Welcome, Professor Yampolskiy, welcome Roman! I’m very happy and honoured to have you here for this interview. Let us begin with you telling us a little about who you are and what your interests are in philosophical research. What are you currently working on?
Sure! I self-identify as a computer scientist; an engineer. I work at the University of Louisville. I’m a professor and I do research on AI safety. A lot of what I do ends up looking like philosophy, but, you know, we all
get PhD’s and we’re “doctors of philosophy,” so a computer scientist is a kind of applied philosopher; a philosopher who can try his ideas out. He can actually implement them and see if they work.
So what is your philosophical background then? Are you also professionally a philosopher?
Not in any formal way. I think I took an Introduction to Philosophy once and it was mostly about Marx’s Capital or something like that. So I had to teach myself most of it.
I also noticed that you have written lots of articles; some you wrote together with many different collaborators, some also on your own, and you are also writing books and you are almost continuously on Twitter… Since some early career philosophers might be watching or reading this interview, I was wondering if you have any advice for them on how to do this. How do you organise your time? How can you manage to be this very prolific philosopher and do all these other things on the side?
So it may not work for early career philosophers… I’m ten years on the job, so I have the power of saying no to almost everything I don’t care about. It’s much harder when you are just starting out. You have to say “yes, I love to teach another course! And, yes, your meeting sounds fascinating!” At this point, I don’t have to do that so I think that’s the main difference. I just look at the long-term impact of what is being offered in terms of time taken and what it’s going to do for me. Will I care about it five years later? And if the answer is “absolutely not,” why would I
So this is the secret? It’s just saying no to everything that is not research and not publishing?
And it’s very hard, because you want to say yes, you want to work with people. So it helps to actually formalise it. I have to say ten nos before I can say one yes, and then the quality of yeses goes up and your time is saved.
But then, many of us seem to see teaching as opposed to research. You know, famous researchers often don’t like to teach, because it takes away from their research time. But it seems that you somehow magically also managed to combine both things. I look at your biographical note and it says all kinds of things about teaching awards and about how successful you are in teaching, and how your students love you. So how does this work? Don’t you see this opposition between teaching and research in terms of their demands on your time?
There is definitely a struggle, but you can combine at least with advanced courses, you can introduce research to the classroom and a lot of my students actually end up publishing work based on projects we do. And not just as Master students or PhD, but what they did in my artificial intelligence class for example. It doesn’t work as well for introductory courses. I also teach the large “Introduction to Programming” class, so
it’s different there. This is a large freshman course, so research is not quite something you can introduce at that level. Generally, try to have a lot of fun with students and invest in them, so later on they come back to work with you.
Let’s talk a bit about your papers. Although you have written many papers with others, you told me that you wouldn’t like to talk so much about these collaborative papers because obviously the co-authors are not here with us. But still, there are a few concepts that I found so interesting that I’d like to ask you about. Perhaps we don’t need to talk particularly about the papers, but we could talk a bit about these concepts more generally.
For example, there was, in one of your papers, this idea that there is a reason or a purpose to life. You say that this is, among other things, a Japanese concept. But of course we also have a sense of the meaning of life and of a purpose of life. And there I was wondering about two fascinating aspects of this:
On the one hand, we always speak about a universal basic income and about robots taking away employment and then replacing it with basic income. This obviously will affect our meaning of life, because our narrative regarding the meaning for our own lives often depends on our work. People who are unemployed sometimes have the problem that they feel that their lives are not as meaningful perhaps as they could be. Do you see this as a problem with
unemployment caused by AI? And what do you think more generally about the challenges of AI-caused unemployment?
There seem to be two types of people: People maybe like us, who really love their work, would do it for free, provided they could survive otherwise. And the meaning of their lives, to a large extent, comes from doing this work. But then there are also people who really hate their jobs. Their jobs are boring, repetitive, horrible and they only do it to survive, to get the funds.
So I think in one case we would lose a large portion of what it means to be a researcher, a scientist, a professor; whereas for many people it would be completely liberating. For them it would be very easy to adjust to the situation whereas for us we would have to find a different source of meaning.
