The Automation of Warfare Friday, Jan 19 2007
friendly ai and warfare 1:05 am
Reading today’s post over at the Center for Responsible Nanotechnology reminded me of this image. The topic at hand was the mechanization of warfare, and the worrisome development of an auto-turret that is accurate at up to half a mile. The image is of Metal Gear D, from the video game series of the same name (minus the D). If human civilization continues as it has, then there are a lot of machines like this in our future. Over at the Lifeboat Foundation blog, I have a few words to say on the militarization of space.
There is a conundrum in the concept of the arms race. The best way to keep the world safe is to have the biggest guns, period. But it’s also an easy way to destroy the world. Keeping the world safe with mere words is impractical. Thus, someone must always have the biggest guns. Anti-authoritarian, well-intentioned people like to whine at length about this. But we have to accept that unless someone keeps the peace, the natural tendency is descent into conflict. All we can do is try to steer things such that the most powerful agent(s) at any given time are truly benevolent.
The best way to accomplish that is not to endlessly shuffle through futile anthropocentric political arrangements, but to actually change the cognitive architecture underlying the most powerful agents to make them more benevolent by nature. You could theoretically do this with enough progress in neuroengineering, but building a Friendly AI just seems easier. Is FAI possible? Yes. Here is a page that argues why.
January 19th, 2007 at 10:32 am
In the auto-turret article much stall is put into the fact that it’s still humans making the firing decisions and identifying military targets. However, under the high pressure conditions of a battle I can imagine such manual targeting niceties being soon dispensed with. Vehicles like this might simply be driven en masse into enemy territory and instructed to exterminate any observed person. Once in the battle zone each machine would either perish or exhaust its ammunition and return to a designated automated resupply point.
January 20th, 2007 at 3:43 am
I’m skeptical of the concept of engineering a “friendly AI.” (Or GAI or whatever they’re calling it now.)
I can’t really put my finger on anything specific but it seems sort of like “raising children who like you and who get scholarships to MIT.” Children are complicated and surprising. They never turn out the way you expect.
What makes us presume we can somehow control the development of superhumanly intelligent organisms to such a fine degree as to ensure they’ll be nice to us?
January 20th, 2007 at 9:28 am
I don’t think that that page actually argues that it *is* possible, just that *if* it were possible, that would be how it would be possible. I don’t think we’ll actually know if it’s possible until we succeed.
January 20th, 2007 at 9:30 am
Pace: did you read the linked article? It answers your objections nicely. Short answer: prove it mathematically. Yes, still really hard. But it has some hope of working.
January 20th, 2007 at 9:42 am
I think the solution is to develop better defense technologies. Definitely not bigger guns. It seems very risk to depend on the benevolence of the ones holding the guns.
The better defensive technologies we have, the more likely we are to protect our freedom.
Think of the ultimate - if defense trumps offense through human-augmentation, upload/backup or shields. In such a scenario, we can maintain individual or small-group autonomy. You can’t be intimidated or coerced if you are shielded from the effects violence.
“big guns” will lead to de-stabilization, either in the case of proliferation or untraceability. Both of these are likely with MNT and advanced biological weapons.
January 20th, 2007 at 9:45 am
“I can’t really put my finger on anything specific but it seems sort of like “raising children who like you and who get scholarships to MIT.” Children are complicated and surprising. They never turn out the way you expect.”
AI and children are totally different- more different from each other than you are from a lizard. We can write an AI from scratch, specifying every bit, while a child’s DNA is already specified.
January 20th, 2007 at 10:19 am
I remain skeptical of the mathematical proof aproach to creating friendly AIs. You can certainly do this for relatively small formal systems, such as those used in aircraft or rocket control systems, but AI is really a different ball game. Once you have a system which can learn from experience it’s very difficult to predict in advance how it will behave in the long run. Furthermore the world itself is quite an unpredictable place, as anyone who has ever done some robotics will be able to tell you. Even if the machines behavior is totally deterministic, with no capacity for learning, the complexity of the external world interacting with the machine can produce surprising duets.
