Professor Emeritus, University of Toronto
Prof. Geoffrey Hinton on Deep Learning and the Importance of AI Safety
Professor Geoffrey Hinton is awarded for his leadership and foundational work in neural network architectures. His 1986 paper with David Rumelhart and Ronald Williams demonstrated distributed representations in neural networks trained by the backpropagation algorithm. This method has since become a standard tool in AI, contributing to advances in image and speech recognition. Professor Hinton’s contributions also include his work on Boltzmann Machines and the improvement of convolutional neural networks (CNNs), particularly in collaboration with his colleagues in both speech recognition and th...
About Geoffrey Hinton
Geoffrey Hinton, the Nobel Prize-winning computer scientist often called a "godfather of AI," has stated in multiple recent interviews that he believes current AI systems are already conscious. He said he rarely discusses this view publicly because it "puts people off from the other safety messages." Hinton described the common model of consciousness as "as wrong as the belief that people were designed by God" and argued that anyone who uses a chatbot regularly knows the systems understand language, calling the opposing "stochastic parrot" argument "complete nonsense." Hinton has also discussed his regret about the technology's trajectory, saying he is "quite unhappy" and that society is not doing enough work to contain risks. He cited potential massive unemployment and the longer-term risk of AI becoming much smarter than humans, noting there are few examples of a much smarter thing being controlled by a much less smart thing. He reflected on his 2016 prediction that radiologists would stop reading scans within five years, acknowledging it was wrong due to the elasticity of healthcare and his incomplete understanding of radiologists' roles. Hinton said he has become slightly more optimistic in the past year or two about the possibility of designing AI systems that care about humans or that act only as oracles, but he cautioned that predicting the future beyond a few years is like "looking into fog."
Transcript (16 segments)
✨ AI-enhanced transcript with speaker attribution
G
Geoffrey Hinton0:00
I'm Geoffrey Hinton, I'm a professor emeritus at the University of Toronto. I spent the last 50 years trying to make neural networks learn, and now they do.
It's a very big prize and I was very surprised that a developing country could give such a big prize. At that point, I hadn't really realized how successful the VIN group was or how rapidly the economy of Vietnam was developing. So it's completely changed my understanding of Vietnam. It's got this very successful group and the economy is developing very fast.
The development of AI, the recent explosion in AI, has been caused by three things. There's the research that went into developing better ways of allowing neural networks to learn, and that's the kind of thing Yoshua and Yan and me and many other people were doing. But then there's the very fast compute that you get from GPU boards and the massive data that you can get. A very nice aspect of this award is that it recognized Jensen Huang for pioneering GPU boards that are good for AI. So it wasn't just that there was a graphics processing unit; it was that there was software for it that made it really good for doing AI on. My group in Toronto was one of the first groups to make use of those boards for AI and got dramatic results.
And then Fei-Fei Li, early in her career, decided we needed a much bigger data set of images in order to be able to see whether the neural networks worked. So she put a huge amount of work into developing a big data set that was crucial in showing that neural nets were good at recognizing objects in images. So that combination of the GPU boards from Jensen Huang and the data set from Fei-Fei was sort of crucial to the modern development of AI, and it's nice to see them recognized as well as the people who worked on the neuronet algorithms.
I remember in our lab discovering that when you use GPUs for training neural nets, they went 30 times as fast, and that was huge. That was like the speed up that you got in computation over a period of 10 years. All of a sudden we were sort of 10 years in the future in terms of how fast you could compute, and that made a huge difference to development of AI. That was very exciting.
Almost everybody in AI believed that the way intelligence worked was you had symbolic expressions in your head and you had rules for manipulating them. So it was like logic. In logic, you take some premises and some rules, you manipulate the premises using the rules, and you get some conclusions. And that was their model for intelligence. It just seemed to me that was hopelessly wrong and hopelessly unbiological, because dogs are intelligent and they don't do that. So it seemed to me that we needed to figure out how you could learn the connection strengths in a neural network, and that would give us a kind of intelligence that was much more like human intelligence and would do things like explain intuition. The logical approach could never explain what intuition was. These big neural nets are intuitive in the same sense as we are.
For me, seeing that that symbolic approach just didn't make sense, I thought that as we made AI more and more like the brain, it would get smarter and smarter, and eventually we'd end up with something as intelligent as people, but it might take another 30 or 50 years. By 2023, I was worried that it might only take like 5 to 20 years and it might be much smarter than people. Yeah, that's when I got scared of AI.
What you've got to realize is that these large language models came from a little language model a long time ago which was using back propagation to predict the next word. That model was not designed for technological reasons; it was designed as a model of how humans learn the meanings of words. So these large language models, just like us, don't store any text inside them. They generate text when they need it. What's inside is just a whole bunch of neural activities and interactions between neurons. That's the same for the artificial neural nets and for real neuronets like us. They process language in much the same way we do. Now many linguists will say that's rubbish, because linguists have an alternative theory which never worked.
