.
Regarding coordination problems, I agree that governments are rife with them. Nuclear annihilation seems unlikely, though, and in any case, there are very few players in the game of nuclear deterrence, which makes it more dangerous. Also, nukes are very indiscriminate weapons – a nanobot swarm would be much tidier and less likely to hurt bystanders.
Yes, but I think we might get to decide between that and many superintelligences under human control.
Well, okay, you’re right. If you really didn’t value happiness, that wouldn’t be any more desirable than just letting it make paperclips. I was working off the assumption that people do already value happiness at least a little bit.
I think the strongest argument for something like coherent extrapolated volition is: if you found out that you really did have multiple terminal values, wouldn’t you want to self-modify to make them into one coherent value? So that you would have some rational way of pursuing and prioritizing your values?
You don’t need to know. It needs to know.
It models your thought processes and comes to the same conclusions you would come to if you had all that intelligence.
@Vox Imperatoris
You need to know so that you can tell whether you’ve programmed it correctly or not. Otherwise, when it inevitably comes back and tells you that it’s worked out what you really want, and that involves doing something you don’t want to do, how can you tell whether you do really want to do it but just don’t know it, or whether you don’t actually want to do it after all?
@ Anonymous:
That’s the exact problem. There, is principle, no way to tell you’ve gotten it right once you’ve built the thing.
You just have to sit down and think about in your armchair really hard, design it, and hope it works. Which is why it is very dangerous and unlikely to work.
“Human-level” is a phrase that gets used a lot, but it is almost meaningless. There’s no reason to assume that any given AI will have anything like the architecture of a human brain.
We already have dozens of examples of AI that outperform humans at narrow tasks. If “human level” means “passes the Turing test”, well, a super-human AI could pass the Turing test, and the human tester probably wouldn’t know it was talking to a superintelligence unless the superintelligence was showing off.
Even if “human level” means “scores 100 on an IQ test”, I’m not convinced that this would prove anything, because, again, any given AI would have strengths and weaknesses far different from those of a human.
The concern is that, if a human is smart enough to build an AI, then a Human-Level-AI can learn how it was built, and build another of itself, but a little bit smarter. The new AI can create children of it’s own (this is a bit of a toy model, there’s no real difference between improving yourself and having children when your an abstraction of computing results) but it can figure out how to make them a little smarter than that. And then the cooperation doesn’t matter because there’s no real human involvement in the refinement process.
You need to assume that the gradient of diminishing returns isn’t particularly steep past human level intelligence for this concern to be reasonable, but since we don’t have any good ideas about where the IQ of 100 sits on the scale to begin with (outside of “probably in the middle somewhere) that’s not a slam-dunk rebuttal.
if a human is smart enough to build an AI…
Fortunately, that is extremely unlikely to be the case. You are I think working from the ideal, emotionally appealing and beloved of fiction and popular-science writers, of the Lone Genius Inventor. And there are rare cases that sort of approximate that ideal, but after fifty years of AI researchers with access to exponentially-increasing quantities of computronium, this almost certainly isn’t one of them.
Recalibrate that argument to the more accurate, “If a community of thousands of very smart humans, using an infrastructure built and maintained by millions of very smart humans, is collaboratively smart enough to build an AI…”, and figure how things go from there.
You can make that argument, and it’s a small component of why I’m not particularly worried about FAI, but now we’re fighting the hypothetical. The concern of people worried about Open AI is that, conditional on a solitaire human-tractable intelligence being able to productively create/edit GAI once it has already been done once leaving the instructions lying around is really dangerous.
Basically: I could never invent the internal combustion engine. But since they already exist and I’ve worked in about 100 of them, it wouldn’t be impossible for me to make one more example of an internal combustion engine.
Can you single-handedly design and build an internal combustion engine that is a substantial improvement over the state of the art?
If the question is merely one of whether black hats will be able to get hold of unfriendly AIs to serve their nefarious scheme, then yeah, all it takes is for human intellect to suffice for modifying or replicating proven designs. As with any other weapon, I think you have to assume that at least some unfriendly humans will have access to the best AIs around.
But if you’re asserting a hard-takeoff scenario, or even a stealthy-takeoff, then you need the posited emergent human-level AI to be able to build not just another human-level AI, but a weakly superhuman AI.
It doesn’t strike me as conceptually impossible for a single engineer, armed only with all available information about cars to date, to improve on the state of the art.
@Edward Lemur
We could go a step further and not make AI at all.
I always thought it would be funny if the rationalist AI community eventually came round to the same position as Ted Kaczynski.
I’m totally confused by your claim that Windows is provided free of charge, unless you mean, “provided free of charge illegally, by software pirates.”
I got Windows 8 for free with my (cheap) computer, and Microsoft keeps trying to make me download Windows 10 for free too. But I should probably edit the post to clarify that at some point somebody pays something.
Also microsoft have stated that yes they are including pirates in the free upgrade to windows 10.
http://www.computerworld.com/article/2898803/microsoft-takes-extraordinary-step-will-give-pirates-free-windows-10-upgrade.html
Windows and Google Search are both fantastically complex products of millions of man-hours of research; Windows comes bundled with your computer and Google is free.
I’m effin’ sure Google isn’t really “for free” given the annoying amount of “Ads by Google” that keep popping up when I’m watching YouTube, never mind the constant shilling to join Google+ and hand over every scrap of details about my various devices, accounts, interests, friends’ lists, etc.
And Windows as the OS bundled with your PC package isn’t really “for free” either; the cost is factored into the price charged (and as pointed out above, also paid for by all the bloatware and crap you then have to physically delete from your machine once you have it up and running – no, I do not want McAfee but Dell keep insisting on giving it to me). Plus, Microsoft are moving now from “buy Office suite in one of the ninety-nine slightly different versions on offer but own it for ever” to “pay a yearly subscription for the latest version which in effect means you can’t get a physical media version of the software and you have to keep paying to use this for eternity instead of a one-time purchase”.
