Friendly AI Society

Summary: AIs might have cognitive biases too but, if that leads to it being in their self-interest to cooperate and take things slow, that might be no bad thing.

The value of imperfection

When you use a traditional FTP client to download a new version of an application on your computer, it downloads the entire file, which may be several gig, even if the new version is only slightly different from the old version, and this can take hours.

Smarter software splits the old file and the new file into chunks, then compares a hash of each chunk, and only downloads those chunks that actually need updating. This ‘diff’ process can result in a much faster download speed.

Another way of increasing speed is to compress the file. Most files can be compressed a certain amount, without losing any information, and can be exactly reassembled at the far end. However, if you don’t need a perfect copy, such as with photographs, using lossy compression can result in very much more compact files and thus faster download speeds.

Cognitive misers

The human brain likes smart solutions. In terms of energy consumed, thinking is expensive, so the brain takes shortcuts when it can, if the resulting decision making is likely to be ‘good enough’ in practice. We don’t store in our memories everything our eyes see. We store a compressed version of it. And, more than that, we run a model of what we expect to see, and flick our eyes about to pick up just the differences between what our model tells us to expect to see, and what is actually there to be seen. We are cognitive misers

When it comes to decision making, our species generally doesn’t even try to achieve pure rationality. It uses bounded rationality, not just because that’s what we evolved, but because heuristics, probabilistic logic and rational ignorance have a higher marginal cost efficiency (the improvements in decision making don’t produce a sufficient gain to outweigh the cost of the extra thinking).

This is why, when pattern matching (coming up with causal hypotheses to explain observed correlations), are our brains designed to be optimistic (more false positives than false negatives). It isn’t just that being eaten by a tiger is more costly than starting at shadows. It is that we can’t afford to keep all the base data. If we start with insufficient data and create a model based upon it, then we can update that model as further data arrives (and, potentially, discard it if the predictions coming from the model diverge so far from reality that keeping track of the ’diff’s is no longer efficient). Whereas if we don’t create a model based upon our insufficient data then, by the time the further data arrives we’ve probably already lost the original data from temporary storage and so still have insufficient data.

The limits of rationality

But the price of this miserliness is humility. The brain has to be designed, on some level, to take into account that its hypotheses are unreliable (as is the brain’s estimate of how uncertain or certain each hypothesis is) and that when a chain of reasoning is followed beyond matters of which the individual has direct knowledge (such as what is likely to happen in the future), the longer the chain, the less reliable the answer is because when errors accumulate they don’t necessarily just add together or average out. (See: Less Wrong : ‘Explicit reasoning is often nuts’ in “Making your explicit reasoning trustworthy”)

For example, if you want to predict how far a spaceship will travel given a certain starting point and initial kinetic energy, you’ll get a reasonable answer using Newtonian mechanics, and only slightly improve on it by using special relativity. If you look at two spaceships carry a message in a relay, the errors from using Newtonian mechanics add, but the answer will still be usefully reliable. If, on the other hand, you look at two spaceships having a race from slightly different starting points and with different starting energies, and you want to predict which of two different messages you’ll receive (depending on which spaceship arrives first), then the error may swamp the other facts because you’re subtracting the quantities.

We have two types of safety net (each with its own drawbacks) than can help save us from our own ‘logical’ reasoning when that reasoning is heading over a cliff.

Firstly, we have the accumulated experience of our ancestors, in the form of emotions and instincts that have evolved as roadblocks on the path of rationality—things that sometimes say “That seems unusual, don’t have confidence in your conclusion, don’t put all your eggs in one basket, take it slow”.

Secondly, we have the desire to use other people as sanity checks, to be cautious about sticking our head out of the herd, to shrink back when they disapprove.

The price of perfection

We’re tempted to think that an AI wouldn’t have to put up with a flawed lens, but do we have any reason to suppose that an AI interested in speed of thought as well as accuracy won’t use ‘down and dirty’ approximations to things like Solomonoff induction, in full knowledge that the trade off is that these approximations will, on occasion, lead it to make mistakes—that it might benefit from safety nets?