January 20th, 2007 at 11:45 am
Bob: I’m skeptical as well, but I think it’s an extremely important line of research: it could be our only hope, if most of the kind of AIs humans might end up developing would undergo a hard, fast takeoff.
But the essay Michael links has a good analogy, in the very first section, that might help you understand why this is even plausibly possible. I will refrain from pasting it here.
January 20th, 2007 at 12:06 pm
We need to solve FAI to survive, so whether or not we’re skeptical about it is kinda beside the point. Even if we successfully navigate everything else, we will perish quite totally if FAI is not solved.
January 20th, 2007 at 12:41 pm
Michael: FAI in the strong sense may not be strictly necessary. If it turns out that fast takeoffs are not possible, and any intelligence takeoffs would take months and years (and while I don’t believe that to be the case, it’s possible) then approaches like Ben Goertzel’s would be fine. And if there’s some inherent limitation to intelligence that puts the practical maximum within a few orders of magnitude of our own (and again, I don’t believe this), it’s conceivable that even a UFAI wouldn’t be so powerful that we would be automatically doomed.
But if you care about your own and others’ existence, you should strongly support FAI research, because even a .001 chance of complete destruction is worth a lot of preemptive effort, and the risk from UFAI is probably much higher than that.
January 20th, 2007 at 4:14 pm
Additionally, it may be that Uploading may be come before strong AI (F or non-F). If that is so, the takeoff would be slow in the subjective time perception of uploads, as uploads will accelerate and improve together with computer improvements.
Since uploads would be human, human social and political systems could stabilize an arms/intelligence race. I’m referring to such systems as adapted to the virtual world.
January 20th, 2007 at 5:44 pm
Michael,
I was reviewing the CRN Scenarios chat today, and noticed you asked for a reference to my comment about violent videogames. Here’s one:
http://www.gamerevolution.com/features/violence_and_videogames
–Nato
January 20th, 2007 at 7:16 pm
“uploads will accelerate and improve together with computer improvements.”
“uploads would be human”
Uploads would cease to be human very quickly if they start self-improving.
January 20th, 2007 at 9:45 pm
Tom’s (of course) right again. And besides, as I’ve stressed, the trajectory unfolding as best I can see (not being on the actual cutting-edge[s], admittedly) is for AGI (hopfully FAGI) and strong **IA** (brain/computer interface and eventually symbiosis) to converge, probably around 2020 (but could be as late as 2030). Systematic research on (meta)Friendliness, though, is still nonetheless very important. All the current approaches (although I myself am most familiar with Ben’s and Eli’s) are important, as we’re still just at the nasent stage(s)….
And while I’m sanguine about uploads (see Marshal Brain’s enthusiastic discussions…!!), I’m a bit dubious about their arriving before FAGI, unless strong IA spins-off upload-capability (which is, of course, not at all especially implausible…indeed, perhaps quite likely…)
And I don’t even think we’ll have a clue about the nature of strong/open-ended/self-enhancing intelligence really, until, that is, we start symbiosis with proto-AGI’s. And while the human component *might* help toward garden-variety friendliness (as distinguished from Friendliness), as nasty and brutish as human beings CAN be, it might work the other way ’round. Which, again, is why research toward understanding how to instantiate genuinely Friendly AGI is so important…I’d much rather be in symbiosis with a genuinely Friendly AGI, wouldn’t you?
And as for the hard take-off or not: I don’t see how a recursively self-enhancing entity could help but zoom off into the wild blue yonder (as Irving Good put it). Once instantiated as a (Friendly, or proto-Friendly) seed AGI, surely it’ll only be days (or hours) until its intelligence is to ours as ours is to a chimp, or dog, or, perhaps, a flat-worm. But if we achieve symbiosis with the proto-Friendly AGI, I think we’ll be able better to “control” (though note the scare quotes), or at least enjoy/appreciate, the Singularity….
