# 2022-09-20 - Peter Werkhoven - Quantum Computing, AI and AGI [YouTube](https://www.youtube.com/watch?v=J8en0igZZmc) Duration: 00:31:39 ## Transcript ### TNO narrator 00:00:00 - 00:01:27 What is the future vision of AI? In 2022, TNO will be 90 years old. A number of experts from the organisation have written their future vision on AI. They wonder where we will be in 10 years' time with regard to artificial intelligence. And they wonder what it means for industry, mobility, sustainability... our health and research itself. The full text is available online. But we were also curious to see how others would think about TNO's future vision. That's why we invited some prominent people from the business world... politics, culture and science to read our story. They will go into the subject a bit deeper in a conversation with TNO. Peter Werkhoven, Chief Scientific Officer at TNO, speaks digitally with David Deutsch... physicist, professor at Oxford and one of the pioneers in quantum computing. In 1997, he put his vision into the book The Fabric of Reality. Together, they discuss the meaning of quantum computing for AI's development and application. Will AI ever be able to generate explanatory knowledge... or to learn ethics from people? Let's take a look at Deutsch's vision of the future. Hi, David. Good to see you. ### Peter Werkhoven 00:01:28 - 00:01:56 Hi. Really good to see you and to have you in the studio for an interview... about digitalisation, about AI, data, the role in human society. And you're one of the persons I love to interview about this topic. Back to your contributions to science in terms of quantum computing, quantum algorithms. What do you think quantum computing will mean for AI? ### David Deutsch 00:01:56 - 00:06:38 The way that I understand it at present... I don't think that quantum computation will contribute to artificial intelligence in the sense of autonomously acting information processing which performs useful functions. Either in the sense of mindless machines which I would call artificial intelligence, AI... or in the future artificial general intelligence, AGI... which is like doing the thing that humans do, thinking, explanatory knowledge creation. That sort of thing. I don't think that quantum computers will play a central role in either of those things. There will be specialised roles. There are particular functions that quantum algorithms can perform... stupendously more efficiently than any classical algorithm. But I think at present it looks as though those will be confined to special purposes. So, for example, drug design or game playing. Those might be quite important, but they are not the generality of either AI or AGI. The difference is, and I often say, that AI, what's currently called AI and AGI... are not only different from each other, they are very close to being the exact opposites of each other. The reason is that current AIs, like an AI that diagnoses diseases... or an AI that plays chess... or an AI that controls a huge factory... those things have objective functions. That is, they have a function that they are designed to maximise. That is why they are used in those particular applications. Or in military terms, you could say the objective is to hit the target. You might say the objective is to hit the target unless something is specified. But a specified thing comes up, in which case don't hit the target and so on. This is, as I said, almost the opposite of what humans do when humans think. For a start, the AI has to be obedient. That is, it has to actually do the things it is programmed to do. Whereas a human is fundamentally disobedient, especially when being creative. When a human plays chess, they are performing a completely different kind of computation. They don't do the same things. They don't investigate the same possibilities... that the artificial chess playing machine does. Because the artificial one is capable of looking at billions and billions of possibilities. Whereas the human can only look at, what, hundreds or something. So they are doing something completely different. Another difference is that the human can explain, can write a book later... having become world champion, can write a book saying how I did it. Whereas the computer programme that beats the world champion can write no such book. Because it has no idea how it did it. It was just following a programme. I was doing this and that and that and none of that is illuminating. And also, third thing, the chess player can decide, I don't want to play chess anymore. From now on I will play Go or from now on I will play tennis. And if commanded to play chess, the functionality will deteriorate completely. So those things are different. ### David Deutsch 00:06:38 - 00:06:56 What we want in an AGI is that it behaves in a way that cannot be specified in advance. Because if you specified it, you would already have the answer. The AGI programme has to give unexpected answers. Answers to questions we didn't even know how to ask. ### Peter Werkhoven 00:06:56 - 00:07:43 That's fascinating. You mentioned a few differences between humans and machines or AI. Do you think the way humans think... You also mentioned thinking, which is not deep learning and pattern matching... that we know at the moment, but that has to do with reasoning. There has been done research that shows that at a cognitive level... humans process information with a bit rate of 50 bits per second, which is not much. Is that the limitation of human thinking? Is that the reason why we think totally different when we play a game of chess or Go? Or do any problem-solving? Or is it that the way we do it is much more efficient and effective? ### David Deutsch 00:07:43 - 00:09:51 First of all, I don't think it's as slow