Thank you for the introduction. Today we are commemorating the 75th anniversary of Turing's paper 'Computer Machinery and Intelligence'. You know who I am. I'm Jose Hernandez-Orallo and this is a privilege to conduct an exceptional fireside chat with Professor Geoffrey Hinton. We will cover a series of themes on the prospects of superintelligence, even if many people don't like this word, but the AGI word is even worse, for about 40 minutes and then we will give the floor to the audience for questions. So a little bit of introduction, even if he doesn't need it. Geoffrey Hinton graduated in experimental psychology here at King's College Cambridge and did his PhD in artificial intelligence in Edinburgh. Then he held several academic positions in the UK and the US, including a part-time position at Google, and finally in Toronto where he has spent a great part of his career. He is now University Professor Emeritus there. His contributions and perseverance have taken the field of machine learning to the prominent place where it is today. He's associated with backpropagation, with solving ImageNet, with Boltzmann machines, and so many other great ideas and achievements. Yet I'm surprised by the impact of an idea, the idea of using backpropagation for learning word embeddings. There are so many papers today trying to get rid of word embeddings in LLMs, but still we are using these embeddings as a first stage of these models. He has received so many accolades that we will lose precious time enumerating them. As everybody knows, he's a Turing Prize winner, which was considered the Nobel Prize for computing, and I was very proud that we had the Nobel Prize for computing until last year. He received the Nobel Prize in Physics for foundational discoveries and inventions that enable machine learning with artificial neural networks. Let's give a warm welcome to Professor Geoffrey Hinton.
Yeah. Welcome.
Maybe they can put my picture up.
There's no need. We have the real you. So, okay. So I said we are commemorating 75 years of the imitation game, but it's been almost the same time for Samuel Beckett's play 'Waiting for Godot', a tragicomedy in two acts. For 75 years we have been shouting 'Think!' to Lucky, but only getting a monologue of gibberish. So Jeff, how many times did you shout 'Think!' to your neural nets? How many years waiting for backpropagation to work so well? These are rhetorical questions. So finally the waiting is over. Turing's goal seems to be achieved. The formerly incredulous society of the 1950s is now talking about machines that can be intelligent or even outsmart people. So it's now high time to revisit the question, not whether machines can think, but the Gallup poll one. And this is the first question to you, Jeff: do you think public opinion has finally admitted that machines can think?
Quite a lot of it, yes. People who use the chatbots, it's very hard for them to live with the story that they're just doing autocomplete and they're just a statistical trick and they don't understand. People who use the chatbots obviously believe that they understand.
Okay. So people are convinced. So they don't think that they are tricks anymore.
I think it's only philosophers and symbolic AI people who just couldn't change their views who believe they don't understand. In fact, I had a discussion with a philosopher. I use the example: you say to a chatbot, 'I saw the Grand Canyon flying to Chicago.' And the chatbot says, 'That's nonsense. The Grand Canyon is much too big to fly.' And you say, 'No, no, I meant, well, I was flying to Chicago. I saw the Grand Canyon.' And the chatbot says, 'Oh, I see. I misunderstood.' Well, if it was misunderstanding, then what's it doing the rest of the time? And when I said that to a philosopher, they said, 'Oh, there's two kinds of understanding. There's the kind of understanding that people do and the kind of understanding the chatbots do.' And that's a different kind of understanding. I don't believe it.
Okay, good. We'll talk about philosophers at the end as well. Okay. You said on some occasions that psychology, your years here at King's College, didn't help you much to answer the question of how the brain worked. You rebelled against this and followed Richard Feynman's motto: 'What I cannot create, I do not understand.' You pursued a career in making machines learn. Do we know how intelligence works better as we are creating more intelligent machines?
I think we do know better. We don't know yet. We don't even know whether the brain does some form of backpropagation. Nobody knows how to implement backpropagation in the brain efficiently. People can have theories about how to implement it. So for small networks, you can get pretty much the behavior of backpropagation, but as soon as you make the networks big, we can't get any theory of how the brain could do that. So that's an embarrassing thing. But what we do know is that if you take a neural net, an artificial neural net, and give it lots of data and get the gradient somehow and just keep updating the gradient based on the data, then it will learn really complicated things. And that's what people didn't believe. That's what symbolic AI thought was nonsense, that you could take a random neural network, give it lots of data, just keep updating it based on the gradient, and it would learn fancy things. And they always said that was nonsense. So we've learned something. That's what the brain's doing. That's what our neural nets are doing. They may be getting the gradient in a different way though.
So maybe one thing that we have learned is that data is very important.