Ciao
January 20th, 2007 at 9:54 pm
And why are we currently in a robotic arms race? Because the world is currently run by narcissistic, control-freak, psychopathic megalomaniacs, whose corporations (*qua* [for legal purposes] artificial persons) themselves manifest all the traits of psychopathy (see the superb documentary, **The Corporation** by Canadians Mark Achbar, Jennifer Abbott, and law prof. Joel Bakan (from the latter’s also superb book by the same title). That’s why we currently have R&D moving toward stuff like right outta RoboCop…Hobbesian psychopathy.
Do we need any more reason to hope that R&D toward Friendliness will prevail…?!?…
January 20th, 2007 at 9:55 pm
Oh, here’s the link…(I fordot!!): http://www.thecorporation.org/
January 21st, 2007 at 4:10 am
Sorry, I still don’t buy it.
Yudkowsky takes a long to time to try to explain away the objection based on the premise that any machine complex enough to support consciousness must be a nonlinear system.
Of course classical nonlinear systems are entirely deterministic BUT they are infinitely sensitive to changes in initial conditions, therefore all we can really do is guess the future states of such systems with reasonable accuracy–but never true certainty.
In other words AIs will be at least as surprising, wonderful, horrifying and frustrating as people are.
I suppose it might be possible to run consciousness on machines that are entirely linear but I’ve yet to see any convincing arguments that this is possible.
Please note that I’m not some crypto-carbon chauvanist. My philosophical positions on AI are purely functionalist and mechanist. I fully accept the premise of strong AI but I also understand and accept the premises upon which chaos theory is based.
Basically chaos places fundamental limits on our accuracy and perfection.
All we can do is try our best to see that the artificial creatures we build generally tend to cooperate or at least ignore us.
In general if parents do their best to raise their children well, generally the children grow to be nice. But sometimes bad apples perversely occur too despite a parent’s best efforts. Perhaps some might object to my use of metaphors here but I think this one does hold some water. Since we haven’t built these critters yet, we are just speculating aren’t we?
But there is no way to ensure total safety. Mistakes happen. Surprises happen.
It just seems to be the way the universe is built.
Attempts at mathematical proof of perfect systems I’m very skeptical of.
January 21st, 2007 at 8:07 am
“Of course classical nonlinear systems are entirely deterministic BUT they are infinitely sensitive to changes in initial conditions, therefore all we can really do is guess the future states of such systems with reasonable accuracy–but never true certainty.”
This is an AI we created entirely from scratch; we actually can control the initial conditions down to the last bit, which is impossible in most other chaotic systems.
“In other words AIs will be at least as surprising, wonderful, horrifying and frustrating as people are.”
Yup. After all, a person’s DNA is written entirely from scratch at birth, which allows for a much greater range of variation than AIs, which all must be %99.9 identical.
“All we can do is try our best to see that the artificial creatures we build generally tend to cooperate or at least ignore us.”
AIs are NOT HUMAN, they DO NOT work like you do. Ignoring someone is pretty much the “default” thing for a human to do; it happens in a large percentage of human-human social interactions. Ignoring someone is NOT the default thing for an AI to do; you have to be very, very careful with the design for the AI to ignore anything. If there’s a %60 chance that another human will ignore you, many people would take that chance; but AIs aren’t human, and the chance of any randomly written AI ignoring you are so low that if I tried to write out the probability I’d probably fill my RAM with zeros before I finished.
“Perhaps some might object to my use of metaphors here but I think this one does hold some water.”
No, it doesn’t. The minute you try to compare AI social behavior to human social behavior you’re committing a giant fallacy. Children are already within the very tight bounds set by evolution on human behavior; an AI is NOT. Children have their DNA determined at birth, and AI does NOT. Those two huge, gaping differences wreck any metaphor of that kind without even going into further detail.