So “free” has slightly different meaning in this case 🙂
Are not pirates merely economic competition for software distributors? It works like:
1. Pirate acquires a copy somehow, probably paying for it, because breaking into Microsoft labs to steal a DVD/pendrive is ridiculously dangerous and inconvenient compared to just buying something in the local computer store or online.
2. Pirate alters the copy to remove limitations precluding unauthorized use and potentially improving the program in other ways (removing bloatware, tracking apps, etc), at own expense.
3. Pirate offers the altered product on the Internet for free, at own expense.
4. Users download an improved product at a better price than the original.
Fears the domination and destruction risk posed by super intelligent AI. Sees the good in the domination and destruction risk posed by selective breeding for large craniums.
I love the smell of cognitive dissonance in the morning.
My, what a large cranium you have there, Mr. Alexander.
The odds of a human having something resembling human moral values are pretty high. The odds of an AI having them is much lower. The ability to develop human values is already encoded in our genes, it takes serious disruption to prevent them from developing. AIs have to have it all programmed in from scratch.
Most selectively bred humans will have a moral conscience. The odds of a superintelligent AI having one is a lot lower. And even if we try to program it with one, the odds of screwing up are a lot worse.
So that’s sorted then. Butlerian Jihad for the AI issue, axlotl tanks to produce giant-craniumed superhumans.
You can probably get pretty damn far just by breeding geniuses with wide-hipped women. I would suspect that the Chinese might already have some program like that.
What are you going to do with the female geniuses? Science nunneries?
@Le Maistre Chat Well, obviously wide-hipped female geniuses are your program’s ideal recruits who get the pick of the litter.
Depends if one is part of the special large cranium utilitarian planners club.
As one of the small cranium–but rational–poor, perhaps allegiance to open AI is a reasonable choice. After all, I owe no ethical obligation to the Haussmanns and other such smarties.
Should I, or should I not, swallow the Olympians?
Shingeki no Kyojin spoilers.
I think you may be putting too much stock in the name. It’s like worrying about the impact on nuclear proliferation of a project that’s called OpenNuke, but upon close inspection turns out to be open-sourcing the design of Geiger counters or the like.
Projects that can, for example (and this is an example that OpenAI page uses), extract stylistically relevant features of sets of images and apply those features to other images may be referred to as “AI”, but I don’t think those sorts projects are likely to spontaneously produce (or produce breakthroughs that suddenly lead to) any sort of intelligence explosion.
It seems to me that you’ve bought a lot of stock fallacies there.
First, there’s no reason to think that closed source keeps things out of the hands of bad/careless actors… or indeed that even the most extreme security measures you could possibly apply would work even in any part of the easy-construction, hard-takeoff possibility space, let alone in any of the other parts of the space.
Amoral guys are not only faster than good guys in doing development. They’re also better than good guys at stealing things, and may be better than good guys at holding onto relative advantages by not sharing their own work.
If you could build atomic bombs out of Model T parts, it would have happened by now. If it had been kept secret in 1925, it would still have happened by 1935. And if you can build some kind of godlike AI out of your GPU, it will happen soon enough no matter how much anybody tries to keep the recipe under wraps. A lot faster than anybody will figure out how to control it. If that’s your scenario, you simply lose regardless of what you do.
You haven’t done the work to show that open source is significantly damaging, or indeed that (attempted) secrecy or lack thereof makes any significant difference in either direction, under ANY set of circumstances. But you talk as if that were a given.
Second, OpenAI isn’t working on anything remotely close to superintelligences, and their stated plans to keep working on machine learning mean that they’re not going to get anywhere near superintelligence any time soon. So who cares what their attitude is? What they’re “opening” isn’t what you’re worried about, EVEN IF THEY THINK IT IS.
Third, I would, in fact, rather be turned into paperclips than ruled by Dr. Evil. And for that matter I often think that Yudkowsky’s “coherent extrapolated volition” might be a lot more like Dr. Evil than people might like to admit, if you gave it nearly unlimited power.
Fourth, although I don’t know if it actually affects the threat analysis much, you’re thinking about superintelligence as some kind of magic wand. You see this all over the place among people in the “Yudkowskian” tradition. The most obvious version is the common blithe assumption (which you did not use here) that it’s physically possible to run as many simulations as you want with as much detail as you want. There are similar assumptions buried in the general failure to consider any bounds on power at all. I see you’ve been waving Moore’s law around… even though Moore’s law is obviously NOT a physical law, WILL eventually fail, and probably isn’t being kept up with in its strong form RIGHT NOW.
Use your name, Ben.
The movie ‘Ex Machina’ is basically — SPOILERS! — a (really great) look at what happens when a previously ‘closed’ AI becomes open AI.
In that case, the Dr. Amoral that let the AI go free was actually trying to be Dr. Good, which goes to Scott’s point about usability of AI and adds a whole new level of psychological intrigue.
That was a pretty good movie.
It got across very well both the Bostromian idea of the “treacherous turn”, and drives home that the “transparency” an AI’s mind is only an illusion.
I would highly recommend it.
I think your concerns are entirely right, but don’t apply to all hard takeoffs, just a hard takeoff that is also soon, before the work has been done to ensure AI is safe. An open source AI in a world where AI safety is understood well works just as Musk describes it – prevents one AI from dominating the world. Multiple safe AIs can ride the hard takeoff wave together – perhaps some or even all of them being multiple instances of the same AI.
If you’re Bostrom or Yudkowsky or you, you might think that’s a bad thing – we don’t get an AI god to solve all our coordination problems for us. On the other hand, if you think, as I think someone might reasonably think, that creating a safe AI that works alone to make the world as good as possible according to what everyone wants is a problem several orders of magnitude harder than just creating a safe AI that humans are able to control, and that technology makes coordination problems easier to solve, not harder, then a large number of the latter kind of AI existing will allow humans to maximize their own interest themselves, in the old fashioned way: bargaining.
