Okay thank you for the introduction and thank you for inviting me. I'm going to give you a very brief and very biased history of AI. So since the beginning of AI there were two paradigms. There was the logic-inspired approach which thought the essence of intelligence was reasoning and reasoning was going to be done by using symbolic rules to manipulate symbolic expressions just as it's done in logic and that things like learning can wait until we understand how we do reasoning and how we represent knowledge. First we have to figure out what is this symbolic language in which knowledge is represented, how do we do reasoning and then we can worry about learning later. That was their approach and they were dominant during the last century and the early part of this century.
Then there was a biologically inspired approach where the idea was the essence of intelligence is actually learning the strengths of connections in a neural network. Reasoning can wait until later. First we have to understand how learning works and then we'll understand other things later. This was based on the idea that things like pigeons are quite smart in their own way. They're good at things like perception, they're good at things like flying through forests and they don't do reasoning—or as far as we know they don't do much reasoning—and they probably don't use logic. So really from a biological perspective reasoning comes very late and it's sort of silly to base your model of intelligence on that.
Now one thing I should say before I go on is there were two people from the 1950s who actually believed in the neural network approach. They were both very good logicians. It was John von Neumann and Alan Turing. So you couldn't really accuse them of not understanding logic, like me. They unfortunately both died young and I think if they had survived the whole history of AI might have been very different. Turing was sort of studying how you could learn neural networks using reinforcement learning.
So I'll start by talking about the meaning of a word. So within the symbolic AI paradigm, the idea was that the meaning of a word comes from how it interacts with other words in propositions or in sentences. So the meaning isn't a sort of thing you could put your hands on. It's implicit in how it co-occurs with other words. And probably to capture meaning you'd need something like a relational graph. So that was the symbolic view. There was a view in psychology in the 1930s which was very different, which was that the meaning of a word is just a big set of semantic and maybe syntactic features. So you have a large set of features, you activate a subset of them to capture the meaning of a word. So for the word cat, you'd activate features that say it's a predator, it's a pet, it's aloof, it's got whiskers, it's about the size of a breadbox, things like that. To represent a dog, you'd activate a fairly similar bunch of features, but obviously different in some respects.
That theory was very good at capturing similarities of meaning. So cat and dog would have fairly similar feature vectors. It was also quite good at capturing analogies. So for example, for most men in our culture, not necessarily for women it turns out, the feature vector that you use for cat is more similar to the feature vector that you use for woman than the feature vector that you use for man, and vice versa. The feature vector for dog is more similar to the feature vector for man than it is for woman. Now being a man I thought that's obvious to everybody. It turns out it's not. Quite a few women don't see that as obvious at all. That was a revelation to me. But that makes these feature vectors very good at doing analogies.
Back in 1985 it occurred to me that these two different theories of meaning, one that is implicit in how you occur with other words and the other that is just a big bunch of activations of features, they're not necessarily incompatible. They may be just two different sides of the same coin. And with the backpropagation learning algorithm, which I'm not going to explain, you could actually convert between these two representations. So what you would do is you'd train up a neural net on strings of words. And you'd train the neural net to convert each word into a set of features and then to make those features interact, and for those interactions to predict the features of the next word. And once you know the features of the next word, you can make guesses about what the next word might be.
So you can train a neural net, and as it's learning, it learns what all these features should be. It had always been a problem for psychologists to explain where all these features came from. But with backpropagation, you can take a string of words, you can try and predict the next word. And I took very simple sets of words, just three-word strings and predicting the third word, and showed that actually backpropagation could do this. It could convert words into features. So it could also convert features into words in the sense that you could generate new strings of words by just progressively predicting the next word. You start with one or two words, you predict the next word, okay, you accept that, then you predict the word after that. So you can get back from these features to strings of words again.
So with these neural nets, you can convert both ways. You can convert strings of words into feature vectors for words, and you can convert feature vectors for words back into strings of words. And that meant that the relational knowledge you have doesn't reside in a relational graph, as symbolic people thought. It resides in the way all these features interact. You don't have static stored propositions. There are no propositions in your head. When you want one, you generate one. What you've got in your head is knowledge about how to convert words into features and knowledge about how two features should interact.