as 50 per second because the brain is somewhat parallel. We don't know how parallel it is. But the total number of operations per second is probably many more than 50. But it's many less than a billion. So we're not in the same league as computers... when it comes to the number of elementary operations per second. So we work not by pursuing every possibility and ticking them off one by one. We work by understanding. So a chess player works by first looking at a chessboard... and tries to understand the situation, understand what is at stake here. And then conjectures... I'm a follower of the philosopher Karl Popper... who taught us that science and thinking in general... is done by a process of conjecture and criticism. No programme at the moment can do that. It can't do anything like that. And as I said, in a very important sense... this is the opposite of what an AI does. Because the way to make a better AI... is to find better ways of pruning the algorithm... so that the programme is unable to investigate... or is forbidden from investigating certain things... which the programmer has decided in advance are going to be a waste of time. It's that the human can create something new... something that the programmer didn't and couldn't envisage. One day we will know how that is done as well... but it is not done in the same way. It's not done by pruning, by preventing the programme from thinking certain thoughts. It's the opposite. It's by allowing the programme to think any thoughts. ### Peter Werkhoven 00:09:52 - 00:09:56 How would you define the notion of understanding? ### David Deutsch 00:09:56 - 00:12:48 Well, understanding to me is very closely related to explanation. So an explanation of something... is a statement of how invisible things affect the visible things. How the things that we don't know about affect the things that we do know about. So in chess, for example, the invisible things are abstract things. They are things like... that knight is dangerous there. It's better to restrict its movement. Now, there's no such thing as dangerous or restrict its movement in the rules of chess. Chess-playing computers don't know about such concepts. They don't use them. So what this means from the point of view of what an explanation is... the human chess player has conjectured that there is an underlying structure... beneath the rules of chess and this particular chess game. And has decided to think in terms of this underlying structure. So the explanation, which may be in words, like I just said... it may be that knight is dangerous and therefore... but it may not be in words. It may be an inexplicit understanding... which is also an explanation, though some explanations are not expressed in words. They are expressed in... mental constructs that haven't been verbalised. And a combination of those two is used by humans for all sorts of things. ### Peter Werkhoven 00:12:48 - 00:12:58 That's interesting. Because you talk about understanding in terms of explanations... or it's very close to explanations... explainable AI is a big discussion at the moment... because many people think that if we design AI for serious applications... high-risk applications, AI should be able to explain itself. And for that purpose we are trying to combine deep learning... the learning aspect of cognition with reasoning. Causal relationships. And within AI at the moment you don't see this... because the nodes have very abstract representations. There are no causal relationships. But if you combine reasoning with deep learning... and you can define causal relationships... within the AI process... Do you think that we will be able to get explainable AI... at a level where humans can understand the explanation? ### David Deutsch 00:12:58 - 00:15:08 Yes, but I don't think that this is a step towards AGI. On the contrary, it's another step away from AGI. The computer programme may be able to say... I diagnosed that this was cancer because... there was this result from the X-ray and this result from... and we'll be able, in a way, that the human can understand. But present-day AI will never be able to say... I guess that there is a new disease here... which operates in the following way, like a human doctor might. AIs in a very deep sense cannot create anything new. They can create new implications of existing knowledge... which has been put into them. They can, just like a programme that calculates digits of pi... can say, I have a world record, I've computed the 10 billionth... decimal place of pi. That is calculating implications of existing knowledge. But it cannot say, I don't want to do this anymore... because there's a better way of calculating digits of pi. ### Peter Werkhoven 00:15:08 - 00:15:12 That's a good one. I think that's a very important notion. So if you talk about generating new knowledge... we have examples where AI discovered new antibiotics. I think it's done by matching chemical properties... by the properties of cells, for example, or bacteria. That's not really explanatory knowledge. That's not really new knowledge. It's just doing the hard work. But how do you feel about AI discovering new laws of physics... based on all kinds of patterns that you feed into AI... of moving objects, very complex patterns... and it comes up with new laws of physics. Not new to humans, but new to the algorithm. Do you think we can classify that as explanatory knowledge? ### David Deutsch 00:15:12 - 00:17:34 No, I think all such cases are going to be cases where... the knowledge has been pre-programmed into the programme or its data... by someone who already knew that explanatory knowledge. They may not have done it consciously... but it's very easy when you're writing the programme... to include knowledge that you