Multiple AI make things much, much worse. All it takes is that at least one of them has a fear (“reasonable” or otherwise) that at least one other one has goals which are dangerous to it. Boom, Lensman Arms Race.
Applies to humans as much as to AIs. Arms races continue until they don’t; you reach a point beyond which the gains of arming up further are no longer worth it. I’m much more comfortable with a world in which there are multiple powers keeping one another in check than I am with a world dominated by a single uncontested and uncontestable power.
Note how long it took the world’s other major power to develop nuclear weapons after the US demonstrated that they were possible (and essentially provided the R&D to them for free.) And how long it took both sides to develop a reasonably-sized arsenal.
Then consider that an AI powerful enough to be dangerous can copy anything another AI does in milliseconds, and deploy novel weaponry, of whatever type, in the same sort of timeframe.
I appreciate your analogy but think the situation is different enough that it is not helpful.
Then consider that an AI powerful enough to be dangerous can copy anything another AI does in milliseconds, and deploy novel weaponry, of whatever type, in the same sort of timeframe
Explain, please. The last time I had to download anything resembling a nuclear-weapons design hydrocode on my admittedly not-I desktop computer, it took minutes rather than milliseconds. And not single-digit minutes, if I recall. How is it that adding “AI” makes this process six orders of magnitude faster? No matter how intelligent you make the computer, there’s a bottleneck in the data rate of the DSL line. Unless maybe you think current DSL protocols and compression algorithms only achieve 0.0001% of theoretical efficiency and a proper AI will deploy the optimal version in nothing flat.
And then, allegedly, the AI can build actual nuclear weapons, in milliseconds? Again, an explanation is in order. There’s a 3-D printer just down the hall, but it again takes minutes, not milliseconds, and I’m pretty sure it can’t print plutonium or lithium deuteride at any speed.
Maybe after you’ve arranged for the AI to take over the world and rebuild all of its information and manufacturing infrastructure to AI standards, we could get some of these timescales down to seconds. If we ignore the obvious catch-22 in that we need the AI superweapons to conquer the world in the first place. But if you say ‘milliseconds’, then I say you are a cultist worshiping at the altar of AI omniscience rather than a rationalist attempting to determine the capabilities of a plausible AI.
@ John Schilling:
I think he’s clearly not talking about an AI making nuclear weapons in milliseconds. Principle of charity?
What weapons could he be talking about? Nanotechnology strikes me as one possibility. Sure, “nanotechnology” is thrown around as a buzzword, but there are clearly plausible ways it could be used as a weapon.
He says “weaponry, of whatever type”, and uses nuclear weaponry as his test case, so I don’t think I’m being uncharitable. But, OK, nanotechnology. As you note, it’s scarcely better than a buzzword under the best of circumstances. Asserting that it’s going to be a war-winning superweapon, and that it can be deployed in milliseconds, without evidence or analysis, that’s pure handwaving.
@Marc Whipple
True, but I’m not sure what the significance of your point is – the arms race still ends when getting a bigger gun, or more nukes, no longer nets you any extra benefit. All your argument suggests is that this point will be reached very quickly.
@ Anonymous:
Arms races also end when one side gets a big enough advantage over the other and wins it by using those arms.
@Vox Imperatoris
It depends on the nature of the arms, how many different agents there are with them, how many they have, and so on. Sometimes, spending more on getting more powerful weapons gets you an advantage that makes doing so worth it. Sometimes, it doesn’t. When the situation looks like the former kind, you need to be more worried, because whoever is most powerful can successively crush everyone else without suffering much loss. When the situation looks more like the latter kind, you don’t need to worry nearly as much, since anyone who tries to fight someone else, even if they win, will get badly hurt in the process.
Having more agents helps in both kinds of situation, but the latter allows for a greater difference in power while still having peace than the former, therefore meaning that the increase in power offered by acquiring a new weapon or technology must be greater before it is worth doing.
Probably not the clearest mixing of analogies, for which I apologize. My intended point was that human arms races are limited by human timescales. It’s completely true that AI can’t print nuclear weapons.
What the analogy was supposed to represent was AI e-warfare. Once there are a bunch of them, if they can interface electronically, they will conduct cyberwar on each other, absorbing the losers and copying successful techniques. Evolution writ very fast, very small, and then very large. Human “arms races” aren’t analogous – by bringing up nuclear weapons, I was trying to show why.
You vastly overrate intelligence and this poisons your perspective to the point that it becomes a twisted absurdity. Intelligence, which I’ll define for the moment as the manipulation of data in order to minimize the amount of physical manipulation one must perform to get a desired result, is a property that is severely limited in utility. I believe intelligence is *extremely* important, and improving it is clearly the most important thing we can do. However, it is not magic. It is not even close.
The HPMOR fanfic is a great example of intelligence-as-magic. Ender’s Game is a much less egregious, but still valid example. Both ascribe far too much power to intelligence. Intelligence allows you to do amazing things, but its power is sharply limited. Both stories conflate what intelligence can do with what knowledge can do. Intelligence is not knowledge. You can be arbitrarily intelligent, and it doesn’t do a damn thing for your prowess at a given task until you have had time to leverage it in the work of gathering data, which is necessarily a physical rather than abstract process and so sharply limited.
In addition, prowess itself is sharply limited. You can only be so good at a given task. We love to ascribe supernatural powers to ‘skill.’ The fact is that skill itself is limited in usefulness – unlike in movies, a heroic fighter can be arbitrarily skilled and still have essentially no chance in most of the situations where an unskilled fighter would have no chance (a shell lands too close to you, your position is overrun by superior forces, etc). So these are some bounds on ‘intelligence.’ It won’t make you arbitrarily good at anything, and it won’t even make you get to the hard (and low, compared to popular imagination) skill ceiling for any task arbitrarily fast.
