The first guest that we have who you already know and who has joined us is somebody who has shaped in a way things that may replace actual intelligence. He's often called the godfather of artificial intelligence and he's got an interesting connection to India with his uncle George Everest who was the survey general of India. He's also the great grandson of Mar, great great great grandson of Mary Bull and George Boole whose work was the very reason why computer science became what it is. His aunt was one of the two female physicists at the Manhattan project, but he charted his own path and I'm sure his family would be infinitely proud. But sometimes he speaks like how Robert Oppenheimer would, saying that he's not very proud of some of the inventions that he's done. But I personally think that there is a lot to learn from his inventions and also from the man. His research helped machines learn from data, the process that we now know as deep learning, the system supports today's leading AI tools including advanced chatbots. And coincidentally, one of his students went on and founded ChatGPT itself. He's a doctorate in artificial intelligence and a bachelor's in arts from Cambridge, but he says that the one year that he spent learning carpentry is the most important skill that he's picked up till date. Let's find out why as we welcome our first guest at a global perspective series that's put together by none other than the Nobel laureates in physics in 2024 for his work on artificial neural networks, Dr. Geoffrey Hinton. Dr. Hinton, thank you for joining us.
Thank you. Let me just correct something you said. It's not my uncle who was surveyor general of India. It's my great great grandmother's uncle.
Okay. And this is some of the dangers with information available on the worldwide web. But he's an uncle in some way. That part the internet got right. But today I must also clarify this interview is going to be supervised learning because I'm prepared, but I'm hoping that if I go a little off that you will guide the algorithm.
So let me start with the first question sir with your permission.
You've studied psychology, physics and AI and in a world that often encourages specialization you have credited your interdisciplinary study for your breakthroughs. What is the importance that you place on interdisciplinary studies with the super specialization?
I wanted to understand how the brain works.
And I didn't deliberately do interdisciplinary studies. I kept looking for a subject where it would allow me to understand more about how the brain worked. So I think you can study a subject or you can try and solve a problem. And if you're driven by trying to solve a problem, you end up having to be interdisciplinary because you may have to learn about different subjects in order to solve that problem.
So in one way or the other you need to study about different things at different points in time in your life is what you're saying. So the good part is that perhaps a suggestion would be start reading about everything when you are pretty young.
No, what I'm saying is what you need to do is focus on a problem that you want to solve and that's different from focusing on a discipline. So I wanted to solve the, I want to understand how the brain works and because of that I ended up studying a bunch of different disciplines.
Got it. So identify a problem, identify an area of interest, then look at studying different, then work on that problem and study whatever you need in order to solve that problem. Understood. Until you found your passion, until you found a problem, then what is the suggestion?
I don't know. I found the problem I was interested in when I was at high school and I've worked on it ever since.
And that's why you're the Nobel laureate giving the answers and I'm yet asking the questions. But let me go to the second part which is very important for our country. We have 86 million internet users now. Every year we seem to be adding more people in terms of those who are using the internet and using artificial intelligence and therefore how important is it for a younger generation those who are studying currently in high school and in college to be able to embrace this technology vis-a-vis some yet saying that AI is not going to be replacing your jobs or going to be taking over the future.
Well, I think it's crucial for people to embrace the technology and even if AI doesn't replace your job, I think for a while people will be having AI assistants. So, a person combined with an AI assistant will be replacing just a person doing the job. And so, I think it's crucial for people to learn to use the AI technology.
Everybody who's joining us today is studying physics or mathematics at a higher level and understands a lot about AI. But just for the benefit of so very many people who are going to be viewing this on the worldwide web and also more so to the schools that we're going to be sending it. Can you explain the evolution of AI because you've seen the growth of AI right from almost its infancy to where it is today. So can you talk about the key breakthroughs that brought us to the ChatGPT era or as it is called and referred to by many?
