Transcription
[Music] I'm Nicholas Thompson. You're listening to The Most Interesting Thing in AI, a podcast produced by Atlantic Rethink, the Atlantic's creative marketing studio in collaboration with PWC, a leader in helping organizations use AI.
In each episode, we're exploring how artificial intelligence is reshaping our lives and redefining what's possible.
[Music] Today's guest is 2024 Nobel Prize winner and AI pioneer Jeffrey Hinton. Back in the 1980s, when machine learning was a fringe academic discipline, Hinton had a conceptual breakthrough. What if you could train a machine to learn in the same connective and worked way that a human brain learns? He and his colleagues developed a technique to do this called back propagation. Later, Hinton developed a probabilistic reasoning model called a Boltzman machine that allowed AI to find deep and often hidden patterns in sets of data. Many of the bright graduate students that Hinton attracted to his lab have gone on to be key figures in the AI industry. And Hinton himself went to work at Google, but he retired last year partly in a wish to talk about the dangers of AI in a free and unconstrained way.
What's apparent in any conversation with Hinton is how he's grappling with both the promise and the dangers of the technology that he helped usher into the world. Our conversation was no exception. We go deep into the idea of AI consciousness and specifically how the act of dreaming is connected to neural networks in all sorts of unexpected ways. Let's get started.
>> Hello, Jeff. How are you? It's great to have you here. Congratulations on your recent Nobel Prize.
>> Thanks for inviting me.
>> All right. What I want to do today is I want to talk to you mostly about AI and consciousness and then I also want to get into dreams and what we've learned about dreams as we've learned more about AI. Sound good?
>> Sounds good to me. First explain as you and others were building artificial neuronet networks what inspiration you took from the human brain and what things that the human brain can do that you weren't able to build into those machines at the beginning.
>> So the main inspiration which also inspired people like Fonoyman anduring was that you have a bunch of relatively simple processes that are connected together. So it's a large number of simple processes and that they learn by changing the strengths of the connections. That's true of the human brain and it's also true of the artificial neural networks. And so you must have a very flexible learning algorithm that can take neural networks and learn more or less anything if it's got the right data.
>> What is it that our brain does that you weren't able to replicate in the neural networks you constructed?
>> I believe we probably haven't got the right learning algorithm. So if you compare a person with GPT4 for example, GPT4 has about 1% as many connections as a typical human brain, but it knows thousands of times more than a typical human. So it's far far more efficient at getting knowledge into the connections. We have very big brains. We've got about a 100red trillion connections instead of the like the trillion in GBT4, but we don't live very long. We only live for about two billion seconds. So, we're very limited by the amount of experience we get, not nearly so limited by the number of connections we have. And therefore, you'd expect to have a learning algorithm that's optimized for getting as much information as you can out of limited experience rather than optimized for storing as much as you can in a limited number of connections.
>> So if GPT 4 or five or six had a learning algorithm that was as efficient as the human learning algorithm, what would be different in the machines?
>> It would learn with much less experience. At present it gets a huge amount of experience because you have many copies of the same model that can share that experience.
>> So it is able to come close to human intelligence in some ways massively exceed human intelligence in other ways because even though it has fewer connections it has far more data and can learn from other systems.
>> It's got far more data. It can have multiple copies that all learn from each other. and it's got the back propagation learning algorithm which is pretty efficient.
>> Based on what I've read and what I've heard from you, your view is very much that there's nothing that our minds can do that cannot be replicated by machines whether it's machines that have invented so far or machines that will be invented in the future. Is that correct?
>> That is what I believe. Yes. As a starting assumption, if your belief is that humans and machines are differently constructed but can end up doing the same things, that means that as a starting point, you are rejecting the notion that there is anything incomprehensible that is inside of our minds. If there is, I mean the famous line, the ghost in the machine, if there's a ghost in the machine, it's a ghost that can be created through algorithms.