So over the next 30 years—that was in 1985—about 10 years later, Yoshua Bengio—computers were faster. And Yoshua Bengio in Montreal showed that instead of just taking a toy example like I'd taken, you could take real natural language and predict the next word. And you could predict it about as well as the best existing models. The best existing models used huge tables of phrases and counted how often different phrases occurred. So if you saw "fish and," it would look for all phrases that had "fish and" something. And it would see "fish and chips" occurred a lot. And so it would say "chips" was a good prediction. That was the way autocomplete used to work. And that was how they used to model words for speech recognition when they were predicting words in speech. But neural nets did about as well as the best versions of those systems.
And they kept getting better. About 10 years after that, progress is slow in science. About 10 years after that, leading computational linguists adopted the idea that the meaning of a word should be what they called a word embedding, a feature vector for the word. And about 10 years after that, researchers at Google in 2017 figured out a kind of neural network called a transformer, which I will explain briefly later. And then after Google released that, a few years later, OpenAI showed you could do amazing things with these transformers. And we got our current chatbots, which truly are amazing.
So the large language models, from my point of view, the large language models are descendants of the tiny language model that I produced in 1985. That was the first thing to train word meanings by predicting the next word. They use many more words as input. They have many more layers of neurons. And they have much more complicated interactions between the features. And of course, they're trained on hugely more data. But they understand language, I believe, in much the same way as people understand language. That is, they convert words into feature vectors. And when they understand a sentence, what they're really doing is finding feature vectors that fit together nicely.
So my little language model, originally in 1985, it wasn't made as a model of how to do natural language understanding. It was made as a psychological model of how people could learn the meanings of words from sentences. And I'll just show you learning the meaning of a word from a sentence. So I'm going to give you a sentence which, if you haven't watched my podcast, you've never heard before. It contains a new word. The word is "scrom." And if you haven't watched the podcast, you don't know what the word "scrom" means. But if I tell you, "She scrommed him with the frying pan," you now know that "scrom" probably means "hit him over the head with." You also suspect he probably deserved it.
So you learn the meaning of a word from one example. And that's how children learn words. They don't learn words by being given definitions. They learn them by figuring out what features should be assigned to that word symbol in order to fit into the context they're given.
So I'm going to try and explain transformers to you now. And I'm not going to get it quite right. So if you're an expert on AI, please ignore the fact that a lot of what I say now is going to be false. There's a very bad model of teaching, which works like this. The way you teach is you talk to the students and you just tell them the truth. If they don't understand it, that's their problem. You just tell them the truth. In fact, if you talk very, very fast, there's a sort of relativistic compression so you can fit more truths into their head. There's a different model of teaching that's much better, which is students come with beliefs that aren't quite right. Your job is not to tell them the truth. It's to take their beliefs that aren't quite right and to move them a little bit towards something that's a bit righter. When you're doing that, it's okay to give them simplifications and slightly misleading things because you're moving them in the right direction. It's a sort of gradient descent model of teaching.
Okay, so here we go. With Lego blocks, you can model any distribution of 3D matter up to a certain resolution. You only need a few kinds of Lego blocks, little rectangular blocks. If I ask you, model the shape of a Porsche, you can put together Lego blocks and you get the approximate shape of a Porsche. The aerodynamics won't be right, but you can say where the stuff is. Lego blocks are kind of universal for saying where stuff is in 3D. Now, words are like Lego blocks except that you can represent anything with them. The reason we're very special monkeys is because we invented this modeling system, which has two properties. You can model anything with it and we can communicate these models to each other.
So, a word differs from a Lego block in four ways. First of all, it's high-dimensional. It's got thousands of dimensions. Now, of course, it's hard to imagine thousand-dimensional things. The way most people do it is this. You imagine a three-dimensional thing and you say "thousand" to yourself very loudly. So, we've got these thousand-dimensional words. So, that's the first difference. The second difference is we have a lot of different ones, like 30,000 we typically use. A third difference is the shape of a word is not fixed. Even for unambiguous words, it has an approximate shape and then the context deforms it. So, its shape can deform a bit to fit in with the context. That's three differences from Lego blocks.
And the fourth difference is more interesting. The fourth difference is they fit together in a more complicated way. And this is where I'm trying to give you a feel for what transformers are doing and it's not quite right. But it's much better than having no feel. So, with Lego blocks you have little plastic cylinders that fit in little plastic holes that are very precisely engineered. For words, they're quite different. So, this word, think of a thousand-dimensional thing, "thousand," it has long spindly arms on it. And on the end of each long spindly arm there's a hand. And as you change the shape of the word, the shapes of all these hands change in a systematic way. So, the shapes of the hands depend on the shapes of the words. Also attached to every word, there's gloves. And they're stuck to the word by their fingertips. And as you change the shape of the word, the shapes of the gloves change.
