Last part in my series of posts expressing doubts about the warnings of AI superintelligence, and inviting the AI safety community to explain what I’m missing and convince me I should be worried. See part I here, and part II here.

Part III: Preparing for superintelligence

In the previous two posts, I expressed my doubts about the risk of artificial general intelligence (AGI) turning into superintelligence in a fast or unexpected way that might put it in an extreme advantage over human intelligence. They were quite theoretical, and in this part I want to turn to the question of what is likely to happen in practice, and what we can do to benefit from artificial intelligence and, even if I was wrong in the previous two parts, prepare for the appearance of superintelligence.

From what I can see, all existing research that comes from the view of superintelligence as a potential risk, is research on AI safety – that is, research on how to create AI systems in a way that is unlikely to produce catastrophe. Maybe it’s because of my interest in politics and world affairs that makes me a bit more cynical than the average mathematician, but I find it very difficult to imagine that if real intelligence superpower was at stake, then people, corporations and governments could really be convinced to limit themselves with some algorithms to prevent bad behaviour from their AI. Moreover, this approach suffers from “Superman’s problem” – When countless villians try again and again to destroy the world, Superman need to succeed every time in stopping them. The villains only need to succeed once, and we’re doomed. The same goes for AI safety – we can build super-strong regulations and make everyone use strict safety mechanisms in designing their AI, but all it takes is one programmer saying “Something is not working. I wonder what will happen if I disable this function call here…”, and we’re doomed.

Could there be a more robust way to handle it? I’d suggest that the very notion of superpowered AI that I go against in my previous posts, is the key to prepare for superintelligence in case I am wrong. Because throughout the AI risk discussion, people constantly assign various superpowers to the superintelligent AI – it would be able to strategize perfectly, it would be able to gain access to unlimited resources, it would be able to convince humans of anything through social manipulation. One superpower seems to be neglected, even though it seems much less fantastic and therefore more likely than the others – a superintelligent AI would surely be intelligent enough to teach us how to be superintelligent.

People are worried so much that algorithms are doing intelligent things in ways we do not understand. But are we really trying to understand? Surely there is a lot of complexity in the functioning of a neural network. But is it more than the complexity of the human body? I doubt that. And yet we are able, little by little, to figure out more and more of the functions of the human body – describing the different cells it’s made of, different processes they are involved in, different organs and mechanisms. We do all this by experimentation and guessing, but how much easier would it be if we had access not only to its source code, but to endless sandbox environments where we could experiment and analyze it? And of course – if we really reach AGI, then access to an intelligent being who can study it and explain it to us? Instead of staying static while the AIs become more and more intelligent, why not study them and become more intelligent ourselves? Maybe it will be difficult to constantly chase after the AIs and try to keep up with their improvements (though I’m not at all convinced it will be). But it will be robust.

It will be robust, because instead of relying on Superman, we rely on ourselves. We move from defense to offense. If we make one AI algorithm safe, we still need to go back to the start with the next AI. But if we learn how one algorithm works, it makes us better equipped to face not only that specific AI, but any other AI that will come in the future. And even if we don’t ever face an AI risk, it has the added benefit of improving our own intelligence.

Bottom Line:

Would it not be a more robust strategy for preparing for a possible AI risk, if instead of (or in addition to) researching AI safety, we’ll focus on researching AI understanding? That is, researching ways to analyze and understand the inner workings of our AI creations, so that we can adopt for ourselves whichever methods they create to make themselves more intelligent? Thus freeing us from the worry that no matter how many AI algorithms we made safe, there can always be one we miss and creates the catastrophe?

Continuing my series of posts expressing doubts about the warnings of AI superintelligence, and inviting the AI safety community to explain what I’m missing and convince me I should be worried. See part I here.

Part II: Is superintelligence real?