>> Yes. Almost everybody in our culture has a radically wrong view of what the mind is. Because if you think there's something special about people that a machine could never have, it gives you a sort of last line of defense when you worry about super intelligent machines. And you think, well, they may be super intelligent, but we've got something they could never have. And that's just not true, as far as I can see. But the emotions, the memory, the imagination,
>> they can have those too. So imagine you were building a battle robot and a lot of people are, a lot of defense departments are. If that battle robot meets a bigger battle robot, its correct reaction is to run away. And so you will need to design it so it has fear. There have to be something in it that acts a bit like the amydala in the brain. And when it's in a situation, it's going to have to assess whether to fight or run away. And if it then has something that does that assessment and comes to the decision it should run away, that's fear.
>> So the emotion that I feel, I'm walking in the woods, I see a bear, suddenly I'm quite scared, I turn around or maybe I hold still. That emotion which I feel has been passed down for millennia through humans, that can be recreated entirely in machines.
>> It depends what you mean by entirely. So with emotions, there's kind of two aspects to an emotion. There's a kind of cognitive aspect and there's a physiological aspect. So when I get embarrassed, my face goes red. And when I see a bear, my breathing changes. Maybe we'll make machines like that in the future. But you could have a machine that doesn't have those physiological reactions, but still has the same cognitive reactions. It still thinks to itself, I better get the hell out of here.
>> right? I mean, and in fact, you could build a machine that is exactly like that. When it sees a larger battle robot or when it sees a bear, it turns red and maybe, you know, it starts to move in equivalent ways to the way we move when our respiration rate increases.
>> You could do that if you wanted to. There might not be much point, but yes, you could do that. And one thing I know from running, which is my most intense hobby, is that pain is often a conversation between your brain and your body, where you feel pain in your quadriceps or your calves, not because there's anything physiologically happening, but just because your body is scared of homeostasis, right? So, it's this very complicated of losing homeostasis. So, it's this very complicated conversation between your body and your mind. But even this complex conversation about this complex thing could be recreated is your argument.
>> Yes. I don't see why not.
>> Let's go back to where we started which is large language models are not exactly like humans. They have been fed different amounts of information. They have different learning algorithms. Is it possible that because of the trajectory they're on and the way they've been built, there may be emotions, sensations, feelings that are hypothetically things you could build into the models, but that we actually never will.
>> I think that's quite conceivable. Yes.
>> we often say there are all these things that humans can do like describe the taste of a bottle of wine or memory of your, you know, whatever your mother cooked as a child. But actually machines will be able to do that once machines are more capable and we have better sensors that are able to understand smell, taste, touch.
>> Yeah, I believe so.
>> And you don't think that any human sensory perception is out of range of machines?
>> Probably not. I mean, there may be some very weird senses we have that we don't know about that are somehow very hard to replicate, but I just believe everything is physical. Um, it's material. We came from dust. We're going to go back to dust. And if you configure dust the right way, you can make people.
>> So then the project of making machines that are fully like the human mind is both a question of getting the right sensory inputs, the right learning algorithms and piecing it all together. Now whether we should do that is one question. Whether we will do that is another question. But your argument is that we can do that.
>> That philosophically it's feasible. There may be big technical problems in doing it. We may be too stupid or too clumsy to do it. But I don't believe there's a reasoning principle. There's not some kind of spooky stuff called consciousness that machines can't have.
>> Now, do you want us to do this? I think we should be very careful because we don't know how we're going to keep control of things smarter than ourselves. When you assess the risk that more intelligent machines might want to take control and then we'd be history. In assessing that risk, I think most people believe, yeah, but we've got something they haven't got. It used to be things like creativity or language. Now it's things like consciousness or subjective experience. And once you believe they could have that too, it gets much scarier.
>> I want to talk more about the implications if we ever did end up with conscious machines. If we ever did have AI systems that most people agreed had some level of consciousness. What are the interesting implications of that?
>> Well, if you believe that they have subjective experience and they're conscious, then you think twice about turning them off and you start asking questions about should they have rights. I've evolved in my thinking about this.
>> So from where to where?
>> becoming more humanist that is selfish. Suppose we made things more intelligent than ourselves that were conscious and we denied them political rights. Is that a justifiable thing to do? And my view now is I used to think no that's not justifiable. That's being excessively humanist. But now I think well we are humans. It's what we care about as humans. And then there's the whole question of whether things more intelligent than us should have rights. And some people will say, well, no, they're not conscious. And because they're not conscious, they don't get rights. It's a decision that we're going to have to make about whether we want to give them rights.