Okay. So AI started in a meeting in 1956 and right from the beginning there were two schools of thought. The dominant one in the last century was symbolic AI and they thought that to understand how people think and to make an artificially intelligent system you need to base it on logic. So in logic what you do is you take strings of symbols and you use rules to manipulate those strings of symbols. So a typical example would be you might have something you believe which is that all men are mortal and you might have another thing you believe which is that Socrates is a man and then there's a rule that from all men are mortal and Socrates is a man allows you to derive that Socrates is mortal. That was Aristotle. That's Aristotelian syllogism. That's a simple form of logic and it's all about manipulating strings of symbols. Now for a long time nearly everybody doing AI was convinced that that was the right way to do AI. And the alternative which a few people believed in from the beginning was to study how the brain works and to try and simulate a network of brain cells. So the people doing logic thought the sort of essence of intelligence was reasoning and we should understand reasoning and we should understand how knowledge is represented in symbolic expressions and the rules for manipulating it. That was the symbolic school. The people interested in the brain said no the central thing is not reasoning the central thing is learning. We need to understand how networks of brain cells can change the strengths of the connections between brain cells so they can learn to perform new tasks. That was a very different school of thought. Now early on two people who believed in neural nets were Turing and Von Neumann but both of them died young and so the symbolic school dominated in the 20th century and it wasn't until we got much faster computers that neural nets started winning. Now neural nets have pretty much taken over and so when people said AI in the last century they meant symbolic AI based on logic. Now when they say AI they mean neural nets and all these large language models are built with neural networks.
So, but when you started out, please again correct me if I'm wrong, but neural networks was not something that a lot of people believed in. You were
Almost everybody, yeah, almost everybody thought it was completely crazy. They thought it was crazy wishful thinking to think that you could get a neural network to learn to do complicated things. And I was told repeatedly that it's been proved that this is a hopeless approach. People tried it and it didn't work and it's never going to work. You should give it up and do symbolic AI.
So in that context, a slightly more generic question. How does one persevere when everybody is saying that what you're working on does not make any sense and all your seniors, how does one persevere just generally in life when one faces such a situation?
Okay, I was quite lucky in that I came from an atheist family and they sent me to a religious school when I was young. So from the age of about seven I was at this religious school and we were taught Christian religion and I just from the beginning knew that it was just nonsense. There isn't a God behind the clouds. God didn't make everything. And so I was in a situation where all the teachers and all the other children believed one thing and I knew they were wrong. And that was a very good start because later on when I believed in neural nets and almost everybody else said it was nonsense, I had the experience of having when I was younger being told that there is a God and then as my friends grew up they came to realize that that was nonsense.
I remember the words of Friedrich Nietzsche who said something on the lines of where there is faith there is no need, where there is truth there is no need for faith and where there is faith there is no place for truth, something something on those lines. I tend to believe that right. So most religions want you to believe something crazy. For example, the Catholic Church wants you to believe that three equals one. It doesn't make any sense. I think reason is what you should use. You obviously have to use your intuition, but you should always listen to reason. And if someone tells you just to believe something that doesn't make sense to you because of faith, you should be very suspicious of that.
And that is perhaps the beauty of science. And therefore we were very keen to start this series and we're very happy to start it with you. But AI is something that you have been responsible for, like I mentioned earlier as well that many refer to you as the godfather of AI. But today you speak of the possible dangers of AI and therefore I wanted to understand what got you to shift from being responsible for the creation of something and today you speak about how it is going to be massively misused. Do you think that there is no possible good use of AI or is it otherwise?