We have certainly learned that. In fact, the first people to really understand that, I think, were Peter Brown and Bob Mercer at IBM doing machine translation. And Bob Mercer was largely responsible for Brexit. He funded Cambridge Analytica which used big data from Facebook. And I think without that maybe we wouldn't have had Brexit. So big data is very important. It's completely screwed Britain.
Okay. There are different kinds of achievements. Okay. On the other hand, and talking about understanding machines, William Thomson, Lord Kelvin, said that when you cannot measure it, when you cannot express it in numbers, your knowledge is of a meager and unsatisfactory kind. We know how learning algorithms work, but we don't really know how big neural networks with trillions of parameters work. So related to the previous question, do you think that mechanistic interpretability, the neuroscience of the artificial, or evaluation, the psychology of the artificial, can really answer this question of how they work, especially these massive ones?
Well, we don't know how we work, right? Mechanistic interpretability has got much further than I expected it to get. I spent many, many years looking at the weights in neural networks trying to understand what's going on. And for the first hidden layer, you can understand what's going on because they have weights to the inputs and you know what the inputs mean. Similarly for things just before the output, they have weights to the outputs and you know what the outputs mean. If you have a deep net though, I think it's very hard to really understand what's going on. And I suspect we'll never really understand what's going on when it answers a particular question because the real answer to how or why did it give that answer to that question is, well, it's got these trillion real value weights. And I think it's a bit like a physicist who, if a leaf falls off a tree, the physicist can't tell you where it'll hit the ground. But there's a sense in which they understand what's going on. It's just they'll never be able to predict where a leaf hits the ground. I think we're like that where the leaf hitting the ground is like why did you answer that question that way?
So, but then we can predict humans even before opening their brains or scanning their brains. We can predict some behavior, some human behavior is highly predictable and some other behavior is not very predictable. Do you think this is going to be the same for machines?
Okay. So the question is maybe we should only use them in the areas where we can predict what they are going to do.
That's not going to be much use for making scientific discoveries. Right.
Okay. No. Because if we don't understand them, that's going to be related a bit more to the questions at the end about the risks of AI. You don't put a lot of hope in that we are going to have a full understanding of how intelligence, either natural or artificial, works, or whether we can predict their behavior.
Well, we're so used to being the apex intelligence that people find it very hard to think of it from the point of view when you're not the apex intelligence. So, imagine this conversation between two cats. Because, you know, dogs have owners and cats have staff. And imagine two cats saying, 'Well, when should we use these people?' It's quite hard to predict what they do. They predictably feed us, but should we use them for anything else when we don't understand how they work? That gives you a new perspective on dealing with super intelligent AI.
Okay. Hope your 2018 Turing Award has been somewhat eclipsed by last year's Nobel Prize. But in your Turing Award lecture, you reminded us that artificial neural nets don't learn the same way that humans learn. However, recently you said that you are convinced that backpropagation is a better algorithm than the algorithms that we have in the brain in humans. And of course both of them are better than some discrete optimization and representation methods based on mutations, something that Turing suggested as a kind of first approach, or evolution itself as a very slow method for learning. But is there room for a variety of algorithms in the future, or is there a trend towards one family of algorithms to rule them all?
So there's a sense in which backpropagation is doing the right thing. It's finding the gradient and going down it. If you ask what these neural nets are, what optimization problem they're solving, it's somewhat different from the optimization problem we're solving. So we're solving the problem of if you have very many connections, like a hundred trillion. So connections are very cheap because keeping a connection going for a lifetime hardly takes any energy at all. But you don't have much experience. So we only live for two billion seconds. So we have a few billion experiences. How do you do optimization when experience is very expensive and connections are very cheap? Backpropagation is good for solving the opposite problem. How do you learn when you don't have many connections? You only got a trillion and you've got lots of data. You've got trillions of tokens of language for example. So they're solving somewhat different problems and that's one reason why you might expect them to be different. I think if the brain could do backpropagation it would be doing it though. I just don't think biology has found an efficient way of doing it, but I'm not sure and it's sort of sad that it's still an open question: does the brain do some version of backpropagation?
So things are converging into very similar architectures. Even in neural nets there were different architectures, supervised and unsupervised, even recurrent architectures, and they're still used for some particular things, but in a way some architectures are becoming really powerful to do a lot of things. Do you think that this trend is going to continue or we're going to see more diversity in the algorithms they use?
I think in digital computers it'll always be based on backpropagation. It will be based on just getting the gradient.
