My blog about philosophy, coding, and anything else that interests me.
02 – On The Alleged "Dangers" Of AI And Disambiguation
Artificial Creativity
References:
- Pessimist's Archive on Twitter: https://twitter.com/PessimistsArc
- David Deutsch, "Creative Blocks: How Close are We to Creating Artificial Intelligence?": https://aeon.co/essays/how-close-are-we-to-creating-artificial-intelligence
- David Deutsch, CBC interview from 2018 about AGI: https://www.cbc.ca/player/play/1251489859794
Further reading:
- “The Bucket and the Searchlight” by Karl Popper
- “The Beginning of Infinity” by David Deutsch, chapter 1 “The Reach of Explanations”, and chapter 6 “The Jump to Universality”
If you have any questions or comments, please leave them here or tweet at me: https://twitter.com/dchackethal
Subscribe to this podcast via RSS Back to all episodesTranscript
This transcript was generated automatically and may contain mistakes.
Welcome, everyone. This is the second episode of the series on artificial creativity. This one is about the alleged dangers of AI, artificial intelligence, and AGI, artificial general intelligence, and also about the differences between the two, as well as how they relate to artificial creativity. It turns out the dangers and the disambiguation are related to each other, and so I decided to tackle both in this same episode.
It’s become very fashionable to talk about the dangers of AI. This is no surprise. People have always been skeptical of new technologies and progress, even though it’s the very phenomenon that solves their problems and makes their lives better. There’s a great Twitter account chronicling these fears called Pessimists Archive. I’ll link to it in the description. It shows how people bemoaned or were even afraid of new styles of dance, the bicycle, headphones, calculators, you name it. You can read newspaper articles from back then. They’re pretty funny. Especially young people were blamed for these developments. And parents were offered advice on how to keep their children from indulging in these so-called fads.
This isn’t to say that there have never been dangerous innovations, say, for example, new weapons technology. But it is to say that most technological innovations offer wonderful ways of improving our lives. And so I’d like to try to convince you that there is a positive, exciting path towards AGI, and that research in this field is not a dangerous but wonderful thing indeed.
Innovations are made in response to problems, things that could be better. Human ingenuity allows us to solve these problems in a never-ending cycle. David Deutsch sums this up nicely in his book “The Beginning of Infinity”. He says that problems are inevitable and problems are soluble. This is part of the human condition. For some, this is the occasion for despair, but problems are the very thing that makes progress possible in the first place. And we should expect new technologies to solve existing problems while creating new, better problems. This is the never-ending cycle of progress.
How do we solve problems? By creating knowledge. This is what science is all about. So, too, we should expect AI to solve some problems while creating new, better problems. If the new problems it created were not better than the ones it purported to solve, we wouldn’t even adopt it. Our ability to try and anticipate new problems is key here, and as we will see, part of the creative algorithm that this series is about.
And this is where the professional worriers about AI come in, the Elon Musks, the Nick Bostroms. In particular, they’re worried about something called superintelligence, a program that can recursively improve itself until it surpasses human-level intelligence, at which point it could outsmart us and enslave us to achieve its goals. Unfortunately, they do not seem to understand what human-level intelligence is. As we will see, such a thing as superintelligence cannot exist. The concept of universality will help us understand this. Ironically, people like Elon Musk are the main worriers about AI, but are at the very forefront of AI research, not AGI research though, because they work on things like self-driving cars. But Musk’s research is safe.
So what’s the difference between AI and AGI? For one thing, AI is an already existing technology and is actively being researched. For example, self-driving cars, automated chess players, etc. AGI, on the other hand, is not yet a thing and is not yet actively researched. Some people think they’re actively researching it, but they’re not really. And I’ll say more about that in the next episode.
In David Deutsch’s article, “Creative Blocks: How Close Are We to Creating Artificial Intelligence?”, as well as in a recent interview with CBC Radio, links to which I’ll put in the description, he disambiguates the terms. AI, in the sense Alan Turing understood the term, used to refer to a program that has the full cognitive abilities of human beings. That is, emotions, thoughts, general purpose problem solving, free will, etc. Over time, this has been watered down quite a bit, and today we use the term AI to describe either programs that don’t need to be programmed explicitly, or, of late and even worse, programs that are sophisticated enough they can do things only humans could do historically, for example, drive cars and recognize faces.
This last meaning is worse because technology’s purpose has always been to do things only humans could do so that humans could focus on other things they would rather do. So nowadays the term artificial intelligence has very little to do with intelligence. Similarly, by the way, machine learning has very little to do with learning but is mostly about applied calculus and linear algebra.
This is why we had to introduce the G in AGI so we could refer to the original thing again, a fully-fledged intelligent being capable of all human cognition. Part of human abilities is, as I said earlier, general purpose problem solving. What we know at this point is that humans solve problems by creating knowledge, which comes in the form of explanations. Humans have a facility that allows them to create new knowledge, and that’s creativity. It’s not just creativity of the kind that lets you write stories or paint a painting, although it lets you do that too. It’s creativity of the kind that can create knowledge in the form of solutions to problems. In principle, it can create knowledge to solve any soluble problem. This is why we are general purpose explainers.