Okay, let's start by correcting something. A lot of people were involved in the development of AI. I was one of them, but it was a lot of people. I think AI is going to have many very good uses in areas like healthcare and education and in many industries it's going to make things more efficient. So I think there are many good uses of AI, but I also came to realize a few years ago that digital intelligence may end up being a much better form of intelligence than what the brain has. So here's something you can do with digital intelligence that you can't do with analog intelligence. Our brain is basically analog. So each brain works slightly differently. If I took the connection strengths in your brain and averaged them with the connection strengths in my brain, I'd get nonsense because your neurons work slightly differently from my neurons, you have connections in different places, it doesn't make any sense. But if I have two copies of the same digital intelligence, they can be running on different hardware. But if I have exactly the same neural network being run on different hardware, then one copy can look at one bit of the internet and figure out how it would like to change the connection strengths to absorb that information. Another copy can look at a different bit of the internet and figure out how it would like to change its connection strengths to absorb that bit of the internet. If they then just average those two changes, so both copies change, each copy will benefit from the experience of the other copy. And if you have say 10,000 copies, then each copy only needs to look at one 10,000th of the internet and then when it averages the weight changes with all the other copies, it will benefit from what all of the copies learn. So if we could do that with people, it would be tremendous. You could go to university and you could have a thousand friends and each one of you could do one course and as you're learning you could average what the connection strength changes with all your friends and at the end you'd have done one course but you'd know what was in all 10,000 courses or all thousand courses. We can't do that. When I want to share information with you I have to do it in a very clumsy way. So what I do is I produce strings of words and your brain figures out how to change the connection strengths. So it might have said the same thing, but there's a very limited amount of information you can transfer that way. So a sentence, for example, only has about 100 bits of information. And so I can't transform more than 100 bits of information to you by telling you a sentence, even if you understood exactly what I said. Now, if you have different copies of the same neural net running on digital hardware, when they average their billions of weights, they're transferring information at billions of bits per averaging. So digital intelligences are millions or billions of times more efficient than us at sharing information if they're different copies of the same model. And that might make them a far superior form of intelligence. It already makes them able to know a lot more. So these chat bots know much more than any one person. They're like a not very good expert at everything. And that's because you can train different copies of the same neural net running on different hardware and they can share what they learn from their separate experiences.
So but then you mentioned this that they can learn and you gave the university example which where you can learn so very many things at one point in time and you did also mention that they're already slightly smarter than human beings in a way.
No, they know a lot more than people. They're not yet more intelligent than people but they will be.
So do you think that there is a possibility of technological singularity coming about?
I don't know if it'll be a singularity. I think it's more likely that they'll get better than us at different times at different things. Like, they're already much better than us at chess and go and things like that. They're much better than us at figuring out how to fold up proteins. They're not as good as us at some things. And I don't think it's going to happen all at once that they suddenly get better at everything. I think different things. They're much worse than us at manipulating things physically. They're getting better, but they're still much worse than us. I don't think it's all going to happen in a singularity.
What impact do you see on schools and on education 10 years from now? Because there are a lot of educators also who are watching this. And how will AI impact education?
So, here's something we know about education in a classroom at a school. If you take a pupil and you put them in a classroom with 20 or 30 other pupils and you compare that with them having their own private tutor, when they have their own private tutor, they learn about twice as fast. And that's because the private tutor keeps them engaged, is saying things at just the right level for them. And when they misunderstand things, a private tutor can figure out what it is they misunderstand and give them just the right examples so they now understand. So a private tutor is just more efficient than being in a classroom. AI will be like a private tutor. And maybe in 10 years time it'll be considerably better than a private tutor. So I would expect in 10 years time using AI school children will be able to learn particularly technical subjects two or three times as fast as they can now in a classroom.
So do you think that brick and mortar schools will exist say a couple of decades from now or do you think that learning
I don't know. There are things the schools provide that are hard to get online, so the social interactions are important. I mean children go to school not just to learn technical stuff but to learn how to coexist with other people. The social interactions of school are very important in developing into an adult and you may need brick and mortar schools for that.
Right. The other area of concern for a lot of Indians specifically is how AI is taking away jobs. A lot of chatter across the world in terms of how this is happening in the west, in India also we see this increasingly. What is your take on artificial intelligence reducing jobs? Is this a real proposition or do you think that they will create enough number of jobs to offset the numbers that they're taking?