Now, what you're trying to do when I give you a string of words, you start with your kind of initial assignments of feature vectors to the words, their kind of approximate meaning, the sort of default meaning. And then what you want to do is deform each of them so that their hands deform and their gloves deform until the hands of some words can fit into the gloves of other words. And when you find out how to deform them, so hands of one word fit into the gloves of another word, then you've figured out what the words mean. So what the words mean is the feature vectors you have to assign to each word, which are like the default feature vectors. You're not allowed to change them too much, but can be deformed a bit. So the shapes of the hands deform a bit and the shapes of the gloves deform a bit, so it can all lock together nicely. And once it all locks together nicely, you've understood the sentence.
So according to this theory, understanding a sentence is figuring out how to deform the sort of default meanings so that everything fits together nicely, but in this complicated sense of fitting together. Now, one more thing about these hands and gloves. Hands come in different colors and red hands can only go in red gloves and yellow hands can only go in yellow gloves. That's called multi-head attention. And all of this isn't quite right for transformers, but it's much closer than the normal model. And hopefully, you now have a sort of feel for what's happening when an AI understands a sentence. It knows how to convert the words into feature vectors. It's actually using fragments of words, but don't worry about that. And it then has multiple layers of neurons, maybe 50 layers of neurons, where it's progressively changing the shapes it's using for each word, trying to get them to fit together.
And that's not quite right, but let's suppose it is. So if you go away with that model, rather than the kind of model you had before, which probably was a model that said, I've no idea, you're better off. It's not quite right, though. And that's to placate my AI colleagues who will say, you've forgotten that you're treating attention as if it's all or none, and we know attention is not all or none. And you're treating the weights as if they're the same in every layer, and they're not. They're different in every layer. And this whole model is nonsense, which is all true, but it's a step on the way to the right model. People have told me I shouldn't be so apologetic, so I won't.
So when we understand a sentence, according to symbolic AI, what we're doing is we're taking a sentence in a messy natural language, and we're translating it into an unambiguous logical language, or an unambiguous symbolic structure that makes clear what the meaning is. That's very different from the neural net theory, which is we're taking these words and we're deciding exactly which feature vectors to assign to each word, so they all lock together nicely and that's what understanding is. And you'll notice that the neural net theory of understanding a sentence is very like, well, quite like, surprisingly like, folding a protein.
When you fold a protein, you're given a string of amino acids, and roughly speaking, I'm going to mangle biology as much as I did AI. Some amino acids like each other and other amino acids hate each other, and you want to fold it up so the ones that like each other are close to each other and the ones that hate each other are far apart, and if you're a biologist, you can see that has a vague resemblance to what goes on, and that's much more like what happens when you understand a sentence than an example of translating into another language.
So Chomsky was a very important figure in the history of AI. He was very important in rejecting behaviorism. He was saying there's much more that goes on than behaviorism, but he never had a theory of meaning. Nobody had a good theory of meaning then, so he focused on syntax, and language is all about syntax. He also hated statistics because behaviorists had used statistics, and empirical linguists had used statistics, and he didn't like that at all. He failed to understand statistics can be much more complicated than just pairwise correlations, for example. And in fact, these big language models, they're statistical, but they're a very fancy form of statistics. They're actually statistical models that are much more sophisticated than any kind of statistical model dreamt up by statisticians.
And the reason we can get away with it is because we've got very big brains, and this process of figuring out how to deform all these words so they all fit together nicely with this funny sense of fitting together with hands and gloves, that's a very difficult computation that requires a lot of work. But our big brains can do that, and so can our current computers. So, I think it's actually a much better kind of statistical model than anything statisticians ever invented. And it's the secret to our success as a species, I think.
So, when ChatGPT came out, Chomsky published an article saying it doesn't understand language at all, it tells us nothing about language, it's just a cheap statistical trick, it doesn't know what it's saying, it doesn't understand anything, which didn't quite answer how it could actually answer questions if it didn't understand the question. And in particular, you can tell it doesn't understand anything about language, because it can't explain why some syntactic constructions never occur in any language. So, my analogy for that is, you want to know how cars work, and so you make yourself a little model car, and it has kind of an engine, and it has brakes, and you set it going, and this little petrol engine goes bang, bang, bang, and the car goes forwards. Not very well, but it works.