One comment about evolution. Evolution is working in a kind of open system where you don't know all the forces that go into turning a genotype into a phenotype. And so you have to use evolutionary methods. The difference between that and a brain is in the brain, the brain has access to all the information required to turn the input into the output. And so it's a known system and that's why a gradient method is so much more efficient. You can get the gradient and the gradient is a much better thing to have than just tinkering. But evolution can't get the gradient. And so what evolution did, not being able to get the gradient, was it evolved a brain that could get the gradient.
Yeah. Evolution in the way we know it with DNA which is discrete. Maybe some other kinds of evolution could have used gradient or maybe there's some kind of a gradient effect in evolution as well. But they are minimal. Okay. So going back to Turing, it seems that machine learning, using very few principles and lots of tweaks over the years but no priors, has proven what Turing said of a child's mind: rather little mechanism and lots of blank sheets still.
Sorry, I want to interrupt when you say no priors.
After you've done a bit of learning, that is the prior.
Yeah, you have very strong prior when you're an undergraduate, you have very strong prior and you come to things with very strong prior. And so, you know, like 10 years ago, people were comparing tabular neural nets with MIT undergraduates and the MIT undergraduates could learn things much faster because they have very strong prior. Now the neural nets, the foundation models, have strong prior and the correct comparison is give them a new task in their context and then they can learn it from few examples.
That was exactly my point actually, what I was going to say after this, that we are not comparing comparable things, so we should compare trained models such as LLMs with different stages of children. So the question was: have we already built Turing style? Is this the LLM the Turing child?
I think Turing would have been happy with what we've built.
Yeah, he might wonder a bit about the efficiency of it, the statistical efficiency. But yeah, I mean it just seems obvious to me now. People sort of wringing, 'Have we really passed the Turing test?' This is just stupid. We obviously passed it.
Yeah. But in a way, once you have frozen the weights, you have an LLM, you can do few-shot learning and that was quite impressive and everybody was impressed that they could do that and you can say they can do something like what a child can do. But still, even if now they start at this level of an undergraduate Cambridge student...
I just said MIT. I just said MIT because most of the experiments were done on MIT students. I've got nothing against Cambridge students. I was one.
Okay, good. No, but then we still have this difference in terms of the number of examples, the flexibility. We are getting somewhere but so maybe Turing will be happy but do you think that we really have this Turing child today?
I mean we haven't got it exactly like human intelligence. But we have got something you can ask it any question, it'll give you an intelligent answer and really that was the test.
Okay, talking about testing, today almost everybody agrees that the imitation game wasn't meant to be a test. It's a fantastic thought experiment. We've been discussing that today, but it shouldn't be used for evaluation as a kind of metric. And you talked recently about a test you improvised about yellow paint fading into white. This was called by some the Hinton test.
Well, it's actually the LeCun test. Yann LeCun gave me that test. He gave me three tests to try out. Yann is a symbolic AI person who's intellectually honest. And so when it started working, he was amazed. And he kept sending me mail: 'How could it possibly do this?' And then he gave me three tests and the most difficult one was that test but without the fading. So I took the most difficult one and made it more difficult and showed Yann that it could solve that.
Okay. So I didn't know that was taken from him. So maybe we shouldn't call it the Hinton test anymore. But in general we have seen a lot of benchmarks. Benchmarks were a main tool for machine learning to progress. And of course I mentioned ImageNet but there are many others and now there are thousands of benchmarks for different things. But do you think that AI benchmarks are enough for evaluating current AI systems or do we need to learn a little bit more from psychology?
I haven't really thought about that issue. I think benchmarks are good. In fact, a lot of the development of neural nets was because you could measure error scores and then mess with things and see how you could improve the error. I think having benchmarks for different things is important. So benchmarks for how long a task can it carry out unaided and things like that and how does that change with time. I do believe that we're creating alien beings. So I believe that the psychology of these alien beings is very important. I don't have much trust in psychologists though.
Okay. Maybe some professors here didn't do a great job.
Well, maybe I could tell you about an experiment I did when I was an undergraduate in psychology.