>> But even if we decide not to give them rights, they may not put up with that.
>> Then we have a whole whole other problem. Let's talk about dreams, Jeff. One of the things that I most want to do in life on my bucket list relates to a conversation we had in 2019 and I was interviewing you at Google and we were talking about back propagation and neural networks and it was at IO and I asked you about dreams and whether there was anything that whether you had a theory of why we dream and you said you had four theories and then you gave two of them. one of which involved Hopfield networks, which is particularly relevant to today since the gentleman you shared the Nobel Prize with. Um, but you gave two of them, but you didn't give all four of them. So today, while I have you here on the podcast, I'd like for all four of your theories of dreams, why we dream, and what we've learned from AI and how it works that may change the way we think about why we dream.
>> Okay, so let me start off before I give you the theories by talking about two phenomena. One of which was emphasized by Francis Crick and Graeme Richard Mitcherson in a paper in the 80s, a widely ignored paper. So we know you dream a lot each night, but when you wake up in the morning, the only dream you can remember is the one that you were in the middle of as you woke up. And that's cuz information about that dream was in fast changing synapses. So the question is why don't you remember any of your dreams? It has to be that the normal way memory works gets turned off in dreams. And Francis Creek's theory was it wasn't just turned off, it was reversed. And so the idea is that you have things you tend to believe. And during the day, what you should do is modify your connection strengths so you're more likely to believe the kinds of things that actually happen. That's what experience is. You get experience with things. The things that actually happen then seem plausible to you. Now, as a result of that, what will happen is some other things may start seeming plausible to you that don't actually happen. And Cicks and Mitches theory of dreams is that what you're doing when you dream is you're just imagining things. You're just imagining what seems plausible to you and you're unlearning those. And if you take this combination of learning based on data when you're awake and unlearning just on your fantasies that it turns out gives you a nice algorithm and it's very important to have the unlearning. If you don't have the unlearning what happens is the whole system goes haywire.
>> So theory number one is unlearning. Now let's go to theory number two.
>> Okay theory number two came later and it was called the wake sleep algorithm. One way to do that is to say we're going to have two systems. We're going to have perception system which goes from data to abstract things. And we're going to have a generative system like computer graphics that goes from abstract things to pixels. Now imagine you didn't quite know how to wire the graphic system, but you didn't quite know how to wire the perception system. But if you had a graphic system that sort of worked not too badly, you could use it to train the perception system. Because what you would do is you'd generate stuff from your graphics system and then you'd have an image and because you generated it, you'd know what features had generated it. And so now you could take the pixels you generated and train something to recover those features. So the idea is when you sleep, you're generating and you're learning the connections in the reverse direction.
>> Oh, I see. Because you just generated the dream, you know what the causes were. Well, you know what the abstract structure was that generated this because you've gone from abstract down to pixels. And so now you can learn the connections in the other direction. That was the idea of the wake sleep algorithm.
>> I see. So the wake sleep algorithm in a way it's running in reverse what we've learned during the day but not for the purpose of learning but for the purpose of strengthening what we've learned. Fascinating. Okay. So that's theory number two. It actually doesn't whichever whether it's one or whether it's two or whether it's one of the two to come or whether it's something else has no bearings on how we sleep. Certainly it means we should try to get a good night's sleep so we can run these processes efficiently.
>> Right? So let's go to theory number three. Theory number three is quite different. There's something called an autoenccoder. So suppose I take an image and I extract some features from it and then I try and reconstruct the image from the extracted features.
>> Mhm. That will be a simple autoenccoder with one layer of features.
>> I see.
>> And it's called an autoenccoder because what you're encoding is the image in such a way you can get the same image back.
>> I see. So you extract the features and then you can recreate the image.
>> and you're running that in a loop-l like process.
>> Mhm.
>> And so another theory of what you're doing in sleep is learning these autoenccoders.
>> So you're learning the different loops that make up the different processes.
>> Right? But it's not that you necessarily go all the way up and all the way down, which would be very likewing all these little loops.
>> Oh, which is why it can feel like when you're sleeping, it's like little random snippets or disconnected snippets, right, from a story because you're
>> learning one loop that is from
>> that is part of one larger story and one loop that is part of another story,
>> right?