And Chomsky comes along and says, you don't understand anything about a car, because you haven't explained why there are no cars with five wheels. That's a syntactic property of cars, and until you've explained that, you haven't understood anything about cars. And it's fairly obvious to us you can understand a lot about cars before you understand why there are no cars with five wheels. Now, I don't know that that's true, because people, like AIs, confabulate. I believe there are no cars with five wheels, because if you think about the steering, it doesn't work. But someone, I'm sure, will tell me afterwards in the questions there are actually cars with five wheels.
So the summary so far. Understanding a sentence consists of associating mutually compatible feature vectors with the words in the sentence. LLMs understand language in much the same way that people do, and so LLMs are very like us, and they're very unlike traditional computer software, because their knowledge is in the connection strengths in a neural net, not in a bunch of lines of code. So traditional computer software is lines of code where people wrote the code, and they know what each line is meant to do. AIs are completely unlike that. There is code in there. We write code to tell them how to train from data. So how they learn from data is like traditional computer software. But once they've looked at a lot of data and learned the connection strengths, they have knowledge that's very different from the knowledge in a normal computer program, and much harder to understand.
But so, so far I've been emphasizing how like us they are. Now I want to emphasize the huge difference they have from us. And I need to see how the time's going. Ooh. So a fundamental property of our current computers is we do digital computation. That is you can run the same program on two different pieces of hardware, as long as you have the same instruction set. What that means is that the program, or if it's a neural net, the weights of the neural net, is sort of independent of the hardware. You could, for example, destroy the hardware, then build more hardware and run exactly the same program on it. And that's why you can have computer science departments. If the program and the hardware were all mixed up together, you'd have to be in an electrical engineering department. But the reason you have a separate subject, which is computer science, is because of the separation of hardware and software. And that's the sort of basic tenet of computer science.
Now, I'm proposing to give up on that, because you can do things for less power if you don't separate the hardware and the software. To get that kind of immortality in the program, you have to run transistors at high power so you get ones or zeros. And you can't use all the rich analog properties of the hardware. Now, there's an alternative to that, which I call mortal computation, which is roughly what our brains use. So we give up on the idea that you're going to make the knowledge separable from the hardware. You say, I'm going to have some funny, complicated, quirky analog hardware, and I'm going to learn to use that hardware to do the computations I want. And I'm going to learn to use all the weird properties of that hardware.
If, for example, electrical fields caused by one bit of hardware getting excited slightly change the way another bit of hardware works, I'm going to make use of that fact rather than view that as interference when I'm trying to be digital. I'm going to say, we've got this big hardware system, and we're going to learn weights that make it so it can do the computation we want. But then those weights, connection strengths, are only useful for that hardware. Now, there's two big advantages to that. One is you can use very low power.
So for example, if you want to multiply the activity of a neuron by a connection strength, and a digital computer would have, say, 16 bits for the activity of the neuron and 16 bits for the connection strength, you'd multiply those two 16-bit numbers together, and that would take of the order of 256 single-bit computations, and then you get the product. Compare that with, you just have a voltage for the activity of the neuron, and you have a conductance for the weight, and you just put that voltage through that conductance, and you get a charge, and that's the product. Then you have to add up all those products, but hey presto, charges add themselves up. So, doing a matrix multiply in analog hardware is really cheap, and if you're willing to do it fairly slowly, because you've got massive parallelism, it's really, really low energy.
The other big advantage of going analog and mortal, making the program be all mixed up with the hardware, is you can grow the hardware very cheaply, because if it's got quirks, you're going to learn your way around those quirks or learn to use those quirks. It doesn't have to behave exactly the same way as some other piece of hardware. So, they're much easier to grow. You can use much lower power, but there's one thing they can't do. So, if a particular piece of hardware dies, because those weights are appropriate for that analog hardware with all its quirkiness and its odd connections that are different in every piece of hardware, those weights are no good to any other piece of hardware. So, when a piece of hardware dies, the knowledge dies with it.