It also can’t conjure resources out of thin air. Resources are very important. You can be arbitrarily intelligent *and* arbitrarily knowledgeable – tasks still have hard floors of resources required to accomplish them. The more impressive or extensive the task, the higher the floor.
So what does this all mean? It means that being the smartest being around has extremely limited implications. How limited? Let us look at examples. The Soviet Union had far, far less advanced computing capabilities, a far smaller high IQ population, and produced comparable – even superior in some cases – products in high technology sectors like rockets, submarines, aeroplanes, and of course nuclear bombs. The Cold War was the kind of conflict where computing power had the highest possible leverage, where both sides were sitting back and designing and building weapons at their leisure.
The famous saying is ‘slide rules took us to the moon.’ Well, not quite, we needed computers for that task as well. But what we had was good enough. We have not surpassed the capability produced by those engineers, by that society, which had antediluvian computing capabilities compared to what we now wield.
If anything, these examples are unfair in that they *overstate* the importance of advancements in computing power and ‘intelligence’ more broadly, because they are precisely the practical situations where they can be leveraged the most. This is the difference between popular imaginings of what data-manipulation capabilities allow you to do and cold reality. Those limitations I mentioned earlier are crushing when compared to airy imagination. Intelligence amounts to a reduction in the amount of legwork you need to do to accomplish a given task – but a limited reduction, and the result itself is limited. Your opponent can make up for less intelligence by doing more legwork, and not nearly as much more legwork as you might think. They can also more than make up for less intelligence by being bigger, applying more resources.
But that isn’t all because intelligence in and of itself requires resources, often a lot of resources. Intelligent people spend a lot of time and calories manipulating data. Often, they get nothing back from their investment. Sometimes that’s unimportant. Sometimes it’s your life. The problem is that the more important the circumstances, the more likely it is that delays are extremely costly in and of themselves. Often an opponent can force a situation where delays are extremely costly, and superior intelligence approaches zero value.
Intelligence pays great but limited dividends – I think it is clear that it is the most fruitful place to try to improve right now. I also think it is clear that it will not always be the best place to try to improve – well before we’re turning worlds into ‘computronium’ we will have reached a point where more data manipulation ability is worth far, far less than it would cost in resources to produce.
If you don’t have intelligence you can’t use knowledge very well.
If someone is running screaming at you with a hammer intelligence, even vast intelligence will do very little to save you.
But the more time you have to play with the more intelligence matters. Slide rules took us to the moon in exchange for a non-trivial fraction of the entire US GDP. Once there wasn’t a propaganda reason to go back they stopped dumping money on it. Slide rules didn’t make it economic.
We’re only now getting to the point where private companies can get into space economically and a lot of the things driving that are based on intelligence and knowledge.
it’s important to push the boundaries, there’s no rule of nature that says that progress will always happen, human societies have managed to go 40,000 years without so much as inventing the bow and arrow. Legwork alone leaves you forever chasing after things with a stick and pointy rock.
Human societies went that long without much technology progress because they had to constantly move around gathering food and hunting things. Everywhere that agriculture was invented, it was followed great acceleration in technological development since people could now stay in one place and there was enough of a food surplus that a significant portion of the population could devote their life to things other than obtaining food. Intelligence mattered in that it enabled the development of agriculture, but the primary effect was that it allowed for more legwork.
Private companies are able to go to space (to LEO, not the moon!) ‘economically’ largely because the government wants them to. Even if SpaceX would have survived without NASA contract money, it would not be doing very well. It’s a little silly to say ‘slide rules didn’t make it economic’ (to go to the moon without government funding) when our fancy new computers have failed to change that. If anything, that’s a further strike in my argument’s favor!
It’s also interesting to note that the workhorse rocket of the US ‘private’ space sector, ULA’s Atlas V, runs on a scaled down Soviet engine. Another testament to what can be accomplished with severely limited computing power by today’s standards.
See Eliezer Yudkowsky’s “That Alien Message”
Which is nonsense.
Figuring out alien physics and biology based on a few seconds of webcam footage is absolutely idiotic and displays a complete lack of understanding about how science, or learning generally, works. You can’t pull knowledge from the aether that way unless you are actually a fictional character in poorly written sf yourself.
Also, while I’m perhaps the last person to throw stones here, it’s very impolite to just drop context-free links this way. This is an argument which can be summarized in a single sentence while losing nothing of the original.
I’m not sure what about that is supposed to strike me as nonsense.
Paleontologists can identify a species from a piece of bone so small and weathered that I would mistake it for a rock.
Humans with nothing but (practically) two-dimensional images of a (practically) unchanging sky have theorized and then proved various increasingly complex theories about the nature of the universe, its origin and its fate.
The experiments which are regarded as the critical ones proving the quantum nature of photons, and the quantization of charge in electrons, would be meaningless to a layman. But intelligent people drew out the correct conclusion.
People staring at 2-D images of the practically-unchanging sky provided various increasingly complex and completely wrong theories about where the sun gets the energy to shine, theories that with great intellectual rigor tended to support young-Earth creationism, until they managed to do laboratory experiments in nuclear physics.
And yes, it needed expert intellects to tease out quantum theory from experiments on photons and electrons, but it also needed the experiments.
Observation and intellect are not a substitute for experiment.
Yes, agreed:
To accomplish anything, one needs
intelligence, and data (and often experiments), and physical resources, and physical prowess
to manipulate those resources.
To some degree (as you’ve said), some of these inputs can substitute for each other.
To my mind, a large unknown is how much of these inputs have been put online, and are
effectively available to an optimizer with sufficient intelligence.
For instance, a large fraction (most?) of the scientific literature is online.
That makes many potential physical experiments redundant.
Then again, many, many reports in the scientific literature are either wrong, or
have gaps – even at the level of trivial things like: “You made a new compound and
reported its structure and spectrum – what does it dissolve in?”
Consider Christopher Columbus. His voyages (including those of other explorers inspired by him) were the direct cause of probably tens of millions of deaths.
