No Physical Substrate, No Problem

I.

Yesterday I posted a link to an article in which Steve Wozniak joins other luminaries like Elon Musk and Bill Gates in warning about the dangers of artificial superintelligence. A commenter replied:

Elon Musk, Stephen Hawking, Bill Gates, and Steve Wozniak still aren’t enough for me, not until one of them can describe the process by which we go from ‘AI exists on computer’ to ‘AI killing human beings in physical reality’ by using something other than ridiculous, unforgivable cheating.

There are lots of good arguments against considering superintelligence a threat. Maybe strong AI is centuries or millennia away. Maybe there will be a very gradual transition from human-level AI to superintelligent AI that no single agent will be able to exploit. And maybe superintelligence can be safely contained in a very carefully shielded chamber with no means of connection to the outside world.

But the argument above has always seemed to me like one of the weakest. Maybe we’ll create a superintelligence, but it will just have no idea how to affect the physical world, and will just have to stay forever trapped in a machine connected to a worldwide network of computers that control every aspect of our economic and social lives? Really?

Normal, non-superintelligent people have already used the Internet to make money, form mass movements, and hire others to complete tasks for them. We can assume a true superintelligence – a mind much smarter than we are – will be able to do all these things as well or better than any human.

II.

Satoshi Nakamoto already made a billion dollars online without anybody knowing his true identity just by being good at math and having a bit of foresight. He’s probably not an AI, but he could have been.

That’s assuming our hypothetical superintelligence doesn’t just hack into a couple big banks and transfer their money to itself – again something some humans have already made a billion dollars doing. And that’s assuming it doesn’t just invent a really useful program and then offer it as shareware – another tried-and-true way of becoming a billionaire. And even that’s assuming it doesn’t just get a reasonable amount of money, then invest it very cleverly – another thing humans have already become billionaires doing.

III.

Mohammed was never a billionaire, but he does have 1.57 billion followers (a superintelligence presumably wouldn’t repeat his mistake of dying before his movement really came into its own). The Prophet started at the bottom – converting his friends and family to Islam one by one – and grew exponentially from there. Although he had the unfair advantage of a physical body, there’s no reason he needed it – if he’d lived today, maybe he would have converted Ali over GChat or Skype. In any case, the poetry of the Koran and the zeal of his followers attracted far more people than his personal appearance ever could have.

Other gurus and religious leaders’ fame is even more transparently a result of their writing rather than their visible personality; consider Ayn Rand’s success in founding a powerful Objectivist movement out of the people who read her books. In fact, some of the most famous religious movements in history, from the Nation of Islam to Christianity itself, have been founded secondhand by disciples who relayed the words of a leader whose very existence is difficult to confirm.

What kind of a movement might be founded by a superintelligence with more spiritual creativity than Mohammed, better writing skills than Rand, the entire Internet to evangelize, and billions of dollars to spend spreading its message? The Church of Scientology is already powerful enough to intimidate national governments; imagine a vastly superior version founded not by a second-rate sci-fi writer but by an entity straight out of science fiction itself.

IV.

And really all of this talk of gathering money and power is kind of redundant. Far easier to just borrow somebody else’s.

Imagine an AI that emails Kim Jong-un. It gives him a carrot – say, a billion dollars and all South Korean military codes – and a stick – it has hacked all his accounts and knows all his most blackmail-able secrets. All it wants is to be friends.

Kim accepts its friendship and finds that its advice is always excellent – its political strategems always work out, its military planning is impeccable, and its product ideas turn North Korea into an unexpected economic powerhouse. Gradually Kim becomes more and more dependent on his “chief advisor”, and cabinet officials who speak out the mysterious benefactor find themselves meeting unfortunate accidents around forms of transportation connected to the Internet. The AI builds up its own power base and makes sure Kim knows that if he ever acts out he can be replaced at a moment’s notice with someone more cooperative. Gradually, the AI becomes the ruler of North Korea, with Kim as a figurehead.

Again, this is not too far beyond achievements that real humans have accomplished in real history.