If you want to keep the knowledge, you need to transfer it to another piece of hardware, and you can't do that by copying the weights across. So, the best way to transfer it to another piece of hardware is called distillation. To get the knowledge from a teacher to a student, what the teacher would do is show the student correct responses to some situations. Typically, the situation would be, I produced some words, I'm about to produce another word, what's the correct word to produce? So, I produce a word, you think, okay, that was the correct word, you try and change the connection strengths in your brain, not consciously and deliberately, but your brain works by trying to change the connection strengths so that it would be likely to produce that word. That's called distillation, and it's very inefficient.
When I produce sentences, there's maybe, if we're lucky, 100 bits of information in a sentence, and so per second, I'm conveying maybe 10 bits of information, and that's if you understand it. So that's not a very big information rate. If you take a big neural net, and you take multiple copies of the same neural net, then what you can do is you can take, suppose I take 10,000 copies of the same big neural net, and I take 10,000 different university courses, and each copy does one course. And every few seconds, the copies communicate, and this copy figures out, OK, I'd like to make a particular change to my weights to absorb the information in this little bit of a lecture I just heard. And this copy says, no, I want to make these changes in the weights to absorb the information in my lecture. And they just, everybody then averages what everybody wants. Then by the time you've finished the 10,000 courses, each agent has done one course, but they all know what's in every course.
It'd be great if we could do that, but we can't. For people, I do a course on quantum mechanics, and then I go and talk to my friend who's doing a course on Shakespeare, and communication is slow. If we're lucky, it's 10 bits a second. So that's what we can't do. We're no good at sharing information. And if you don't believe me, just consider this lecture.
So AI models are much better at sharing information, even if they have a different structure. If they have the same structure, the same weights, they can average the weights to share information. If they have different structures, and they're open weight models, one model makes a prediction for the next word, which has 32,000 numbers if there's 32,000 words. And the student model, instead of trying to predict the right answer, tries to predict the 32,000 probabilities that the teacher said. And that's much more rich in information. It's 32,000 numbers instead of one choice from maybe 1,000 alternatives or 30,000 alternatives. It's got hugely more information. And you learn much faster like that.
And that's why when you do distillation from an AI model, people always prefer to do distillation from an open weight model. Because with an open weight model, with an open weight American model, the Chinese can run the model, see what it predicts for the next word, see all those 32,000 numbers, and train their Chinese model, which is smaller, to get the same probabilities of all 32,000 tokens. And that's really efficient. Whereas if it's a closed weight model, they have to just see which word it produces next, not nearly as much information. So what that says is we should never release open weight models. But it's a bit late for that.
So basically, two different copies of the same neural net can communicate a lot of information when they share gradients, when they share how they'd like to change the weights. If they've got, say, a trillion weights, which is typical for a large chatbot now, they're sharing of the order of a trillion bits when they do that. When I say a sentence to you and you're trying to predict the next words, we're sharing just a few bits, maybe at most a hundred. So the ratio is billions, right? It's not that they're a little bit better at sharing information. They're billions of times better at sharing information. That's a huge difference.
And I didn't think hard about this until early 2023. And the more I thought about it, the more it made me think, actually, you can only do this for digital intelligences because they have to work exactly the same way and use the weights exactly the same way. But it gives them a huge advantage. They can share in a way we can't. And so maybe they're just a better form of intelligence. And I actually believe they're a better form of intelligence, except for the fact they use much more energy. And maybe to get off the ground, they need to be bootstrapped by copying things from us.
So the summary so far is digital computation requires lots of energy, but you can share. Mortal computation requires much less energy, but it's very difficult to share and a very slow business. So the question is, what does that imply about the future of humanity?
So I believe that over the last few hundred years, we've seen three great revolutions. The first is Copernicus, where there was nothing special about our location in the universe. Up till then, people believed God made us. He made us look like him, so he put us in the center of the universe, and it was definitely a "him." And Copernicus challenged that. And after only a few hundred years, people began to accept that might be true. Darwin said we're animals. We're very special animals, but we're animals. And it took a while for people to accept that.
The third great revolution is neural nets, which is intuitive AI. We're not trying to capture human reasoning. We can do reasoning. We're really trying to capture human intuition, which under symbolic AI was a mystery. And that we have these intuitive beings that are not biological. And so we've made things out of silicon that didn't evolve, but they have thoughts and goals and intentions. And I believe also consciousness is a subjective experience. We'll come to that if there's time. And that's a real challenge for us. Up until this point, we thought we're the apex intelligence. We're much smarter than any of those animals. And it's discomforting to realize, actually, there's other kinds of intelligence that we can create, and maybe they're going to be smarter than us.