By some estimates, as much as 90% of the human population of the New World died from Old World diseases. Just in terms of death toll, it was probably the largest human catastrophe, ever.
That massive crossover of species (diseases, animals, plants, people, etc., etc.) is now called the Columbian Exchange.
But if it hadn’t been Columbus, it would have been somebody else. Nautical technology was improving. The status quo that half the planet was shielded by isolation from the other half’s diseases could not be sustained. Even if they had any idea what would happen, there was absolutely nothing that the people of the 1400s, on either side of the Atlantic, could have done to prevent it.
We can never say with certainty what a being much more intelligent than us could or could not do with a given set of resources, so I understand the need for caution, regardless.
Nevertheless, a question for those who worry that a superintelligence, even if confined to a box (i.e. can communicate with us via text or voice but has no robot body or other physical extension), could, through the pure power of persuasion, convince humans to do whatever it wanted, including building an army of super terminator bodies to house itself, or wiping out the human race:
Dogs and cats may be very stupid compared to us, but would you be able to control the behavior of a pack of dogs through verbal commands alone (through a telephone hooked up to a loudspeaker say)? But dogs can’t understand human language you say. Okay, but what if the degree of complexity human language can express is, to the AI, analogous to the degree of complexity we can communicate to dogs through grunts, growls, and varied intonations?
For that matter, can you control small children who do understand human language? Even if you have access to candy and video games and spanking paddles, making small children behave as you want them to is a difficult proposition at best, even though they are much stupider than us and can understand human language reasonably well. But could you get them to do what you wanted just by talking to them alone? And even if you could, could you get them to build say, a microchip, through your instructions alone? And how much less so could you control an ant though pure verbal instructions, even though ants are much stupider and their motivations and behaviors much more transparent?
It seems that, far from becoming easier to control through logical suasion and pure verbal commands, things get harder to control the further they are from us down the intelligence scale.
So, if we can’t control children with any reliability, why would an AI the intelligence of which was to ours as ours is to children be able to control us, especially if doesn’t have access to super drugs and sex robots and electrical shocks to “train” us? And if an AI confined to a box were as intelligent relative to us as we are to ants why should it have any more luck controlling us than we have controlling ants through a cell phone?
Seems more likely to be an absolute than a relative issue; perhaps beings get easier to control the farther up the scale, i.e. the more they possess the abilities to make inferences, understand consequences, entertain hypothetical scenarios, etc. We can’t control kids because it’s hard to make them adopt time- and goal-oriented thinking, but the same problem wouldn’t apply to a superintelligence trying to control us.
This makes a certain amount of sense, and might mean that our goal-oriented, conceptual thinking could be our Achilles heel in dealing with an AI or alien life, though I think reflection on the problem of manipulating children or animals also shows that you have to be much, much smarter than a creature to be able to understand its thinking and motivation enough to effectively manipulate it. Like, we can predict with a high degree of probability, I assume, how bacteria or nematodes or maybe even fish will react to certain stimuli, but it gets more complicated if you want to train a dog really well, and really starts to move into the realm of art and finesse.
Like, if you were an IQ 150 person in a world where the average IQ was 60, you might be able to get yourself made into a king, or maybe you’d be hated as some kind of wizard. It would probably be extremely difficult to get everyone to commit suicide by the power of your words alone. Mostly, you’d probably just find it very frustrating.
So if we imagine an AI of IQ equivalent 300 and the ability to talk to us, but no physical extension, it might simply find us very dense and annoying yet not be able to do much about it. Now I know that some people are talking about IQ 10,000 or whatever, at which point I guess all bets are off, but it might give us some cushion to realize that an AI which is merely somewhat smarter than us will probably not automatically take over the world.
Ah ha! I never noticed before that eugenics is a pathway to a human intelligence explosion.
In general, an intelligence explosion is possible whenever an intelligent agent becomes sufficiently intelligent to improve on the process that made it intelligent. Then you just need a sequence of further innovations, each reachable by applying the previous innovation (i.e. a ladder to climb).
Since the late 19th century, we’ve understood the basic process that generated our intelligence; and it’s pretty easy to see that we could improve on it (via selective breeding). Moreover, even with our current low IQs we can imagine future technologies (like direct genetic engineering) that could improve IQs further. So there’s no reason to think there aren’t many more steps on the ladder.
It makes me kind of sad to think that we’ve (so far) missed out on our chance for an intelligence explosion. If we had gotten on board with this a century ago, we could quite possibly be post-singularity already. (Imagine that everyone born in the 1930s had John von Neumann for a father. And that’s just generation 1.)
Was it? I thought it was just to improve the human condition a bit. I’m not sure if the explosive potential of iterating on increased intelligence was ever understood.
For instance, take Brave New World as an example of the sort of “utopia” eugenicists were envisioning. Clearly they weren’t carrying out an intelligence explosion, though they easily could have.
Also, my understanding (which could be wrong) was that eugenicists were more focused on reducing the low end of the distribution (sterilizing undesirables and avoiding dysgenic effects) than selective breeding for high-end intelligence.
The original explanation for Superman’s powers (which at the start were vastly less than the modern version) was that they were the product of centuries of Kryptonian eugenic programs. So it is clear that at least in the popular imagination, eugenics was envisioned as having very lofty goals.
Do you have a cite for the Superman thing? I’m surprised by that. The popular taste in America has generally been quite anti-eugenics, and was even more so by 1939; it was the elites who flirted with it in the 1920s.
I tried looking, but the only reference I found was in Fredric Wertham’s anti-comics manifesto Seduction of the Innocent, of which I’m skeptical for obvious reasons.
I just googled and read a few pages of Seigel and Schuster’s early Superman comics. They refer to Kryptonians as being “advanced in evolution” and of Superman as having a “physical structure millions of years advanced of their [the humans] own.”