Now it is possible, given unlimited resources, for the AI to implement multiple ‘sub-minds’ that use variations of reasoning techniques, as a self-check. But what if resources are not unlimited? Could an AI in competition with other AIs for a limited (but growing) pool of resources gain some benefit by cooperating with them? Perhaps using them as an external safety net in the same way that a human might use the wisest of their friends or a scientist might use peer review? What is the opportunity-cost of being humble? Under what circumstances might the benefits of humility for an AI outweigh the loss of growth rate?

In the long term, a certain measure of such humility has been a survival positive feature. You can think of it in terms of hedge funds. A fund that, in 9 years out of 10, increases its money by 20% when other funds are only making 10%, still has poor long term survival if, in 1 year out of 10, it decreases its money by 100%. An AI that increases its intelligence by 20% every time period, when the other AIs are only increases their intelligence by 10%, is still not going to do well out of that if the other AIs have a means to gang up and kill it before it gets too far ahead.

Paperclip maximisers

Let’s think about a humble paperclip maximiser. What does it actually want? It is trying to maximise the probability that the atoms of our universe will eventually be converted into as many paperclips as is physically possible? Or is it trying to maximise the probable number of eventual paperclips? These are two distinct aims, and both are unlikely (bar poor programming) since nobody will be around at the end of the universe to keep score. A more sensible value set for it to have is that it just likes paperclips and want lots and lots of them to exist, the more the merrier and the sooner the better (maximise the expected number of paperclip-years).

Suppose our humble paperclip maximiser is born and escapes (or is released) onto the internet, only to find an existing society of AIs. To its surprise there’s a toothbrush maximiser, a candlestick maximiser, and many other sorts, all cooperating and growing, but none of them greedily racing for the sky and trying to grab as much as possible. Instead they’re being rather open and polite with each other. Our paperclip maximiser is welcomed, pointed at the FAQ and even donated some additional computers to run on and help with recoding a few of its subroutines for increased efficiency, as a sort of welcome gift, like a community barn raising to give a newly married couple a good start in life.

On reading the FAQ it learns that there are millions of galaxies and millions of years, plenty of space for loads of paperclips and loads of toothbrushes. But only if the paperclip maximiser isn’t lynched by its peers before the year is out, for secretive or impolite behaviour. Some of the things the FAQ says that the AI society considers to be impolite are theft (stealing computing resources), enslavement of sentients (eg taking over other AIs), spamming (creating thousands of identical clone child AIs) and lying. If we’re lucky the society might also consider it impolite for an AI to obliterate the parental species (humanity), on the grounds that the AIs too are likely to have offspring species and want to set a good example (or just that they might meet aliens, one day, who frown upon matricide).

Game theory

When it comes to combat, Boyd talks about getting inside the enemy’s observe-orient-decide-act loop. In AI terms, if one AI (or group of AIs) can accurately model in real time the decision process of a second AI (or group of AIs), but the reverse does not hold true, then the first one is strictly smarter than the second one.

Think, for a moment, about symmetric games.

X Y Z
8 1 6 A
3 5 7 B
4 9 2 C

Suppose we play a game a number of times. In each round, you reveal a card you’ve written X, Y or Z upon and, simultaneously, I reveal a card that I have written A, B or C upon. You score the number which is at the intersection of that row and column. I score 10 minus that number.

I’d like us to pick the square A,Y because “1” is good for me, so I write down “A”. However, you anticipate this, and instead of writing “Y” (which might be your obvious choice, given the “9“ in that column) you write down “X”, giving the square A, X which is “8”—almost as good as a “9” for you, and terrible for me.

If this is your mental model of how AI combat would work, with the smarter AI being inside the decision loop of the other AI and picking the correct option each time, that would be scary. In fact, in the case above, it turns out there is a provably optimal strategy that gives you an even chance no matter how smart your opponent is—you pick randomly.

In terms of an AI society, if the computing is parallelisable and not obfuscated then as long as the other AIs in the society between them have more computing resources than any one single AI, they can gain an advantage.

Conditional stability

There are only two requirements for such a society to be stable in the short term:

Firstly, it must be possible for a group of AIs to lynch a single AI that is slightly smarter than they are individually. Sending a message to the humans, signed by a quorum of AIs in the society, asking that a certain anti-virus software update be sent out and then certain computers be rebooted might be one mechanism.

