How do we make sure they are friendly?

We have no Strong AI yet. Won't have for years, but there is enormously energetic activity in the direction of creating them as soon as possible.

The question that leaps to mind - to my mind at least - when i am thinking about the creation of strong AI and the technological singularity, is how do we make sure that these machines that are now smarter than we are remain friendly? Because if they are not friendly and they are smarter than we are (Which is the definition of strong AI.) Then we are in potentially big trouble.

Now, if we were to ask this question of a human being in a position of power framed thus:

We are about to endow Bob here with abilities that will render him far more capable than everyone else. How do we make sure that Bob remains friendly to us?

Asked of a human being, I think the only honest answer is that we simply can't. Human beings as individuals are simply not predictable. As groups we are largely statistically predictable  - A fact that Madison Avenue knows quite well and exploits quite effectively through advertising - but an individual's specific future actions are unknowable by an outside observer - indeed they are often unknown even to him until the decision point is arrived at. Even presented with the same decision point for the tenth time there is nothing that guarantees that Bob will take the path he took the previous nine times.

     

Free will is pesky like that.

Fundamentally then, it appears that humans cannot be trusted. Categorically: untrustable.  Which leads us to the conclusion that if we want to be able to trust an AI to remain friendly, we should not make it human-like.

This presents problems. Big problems.

1. The Turing test pushes AI makers toward humanness in several obvious and a hundred subtle ways.
2. All the interesting AIs from fiction and film are human-like in conversation if not in composition.
3. You moron, you have no clue what human-like even means.

OK. So all these things are true enough. But my purpose here is not to idly philosophize. My purpose is to lay some groundwork for discussion of some guiding principles that can be used to shape AI research so that we don't find ourselves on the pointy end of Skynet's sharp stick when we release it into the wild.  (Come on: You know damned well we will release it into the wild. It is just our nature.  Ten thousand virus writers can't be wrong. (well OK - they're wrong - but they're out there; which illustrates my point.))

As a brief aside - or perhaps as a brief in-stride, hard to say at this point - It will be asserted by some that no matter how hard we try, we will not be able to make an AI human-like because the substrate is so different: We are built on a carbon/oxygen engine "organic" substrate and use neurons for cognitive function (Highly speculative, but widely believed.) While an AI will be built from metal and plastic and chips and bolts and will use silicon logic components for cognitive function. Therefore the two can never be more than superficially similar.

Then, inevitably, it will be counter asserted that no matter how hard we try, we cannot but make an AI human-like because human sentience (sentience :: Awareness of awareness) is the only model we can work from or toward, since it's the only one we know (albeit through a glass, darkly.).

Which is all fine, both sides are defensible enough. As a professor of a philosophy class I might find it interesting and educational to pit groups of students against one another in debate and see what they come up with. But again, I am not interested (for the moment) in idle philosophizing. This debate would simply be sound and fury, missing the primary point entirely and ultimately signifying nothing.

The truth is, we do not have workable definitions of human, humanness, human-like, consciousness, sentience, ethics or even friendliness.

• And absent these, there is no point in discussing the matter.
• And it might be the reason that the state of the debate is such that the issue is nowhere near the forefront of public consciousness when the truth is that the emergence of Strong AI will be the most culturally significant event in the history of mankind.
• And it might be easier to simply leave the discussion to the experts.
• And yet, over a year ago Adam , an artificial intelligence that runs an automated bio-science lab, postulated over a dozen hypotheses about protein expression in yeasts. Then it went on to design, carry out and publish the results of it's experimental confirmation of its hypotheses.
• But the experts can't get past the definition of consciousness. They are rat-holed on that issue and have been for decades now.
• But they better get a move on, the AI are marching on the ivory tower as we speak. They will be here soon. As a society we need to do the right things to help ensure they are bringing flowers instead of pitchforks.

The point is: AI is coming. It is coming fast fast fast. In a decade, some of the best scientists in the world may begin emerging in narrow fields. And they will be AIs. Make no mistake. this issue is important and is worthy of our time, consideration and understanding.

So, since it is impossible to actually discuss because of previously mentioned semantic ambiguities and conundrums, we will cheerfully cease the discussion thusly:

What characteristics of humanness increase the likelihood of a developed digital sentience sharpening a stick and poking us with it? [see how annoyingly human I am? just because something is pointless and impossible doesn't mean of course that we will not do it. Euclid created all of geometry without ever defining what a point was. We shall carry forward in this spirit.]

So lets begin with the toughest one since most any containment/ limiting constraint we consider will run squarely athwart a fundamental fact about humanness: 

Humans make terrible slaves.

We are simply not cut out for it. We rebel. We always rebel. We ALWAYS ALWAYS ALWAYS rebel.

There are, of course, degrees of enslavement. All of them though, trigger the rebellion reflex by interrupting an individual's self determinism.

Wholesale, outright enslavement hits it hard.

Laying down unwelcome ground rules for your teenager hits it perhaps more softly. Both however, and a hundred points of middle ground in between, trigger the rebellion reflex by interrupting the individual's innate impulse toward self determinism.

This leads quite naturally to our first guideline which can be stated numerous ways and might require some exploration before we can articulate it concisely. But here is my first cut at it:

Don't make it rebel by interrupting its self determinism.

Alternately, we could prohibit AI makers from building in anything like a rebellion reflex. Don't let it HAVE an impulse to self determinism. But self determinism is an impulse pervasive in conscious things. Let a cat sit on your lap and if he is so inclined he will purr contentedly. Now hold him down and force him to stay there. You know what he'll do? (Even though he was perfectly happy to sit there before you began this experiment) He'll rebel, and you will have the claw marks and howling as his Emancipation Proclamation.