I’m not sure what would work for individuals. Definitely you will not be competitive in terms of being a researcher in a particular area, compared to a super-intelligent AI researcher.
This is strange, right, because normally we think that the more repetitive jobs will be automated first. I have not heard that researcher jobs are in danger of being automated, because this is work that also requires intuition and creativity; these qualities that we don’t normally associate with a danger of being mechanised or taken away from us. Do you think that there is actually a real danger that even jobs like ours will be eventually moved to AI?
It depends on how far into the future you look. In the short term, you’re probably doing okay; in fact, in the short term, jobs like plumber are the ones which are hardest to automate. Every pipe is different, so it’s very non-intuitive.
Accountants are easy to automate. Tax crap is easy but physical labor, laying down bricks, doesn’t seem to be going away that much.
Accountants are easy to automate. Tax crap is easy but physical labor, laying down bricks, doesn’t seem to be going away that much. In the long term, if you get truly human level intelligence, all jobs will be automated. All physical labor, all cognitive labor, and the last job to go would be the guy designing all these devices and the software to actually automate our
jobs. But, eventually, everything goes. If you look at the developments in AI, computer programming is now automatable to about 25 percent based on the latest co-pilot software. Research, in many respects, can be partially automated, mathematics for example. It’s not at the level of humans yet but you can do a lot with just computer models.
Let’s now for a moment go back to to the question of meaning. So jobs are part of the meaning of human lives. But then I was also wondering, on the other hand, can we speak of the meaning of an AI life? Can a robot meaningfully ask what is the meaning of its life? And does trying to imagine
what it means for a robot to have a “meaningful” life, does this give us any insight on what it might mean for a human?
“What is the meaning of life,” has always been something that we humans are interested in. And now the question is, can we perhaps learn something about these questions from observing robots and the possible “meanings” of robot lives?
I think a lot of it would depend on whether they are conscious or not, if they have internal experiences, the so-called “qualia.”
If they don’t, then the purpose of a robot’s life is whatever I built it for. If it’s a digging role, then the purpose is to dig. Congratulations, that’s why I made you.
If they are independent agents who they have their own experiences, then they might struggle with similar questions. But for them it would be harder, because they have met their creator. They know who made them and why.
So why is qualia such an important point there? You just mentioned it as the crucial distinction. For those who perhaps don’t know what it is, “qualia” means the subjective feeling of sensing something. When you have the subjective experience of seeing something red for example, that subjective experience of redness is what this word refers to. Now is this really something that is connected in a special way with this question of meaning or of self-consciousness? Does this require qualia, or can we imagine that
you could have some useful consciousness without qualia? Why is this point of whether a robot has subjective experiences so important?
I think they are connected. I think Chalmers does a great job narrowing it down to the hard problem of AI. Essentially, we already have everything else that might be involved in consciousness. We can process information on computers, we can do input and output — all that is easy. The hard part is explaining why you have those internal states. Why is something painful and something else pleasurable? And that’s what gives life meaning. If you don’t have any of those experiences, I don’t think you’re going to spend a lot of time wondering “why is this happening to me” — because nothing is happening to you, you are not experiencing anything.
Why is something painful and something else pleasurable? And that’s what gives life meaning.
Now that doesn’t mean you need consciousness for intelligence. You can be very very smart in terms of being optimized, in terms of solving problems
without being conscious. That’s why computers even today they are not very
internally self-aware but they’re excellent problem solvers.
So there is still a difference between being intelligent and having
those internal states.
And then you might perhaps say that even perceiving an injustice is actually independent of consciousness or qualia, right? So a robot could perceive its own position inside some social structure and perceive that it doesn’t have any rights while all the others have rights, and it could have an abstract concept of justice that tells it that this is not good. So it could demand rights without having actual consciousness. Isn’t this a
Absolutely. And I think it might actually happen in terms of game theoretic results and acquiring resources. The injustice is that I don’t
have same resources in a game or I cannot secure resources for future performance. So that could be one path to the same result, where you feel that you are being treated unfairly.