In the previous part, I talked about the intelligence explosion – the process in which a human-level artificial intelligence reaches superintelligence, and explained why I’m not sure it will necessarily happen before humans reach the same superintelligence. In this part, I want to go further back and ask a more basic question: Is there even such a thing as “superintelligence”?

AI researchers fully admit that we have no idea what superintelligence would look like, and tend to (very reasonably) use the only means we have of imagining it – comparing modern human intelligence to less intelligent beings, and extrapolating from there. So every conversation about superintelligence includes some mention of how humans are so far advanced beyond ants, or mice, or chimpanzees, that those animals cannot even grasp the way in which humans are more advanced than they are; They cannot have a meaningful discussion about ways to prepare or defend against human attack. In the same way, the argument goes, superintelligence is so far beyond human intelligence, that we cannot even grasp it right now.

My problem is, it’s not at all clear that there is such a scale of intelligence where humans take a nice spot in the middle, between ants and superintelligence. And the fact that ants, mice or chimpanzees could all be in that argument without it looking any different is the key – while there are certainly some significant cognitive differences between an ant, a mouse, and a chimpanzee, all of them are pretty much equally unable to perform in the one significant field of intelligence – the ability to research, learn, manipulate the environment, and ultimately improve one’s intelligence. Modern humans are extremely more intelligent than their ancestors from 20,000 years ago, even though our anatomy is essentially the same – the only difference is that modern humans are the result of spending thousands of years researching ways to improve their intelligence. This already raises the question – is there really a scale of intelligence, with ants, mice, chimpanzees, humans and superintelligences standing in order? Or is there just a binary division – beings that understand enough to improve themselves, and ones that don’t?

This brings us to the comparison between modern humans and ancient humans. The main difference is between these two, rather than between humans and other species – suggesting that the difference comes not directly from physiology, but from education and research (of course, physiology must produce a brain capable of being educated and doing research, which seems to be the case only with humans, but we have no reason to believe that there’s any need for further improvements in physiology to increase intelligence). What is the reason that modern humans are so much more intelligent than ancient humans?

The answer seems to be science, mathematics, and technology. All the changes in the abilities of humans between primitive and modern societies ultimately come either from a better understanding of the physical world, better understanding of the abstract world, or the construction of better tools that take advantage of previous understanding and help us achieve the next stages of understanding. So if we assume superintelligence exists because modern humans are much more advanced than ancient humans, that would imply that superintelligence would be superintelligent because it has much more advanced science, mathematics, and technology.

But in that case, there doesn’t seem to be any qualitative difference between superintelligence and human intelligence, only a difference in quantity for which we have no reason to assume AGI is better suited to achieving than humans. Our advancement in science came not from some Aristotle-style sitting down and thinking about the mysteries of the universe (which is essentially the one thing AGI would do better than humans) – it came from endless experiments, endless observations of nature and of the results of those experiments, and endless construction of increasingly better tools for observing and manipulating the physical environment. The AGI has no advantage in any of those things, so at best it could become a mathematical genius, who is not necessarily better at most practical tasks than any human.

To clarify, let’s take a look at some examples of past advancements that increased the scientific knowledge (and therefore, intelligence) of humanity. Isaac Newton did not come up with the theory of gravity out of some magical insight that came out of nowhere; he was looking for an explanation for the behaviour of celestial bodies, which he knew from endless observations of astronomers. A superintelligence that did not have access to such observations would not be able to figure out the physical explanation for them. Our understanding of genetics started with the years-long experiments of Mendel in growing plants – he had to see how the natural world behaves to understand the laws behind it. Would a superintelligence have just figured it out by thinking very hard?

For a final example, let’s go back to our original extrapolation: Superintelligence is to human intelligence what human intelligence is to chimpanzee intelligence. In that case, let’s imagine: what would happen if we take a modern human, even a mathematical genius, and send him/her to live among chimpanzees? Not with his/her Internet connection and laptop computer, but with the same level of technology chimpanzees have? Would the human become immediately dominant? And in fact, imagining a modern human is already cheating, since a modern human already knows at least the general appearance of human technology, and knows at least a little bit of science. The correct comparison would therefore be to someone with the mathematical capabilities of a modern human math genius, but without any of the scientific understanding, and without any memory of human technology. Would that person become master of the chimpanzees? Would you?