It’s not clear to me at the moment whether this creative ability automatically entails the ability to feel and think. I go back and forth on that. But I think the most interesting and important ability of an AGI is to create new knowledge. We can grant for the moment that something is only an AGI if it can also think and feel, but my guess is that neither thinking nor feeling are all that useful if you don’t know how to understand your thoughts and feelings, and for that you need creativity. In any case, creativity is primarily about understanding things and less about doing things, which is what software engineers refer to as causing side effects.
AI and AGI are qualitatively different and almost opposites. AI is narrow and problem-specific. For example, a self-driving car can solve the problem of how to drive a car from A to B without human guidance, but nothing else. And most, if not all, knowledge of how to do that is created before runtime by the programmer. This is why the claim that AI programs do not need to be fully specified is misleading. AGI is a general purpose problem solver. It has a vastly bigger repertoire. So vast, it’s inexhaustible. There is no soluble problem a general purpose explainer could not in principle solve, and its knowledge is created largely at runtime by itself. Due to this general purpose, the terms AGI, general purpose problem solver, general purpose explainer, artificial creativity, and even person as I will elaborate later on, are all equivalent, and I will use them interchangeably from here on out.
This is why I said that superintelligence cannot exist. Once you have reached the level of general purpose problem solving, your repertoire is already inexhaustible. Where else could you go? A popular metaphor people from the superintelligence camp like to invoke to illustrate the danger is that of our relationship to ants. An ant lacks the ability to understand our goals and defend itself against us should those goals impinge on its ability to live. Likewise, a superintelligence is said to see us as a kind of ant equivalent, and we couldn’t even hope to begin to understand its goals nor defend us.
Humans and ants are indeed qualitatively different. Humans are general purpose problem solvers and ants are not. But AGIs and humans both share that universal capacity. They are not qualitatively different. They’re the same. This, by the way, is the reason racism, sexism, and other kinds of prejudice don’t make sense. Because in regard to their ability to create explanations, all people are literally the same. There is a one-to-one correspondence between the sets of explanations any two people can create.
Two sides notes here, though. The fact that this is an excellent tool against discrimination doesn’t make it true, but it is true nonetheless. Also, with this equality in mind, the qualifier “artificial” seems misplaced and even discriminatory. There is nothing artificial about artificial creativity. It’s the same as any other person’s creativity.
Okay, with that, now that we can properly disambiguate, it seems that we’ve merely split the problem of the dangers of AI into two problems, the dangers of AI and the dangers of AGI. So I will summarize them both. AI is not dangerous because, by definition, its repertoire is exhaustible. There is a finite list of things it can do. If it ever does something we don’t like, we as general purpose problem solvers can think of something not yet on the list, which constitutes a problem it cannot solve. And this is how we could always keep it in check, no matter how fast it is. It can’t outpace us.
AGI could be dangerous, but only with a few caveats. It’s a general purpose explainer, just like people are. This makes it qualitatively no different from people, at least in its creative capacity. And so there’s nothing an AGI could do that a human couldn’t do, and vice versa. Therefore, warning about the dangers of creating an AGI is a bit like warning someone about the dangers of having a baby. In both cases, you’re creating a new person, who, in a healthy society, will most likely grow up to be a healthy individual, not a destructive monster.
I said earlier that one of the differences between AI and AGI is that AGI creates most of the knowledge it needs at runtime. That knowledge is not created by the programmer. What we need to understand, then, is how knowledge is created in the first place. This is an epistemological question. Epistemology is the study of knowledge and how it’s created. In short, it asks the question, how do we know? Science and philosophy are disciplines that strive to know, to understand the world better. Our best theories of how we know were created by Karl Popper and David Deutsch.
Therefore, we can and should read their works as contributions to the question of how we can build artificial creativity. AGI research should be understood as a continuation of and a contribution to the millennia-old cosmological effort to understand reality. I hope this convinces you that there is a positive, exciting path towards AGI. Building the creative algorithm will be a huge technological breakthrough with countless benefits for humanity, and I cannot wait for it to happen.
Thank you very much for listening. To learn more about how we know, I recommend you read “The Bucket and the Searchlight” by Karl Popper, as well as the first chapter of “The Beginning of Infinity” called “The Reach of Explanations”. To understand universality better, I recommend chapter six, “The Jump to Universality”. There are other highly relevant chapters in the book, which I will reference later as we get to those topics. In the next episode, we will explore the current state of AGI research, and specifically why some of the most prominently featured AGI research is sadly particularly ill-suited for the task. And we will explore what may work better.
If you have any questions or comments, please leave a note or tweet at me. I’ll link to my Twitter profile in the description. I’ll see you in the next episode.
What people are saying