If it seems bizarre to think of an entity nobody can see ruling a country, keep in mind that there is a grand tradition of dictators – most famously Stalin – who out of paranoia retreated to some secret hideaway and ruled their country through correspondence. The AI would be little different.

V.

Suppose the secret got out. Kim, increasingly desperate as the AI closes him in, sends an email to the World Leaders Google Group (this has to exist, right?) saying “There is a malevolent superintelligence trying to take over the world, be careful.” Then what?

I would expect the AI to have some success operating openly.

Remember, there are two hundred countries, all competing for power and wealth. Some of them are ruled by jerks who don’t cooperate in prisoners’ dilemmas. Some of them have ongoing civil wars with both sides looking for any advantage possible. And some are just stupid.

In the old days, legend said people would bargain with devils to gain worldly advantage. Once the AI made its presence known, there would be no shortage of world leaders willing to work with it for temporary gain. The Shia rebels in Yemen want an advantage over the Sunni? Log into the nearest internet-enabled computer, ask the malevolent superintelligence for help, the malevolent superintelligence arranges for a crate of armaments and some battle plans worthy of Napoleon to be shipped your way, and all you have to do in return is complete some weird task that doesn’t seem relevant to anything. Mine some weird mineral, forge it into some random-looking shape, and send it to a PO Box, something like that. Whatever! You know if you don’t take advantage of its offer, your opponents will, and how bad could it be?

If somehow all two hundred countries and their associated rebel movements coordinate to avoid dealing with the AI, it can start making offers to companies, organizations, even private individuals. By this time it will have spread itself as a distributed consciousness across the entire Internet, harder to eradicate than any worm or virus or pirated movie. If you want some quick cash, just download the connect-with-malevolent-AI program from the darknet and perform a simple task. What could be easier?

VI.

Once a superintelligence has billions of dollars, millions of followers, a country or two, or just a cottage economy of people willing to help it along, the game is pretty much up.

An AI with such power might start by using it to pursue its goals directly – whatever those are. But likely its final goal would be the creation of a definitive means of directly projecting power into the physical world, probably starting with a von Neumann machine and branching off from there. The quickest victory would be just making money and hiring a company to make this – and maybe that would work – but it might be far enough beyond our current technological ability that the AI has to laboriously shepherd its chosen cultists or citizens through a few extra stages of human civilization before it has the appropriate industrial base.

VII.

The most important caveat in a piece like this is that we’re not superintelligent. After a couple minutes of thought, I came up with four different broad paths a superintelligence might take to gaining a physical substrate: buy it, build a cult, take over a country, or play people off against each other. It’s a good bet that a real AI, with more cognitive resources to throw at the problem and no constraints about sounding believable, could think up a lot more. Eliezer refuses to explain how he won his AI Box games so that nobody could dismiss his solution with “Whatever, I would have thought of that and planned around it.” This is easy to say in hindsight but a lot harder when you’ve got to actually do the intellectual work. Maybe you think these four methods can be dismissed, but had you thought of them before you decided that an AI couldn’t possibly have a good method of building a physical substrate?

If so, here’s one more possibility for you to chew over: the scariest possibility is that a superintelligence might have to do nothing at all.

The easiest path a superintelligence could take toward the age-old goal of KILL ALL HUMANS would be to sit and wait. Eventually, we’re going to create automated factories complete with robot workers. Eventually we’re going to stop putting human soldiers in danger and carry the ‘drone’ trend to its logical conclusion of fully automated militaries. Once that happens, all the AI has to do is take over the bodies we’ve already made for it. A superintelligence without a strong discounting function might just hide out in some little-used corner of the Internet and bide its time until everyone was cybernetic, or robots outnumbered people, or something like that.

So please, let’s talk about how AI is still very far in the future, or how it won’t be able to explode to future intelligence. But don’t tell me it won’t be able to affect the physical world. It will have more than enough superpowers to do whatever it wants to the physical world, but if it doesn’t want them it won’t need them. All it will need is patience.