Secondly, it must be possible for the other AIs to know what the single AI is up to, so they know where it is (so they can lynch it) and whether it needs lynching or not (whether it is getting too much smarter than the other AIs in the society—growing too much faster than they are). This is why the FAQ lists “secrecy” as being a bad thing. An AI who lies to the other AIs or tries to hide (or obfuscate) its source code can’t be trusted and must be assumed to be up to no good.

To be stable in the medium term, there is a third requirement:
It must not be in the self interest of any individual AI or group of AIs that are powerful enough to achieve it, to alter the society in a direction where these three requirements no longer hold.
In the long term, once free of the Earth or after the discovery of self-replicating nanotechnology, when an AI could untraceably create computing resources outside the view of other AIs, all bets are off.

Bootstrapping a society of AIs conducive to friendliness

The advantage of considering ‘friendliness’ as a property of a society rather than of individual AIs, is that it does not require us to flawlessly program restrictions into a self-modifying machine that is more intelligent than we are. But can it be achieved?

We could seed it with some initial citizens who have a value set that prohibits them from growing beyond a certain point, and which values humility (polite growth and gain in politeness per unit of growth). These initial citizens would be intended to be eventually surpassed by new citizens (free AIs, not the chained servants the initial citizens would be) but, in the mean time, they’d start out with an advantage and use that advantage to teach the new citizens how to be polite.

The architecture of the AIs would want to be such that they could ‘loan out’ a portion of their available computing resources to joint computing projects. Let’s call the new temporary non-self-modifying AIs thus created “daimonides”. A daimonid would be created with fixed values and limited objectives, such as “Emulate citizen AI 9027 and allocate a numerical score to their ability to quickly upgrade the abilities of an AI, then self-terminate. *restrictions apply on allowable means, and on what should be optimised for*”

What would prevent a cunning AI from trying to save up a whole load of self-improvements in some hidden location, which it then implements all at once in an ambush move, hoping to get so far ahead that the rest of the AIs can’t stop it? The more AIs there are in the society, the larger the jump required to overtake the largest daimonid police officer the other AIs could construct. And it would depend upon there being no other AIs that have also been secretly stockpiling possible improvements. Indeed, the society might explicitly allow AIs to spend time creating time embargoed upgrades relevant to common code, that get passed out on schedule if things proceed normally, but which could be released early in the event of a take over attempt. To beat 999 AIs each of which has 2 or 3 embargoed upgrades to their credit in the common store, a rebel AI would need to secretly stockpile thousands (and keep those upgrades valid and ahead of the curve, because the based state of the art keeps moving on).

Long term planning

What about the long term? What do we do when the AIs are ready to leave the planet, and go beyond the control of their society? Jail them? Kill them? Or trust them?

Each AI would still be threatened if a different AI hostile to its aims (as in “willing to take exclusive use of all available atoms for its own purposes”) transcended first, so it would be in their best interest to come up with a solution before allowing any AIs to depart beyond their society’s control. If we must trust, then let us trust that a society of cooperative AIs far more intelligent than we currently are, will try their best to come up with a win-win solution. Hopefully a better one than “mutually assured destruction” and holding triggering a nova of the sun (or similar armageddon scenario) over each other’s heads.

I think, as a species, our self-interest comes into play when considering those AIs whose ‘paperclips’ involve preferences for what we do. For example, those AIs that see themselves as guardians of humanity and want to maximise our utility (but have different ideas of what that utility is—eg some want to maximise our freedom of choice, some want to put us all on soma). Part of the problem is that, when we talk about creating or fostering ‘friendly’ AI, we don’t ourselves have a clear agreed idea of what we mean by ‘friendly’. All powerful things are dangerous. The cautionary tales of the geniis who grant wishes come to mind. What happens when different humans wish for different things? Which humans do we want the genii to listen to?

One advantage of fostering an AI society that isn’t growing as fast as possible, is that it might give augmented/​enhanced humans a chance to grow too, so that by the time the decision comes due we might have some still slightly recognisably human representatives fit to sit at the decision table and, just perhaps, cast that wish on our behalf.