January 22nd, 2007 at 2:46 am
My impression is that FriendlyAI-theory just replaces the Asimov’s three laws with one law: be “friendly”. Then you throw in bunch of stuff like friendliness optimization and other jargon to make it look professional.
Formally proving that an AGI architecture is Friendly is probably in practice impossible.
January 22nd, 2007 at 4:14 am
Tom wrote, “AIs are NOT HUMAN, they DO NOT work like you do.”
Right. But that’s not really what I’m saying. I’m saying there is a general class of machines capable of consciousness. Human brains and any AI that is capable of consciousness fall into this category.
Now, in this category of machines there is a lot of variation but there are is commonality too. One thing these machines all have in common is that they are nonlinear dynamic systems. Thus they are surprising even if they are entirely deterministic.
Even if we control the building of these artificial brains to exacting detail, but never infinite precision (Remember that.), once we place them in the complexity of the real world all kinds of interesting things happen.
We can work to make sure these new organisms are generally friendly towards us but we’ll never assure that with absolute certainty. Such perfection seems to be forbidden. The quest for perfect safety leads to diminishing returns eventually.
I agree with skeptic, the FriendlyAI project can be, at best, something like Asimov’s laws–brittle and imperfect.
But this really isn’t that bad. Most human beings are mentally healthy and generally cooperative. Our society so far hasn’t destroyed itself. And this is the result of a blind process of variation and selection. Engineering, on the other hand, might do better than that or at least do it more efficiently but it won’t be perfect either.
January 22nd, 2007 at 12:40 pm
“Thus they are surprising even if they are entirely deterministic.”
Yes, that’s kind of the point of having an AI with superintelligence: it will surprise us with what it is capable of. However, we don’t have to prove that the solutions it generates will do XYZ; we just need to prove that it will maintain Friendliness.
“but we’ll never assure that with absolute certainty.”
Absolute certainty is indeed impossible, but without a mathematical proof that the system will remain stably Friendly, the odds are near-certain that it’ll become UFAI at some point. Even if the odds are only one-in-a-million on each successive self-modification, after a billion of these modifications we’re pretty much doomed.
“Most human beings are mentally healthy and generally cooperative.”
More anthropomorphicism. Most humans are mentally healthy by human standards; most AIs are NOT.
“Our society so far hasn’t destroyed itself.”
If we had the intelligence and technology of a superintelligent AI, you can bet we would have destroyed ourselves by now. Think about it- everyone, every minute of every day, having their fingers on the “blow up the world” button. We have a hard enough time keeping the US and Russia together; how are we going to manage six billion “nuclear powers”?
October 3rd, 2007 at 9:07 am
I’m not going to debate whether or not I think automated weapons pose a threat to the world. Speaking as an Armament Technician for the United States Army, I can say in all honesty that I believe having superior firepower is probably the most vital factor in determining the victor in today’s warfare. This truth has only become recent, as better tactics and training used to have a lot more significance.
Natrually, I am well aware that hundreds if not possibly thousands of human beings are dying because of the work that I do. Afterall, I basically give life to weapons so that those weapons can take lives from human beings.
But it’s a necessary evil because simple ideals, morals, and principles do not win wars.
Taking into account that human beings will inevitably die as the result of such weapons, it really doesn’t make a difference if some sort of Artificial Intelligence is the one pulling the trigger, the danger remains. A rifle is afterall, only a machine in the end.
I am a long way from designing advanced mechanisms that have a more efficient lethality rate than what we currently have, as I am still working on my degree in Mechanical Engineering. But I can tell you that if I had the knowledge and the experience necessary, I would certainly do my best to contribute to the “automation of warfare” if it would benefit our success on the battlefield.
War isn’t about what’s ethically right or wrong. It’s about defending your nation and developing the most efficient killing machines before the enemy gets the idea first.
August 6th, 2008 at 3:04 pm
Lesbians licking pussy….
Asian lesbians licking pussy video. Guide to licking a pussy. Pussy licking. Free hot girls pussy licking. Pussy licking stories. Licking pussy….