>> And so when you sleep, you're reinforcing these processes.
>> Yes. The fourth theory is a variation of Boltzman machines. But in Boltzman machines, you had this beautifully simple learning algorithm, but it only worked, it only applied if you let the network settle for a very long time. So the idea is you put in an image, you let the network settle for a long time, and then when two neurons are active together, you increase the strength of the connection. And then when you're asleep, you don't put in any image. You let the network settle for a long time, a very long time. And then when two neurons are active together, you decrease the strength of the connection. But the problem was it had to settle for a long time for all that to work.
>> Why does it have to settle for a long time for it to work?
>> Because the math doesn't work otherwise. Basically, what happens is when you let a system settle to thermal equilibrium, the math suddenly becomes nice and simple because you get what's called a bolt of distribution. And so when we sleep, all of this information processing is basically moving to an equilibrium.
>> Well, that's what you'd have to do, but it's too slow. So the last thing I worked on before I retired was trying to get the same kind of algorithm, but without this settling.
>> Mhm.
>> But then you have a much more kind of ad hoc energy function that isn't nearly as beautiful. So the idea is throw away the beautiful bit and see if you can still make something work that is trying to make things that you perceive be plausible and things that you generate be implausible. So we have these four theories of sleep. Theory number one I have the deepest emotional attachment to. It's the best story, right? Sleep is actually about forgetting.
>> Yes.
>> But theory number two sounds like it's right. If you were to ask me which of these four theories sounds most correct, I would say it's this process of top down, bottom up learning, syncing them together. And what do you believe?
>> My best bet is theory number four, which is not an elegant theory, but if you could get a good generator, it's still got the aspect that it generates and it tries to make whatever it generated be less plausible and it perceives and it tries to make whatever it perceives be more plausible. In a way, it's a little bit like theory number one. So, what you're saying is that while we sleep,
>> oh, it's just like theory number one as with respect to wake and sleep. It's just the energy function is completely different. You can go for something that doesn't need to settle down. So, it could be done in biology. You don't need to settle down for ages,
>> but a lot of the nice guarantees that you get and the beautifully simple math that comes out of thermal equilibrium doesn't apply anymore.
>> Right? So now we're fairly deep into this artificial intelligence explosion revolution. Has anything that we've learned or anything that we've seen in the last 3 years changed your perception of the sleep question?
>> Yes. I think for a while towards the end of the 201s I believe quite strongly that the brain must be doing some form of back propagation because back propagation worked so well. Then once we started getting things like GPT2 and beyond, I started believing the brain can't be doing something like back propagation because back propagation works too well. And it was partly seeing that back propagation seemed to work much better than anything we had in the brain. It could pack much more information into far fewer connections. That was part of me coming to believe that these things were going to get smarter than us.
for the thought experiment. If the people building AI are making decisions or if there is a way to structure society so that the most moral ethical, the most good guy men and women were making decisions over AI. Would you have faith that super intelligent AI could be built that would not be an existential risk to humanity?
>> It would improve our chances, but I don't think it would guarantee it. For me, the key question is, is there any way to make it safe? is any way to be reasonably sure you can build super intelligence and have it not take over. The big companies aren't going to put enough effort into safety. They're going to be driven more by profits than by safety. What government can do is insist they do more work on safety. They can somehow try and force the big companies who have the resources to put more of those resources into safety. And so we need to get as close to that as possible through regulation, through education, through moral pressure, while also taking other steps to limit the risks of AI.
>> Yes.
>> Tell me what you most hope to see happen in the next year with AI.
>> I guess it'd be really nice if there was legislation that forced the big companies to pay more attention to safety. That's what I'd like to see in the next year. Yeah.
>> It was a great pleasure to talk with you about consciousness, subjective experience, and dreams.
>> Okay. Thank you so much, uh, Nobel Prize winner Jeff Hinton.
>> Thank you. That concludes my conversation with Jeff Hinton. I'm Nicholas Thompson, and you've been listening to The Most Interesting Thing in AI. Join us next week when we speak with Eva Galprin, director of cyber security at the Electronic Frontier Foundation about the risks posed by AI deep fakes. If you enjoyed what you just heard, like and review it on Apple Podcast. Help spread the word about our series to more listeners like you.
[Music]