So some economists and a few people in AI think AI will create new jobs to replace the jobs it gets rid of. I don't believe that. I think it's going to remove lots and lots of jobs and not create many new jobs. So I think it'll cause mass unemployment and the kind of jobs it'll get rid of are early on it's going to get rid of jobs in call centers. So India actually has a lot of call centers because you speak English. So the west often uses call centers in India. And I think those jobs are very unsafe. The people who do those jobs are often not very well trained and very badly paid and they're actually not very good at answering the questions. If you call up with a technical question, they often don't know the answer. AI fairly soon will be considerably better at answering the questions and so I think those jobs are very unsafe. Similarly, paralegals, people who look for similar court cases for lawyers, that can be done by AI. Those jobs are very unsafe and they're going to disappear over the next few years. Already in the west if you train as a lawyer it's now hard to get an initial job because the kind of boring work that was done by junior lawyers is now being done by AI. Similarly with programming, junior programmers who are just doing kind of routine programs to do straightforward things, that can all be done by AI. For senior skilled programmers who are designing complicated systems, they'll be in business for a while. But routine mundane programming is going to be done much more by AI. Now it may be that a person with an AI assistant can be 10 times as efficient and what we get is 10 times more programs. I mean we could have a lot more software. So you have to distinguish between areas where the market is elastic like in healthcare, if we could all get a lot more healthcare that would be great, we're not going to get unemployment in healthcare because the market can absorb lots more healthcare, but in things like call centers I don't think the market's that elastic.
So things such as professions such as law even perhaps authors, people who write literature for them also. So AI has come up with many tools where it's helping you write but at some point in time I feel that you know it's taking away a percentage of your cognitive ability and then you're just outsourcing your work to AI as such. So to that end do you think that the more reliant that we are on artificial intelligence it will slow down the brain or is that again just an assumption?
No, I don't actually believe that. I think when we get more reliant on artificial intelligence, we won't be so good at doing some of the things we did before. So, let me give you an example with neural nets. It used to be in the early days of neural nets, when you wanted to figure out how to change connection strengths, you had to write code that figured out the gradient. So, the standard way to do learning is you get the neural network to give you an answer like predicting the next word and then you compare that answer with the correct answer. Then you send information backwards through the neural network in order to compute a gradient, that is in order to compute how changing one of the connection strengths would change the probability of giving the right answer. Now in the old days when we did neural networks we had to do all that differentiation by hand and so we were all quite good at differentiating things. But we didn't have very complicated neural networks. Then people introduced automatic differentiation where the computer software that simulates the neural net does the differentiation for you. Then we could have much more complicated neural nets but the people now are not so good at doing differentiation. If you ask them to differentiate something they say oh the software does that. I don't know how to do that. That's not really a problem so long as the software is doing it. They don't need to know how to do differentiation. It's a bit like when we have pocket calculators. When pocket calculators first arrived, people said, 'Oh, that's terrible. Students won't know how to multiply numbers together.' And I think probably now if you asked young people in school, 'What's 12 times 11?' Many of them won't know anymore. They don't need to know because they have a pocket calculator. When I was at school, you needed to know what 12 times 11 was, but now you can just ask your pocket calculator. So, you will lose some skills, but it's not making you stupid. You just lose the skills you don't need anymore.
So it's not necessarily making us stupid which is why but there must be a fine line that perhaps one must draw in terms of outsourcing their homework to ChatGPT on a lighter note. But
Well that's just, I mean if you just cheat on all your homework you're not going to learn much.
But I must also ask you what are the skills that one should focus on in the era of software such as these, in the era of artificial intelligence. What should a high school student today focus on in terms of skills that will help them in the long run?
I think it's a very good idea to learn math and physics. You can learn those much better when you're young. It's hard to learn those when you're old. And they're a very good basis for many things. So it teaches you to think, it teaches you to sort of understand models and I think it's a very good thing even if you're never going to use the math or physics. It's going to be useful in training you how to think properly. So I would advise people to keep on learning math and physics as well as learning about AI.
The other thing that one sees because of the misuse of AI is how there are deepfakes and how there are materials which are circulating which cannot be relied upon because the information is not necessarily authentic in nature. How real is this threat and what can a young person do in order to discern between what is fake news and what is actual information?