So it's not about developments in AI; it's about us actually getting a proper model of what consciousness is. No, the big issue about whether they're conscious authentically or have subjective experience is not to do with scientific progress in AI; it's to do with the fact that almost all people in our culture have a very crummy, incorrect model of what the mind is and what consciousness is. Many people say, 'Well, they're definitely not sentient,' and then you say, 'Well, what do you mean by sentient?' and they can't tell you. So it's very funny to be confident something's not sentient when you don't know what sentient is. Let me give you an example of an AI having a subjective experience. Suppose I have a multimodal AI which can point, and it has a camera, and it can talk. I train it up, then I put an object in front of it and say, 'Point at the object.' It'll point at the object, no problem. Then I put a prism in front of its lens when it's not looking, and then I put an object in front of it and say, 'Point at the object.' It'll point off to one side. Then I tell it, 'No, that's not where the object is. I messed with your perceptual system using a prism. Actually, the object's straight in front of you.' And the chatbot says, 'Oh, I see. The prism bent the light rays, so the object's actually there, but I had the subjective experience that it was there.' Now, if it says that, it's using the phrase 'subjective experience' exactly the way we use it.
There are many different risks of AI. It can do wonderful things, and it will be incredibly beneficial in areas like healthcare, tutoring, designing new materials, or designing new drugs. It's going to be wonderful for things like that. It's going to increase productivity in almost every company, because anytime you want to predict something from data, AI will help you make better predictions, particularly if you've got a lot of data, and most companies do. But there's a whole bunch of risks associated with it. There's some short-term risks and there's one particular long-term risk. Many people have talked about the short-term risks. I think those are very important. I've spent most of my time talking about the long-term risk because some people think that's just science fiction. The long-term risk is that AIs will get smarter than us and take over, they'll take control, and we have no experience of things smarter than us. We don't know what it's going to be like, and we don't know whether we can keep control. If we can't keep control, that might be the end of humanity. If we can keep control, it might make life very easy for everybody. I don't know.
In those circumstances, you might think it would be a good idea to just pause the development of AI, but that's not going to happen. There are so many good things it can do and so much money to be made by developing it that the big tech companies are going to develop it very fast whatever we say. So it's not an option to stop the development. The only option is to try and figure out how we can keep it safe when it's more intelligent than us.
In healthcare, for example, it's going to be good for designing new drugs, it's going to be good for making better diagnosis. In North America, for example, about 200,000 people a year die from bad diagnosis, and AI is already better than doctors at diagnosing in difficult cases. AI combined with a doctor is much better than doctors, and AI is improving all the time, so there's going to be far fewer deaths from bad diagnosis. It's going to be very good for intelligent tutoring. Data suggests that if you give a child a tutor, they learn stuff about twice as fast as they do in the classroom. An AI will be able to, not yet but it will in a few years, have intelligent tutors who understand what it is that the child doesn't understand and can give them just the information they need to correct their mistakes. That'll be great for everybody, including adults. Already, when I want to know something, I ask GPT-4, and it normally gives me a very good explanation.
There are optimists like Yann LeCun, my friend Yann LeCun, who think it's going to be easy to keep it safe because we make these things, so we're going to be able to make them so they don't want to take control even if they had the ability to. And then there are pessimists like Eliezer Yudkowsky, who think it's almost certainly going to take over and we should start bombing data centers now. I'm neither of those. I'm in between. I think it's very, very uncertain. And if something's very, very uncertain, it pays to think it might happen but it might not happen. And what we should do is work very hard to make sure it doesn't happen. We don't know if we can prevent AI taking over. We should be doing a lot of research now on whether we can. I'm worried. I don't think doom is inevitable, but I don't think we should ignore the possibility. I think we should be working very hard now to avert it.
Hopefully, a lot of the smartest students will go into AI safety, will go into figuring out ways to make it safe and to deal with all of the various threats, both the long-term threat that I just talked about, the existential threat of it taking over, and short-term threats like it facilitating cybercrime. Last year, for example, the increase in the number of phishing attacks was 1,200%, and that's because these large language models can make phishing attacks very convincing. It's going to be very bad if they make lethal autonomous weapons, which they're already trying to do, and they'll succeed in a few years' time. There's going to be very nasty things happen with those. And all the regulations at present by governments don't regulate those. The European regulations, for example, have a clause in them that explicitly says none of these regulations apply to military uses of AI. So we're going to see very nasty military uses of AI, and I don't think there's much stopping that. Maybe after they've done horrible things, we'll get something like the Geneva Conventions for chemical weapons, which basically worked.
There's also fake videos destroying democracy, targeted fake videos. Initially, I thought we could maybe mark fake videos as fake, insist you had to be marked as fake. Governments do that with currency. You're allowed to print things that say they're a dollar bill as long as they say they're fake. So in Monopoly, you have Monopoly money, but it's clearly fake, that's fine. But if you start printing dollar bills and you pretend they're real, that's a very serious crime. And it's a crime even to pass them, even if you didn't make them. If you know they're fake, it's a crime to pass them to anybody else. So maybe governments could have done that with fake video, but it's very technically difficult. And it's probably easier to have some way of marking real videos as real rather than fake videos as fake. So for example, if you have a political advertisement, you could have a QR code in it that takes you to a website, and if that website is the website of the campaign and it has the identical video, all of which could be checked by your browser, then you know it's real, and if it doesn't, you know it's not real. So you can do things like that, and people are working on that now.
The advice I give to young people is, first, work on AI safety. But that's not the only thing you should work on. My experience has been that the way you do the best research is by pursuing something you're really, very interested in. Curiosity is what drives really good research. In particular, you should look for something where everybody else seems to have one approach and you think there's something wrong with it. You just have a gut feeling that there's something wrong with the way they're doing things, and you need to explore that. Often it turns out your gut feeling was wrong, but sometimes your gut feeling is sensible: people are doing things wrong, and if you explore it long enough, you figure out what it is they're doing wrong and you figure out how to do it right. That's where a lot of the very best research comes from. So if you can do that, you might do some very good research.