know. For example, knowledge of how to partition the possible data... into what is relevant and what is irrelevant. Or even within the relevant things, what the types are. For example, as you mentioned... all the positions, for example, is a category that you would feed into the programme. So you're already telling the programme... to find an answer in the form, all the positions obey such and such. But that's not what physicists do when they discover a new law. They discover the concepts behind all the something. So you don't know what to ask until you have an explanatory theory. So you already feed AI by the type of question you ask. And this brings me... And then it's... Sorry, as you said, it then does the hard work. But the hard work is not the discovery of a new law. ### Peter Werkhoven 00:17:34 - 00:17:48 No. I think this is a really crucial part in the development of AI... because many people say AI should do what humans do... and that's showing moral behaviour. We don't only follow the rules, that's what you said a few minutes ago. We also cross the lines. We make trade-offs... in how we relate to each other, how we make decisions in critical situations. And later on the judge may judge whether this was a good trade-off. But we do it, we cross the lines. Which you could call moral decision-making or ethical decision-making. Now, a very important question is relating to what you just said. Do you think that AI, and not necessarily general AI... but whatever form of AI... will be able to learn moral behaviour... from its interaction with human beings? ### David Deutsch 00:17:51 - 00:19:46 Well, only in the doing the hard work sense. The real content of the moral behaviour that it executes... will have been put in by the programmer. And it's a bit like training a dog. So a dog can't create explanatory knowledge either. A dog can be trained to follow certain moral patterns. And it's actually amazing what dogs can be trained to do. I saw a video recently where a dog was trained to intercept a knife. Somebody attacking somebody with a knife. And the person then did many things that to a human... look a bit like attacking somebody with a knife. And the dog just sat there and just smiled and did nothing. And then when they really did attack the person with the knife... the dog intercepted it much faster than a human can react. Machines can no doubt do that sort of thing even better than dogs. But that's not moral behaviour. That is just obeying rules. When it comes to moral behaviour, we will need to have an AGI. And then we will have to lose the feature of obedience. So an AGI that is constrained to be obedient is a slave. And having slaves and basing an economy on slaves... is a way to total disaster. ### Peter Werkhoven 00:19:46 - 00:21:43 And if you talk about self-driving cars, for example... or autonomous responses to cyber attacks... which is also intelligence software... Do you want these systems, self-driving cars, response systems... to be slaves or do you want them to be disobedient... and do whatever they like in those roles? ### David Deutsch 00:21:43 - 00:23:09 So I keep having to stress that we have no idea how to make AGIs at the moment. No. So it's not whether I want it to be a mindless entity. It is a mindless entity and cannot make moral judgements. In fact, it can only execute other people's moral judgements. So in the case of self-driving cars... I think it's an incredibly difficult problem and I have no idea when this will be solved. I think perhaps people are being too optimistic at the moment. But in principle, I don't think it's a great problem. It is very similar to guide dogs for the blind. So the guide dogs for the blind also have to assess traffic... and assess other people and assess certain types of threats... and ignore others and so on. And again, it's amazing what guide dogs can do. And I think autonomous cars are probably a harder problem than that... but it's really the same kind of problem... and doesn't require anything fundamentally different from training a guide dog. Yes, sometimes even a non-autonomous car can kill people because of a design flaw. Somebody will press the brake under circumstances... where the anti-skid system should not have been engaged... and it was engaged and therefore kills someone. Now we say that that's obviously a matter of design... or maybe it's not even a design flaw, it's a design trade-off. It's saying that 999 times out of a thousand... that way of judging the situation will be the right one. And the thousandth one kills somebody just in the way that... sometimes people are killed by their own pressure cooker. ### Peter Werkhoven 00:23:09 - 00:23:49 I think humans in very critical situations... for example car driving and accidents, they just panic. They don't show any moral behaviour at all. But still, you know the moral machine experiments of MIT... they try to let people decide between two different situations... to see whether gender, for example, and age of humans... determine the value of life and what should the car do. I don't think this is a very representative experiment... to progress on self-driving cars. But the question is, do you think there is any potential... in trying to make human moral decision-making... the basic attributes and their values... to make them explicit, quantitative, understandable for machines... so we can make self-driving cars showing moral decision-making? Do you think that's...? ### David Deutsch 00:23:49 - 00:26:58 In principle, yes. In principle, it's always worth trying to make our inexplicit ideas explicit... so that they can be more efficiently criticised and understood. But I think it's very dangerous to think that one can do this... in any kind of complete way. The conscious