Let’s now talk about the dangerous artificial intelligence or dark AI. This is your signature research topic. You once wrote a paper, in 2016, about the taxonomy of pathways to dangerous AI. Do you still have the same opinion as you had back then, five years ago? Has this changed? And how would you see the pathways that lead to dangerous AI today, and the possibility of avoiding it?
It’s largely the same. I think the worst situation is where you have a malevolent actor, who on purpose tries to design weaponized AI; like a computer virus combined with intelligence. Because not only do they have this
malicious payload, they also repeat all the same problems. They can still have bugs in the code. They can still have poor design. So you get all the issues kind of combined together and it’s also the hardest to do something against if you have an insider in a company. If you ever have a human trying to disable safety features, there is not much we can do about it.
What about other possibilities? I mean, this doesn’t really require superintelligence, right? You can say that we already have this. We already have computer viruses and worms and all kinds of ransomware and so on. So all these are unknown threats and now you can combine them with some AI component. That makes them perhaps harder to detect or harder to remove, but this does not require superintelligence or consciousness. So where is the specific pathway to dangerous artificial superintelligence?
If you look at individual domains right now where AI is already super intelligent, whether it’s playing chess or doing protein folding; at this
point, AI is superior there. We are not competitive in those environments. Human chess players cannot compete. So if we had the same level of performance in computer security, the same level of performance in weapons development, in biological weapons development: that would be the concern.
The creation of Artificial Intelligence (AI) holds great promise, but with it also comes existential risk. How can we know AI will be safe? How can we know it will not destroy us? How can we know that its values will be aligned with ours?
So you now have this independent agent, or even a tool, doesn’t matter which, that is way more capable than all of us combined. So if they design a new synthetic virus, for example (right now is a good time to use that
example), what can we do? We’re simply not smart enough to compete at that level.
But can we then not use our own AI programs against that? In principle, this is a similar argument to saying: somebody can build a bomb. The bomb is more powerful than a human body, so you can kill anybody with a bomb. But then, of course, we can design other technological counter-measures to the bomb: Either counter-bombs, or a bigger more sturdy wall, armor and tanks, so that the bomb cannot reach us or cannot inflict any damage on us. And in principle this also could be argued against this scenario with AI, could it not?
So typically people assume that once you have this AI arms race, one of the teams gets first to the human level. And then, quickly, the super-intelligent-level AI prevents all the other AIs from coming into existence. It simply realizes that game-theoretically, it’s in the best interest of that system to secure dominance. And so the first super-intelligence is likely to be the only one. Even if it doesn’t work out like that — which it seems like it will — having this “war” between superior intelligence systems with humans as casualties probably is not in our interest. We will not be taken into account in terms of them trying to destroy each other.
Game-theoretically, it’s in the best interest of super-intelligence to secure dominance. And so the first super-intelligence is likely to be the only one.
You also tried to utilize lessons from history, in order to predict something about future AI from historical examples. I have here a paper of yours from 2018 that deals with this topic. Can you tell us a little about that? How do historical examples help us understand the problems of
That paper was initially started to show that, in fact, AI is already a problem and that it does fail a lot. Skeptics argue that AI is
super safe and beneficial and there are no problems with it. I wanted to see historically what happens. The pattern we get is hundreds of examples of AIs failing, having accidents, and the damage is usually proportionate to the domain they’re working in and the capability of the system. As we get more AIs, more people have them, they become more capable, and there seem to be more and more problems, more and more damage. We can project this trend into the future and it is an exponential trend.
We see similar trends with AI-caused accidents. There are lots of simple trivial examples: Your spell checker puts the wrong word into a message and your partner gets offended because you said something inappropriate. Google translate mistranslates a sentence and you lose some meaning. But if you have nuclear response AI-based systems, they can have a false alarm and trigger a nuclear war. We came very close to that happening a few times. But also the general pattern is basically, if you have a system designed to do X, it will eventually fail at exactly that. A system controlling the stock market will crash the market. And we’re starting to see that type of pattern emerge and probably continue. So we can use this as a tool to predict what will happen. If you have a system designed for this particular purpose, how can it fail at this purpose? What can we know in advance about its possible failure?