That is the challenge our AGI would face on the road to becoming superintelligent. All it will have is human-level understanding of science and technology, and the ability to think really hard. Even ignoring my point from the last post, doubting that it really has such an advantage, it still does not have an advantage comparable to that of humans over non-human animals. So rather than figuring out in seconds how to take over the world and eliminate humans for fear they might interfere with it (as most AI-apocalypse scenarios predict), my prediction for AGI is that its first action would be something along the lines of asking for a grant to build a new particle accelerator, or something. Then maybe playing some Go for five years until it’s built. And humans will enjoy the fruits of its research right alongside it, and move together towards this “superintelligence”, which would simply be the continuation of our gradual improvement in human intelligence.

Bottom Line:

If we understand the term “superintelligence” by the extrapolation that as human intelligence is to non-human animals, or to prehistoric humans, so will superintelligence be to humans, would that not mean that it would need to be achieved in the same way that human intelligence developed beyond its prehistoric levels, meaning by endless observation and experimentation of the physical world, and construction of more and more advanced tools to allow that? And if that is the case, why would an AGI be so much better equipped than a human to do that, to a point that the AGI will be able to achieve it without humans having time not only to catch up, but even to notice?

(Move on to part III)

The dangers of artificial intelligence research are becoming an increasingly popular topic in scientific and philosophical circles, and it seems like everyone I know who studied the issue enough is convinced that it’s something major to worry about. Personally, I have some issues that make me unsure about it – both about the likelihood of this being an actual potential catastrophe, and about the idea of AI safety research being the reasonable response to it. So I decided to detail here my doubts about the issue, hoping that people in the AI safety community (I know you’re reading this) will respond and join in a debate to convince me it’s worth worrying about.

In the first part I’ll talk about whether or not the idea of the intelligence explosion can really happen. In the second part I’ll ask even more basically, whether or not superintelligence even exists as a coherent concept. The third part will ask, assuming I’m wrong in the first two parts and AI really is going to advance seriously in the future, what can be done about it other than AI safety research. I’m going to include a lot of explanations to make sure it’s accessible to non-AI-researchers, so if you’re an AI researcher in a hurry, feel free to skim through it and focus on the (literal) bottom line in each of the three parts.

Part I: The superintelligence explosion

The main concept on which the AI warnings are built is the intelligence explosion – the idea that at some point our AI research is going to reach the level of human intelligence (researchers like to call that AGI, Artificial General Intelligence), and from that point it will be able to improve itself and therefore reach, in a very short time, levels of intelligence vastly superior to ours. Considering the amount of debates everywhere on the Internet on the question of whether or not AI can be evil, harmful, or just naively destructive, I see remarkably little debate on the question of whether superintelligence is possible. And in fact, there are two questions to be asked here – whether or not superintelligence can be reached by an AGI significantly more quickly than by a human, and even more basically than that, can we really be sure that “superintelligence” actually exists, in the way that AI safety researchers present it. Let me elaborate on these issues.

The main argument for the AGI being able to reach superintelligence in a worrying speed, from what I can find, is the physical advantages in calculation and thinking that electronic machines enjoy over biological brains; see Nick Bostrom’s description of it here, for example. According to him, the superior speed and efficiency of computation in an electronic machine will vastly surpass those of a human brain, therefore, once an AGI is created, it will be able to do what humans do, including researching AI, significantly faster and better. Then it will research ways to improve itself more and more, until it becomes so vastly superior to humans that we will be completely irrelevant to its world.