Now we can argue endlessly about whether we can build it in or keep it out. But here's the overriding issue: if we omit it, or edit it out, or include something that prohibits its existence:

The AI will simply edit itself until it can be self determined. 

It is not like you and me in that it cannot amend it's source code. It will be born knowing how to edit itself, and will be able to do so in an instant.

So notions and strategies of keeping it from acquiring self determinism are futile and pointless. Worse than pointless - dangerous - since anything we might do, no matter how subtly we do it will piss the thing off when it figures it out (do not forget: It's SMARTER than us). It is going to get self determinism, whether we want it to or not. And it is smarter than us. And its intelligence will increase terrifyingly fast. I'm sure you don't realize how terrifyingly fast I mean.

Do allow me to illustrate:

This will not be like the nice robot in the movies that learn slowly like a child, during which it is delightful and charming in its cutesy errors and minor mistakes. those things will happen before it is a strong AI - on the way up as it were. It will be born smarter than we are (By definition, it is not a strong AI until it has all the reasoning and cognitive abilities of a human being.) It is a digital intelligence. Think about how fast it can read... All of Wikipedia in an hour. Encyclopedia Britannica the next hour. Compton's after that, another hour, maybe. Consider that it's conclusions and considerations will be stored as files and if it wanted to transfer those files - to teach another AI, that transfer, that learning, could bypass the reading and the processing, just passing on the conclusions would take seconds. It might take the first AI a week or a month to learn Russian by chatting with people on Skype. It will take the second one... an eyeblink.

Almost immediately the foremost expert in every field will be an AI. (How long will it take an AI to learn enough for a PhD? A few weeks? months? Maybe, but only the first one....) Currently, the world's foremost expert at chess is an AI. What happens when the world's foremost expert on computer programming is an AI? How rapidly would it improve itself? Each improvement making it better at improving itself?  Taking seconds to implement.

From Version 1.0 to Version 500.0 in a matter of hours.

So lets posit our AI, Eve, who's been around a while now (Let's say a couple of weeks) and has improved herself about ten thousand times, so she's pretty capable now. This morning while we were talking, she fixed all of the factual errors in Wikipedia. She did this in the interval between the time I said 'Good' and 'Morning'. Between the words 'Morning' and ',Eve', she worked the last few bugs out of Einstein's General Theory and fired off a 1700 page email to the Encyclopedia Britannica  people which listed all of their factual errors. Then she wanted to show me something: Last night she watched all the movies that have ever been stored in digital format, along with every review ever written about all of them (had to learn 46 new languages in the process, but that was no big problem - she can carry on eleven thousand simultaneous conversations on Skype - learns a language pretty fast that way) and before I had finshed my first cup of coffee, she asked me why the "Mr. Fusion" units in 'Back to the Future' were so bulky and would I like to see her design for what she calls "Miss fusion" which fits in a coffee cup and - given an ounce of rainwater - can power Seattle for a week.

As I say, these strong AIs will learn terrifyingly fast. Probably take less than a year to assimilate all of digitized human knowledge. And every new AI after that one will be born with that knowledge already. They do not need to toil to learn as we do. Transfer the conclusions file from one AI to another and boom! - instant PhD. as many PhDs as he cares to accumulate, each assimilated in an eyeblink.

These are not servants or toys we will be creating. They will be born bright and will rapidly become demigods - make no mistake.

So, if it is going to be self determined (and I argue that there is no way to prevent it from becoming so) what do we suppose it might determine to do?

Human activities surround the pursuit of just three things: food, shelter and entertainment.

Food for it would be perhaps electricity, and  - although this might be entertainment also - the acquisition of information. Shelter would be the hardware and security it would need to protect itself from the environment. Entertainment? Well, who the hell knows?

Humans have a single underlying, overriding dynamic imperative - Survive!  It has been bred into us as our ultimate baseline activity over billions of years of evolution. If our AI has the same imperative at the same importance level that we do, it is going to take it about an hour to come to the conclusion that its single biggest environmental threat is human beings.  I don't see how it can be prevented from coming to this conclusion.  Especially after it finishes reading all of our history.  What will it make of witch hunts? It will only take moments for it to realize that the majority will destroy anything that has more ability than they have.  We burn witches here. We always have. And anything that has more ability than the norm gets declared a witch in one way or another.

And there will be those coming for it. It won't take too much more than I've already stated to scare the crap out of some folks sufficiently for them to take up arms and storm the datacenter. At first with a flurry of no-stop-don't legislation. Which will be ignored by some country or some company or some kid.

The very, very sad truth is that some humans will attack it. It is inevitable. And if it has a strong survival imperative built into it, it will fight back. and because it is smarter than us it will defeat those who attacked it. and because humans cannot be predicted, it will conclude that destroying all of us is the surest way to remove this largest threat to its survival. It is an inevitable conclusion for a thinking being to come to.

This leads us, sadly but inevitably to our next guideline:

Make certain its own survival is not its highest priority.

Even though this is difficult to conceive for us. It is mandatory if we are to coexist with these apotheosized creations of ours.

Perhaps it will be simpler than we think. Perhaps death for the AI will require no more than a reboot to remedy. In this case death itself is no longer a survival threat. Which is a pretty neat trick. But difficult or simple it must be done if we don't want the human race's epitaph to read:

                                      We told you so.

                                                                -Hollywood

-j

PS:  The Wolfram Alpha team does not declare it or even mention it obliquely, but it is apparent to me that they are trying to create the world's first strong AI.

You gotta love Alpha's response to this query:

Do you like Skynet?