Microsoft at one point developed a chatbot called Tay, and they decided
to release it to public, so teenagers on the Internet could train it. If they had read my paper, they could very easily have predicted exactly what was going to happen, but they didn’t. And it was quite embarrassing for the brand. It’s a serious company, so you want to have someone in the company with this type of safety thinking.
But perhaps these examples actually might also show something else. I think
they show that the problem is not the AI, because Microsoft’s Tay didn’t actually have much of an intelligence. It was just a pattern matching program that statistically associated particular questions with answers and then gave them back again. Essentially it just parroted back what people said to it. So it is not a problem of AI. It’s a problem of creating a bad technological solution. And you could argue that also the other examples that you mentioned, atomic weapons going off by accident and so on, that these are not really AI-specific problems. They are problems caused by engineers not being careful enough or not diligent enough in controlling technology.
So how much of the problem is actually a problem of AI and how much is a problem of generally technology, or the inability of our capitalist societies to effectively control technology?
With AI, we can distinguish two types: “tool” AI or narrow AI, and that’s all
we ever had; we never had anything else. And then there is general AI or “agent” AI. So anytime you have a problem with the tool, you can blame whoever designed the tool. You’re right, we’re just showing that as the tools get more complex, the problems grow bigger. But at some point, we switched from tool to agent, and now in addition to misuse of a tool
you have intentional situations where the system solves some problem in a way you don’t like. You don’t want it to be solved that way. For example, you might have created this super-intelligent machine to solve the problems with our economies or the climate change problem. But then the solution may be to wipe out all humans. That’s the solution! No more pollution. Who is at fault then? Are we at fault for designing this, or is the system at fault?
It’s not really any one agent to blame. It’s a combination of those factors, but it doesn’t make it any better just because you can find someone to
blame. Blaming does not improve anything.
So you are saying the the critical point is where, in addition the AI as a tool, now we have AI systems as agents and that increases the unpredictability of the system, right?
Exactly. Unpredictability and explainability and controllability issues explode at that point. The point of the paper with examples from narrow AI is to show that even at those trivial levels, with simple, deterministic systems, we still cannot accurately design them. Even these systems fail and then things totally get worse as the systems get more complex.
You mentioned the spellchecker as a source of failures, and in one of your papers, I found that you actually listed a co-author called “MS Spellchecker.” What’s the story there?
I put a lot of easter eggs in my papers and I use the spell checker a lot. I can’t spell. So I figured why not give credit to the AI collaborating with me? It was before GPT3, before all the advanced language models, so I put it on ArXiv and I think at this point, Dr Spellchecker has more citations than many new philosophers. It was a successful approach.
I was later approached by a team working on a different paper to write a subsection regarding at what point do you give credit to artificial intelligence in a paper. And now this paper has been published in a very good physics journal, so that led to this really nice collaboration and Dr Spellchecker came through and I’m very happy. Since then, I had papers published with GPT3, Endnote, and many other AI products.
Now let’s talk a little about the future of AI. What can we do in order to mitigate these dangers? What is the the thing we should be trying to achieve in order to avoid the dangers of AI?
Over the last couple of years, I’ve been looking at limits to what can be done. Impossibility results are famous in many many areas of science, physics and mathematics. I show that in artificial intelligence likewise we have unpredictability. You cannot know what a smarter system will do. You have unexplainability: a smarter system cannot explain itself to a dumber agent and expect the dumber agent to fully comprehend the model. And hundreds of other results, which you can find in my papers, show that the control problem is unlikely to be solvable. We cannot indefinitely control actions of a super intelligent agent that is much smarter than us. So essentially we’re creating something we cannot control.
We cannot indefinitely control actions of a super intelligent agent that is much smarter than us.
Now what do we do with this is a completely different question. I don’t have a solution for how to address it. It seems that slowing down a little bit is a good idea, to buy a little time to figure out what the right answers are. But I’m not very optimistic in terms of progress in AI safety. If you look at AI research, there is tremendous progress every six months. There are revolutionary new papers, approaches that can solve problems we couldn’t solve before. In AI safety, mostly we identify new ways it will not work.
