This article is a continuation of what I have written here.
Firstly, when I red team the possibility of anybody supporting the use of large IEDs or EMPs and nuclear airbursts to stop AI that will probably enslave or annihilate humanity, it depends of course on whether he or she 1) supports certain interpretations of the Just War doctrine, and 2) accepts a very negative perception of AI. Here, in this article, I'm simply analyzing what can follow logically from 1) and 2). It does not mean that you or I have to accept these logical conclusions, because if you don't agree with the premises, you will reach other conclusions that also are within the ballpark of logic. Personally I'm a pacifist, so even though I can recognize the logic of the violent perspectives I describe below, it doesn't mean that I will base my decisions on these viewpoints.
There are two main perspectives that can justify the use of nukes and IEDs to stop deadly AI:
The most valid argument, in my opinion, is that AI systems are now being used to expand the panopticons we already have in the West and East, so IF you are opposed to authoritarian value lock-in caused by AI-enhanced surveillance regimes, and IF you also have the same attitude as many in the West had during the Cold War, when they said "better dead than red", THEN it's within the ballpark of reason to use any means necessary, including IEDs or nukes, to prevent the growing power of today's AI panopticons.
Our unprecedented surveillance regimes will soon become truly omniscient when insect drones and smartdust have proliferated enough to monitor every square meter on Earth. This development is inevitable if AI corporations and robot companies are not destroyed.
If Stalin had LLMs in 1946, then it's very likely that the Allies would have executed Operation Unthinkable, so IF you have the same attitude as Churchill back then, it will be within the ballpark of reason that you should also use any means necessary to destroy AI companies in the West and East, in order to prevent that surveillance regimes today continue to grow.
We have also seen that OpenAI wants their AI to influence politics, so if we don't destroy OpenAI right now, it's very likely that we get an unavoidable value lock-in that lasts forever, as long as humanity exists. China already suffers under the value lock-in of the CPC. To prevent that something similar happens in the West, it's safest to destroy Big Tech as quickly as possible.
But don't we need AI to defend us against the totalitarian AI of China? I have discussed this here. Let me just add: AI will enslave us all if we don't stop it, which means that it basically doesn't matter if we are imprisoned by the AI of Big Tech in China or by the AI of Big Tech on the West Coast of the US. Basically they are all part of the same Pacific Rim AI.
We know that many owners of Big Tech companies are evil scumbags, and it's well-known that some of these businesses operate unethically, in addition to cooperating with the CPC, a party that has killed more people than the Nazis, so IF you accept a Cold War interpretation of the Just War doctrine, it's within the ballpark of logic to assassinate Big Tech leaders and kill the most talented AI researchers, similar to how Israel has killed nuclear scientists in Iran.
This is not preemptive warfare, because extreme AI surveillance regimes already exist in the West today. If you're a supporter of constitutional democracies, it's your duty to combat these AI panopticons.
The second main perspective that can justify the use of nukes and IEDs to stop deadly AI is the so-called "doom" scenario described by Eliezer Yudkowsky. He writes in Time: "Make it explicit in international diplomacy that preventing AI extinction scenarios is considered a priority above preventing a full nuclear exchange, and that allied nuclear countries are willing to run some risk of nuclear exchange if that's what it takes to reduce the risk of large AI training runs."
I'm not (entirely) convinced that the human extinction scenario is inevitable. I'm not a "doomster", since I'm not sure if Yudkowsky or Yann LeCun are right in the debate about the inevitability of humanity being killed by a more or less superintelligent AI, if we are able to develop something like that in the future. It depends on the specific AI tech which is created, how it's designed to work, and how "superintelligent" it actually is.
However, IF you think it's a 16.66 % risk that a superintelligent AI will annihilate humanity or kill billions of people, then it can easily be argued that AI companies today are forcing each of us to play Russian roulette with one bullet in the chamber. If a person forces you to play Russian roulette, you have a right to kill him. And you have this right even if he pays you 10 million dollars if you survive. The fact that AI development can perhaps be very rewarding for humanity if we are lucky enough to survive doesn't justify that AI companies force all human beings to play Russian roulette today.
We have often seen the incompetence of governments and corporations, so it's not reassuring at all when some AI companies demand that governments must regulate the development of superintelligent AI. Firstly, if governments manage to control an ASI, we end up in a panopticon, as described above. Secondly, if we actually manage to develop a truly superintelligent AI that's 100% autonomous, we will not be able to control it, per definition, since it will be like chimpanzees trying to control humans. Trying to regulate a genuine superintelligent AI is similar to putting a demon or god in a kindergarten and claiming it's safe because a committee has been created to regulate the demon or god. As Geoffrey Hinton has pointed out, it's not logically possible for humans to have power over something that has much more power than humans. Can humans stop a "godlike" power? Of course not.
This does not mean that LLMs in 2023 are a mortal threat to humanity. But researchers like Yann LeCun are working on developing new AI systems that are much better than LLMs. It may take years for this to happen. Maybe it will take 20 years. But it will perhaps only take 2-3 years now that all great powers in Cold War 2 are working extremely hard to win today's AI race, because they know that the nation which controls AI will control the world.
If Japan in 1934 had known that America would develop nukes in 1945, they would in the 1930s have tried to kill the main nuclear physicists in the West.
The Just War doctrine doesn't allow preemptive war, but this doctrine was not created to deal with threats that can annihilate humanity. In the AI safety debate it's therefore reasonable to modify the old and outdated Just War doctrine. A preemptive War on AI is therefore in the ballpark of legitimacy if there is at least a 16.66% risk that superintelligent AI will annihilate humanity.
You may disagree with the arguments presented above. That's fine. I'm a deontological pacifist, so I reject violence in principle, partly because I have no kids, and I'm an old Christian who believes in an afterlife, so what happens here on Earth is basically not my problem. Critics of the above arguments have a problem however, because they can't refute them basically, since many opponents of AI have good reasons to view these arguments as being within the ballpark of legitimacy.
Henry Kissinger has said that in politics you often haven't the luxury to postpone making a decision until you know with certainty that your decision will be successful in the future. Because it will often be too late to make effective decisions if you wait until you know with a high degree of certainty that your policies will actually work.
Many important decisions in politics are a gamble. Do you feel lucky?
Maybe the two main perspectives described above are wrong. However, IF you believe that one of these viewpoints are correct, and IF you have the same attitude as a Cold War warrior, then you should act against AI corporations now, with the same determination as Eisenhower in WW2. In that case, there's no time left for academic discussions about AI safety, because AI develops so fast that we can't afford to waste time on endless debates. This is a personal choice. There are no (trustworthy) authorities or experts that can hold your hand and offer reassurances that you are doing the right thing, because this is truly an unprecedented situation that humanity has never dealt with before, so there are no (reliable) authorities, no experts who scientifically know what is right and wrong in this case.
Even if somebody actually knew with a high degree of certainty that their AI policy is the correct one, it will not help much in the arms race of Cold War 2, since the international system is anarchic, so AI will basically remain the Wild West, beyond regulation despite how much governments try to monitor things like GPU clusters, because 1) a small nuclear reactor inside a mountain can hide power supply to AI research, and 2) we already have Guanaco 65B that can run on a single GPU. It can easily be argued that talk about AI regulation is security theater, a way for AI supporters to bog down critics in endless debates while the power of AI grows exponentially. So act now or accept defeat.
No comments:
Post a Comment