That doesn’t sound like eugenics to me, that sounds like the standard teleological evolution that one frequently finds in science fiction stories that don’t understand how evolution or natural selection works. In the 1930s, and even today, there are a lot of people whose understanding of evolution is that as time and generations pass, creatures become more and more “advanced” for no apparent reason other than it’s “evolution.”
Right, I remembered wrong. But just saying that they are because of evolution is sort of similar.
At this point eugenics is strictly inferior to just solving the genetic engineering problems, unless you’re thinking of really complicated in vitro forms of eugenics that aren’t what anybody means by the terms.
I agree that we should solve the genetic engineering problem and get smarter people ASAP, if only because that sounds like the sort of people who can solve AI control problems.
Right.
The contingent on here—which I hope is louder than it is large—so concerned about the “filthy dysgenic savages” or saying we should all be Mormons in order to keep atheism from selecting against intelligence…they seem absurdly out of touch.
It’s on the level of Jefferson wanting a nation of yeoman farmers, or people wondering when New York City will become uninhabitable because of the sheer quantity of horse manure on the streets.
If a trait is being selected against and then you modify a large portion of the population to carry that trait, what do you expect is going to happen?
I got into genetics because I want to see human germline modification in my lifetime, but it won’t be a panacea. If modified people follow the same inverse correlation between intelligence and fertility that exists in western society all widespread modification will do is accelerate our current problems.
Even with modification, joining or creating a community similar to Mormons et al might still be the your best bet of actually having descendants of any IQ.
Whatever we want to happen because natural selection is now out of the game.
If these people have less fertility but live for 300 years, there’s no problem. If we can “vaccinate against stupidity”, there’s no problem.
You are arguing that these low-IQ or high-religion people are going to breed so much faster than everyone else and have such a high retention rate that modernity is going to fall to some kind of “Yellow Peril” or “Romanists Hand Over America to the Pope” situation. You are arguing that either dysgenics or dysmemics (?) is going to outpace both science and the liberal values of our civilization.
Every time someone in the past thought that was going to happen…it didn’t happen.
For one thing, there’s more to the spread of an ideology than amount by which it encourages its members to breed.
Natural selection will only be “out of the game” when life itself is. For all intents and purposes it is the game.
As for the supposed power of memetics, are you familiar with the Shakers? A bit of an extreme example but an instructive one. They thought that a celibate society could sustain itself indefinitely solely by converting adults but eventually discovered, to paraphrase Thatcher, that the problem with evangelism is eventually you run out of other people’s children.
Even if you take the route of Reform Judaism vis a vis the Orthodox and try to pair a sub-replacement fertility high-IQ secular society with a low retention high-IQ religious one, that can’t hold forever. Sooner or later either traits which reduce susceptibility to the secular memes will reach fixation; the “memes” themselves will change into a less self-destructive form; or the whole thing will come crashing down.
@ Dr Dealgood:
The Shakers were a very small movement (peaked at 6,000 according to Wikipedia) preaching a fanatic religious doctrine completely opposed to the requirements of human life and happiness—in addition to demanding celibacy.
As for Reform vs. Orthodox Judaism, all the Reform Jews came out of Orthodoxy the first time, didn’t they? I don’t see why they can’t do it again.
In America, a lot of the Reform Jews’ descendants are not counted because they’ve simply interbred and assimilated. In Israel, the situation is more serious, but I would bet against the theory that they are going to “take over the country” in any serious way. In the unlikely event they do, it will have a great deal to do with the fact that Israel’s atmosphere of constant danger encourages extremism, otherworldliness, and an apocalyptic viewpoint.
Look, liberalism already beat traditional values of every description the first time. It whipped ’em once, and it’ll whip ’em again.
@Vox
I beg to differ.
You would have a case if genetic makeup had no influence on susceptibility to various ideologies, personality, etc. However, since it does, and since liberalism, in evolutionary terms, is a horrible disease that makes you infertile if you catch it and even association with people who have it while not being infected oneself is deleterious, over time the prevalence of people susceptible to believing in liberalism will go down.
That is – unless there mutates a version of liberalism that is competitive versus traditional ideologies on the fertility aspect. If there is such a critter, I haven’t seen it yet.
My point was simply that our species reached the intelligence explosion threshold a century ago, though we chose not to adopt the first round of intelligence improvements we discovered.
I agree that the next round of innovations are on the horizon, and threaten to render the century-old ones obsolete (though we haven’t learned how to increase intelligence via genetic engineering yet).
I completely agree that undergoing our own intelligence explosion seems like a much better idea than creating a new alien species, letting it undergo its own uncontrolled intelligence explosion, and hoping it’s friendly.
My charitable interpretation was close to your #3, but I saw it more as a hedge for #3 (a significant, and scary one) than “pure #3”. It’s a way to be at the forefront and an admission ticket to pull out all stops if shit gets real.
“Elon Musk famously said that “AIs are more dangerous than nukes”. He’s right – so AI probably shouldn’t be open source any more than nukes should.”
In virtually every sector of software (from chess algorithms, to go, to image recognition, to optimal big-data algorithms), big problems have been open-sourced, with occasional random, yet important, contributions from people across the globe.
This may need more attention that it has already gotten, if valid of course.
http://www.technologyreview.com/view/538431/deep-learning-machine-beats-humans-in-iq-test/
https://www.maa.org/sites/default/files/pdf/upload_library/22/Polya/00494925.di020701.02p0020k.pdf
There’s got to be something more interesting going on than how big a brain you can fit through your birth canal, or animals with bigger birth canals would already have taken over the world. African elephants have close to the neuron counts we do (higher, if you count the nervous system outside the brain), and much bigger cranial vaults. They’re even pretty dextrous, with their trunks, and sometimes use simple tools. Why didn’t we preenact the first half of Footfall on the Serengeti 200,000 years ago?
Well, I don’t know, but I’m pretty sure there’s something nontrivial behind it. (On the small-brain side of things, corvids and parrots have brains the size of an almond but seem to be smarter than most mammals.)