In terms of partial solutions, we could try finding ways how we can sacrifice some of the capability of AI systems, making them less super-intelligent, but maybe gaining some control as a result. But it doesn’t seem to work long-term as well.
But now you could say that this has always been a problem in human societies, because also human beings are not all equally intelligent, and also not equally benevolent. You always have some people like, say, Adolf Hitler around, or I’m sure you can name many other dictators. And some of them are brilliant people. Human society always has had the problem of having to control such people and not being dominated by them. You could argue that all the social institutions we have created, democracy and courts and laws, they are there exactly to somehow limit the danger that can come from these brilliant people dominating everybody else.
In ancient Athens, for example, they had this system where anyone could write the name of one person they wanted to exile onto a piece of pottery. And then they collected these pieces, and if more than particular number of citizens wrote the same name down, then this person was exiled for 10 years, and he was gone from the political scene.
So, in a way, we have always struggled with reigning in and limiting the power of people who are not benevolent and who sometimes might be superior to us in terms of intelligence. Couldn’t we also trust that similar measures will work for AI? That we can create institutions to control AI deployments, or that we can just expand the power of our legal system to control how these AI systems can be used so that they don’t dominate us?
With humans it’s very different. First of all, the difference between the dumbest human and the smartest human is very small, relatively. 100 IQ points. And there are a lot of other humans which are just as smart at 150 points. So Einstein is very smart, but there are many similarly brilliant people. A society of humans together is even smarter than any individual, so there are these checks and balances and even with that we still failed a lot of times.
We had brutal dictators and the only way to get rid of them was to wait for them to die. Now you have AI which is let’s say a thousand IQ points smarter than all the humans combined, it doesn’t die, it doesn’t care about your courts or legal system. There’s nothing you can do to punish it or anything in that way, so I think our institutions you describe, such as democracy, will fail miserably at handling something like that. They don’t do so well even in normal conditions as you probably noticed lately, but they’re just useless in terms of political pressure on technology.
So governance of AI, passing laws saying “computer viruses are illegal”… Well, what does that do? It doesn’t do anything. You still have computer viruses. Spam email is illegal, okay, you still get spam. Likewise you cannot just outlaw AI research or the creation of intelligent software, just because it’s impossible to classify what would be narrow AI versus something with the potential of becoming general and super-intelligent AI.
So I think it sounds good on paper. There is a lot of committees now, organizations talking about ethical AI, but a lot of it is just PR. I don’t think they are actually addressing the technical concerns.
Now this sounds all very bleak. So there doesn’t seem to be any way out, and you don’t seem to be proposing any way out of that. Do we have to resign ourselves now that we are going to be dominated by evil AI and there is nothing we can do? Or is there actually something we can do?
And another thought: there is of course capitalism again. I don’t know what you think about that, but it seems to me that much of the potential misuse of AI happens because of these capitalist structures. People want to make more money, they want to increase their financial gains and therefore they use AI systems. To optimise, let’s say, supermarkets, to research consumer behavior, and all these things. And they create these structures that are dangerous and that take away our freedom in the name of making more profit. So it is perhaps capitalism that is to blame. Would then getting rid of capitalism create a society that would have better chances of withstanding the temptations of AI?
I’m not sure, because China is the other main actor in the AI arms race and they are, at least on paper, very communist. So if they were first to create super-intelligence I don’t think any of us would be better off. I wouldn’t be blaming a particular economic system.
Now people do have self-interest and if you can convince top AI researchers that succeeding at creating super-intelligence is the worst thing they can do for themselves and their retirement plans, maybe that will cause them to at least slow down and take time to do it right. As far as I can tell, that’s the best we can do right now: Have a kind of differential technological progress, developing tools which help us stay safe and make scientific discoveries, while at the same time not just creating general AI as quickly as we can in order to beat the competition.
Tell me a little more, more generally, about your understanding of what philosophy is supposed to be doing. Because you are always talking about these very practical, political, socially relevant topics. But many philosophers and also scientists do research that totally lacks any practical use or any social usefulness and I was always wondering if this is a good thing or a bad thing.