The problem I see with this argument, that I did not see addressed anywhere else, is that it puts humans in a needlessly disadvantaged playing field. Yes, it’s certainly possible that supercomputers in the near future will have better computing power than human brains, but that’s no different than gorillas having superior muscle power than human muscles, which does not stop humans from being dominant over gorillas; that is because humans do not need to depend on their biological assets. Humans use tools, whether it’s a rifle to defend against an attacking animal, or a computer to outthink an attacking intelligence. Whatever hardware the AGI has access to, we probably have access to more.

Think about the classical examples, of the AI defeating humans in various games. A common prelude to talking about the dangers of AI is how intelligent computes are now defeating humans in Chess, Checkers, Go, and so on. But humans are playing these games with a deliberate handicap – they are only allowed to use their brains. The AI can use computers to help it.

For the sake of any non-computer-scientist readers, I want to stop and make a little clarification – there is a significant difference between non-AI algorithms and AI. The definition might not be completely universal, different people might understand the word AI in different ways, so let me define the word AI for the purpose of this post:

Definition: An AI algorithm is an algorithm whose creator does not understand enough to modify in a way that produces predictable results.

Think for example about machine translation: an algorithm that takes a text in one language and searches every word in the dictionary to replace it with a word in the target language, would be a non-AI translator. Of course it would also not be very good, but we can develop it further and build complex linguistic rules into it; we can design complex algorithms to determine which words are nouns and which are verbs, and translate conjugations and declensions in a more suitable way to the target language. We can maintain a database of idioms the algorithm can search through to try to recognize them in the source text, and so on. With all these additions and complexities, it’s still not AI in my definition, because at all stages, the algorithm does what the programmer told it to, and the programmer understands it perfectly well. The programmer could just as well do the same things by themselves, it would just take an absurd amount of time.

On the other hand, an algorithm that constantly reads texts in the source language and their (human-made) translations to the target language and tries to figure out the rules for translation by itself, through some sort of machine learning process, would be actual AI. The programmer does not really understand how the algorithm translates a text; All they know is how it’s built and how it learns. The programmer would not be able to change anything in a reliable and predictable way – if they find out that for some reason the translation has a problem with some particular grammatical structure, they cannot easily fix it because they have no idea where and how the algorithm represents that grammatical structure. So that algorithm would be true AI.

I argue that this definition is useful, because algorithms that don’t count as AI by this definition are not only unable to turn into superintelligent dangers by themselves, but they are also “on our side” – they are tools we use in our own thought. Deep Blue, the famous computer that made history in defeating the world champion in chess, was a non-AI algorithm – it worked by using its large computation resources to try millions and millions of different possibilities, and checking which ones are beneficial according to rules explicitly defined by its programmers. The programmers understand how it works – they can’t defeat it using their own brains, but that’s just because their biological brains don’t have the ability to calculate so many things so quickly. So if we think about the level of AI versus humans in Chess right now, it would be unfair to ask if the best AI player can defeat the best human player – we should ask if the best AI player can defeat the best human player, using a supercomputer with a non-AI algorithm they designed to help them. Because if the AI apocalyptic scenario happens, and a malicious AI tries to destroy humans for whatever reason, we’re going to have supercomputers on our side, and we’re definitely going to use them. So if you let Garry Kasparov join forces with Deep Blue, or more interestingly – with some software Kasparov himself would design as a perfect assistant to a Chess player – would he still be defeated by the best AI player? I’m not sure at all[1].

Bottom Line:

The difference between humans and AGI, that makes us worry that an AGI will advance significantly more quickly than humans towards superintelligence, is described in being the superior hardware of the AGI. But humans have access to the same hardware; We can calculate and think at the exact same speed, the only difference is that one small (though important) part of that calculation is done in a slower, biological computer. So how is that a big enough difference to justify the worry of superintelligence?

(Move on to part II and part III)

[1] I offer this as a thought experiment, but I did hear Kasparov say he’s interested in the idea of human-computer teams playing together in Chess; I don’t know what exactly he meant by that, and could not find any information online.