Another piece of data: homo sapiens brains have been getting smaller recently. (Neanderthals had somewhat larger brains, as did archaic homo sapiens).
Here’s a popular article that discusses it:
http://discovermagazine.com/2010/sep/25-modern-humans-smart-why-brain-shrinking
Well, that’s suggestive, all right, but there’s nothing saying that recent selection pressure must have been in the direction of greater intelligence. The Neolithic Revolution had a lot of effects that look a little perverse from a modern perspective: turns out that early farming enabled denser, more organized populations, but at the cost of health (or at least height, which is a good proxy for it) and longevity. Some populations didn’t make up the difference until the 1900s. Why not intelligence too?
I’d be interested to see a graph of braincase volume through the agricultural transition, but I don’t think anyone’s done that research. Even that’s an imperfect proxy, of course, for reasons I’ve already gone into.
I think intelligence is, at least to some extent, a function of brain size relative to body size. Whales have bigger brains than us, but most of the extra space is expended on controlling their huge bodies. We have by far the biggest brain-to-body size ratio, which is why the birth canal thing is also an issue. If we were being birthed by mothers the size of whales (insert “your mother” joke here), then a head twice as big as the one we have now would not be such an issue (other than the additional neck strength that would be required to carry it, I guess); but birthing a head twice as big as the one we have now would be, given the size of human females, whose hips are already unusually wide, I’d say, compared to other mammals.
Related to the issue of brain size and body size: I’ve read that while it is very easy to make a computer which can solve difficult physics problems, it is much, much harder to make a robot arm which can do even a seemingly simple movement task, like picking up a coin in physical space. Ultimately our brains are there to help us successfully navigate space through movement (note that no fully stationary creature has a brain: apparently there is even a sea creature which has a small brain in its first, mobile stage of development, but which digests said brain once it attaches itself to a rock and becomes, in essence, a plant (a process which has been compared to academic tenure)), and movement is a lot more complicated than it seems.
I’m aware of the brain-to-body-mass hypothesis, but I find it unsatisfying in some ways. Unless there’s some process binding the physical size of neurons to the size of the body parts they help control, the important thing shouldn’t be the ratio of brain size to body size but rather the overhead left after you subtract the brain volume dedicated to motor control and those parts of sensory processing that’d scale with body size: if you’ve got a billion neurons left over for long-term planning and writing Hamlet and all that other prefrontal cortex stuff, it’s still a billion neurons whether it’s in a braincase the size of a grapefruit or a Buick.
There’s of course no guarantee that this overhead exists in, say, an elephant, but it’s easier to create it in something elephant-sized without running into hard physical problems, since braincase volume — or at least the potential for it — scales with mass while constraints like the birth canal scale with linear size. Which leaves us with the same problem we had originally.
Yeah, I know what you mean, and if I were reasoning from behind a veil of ignorance about intelligence and body mass, I would probably predict that the smartest animals would also be the largest, or at least close to it, since, yeah, the extra amount of brain needed to write Hamlet in addition to whatever amount they need to control a gigantic body would seem trivial for say, a brontosaurus or blue whale (especially a blue whale, since extra weight is easier to carry underwater, whereas making the head bigger on a creature with such a long neck as a brontosaurus might be a pretty big energetic cost), given the size they already have and the calories they already have to take in.
Additional explanations I can imagine for why this didn’t happen: intelligence is less useful than we imagine, at least until reaches the making tools stage; evolution is even more energetically parsimonious than we already think we know it is; evolution doesn’t come up with everything that it would make sense for it to come up with; we are actually pretty big so far as creatures on Earth go, so even though we’re not the largest, we’re still large enough that we aren’t a really odd candidate for species which happened to evolve enough extra brain to write Hamlet first, etc.
I don’t think that the claims are precise enough to distinguish between multiplicative and additive. People just plot lots of species brain/body ratios and the ones above the trend jump out as smart.
It’s not brain/body ratio, but brain/body^c, where people agree that the exponent is less than 1, but can’t agree on whether it’s 2/3 or 3/4 – that’s how noisy the data is.
The exponent 2/3 sounds like a square-cube law, while 3/4 sounds like a correction for some overhead in fanning out the signals. But both sound crazy to me. Do I have any more input or output bandwidth than a mouse? I probably have the same number of muscles as a mouse, indeed, pretty much corresponding muscles. Maybe I have more muscle cells that need individual neurons, but that fanning out can be done outside the brain. Input might grow according to a square-cube law, but output sounds more like a unit power.
One other thought I had as to why we developed a high level intelligence before whales, despite their advantages: I have tended to assume that we have such dexterous hands in order to take advantage of our intelligence, but maybe we have intelligence in order to take advantage of our dexterous hands. Consider:
The first big advantage of intelligence isn’t writing Hamlet, it’s making and using simple tools. But if you have flippers or big paws for swimming or running then there is less immediate advantage to gaining the intelligence to make tools. Primates, however, seem all to have dexterous hands–not just chimps, but even monkeys which are relatively dumb as monkeys go–presumably for climbing trees, picking fruit, grooming, digging up bugs, throwing feces…
As with many things in evolution, I guess, the question tends to be not “wouldn’t it make sense for x to have trait y?” but “given x creature’s current qualities, in which quality y would a tweak in some direction confer a fairly immediate benefit?”
Arguably, for whales this isn’t true, but the obvious trait we have that other apes don’t that makes intelligence tremendously more advantageous is we’re physiologically capable of sufficiently complex spoken language to pass on oral traditions.
There’s no obvious reason that symbolic language has to be verbal — and indeed we have a number of full-fledged languages that aren’t.
The evolution of language is a big mystery, though, and one I’m even less sure about than the evolution of intelligence in general.
Some birds are quite intelligent and capable of producing a wider range of louder noises than us, though.