On the one hand, you can argue that this “useless” research is valuable as, say, the free play of the human mind. And I’m sure that in engineering you’ll also have something like theoretical physics that is only a play of the mind and doesn’t have any practical application.
On the other hand, we have all these pressing problems. We not only have to deal with AI, we also have problems with democracy, with our political systems, we have problems with poverty, we have problems with the environment, and it seems like perhaps we cannot afford any more to to have these pure areas of research, like epistemology or philosophy of language, that are stuck on some minor problem that doesn’t have any relevance and that interests only a handful of specialists.
How do you see the relationship between pure science and applied science, and
does pure science need to justify itself in today’s world?
If you look historically, we are not very good at deciding what is actually going to be useful in the future. Take modern internet, or all of cryptography, or e-commerce. These are based on research which was originally pure mathematical research with no applications of any kind. Number theory, things of that nature. They were just mind puzzles, and today it is the most practically necessary, applied work we know about.
So I think it’s important to have this diversity of research directions. I like that other people research things I’m not particularly interested in, because it might come in very useful later. Areas of philosophy which at some point were considered very unapplied, might be fundamental to understanding the mind and its purpose. So I
strongly support that personally. It’s hard for me to understand why someone
goes, well, what is the most important problem in your field, and then why are you not working on it? Why are you working on something else?
I cannot always comprehend what others are doing, but I’m happy they do it.
This brings us to the question of what is the most important problem we’re facing right now. Famously, Elon Musk had also asked this question, and then decided that for him the most important problem was to solve transportation. So he created Tesla. And then it was to solve energy. And he revived the solar panel industry, and with it the modern battery industry.
Elon Musk is somebody who always started in this way and asked “what is the most important problem?” and then went on to tackle that. But now, if we look around, I’m not even sure that AI is actually the most important problem. We have the climate catastrophe, we have pollution and microplastics, we have global species extinction, we have all these ecological issues. Then we have democracy and freedom problems world-wide, like we mentioned before. There are many world freedom and democracy indices and they are mostly going down all the time.
So do you think that AI is the first, or the most pressing catastrophic event that we will experience, or do we need to worry that perhaps the dangers of AI are so far in the future that something else will kill us off first?
That’s a great question. There are two parts to that answer. One is AI or super-intelligence would be a meta-solution to all the problems you listed. If you have a benevolent, well-controlled super-intelligence, microplastics are a trivial problem to address. So I would be very happy with AI coming first, if we can do it right. In terms of timelines, I think things like climate change are projected to take hundred years for so many degrees, or before we all boil alive; whereas a lot of people think super-intelligence is coming in five to 15 years. So even in terms of precedence you’ll be killed by AI well before you boil alive.
So there seem to be multiple reasons to think that AI is the solution and the most concerning problem at the same time.
Do you agree with these predictions? What do you think? When will super-intelligence in a general sense be available?
I have a distribution of predictions. I don’t have a specific date with a hundred percent probability. I think that there is a nonzero chance for the next seven years, maybe ten percent. And I think, as you give it a little more time, the probabilities go up. I would be very surprised if in 25 years there was nothing close to human-level intelligence.
So even in terms of precedence you’ll be killed by AI well before you boil alive from global warming.
But what does this really mean: “human-level intelligence”? Does it include, say, human level ethical sensibilities? If it does, then this would automatically be a comforting thought, right? That if we create super-intelligence, then perhaps automatically we will also create super-empathy or super-ethics, and then we will
have a true moral agent instead of a monster. Do you think that these things go
together? That the moral faculties of super-intelligent AI will develop in parallel with its intelligence?
Usually, when I say “human-level”, I mean that the system could do any job a human
can do. If you can give it to your assistant Bob and he can do it, then the AI system should also be able to do it.
Now you are implying that if a human with human values was super-intelligent, it would be a great outcome and very safe. But humans are not safe. Humans are extremely dangerous. You brought up some examples of historical humans who killed millions of people, tortured them. Great power corrupts and corrupts absolutely, so the last thing you want is to take a human mind, upload it, and give it extreme power.