So you had to do an experiment and my experiment, which I did with another student whose name I've forgotten, was on children between the age of three and five trying to show that they changed from attending more to color when they're three to attending more to shape when they're five. And the way you do the experiment is you train the children on three objects. And the training objects are things like two yellow circles and a red circle. And the child has to learn to point out the odd one out. Or you give them two yellow circles and a yellow square. And they have to point the odd one out. Once they've learned to point the odd one out, you then give them a yellow circle, a yellow square, and a red circle. So, if they're going to use shape, they pick the yellow square. And if they're going to use color, they pick the red circle. And it was all going fine until I gave it to a 5-year-old who happened to be the grandson of J.D. Bernal, who was a famous scientist. And he looked at it and he pointed at the red circle and said, 'You painted that one the wrong color.' Now, the model that we were testing was a really dumb model. It was that kids' attention gets grabbed by stimulus dimensions. That's how psychologists described it at the time. But kids are agents. They're doing an experiment. They know they're doing an experiment. They wonder what the point of the experiment is. They think they've got the point of the experiment and then you show them something where their theory of what the experiment is doesn't work and they think maybe you made a mistake. And can you predict the response? 'You painted that one the wrong color.' Well, now we could take a trained multimodal chatbot and we could train it up on the training examples by showing it these white cards with objects stuck on them and telling it the right answer. And then we could give it a test item. And there is some chance, just a small chance, but some chance that a chatbot would look at it and say, 'You painted that one the wrong color.' In other words, the real predictions, the real behavior of real children, we're getting much, much closer to being able to model it. Whereas in 1970, when I did that experiment, there was just no chance that you could have a model that would say 'You painted that one the wrong color.' But now we have a model that might do that.
There's a lot of discussion of using LLMs as human models, people in favor, people against.
Yeah. Let's move a little bit more to the imitation game. Adversarial behavior. And the other day at another event someone told me that Turing didn't like theater. But in the coffee break, Cambridge professor Sarah Dillon told me that that was wrong. So Turing actually did like theater. So let's do a little bit of theater of imitation. The imitation game is a thought experiment. A game designed to make the substrate of intelligence not matter. It does not matter whether you're a man or a woman or a machine. A person is a mask in classical theater. The essence of intelligence is our capability to play other roles, to imitate other minds. So with enough data, how good can future intelligence become in imitating each and every one of us or even all of us at the same time?
Sorry, I didn't really understand the question. How could you imitate all of us at the same time when we all do different things?
Language models are able to. So following with the previous question, LLMs, for instance, they are good at imitating human behavior in some way.
Okay, I see what you mean. What you mean is at any one time it imitates one of us but it can imitate any one of us.
The question is whether they can do that with information about all of us from our memories and the information that we have made available.
That's sort of what they're doing now because they read half a document and from the half document they've read they understand a lot about who wrote that document, the personality of the person who wrote that document, and then they presumably in some sense take on that personality so as to be able to predict the next word. So they have to be able to imitate the personalities of all the different people on the web.
So they are the best actors that we have, even in some cases better than humans as actors.
So, but do you think that by imitating humans, and LLMs are very good at imitating humans, and maybe at the earliest stages in the history of machine learning imitating human learning, so maybe we try to imitate humans and that question was around today as well. Are we limiting what AI could potentially do?
If that's all it does, maybe we are. Although just by imitating humans, I think you can get much smarter than humans. Many people think that when you train a neural net, it can't do better than the people who produce the training data. And that's nonsense. You can train a neural net and it can do much better than the intelligence that produced the labels it's trained on when you're doing supervised training. So I have a little example I like: you take the MNIST data set of handwritten digits and there's 10 possible answers. So what you do is you corrupt the data set by making half of the answers be wrong. So you take half of the examples and you make it one of the other nine digits instead of the correct digit and then you train a neural net on that. And many people think, well you get like 50% correct because it's 50% correct in the training data. No, you get like 95% correct. You don't get 99% correct, you get like 95% correct. Because the neural net sees structure and when there's exceptions to that structure, it just in effect thinks they're exceptions. This is why your students can be smarter than you. When you say something intelligent and sensible, your students listen to it and absorb the information. And when you say something daft, they just ignore you and then they end up smarter than you. So just on human labeled training data, it can end up much smarter than the people who produce that data.
Okay, that's hope in the first part of the answer. The second part, so you've been maybe in a more extreme way than Turing himself. You've been talking about the argument from disability or consciousness, subjective experience, that you are saying that once we understand that there's no inner theater, it may be the case that consciousness doesn't play any role.
But there is... I didn't say that. You're putting words in my mouth. I didn't say consciousness doesn't play any role. I believe that people have unconscious processing and conscious processing. That is when they're slowly and deliberately working through something. Like if you're playing a game of chess and you think 'if he goes there, I'll go here and then he'll go there and oh whoops.' That's conscious deliberative processing, right?
opposed to just looking at the board and saying a knight has to go there, just because it looks like that's the perfect place for a knight. When you're not sort of thinking through the moves ahead, you're just saying that's where knights need to go.