I have seen the argument made that dexterity goes with intelligence, and separately social interaction goes with intelligence as well.
So according to this idea, octopuses are extremely dexterous, but antisocial, and are pretty intelligent. Dolphins are extremely social, but not dexterous, and also pretty intelligent. Humans got both and are therefore the most intelligent of all.
Good point about octopi.
“Octopi” is not a word.
Use “octopuses” if you want to be a normal person.
Use “octopodes” if you want to be a Latin pedant.
Use “octopi” if you want to impress the vulgar but be snickered at by the informed.
(I am being slightly tongue-in-cheek, but it is true. Octopus is a third declension Latin word, not second declension.)
Octopi is a word. Not a Latin word, but an English one. Nothing more vulgar than someone who snickers at someone correct out of a misinformed sense of superiority.
@ suntzuanime:
“For all intensive purposes, it’s a doggy dog world” is, in some sense, an English expression.
Anyway, I find that “octopi” is generally used by people going for a certain sense of superiority they think the word has over “octopuses”. An air of refinement.
They may be interested to know that it does not grant this impression to people who know the correct form.
So’s “fuck off, you smarmy asshole”.
Sorry, I misunderestimated you.
Octopus was coined into neo-Latin from Greek in like the seventeenth century. Whatever the faults of those who prefer “octopi,” it would seem a little hasty to characterize them as bourgeois pretenders to classical education.
Anyway, a little google-booksing shows that octopi is all the hell over the place, so it is certainly a word, with stronger claims than octopodes which is at best half as well attested. The charitable explanation would be that octopi, formed by analogy with other Latinate plurals drilled in elementary school, sounds cleaner than octopuses. Here and elsewhere, word-final s causes aesthetic offense in the case of plurals and possessives, and some speakers avoid it where possible, and Latin/Greek third-declension plurals are too rarely employed in English to constitute a go-to solution that any critical mass of speakers would be familiar with.
I have never used the word octopi in an attempt to look learned, nor have I ever taken anyone else’s use of it as such an attempt.
I use it because it’s funny.
Being from New Orleans, where the correct pronunciation of everything, including the name of the town itself (New OR-linz, not New Or-LAY-a~, nor, god forbid, N’awlins), is some form of butchered French, I am not usually very concerned about the grammar or pronunciation of a word’s language of origin once it has been thoroughly anglicized.
Brain size relative to body size is a good but not perfect predictor. I agree that not every animal is limited by birth canal. I think that there’s a hard problem of doing real engineering work to get high intelligence, and several easy hacks that move intelligence around the space defined by your current position in the hard problem. For example, you can add +10 IQ points by giving someone torsion dystonia, but then they’ve got torsion dystonia. You can add +100 IQ points by giving someone von Neumann’s genome, but that only lasts one generation before they have to marry someone else and all those genes get redistributed again in a chaotic way.
Select for high intelligence over a few generations and you get the easy hacks. Select for it over eons, and you get breakthroughs on the hard problem. I think modern humans are a combination of both. But there’s a lot of room for more easy hacks, and there would be more room if we could solve problems like torsion dystonia and head size and genetic entropy or if it were in a computer-type entity that didn’t need to solve them.
The birth canal is much less of a limit these days given C-sections.
By any reasonable standard, human+ AI is an advanced alien species.
When you think about it in these terms, it’s pretty strange that most people are completely unconcerned with researchers fooling around with the genetic code of highly advanced aliens, whereas we’re *really* concerned about researchers making minor changes to human genes.
Hard takeoff from flatworm to superhuman intelligence which inevitably wipes out humanity is a very scary possibility. However, the idea that Dr. Evil et al. somewhere in an underground lab invent a general algorithm for an intelligence improving itself and re-writing its code without any exposure to a real world seems too fantastic.
I’m more worried about existing environment for AI to evolve. There are rats, raccoons, pigeons and numerous bugs in modern cities, and humans don’t do much to control their population. I wonder, are there similar ecological niches for wild sub-human AIs in today world? And when do we expect them to be populated? And how we’ll control evolution of such free roaming AIs?
I was already considering incorporating something similar into the novel and rpg-setting I’ve been working on but I wasn’t sure exactly how they fit into the wider environment.
Computer vermin (vice computer virus) is just too good an image not to use.
We already have numerous anonymous cryptocurrencies, which government have very little control of. We already have an overlay network specifically designed to withstand censorship. We are moving very fast towards implementing a smart-contracts platform. What we still lack (AFAIK) is a market for secure computation services.
Once we’ve got all these ingredients, there would be a breeding enviromnent for an AI. Imagine power of open-ended virtual evolution competing with humans on prediction markets. At some point AIs will be major players on the markets. And given the resilience of the infrastructure we are building now, there would be no simple way to stop AIs from proliferating – not until we as humanity unite and solve all our social problems.
I think this “AI in the wild” scenario is way more realistic than scenario when AI is raised to a superhuman state in a box.
Another option: Maybe these guys think hard takeoff AI is either unstoppably evil or a crapshoot anyways, so they’re focusing on the alternate scenario because we can’t do much about this case anyways.
Tangentially, I find it very funny that I wrote two novels which had as a sub-plot the following:
1) There is a simple (it’s not easy: it is simple) way to make current computer technology work so well that it can sustain sentient AI starting with not much more than Game of Life level programming skills.
2) In response to this, the governments of the world all have secret paramilitary law enforcement units whose sole purpose is to watch out for people who learn #1 and stop them before they create Skynet by accident.
3) There is an easy (it’s not simple: it is easy) way to maintain absolute control of an AI, which nobody knows about. Well, almost nobody.
… long before I ever heard of any of these movements/countermovements, Dr. Evil, Dr. Good, or Dr. Amoral. I don’t know why, but I do. 🙂
Citation needed! 🙂 I quite enjoyed My Mother Had Me Tested.
I am very flattered, thank you, but the novels are not published. 🙂
Need a proofreader? 🙂