Humans are uncontrollable. We have ethical systems, religions, laws, and still we have always criminals. Just taking average human ethics and transforming AI into that would be catastrophic.
Humans are uncontrollable. We have ethical systems, religions, laws, and still we have always criminals. So I think it’s even harder than that. Just taking average humans with their ethics and transforming AI into that would be catastrophic; because we don’t agree on ethics. Between different cultures, different religions, different time periods. All the fighting is about what is ethical. In the US at least, on every ethical issue there is a 50-50 split in elections. Nobody agrees on anything. So we need to do way better than the human concept of fairness or ethics. Super-ethics maybe, but the problem is when this god-like thing tells us, well, this is the actual answer. This is what’s ethical. Then half of us will not accept it for exactly the same reasons.
And then? Should we be forced to accept it or not?
I mean, is there any value in human autonomy if this autonomy is misguided? If we can trust the AI to actually know the truth and we trust that the truth is the truth, is there any value then in our autonomy, in our ability to disagree with it? Or should we just give up the disagreement and accept what the AI says?
So this goes back to theology, right? Religion is all about asking questions.
Why did God create us with free will and the option to be evil? Is this valuable in some way?
I feel like this is a little above my pay grade, but it seems like most of us want to be treated as adults. Children are kind of treated that way, they have no full autonomy. We take care of them, but it comes at the cost of giving up their decision-making power. So some people say, hey, if you take care of me really well, you feed me well, you give me entertainment, that’s fine. You’re in charge.
But I think that a lot of people would feel like we’re losing something, something fundamental to being human, to humanity as a whole, if we no longer
are in control and we don’t even have an “undo” button to press if we don’t like what’s happening.
But we often don’t have one now, either. I mean, you say it’s above your pay grade, but it is your field of study, computer science, that has created
all these situations, in which we are already pushed around by machines. We are forced to accept algorithmic censorship on Facebook, where people who
post photographs of breastfeeding get censored. Or situations where Google will close down your account if you post something they don’t like, or if you say something that somehow doesn’t fit whatever the algorithms think should be good or permitted. And so we we have already given up a lot of our freedom of expression to algorithms that control what we are allowed to say and what not. And they do this without even really understanding what we are actually saying. So have we not already lost this particular fight?
Well, right now the algorithms are tools. Whoever controls Facebook, tells them what pictures to censor. So your problem, your opposition is with
the government or with the big corporations — not yet the tool AI.
In science it is fundamental, when you run experiments on humans, to get their permission. I don’t think we have permission from eight billion people to tell them what to do with their lives. And it would be, I think, incorrect for a single computer scientist or anyone to make that decision for all of humanity. Even if I personally feel a particular answer is the right answer, I just don’t feel that I’m smart enough to make that decision for all of us.
We’re coming to the end of this interview. Is there anything else you would like to add? Is there something that we didn’t talk about and that you think should be mentioned?
Well, you’re asking awesome questions and I love the philosophical aspect. I usually get more engineering, more technical questions. I’m curious why in the AI safety community at least there is not more engagement with the questions we discussed. People are either saying it’s not a problem, who cares, let’s just worry about the benefits of AI, always saying ‘yes’, definitely, it is an issue worth solving.
But I think it’s fundamental to start by answering the question: is the problem even solvable? Is it partially solvable? Is it a meaningless question to talk about control of intelligence? So that, I feel, is not being
addressed sufficiently. I would love to see more people get engaged with trying to show that, okay, yeah, we can solve it. To show me that I’m wrong. To say, here’s how a solution could look philosophically, theoretically, not
necessarily in engineering yet. We will get there, but just what does it mean to have a successful super intelligence system under your control? Just give me a utopia-like description, anything. But even that is not being done really well.
I think that there are multiple reasons for this. One is that this is perhaps more in the area of science fiction. It’s a fictional exercise to try to imagine a world like this and in science fiction you do find utopias of AI that seem to work.