Okay. So, I think it's a very real threat. It's very hard now when you look on the web to know what's true and what isn't. For a long time I thought we should be making AI systems that can recognize when a video is fake. But it turns out that's very difficult. And I think it's more likely that we'll make systems that can recognize when a video isn't fake. So, right now, if I want to know what to believe, I read the New York Times and I look at the BBC website and I trust them because they typically insist on several sources saying the same thing before they'll publish it. And they're quite careful. What we want is a way of verifying that videos are real. And for things like political videos, that's not that hard to do. If every video started with a QR code and the QR code took you to a website, websites are unique. So if it's the website of a political campaign and the identical video is there, you know it's a genuine video. So it wouldn't be that hard to have a system where browsers can just check whether the video the QR code takes you to the right website and that website has the identical video. Then you know this video really came from that campaign. So for things like political advertisements it's not that technically difficult to verify them.
But will this kind of regulation come about considering all nations are essentially in the race to develop the technology and they want to be first in that? Do you think that there is enough importance being given to put in such frameworks in order to help systemize AI and make it more accessible in a responsible manner for young people?
No, I don't. I think the countries are competing and within countries particularly within America the big tech companies are competing to get there first and they're not putting enough thought into the dangers. Also governments aren't putting enough thought into regulating AI. It's sort of like the early days of big oil when the environment wasn't protected at all and they just caused lots and lots of pollution and the companies will resist regulation and it's the job of the public to insist that the politicians regulate AI. We need to educate the public, which is what I'm trying to do, so they understand that AI needs to be regulated. And they provide a counterweight to the big tech companies who are all busy saying, 'Oh, don't regulate AI, you'll slow it down.'
So is this without jumping into the speculation part of it which is there on the worldwide web but could you tell us if you were comfortable, was this the reason why one of your students who was responsible for OpenAI, was that a reason why he left OpenAI in terms of not having enough safety protocols that are in place on platforms such as ChatGPT?
He hasn't told me that directly. This is Ilya Sutskever you're talking about. He hasn't explicitly told me that but I believe that is the reason he left, that they weren't putting enough work into safety and he's now set up a company which is aiming to develop super intelligence safely.
And you said you must develop super intelligence safely. Whilst you're developing super intelligence there is a conversation on consciousness also. Do you think that AI would possibly ever become conscious as such?
Okay, so this is a very difficult subject and it's, I think the problem is not a technical problem. The problem is that almost everybody has a completely wrong view of how the mind works and what the mind is. And so I'm going to spend several minutes now trying to convince you that existing multimodal chat bots, so a chatbot that for example has a robot arm and a camera and can talk, those things are already have subjective experience. So most people think in the following way. They think the mind is a kind of theater and there's things going on in this theater that only I can see and that those are subjective experiences and other people can't directly see my subjective experiences and they think subjective experiences are things like a photograph is a thing. So if I take a photograph of something you can ask what's the photograph made of then where is it? Where is the photograph? And people think it makes sense to ask that about experiences. And I think that's just a misunderstanding of how the language works. So I'm now going to try and give you an example of that. Suppose that I drink too much and I tell you I have the subjective experience of little pink elephants floating in front of me. Most people would interpret that as there's something like a theater which is my mind and in this inner theater there are little pink elephants floating around and only I can see them. And if you ask a philosopher what are those pink elephants made of, the philosopher will tell you they're made of qualia which is just made up stuff to explain what little pink elephants are made of. Let me give you a completely different way of thinking about a subjective experience. So, sometimes my perceptual system goes wrong and I know it's gone wrong. And if I see, if I think there's little pink elephants floating in front of me, what I know is my perceptual system is lying to me. There aren't really little pink elephants. And I indicate that to you by saying it's a subjective experience, not an objective experience. And then what I do is I describe a state of the real world which if it were true would explain why my perceptual system is saying what it's saying to me even though I know it's not true. So I'm trying to tell you how my perceptual system is trying to mislead me by telling you a state of the world which would make my perceptual system be behaving incorrectly.