and explicit part of our thinking... is a small layer on top of the important part... which is the inexplicit part. And that will always be so. We can make some of the inexplicit part explicit... but there will always be more underneath... that we can't yet express in words. And if we take the part that we express and take it too seriously... like regulators in some industries sometimes do, for example... then we end up with a ridiculous implementation... that obeys what we think we know... but actually in practice will completely go against common sense. Also, I think these experiments you speak of... I don't think that much can be deduced from those... because the participants in an experiment are in an artificial situation. They have gone to this experiment because they've been invited... and then they go in the door where it says something above so-and-so laboratory... and then they go up the stairs and into the room and so on. And this is a completely different experience... to getting into your car to go to work today... and then unexpectedly finding a sheep in the road. It cannot be compared. So I think the outcomes of such experiments are not helpful. These things can't really be experimented on at all. You can train a guide dog to do that job reasonably well... in fact so well that everybody is impressed with how well it can do it. I think the same will be true. The ultimate limit of self-driving cars will be that limit too. They will do it very well. Sometimes they will make a mistake. But in general they will do it so well that people will say... Oh my God, I'm impressed with how it dealt with that situation. And this will be done mostly by feeding in rules... just like with the dog. The rules contain both explicit and inexplicit knowledge. The rules automatically contain knowledge that we don't know how to express. But writing down the rule is a way of expressing it. And then once you've done this with a few dogs you know the pitfalls. You know that sometimes unless it's been trained to do so... the dog will make a mistake of a certain kind and then you know that you've got to do better. But I'm sure that the guide dog for the blind... looks after its blind customer better than say a human would... because a human is much less devoted to the task. Sooner or later the human will make a mistake. I think the guide dog is a great metaphor. ### Peter Werkhoven 00:26:58 - 00:29:31 What would you say is the greatest challenge for humans... to use the potential of AI and not to be the victim of AI in the coming years? ### David Deutsch 00:29:31 - 00:30:04 So with AI it's important to realise that... AIs are fallible and not entirely controllable by the rules... simply because we do not know what rules to give them. Just like dogs are fallible. So when we let's say introduce a machine to do surgery... or to diagnose disease... we must expect, although we will have built in lots of inexplicit knowledge... as well as explicit knowledge into the rules... we must expect that the first one will not be as good as the tenth one. And therefore it's important to supervise the first one... in the same way that we supervise a human surgeon... on the first time they remove someone's appendix. There has to be a more experienced surgeon standing by... and looking out for mistakes. And if that surgeon is experienced in training young surgeons... then that surgeon will get experience of what kind of mistakes a young surgeon makes. And it should be the same. We mustn't ever rely on an AI being right when we don't know. We must be aware of saying, well, it's doing so and so. And it's usually much better than a human. So it must be right. It must be supervised in that sense. But then, once a technology is mature... and these mistakes have become far less than human mistakes... then we can begin to rely on it in the same sense that we rely on... an elevator, an automatic elevator, not dropping us to the ground. You know, the very first time a steam engine was demonstrated to the public... it killed somebody. I mean, a steam locomotive. It killed a member of parliament. Because people were not used to the situation of the locomotive... passing another object rapidly. Where rapidly meant like 15 miles an hour. And this MP, I forget who it was... but anyway, he lost a leg. And then later the same engine took him to the hospital... much faster than he could have got there any other way. But he still died, unfortunately. Otherwise it would have been a much better story. ### Peter Werkhoven 00:30:04 - 00:30:25 But one burning question is this large community of applied researchers... that's working on making AI applications in a responsible way... for all these domains, mobility, energy, human healthcare, whatever. What advice do you have for this community? ### David Deutsch 00:30:25 - 00:31:15 To the extent that they think that they're making AGI, they're not. And they're going down a blind alley and there's a danger of fooling themselves. Not to mention wasting their effort. In regard to making AIs, the sky's the limit. I mean, every year I'm amazed by the power of AIs. And I think the dangers are not apocalyptic. The dangers are similar to the danger of any new technology like the steam engine. We have to be wary. We have to make sure that... in applications that look as though they don't require supervision. In fact, they do until the technology has matured. ### Peter Werkhoven 00:31:15 - 00:31:18 Thank you so much, David, for your... ### David Deutsch 00:31:18 - 00:31:20 Well, it's been fun. ### Peter Werkhoven 00:31:20 - 00:31:28 For your explicit thinking, but also for your implicit wisdom... that you brought into this conversation. Thank you very much. ### David Deutsch 00:31:28 - 00:31:29 Well, thank you.