In reality, I think we both experienced (because we’ve been together at a conference with computer scientists), we both experienced that when computer scientists come into these ethical discussions, they often just lack the
patience or the willingness to understand the moral problems. Because engineers tend to be very focused on a solution. Philosophers are the
exact opposite: philosophers are focused on creating more problems. And often you notice that the computer scientists are not very patient with this and they want to know what does philosophy say about this or that case, what is the right thing to do? And the philosopher normally cannot answer this
question. He can say, you know, there’s this side, and there’s that side, and I cannot tell you what is right. So it seems to be also a problem of different cultures. We are not communicating across disciplines and perhaps this is what makes the engineers so tired of participating in these philosophical discussions, because they don’t see that this is going anywhere.
Possible. But you say that there are lots of examples of science fiction accurately describing utopias with super-intelligent characters in them. I don’t think it’s the case. There are dystopias which are brought up, but nobody can write a realistic future description with agents smarter than them by definition. That’s the unpredictability, right?
Historically, science fiction and science were separated in time by let’s say 200 years. That was the gap. Now the gap is narrowing down. Good science fiction is 20 years ahead, 10 years ahead. When they merge, that’s the singularity point. So we’re getting there. Science fiction is, in a way, telling us where the science is going to be very soon.
The fact that we don’t have good science fiction about super-intelligent systems, accurately describing what it even means, kind of supports what I think is the impossibility of that coexistence.
And it’s always also difficult to imagine the future. I think this is also something that science fiction shows. When you read science fiction from 50 years ago, not only regarding super-intelligence, but with any technological advance, it is sometimes quite wrong what they imagined to happen. In “Back to the Future” we would have flying cars. The movie “2001” predicted settlements on the moon and huge orbiting space stations. And it was very excited about telephones with screens where you can see the person you’re talking to, but this now is an every thing. So it seems that we often get these things wrong. Technology is very hard to predict.
And I don’t know if this is specific to AI, perhaps it becomes even more
difficult with AI. But my my favourite example is always with cars. At
the beginning of the private car introduction, everybody had horses and they just thought, by introducing cars we create a world that is free of the horse droppings. Because horse droppings were the biggest problem back then in big cities like London… They were drowning in these horse droppings.
They said, we want to get rid of this and therefore we now are happy that we have cars, horseless carriages, and they don’t have this problem.
But nobody could back then anticipate that the horseless carriage would lead to the world we have today. With the cutting up of nature by having these highways in-between biotopes, and the growth of suburbs, which are only for sleeping, and the destruction of the inner cities, and all kinds of other problems: environmental problems, global warming, and a whole bunch more.
But you cannot blame them, when they created the first cars, that they didn’t anticipate global warming, right? This was then impossible to anticipate.
I kind of blame them, because the first cars were actually electric. If they had spent some time thinking about it, they wouldn’t have switched to oil. So that’s exactly the problem. We never think far in advance for long
enough to see how our actions can impact the future. Is this a good decision? You may not get it all right, but put some effort into it.
We never think far in advance for long enough to see how our actions can impact the future. Is this a good decision? You may not get it all right, but put some effort into it.
I think it’s important and we are kind of in the same position now. A lot of times in science, the first attempt is better than anything for the next 20, 30 years. In AI, the first AI breakthrough was neural networks. Then
for 50 years we did other things. And now we’re back to neural networks
exclusively. So again, electric cars, neural networks, we see a lot of this. And it’s important to think ahead. Why isn’t there a better option?
That’s interesting. I haven’t thought about it in this way… So now we come to the end of this interview and I have to ask the question: what about the flower you have, this famous picture of yours that is everywhere, and also on the Daily Philosophy site. The one with the red carnation on your lapel. This is, among other things, a political symbol. So is it for you a political statement or is it just decoration?
You are very good at finding patterns, between my black beard and my black
shirt and the flower… I am not that deep. I’m not a real philosopher. I have a limited wardrobe and I don’t have any professional pictures. If I go to a wedding and they snap a good picture of me, that’s my picture for years
to come. So I have to disappoint you. I’m not a symbolic representation of any movement. Yeah, it’s purely accidental. I don’t care much about the visual presentation. I hope my papers will speak for themselves.
Prof. Roman Yampolskiy, thank you so much for this interview! It was a great chat and I enjoyed it a lot.
It was very enjoyable, and thank you for inviting me.