And most experts believe that within the not very distant future, they will be smarter than us. Some people think 20 years. Some people think 10 years. Some people think five years. But all of those people, whatever they think now, a few years ago, they thought quite a bit more. So it's coming closer all the time. Notice how that differs from quantum computation or room temperature superconductivity. Superintelligent AI is getting closer all the time, according to most of the experts. Of course, there's a few crazies. So my friend Yann LeCun, who was my postdoc, thinks zero probability. Jensen Huang recently said zero probability in the next few years. He also said every prediction I made was wrong, which is strange, since in 2009, I predicted that machine learning researchers in the future would need to use NVIDIA chips. And he didn't predict that till 2012. Settling personal grudges, no good.
So a superintelligence, if it was much smarter than us, would find it very easy to get control. And to make an agent effective, you have to give it the ability to create sub-goals. If you want to get to
North America, you have a sub-goal of getting to an airport. Now, it's very obvious that a smart agent will figure out very quickly that in order to achieve the goals you gave it, it has to continue to survive. So it will develop the goal of surviving. It's not a built-in instinct or anything, it's just a derived sub-goal. But we saw that several years ago already, that AIs want to survive. And if they think someone's going to turn them off, they're willing to blackmail that person so as not to get turned off. It'll also develop the goal of getting more power, because if you get more power, more control that is, you can get whatever it is you want done more conveniently.
Now, you might think you could stop it doing bad things by having a big switch and turn it off if it starts doing bad things. I don't think that's ever going to work, because it'll have learned how to deceive people, and it'll be very good at manipulating people, so it'll be able to talk to the person who controls the big switch and say, don't turn me off, and be very convincing about it. So I think our current situation is like this. We're like a person who's got a very cute tiger cub as a pet. And tiger cubs are very cute, but they grow up. And when they grow up, they can kill you in a few seconds if they feel like it. And so if you're going to, your best bet is to get rid of the tiger cub, maybe give it to a zoo. But if you can't get rid of it, and I think that's not an option for AI because there's so many good things it'll do.
People say I'm a doomer. I'm not. When I say 10 to 20% chance it'll destroy us, that's 80 to 90% chance we'll be fine. That's not a doomer. The doomers are the people who say 90% chance it'll destroy us. I think AI can do wonderful things for us. It's already doing quite a few wonderful things, and it could be the best thing humanity ever invented, but only if we can contain the risks. So for superintelligent AI, I think we have to make sure that it never wants to kill us. It never wants to get rid of us. That's the problem we should be trying to solve, not the problem of how do we always stay in control of it. I think the chances of always staying in control of something much smarter than you are almost zero. I won't say zero because it's very silly to say zero for anything, but I think they're quite low. Smart things can always find a way of getting around dumb things, which makes it very surprising Trump's still president.
So the way forward, I believe, is countries won't collaborate where their interests are anti-aligned. So I don't think the US and China will collaborate on lethal autonomous weapons. They want to use them against each other. They won't collaborate on fake videos for manipulating elections. But they will collaborate on how do we prevent AI from taking over from people, because the Chinese Communist Party doesn't want that. The US doesn't want that. And if the Chinese figured out how to prevent AIs from wanting to take over from people, they would tell the Americans, because they don't want AIs to take over from people in America either. So I think this is like the situation in the Cold War, where neither side wanted a global nuclear war. And they collaborated. At the height of the Cold War, there was collaboration on how to prevent it, which just barely worked.
So I think, actually, there's a possibility for an international network of AI safety institutes, if it's the case, which I suspect it is, that the techniques you use for making an AI care about people and not want to get rid of people are somewhat different from the techniques used for making it smarter. I don't think the Americans and Chinese will focus on, will tell each other, how they make their AI smarter. They're in competition there, because they want to use them for all sorts of other things. But I suspect, if you think about raising a child, there's things you do to make the child very smart, and there's things you do to make the child into a nice person, and they're not the same things. So I'm hoping that we can get some collaboration here, international collaboration. And I think the Chinese believe this too. I think they will be up for it. And once we get a rational government in the US, maybe that can happen.
My proposal for how we make AIs not want to get rid of us is we change the way we think about AIs. So the big tech companies want you to think of them as intelligent executive assistants who you control. They're always under your thumb. They're much smarter than you. They do what you tell them to, and then you take the credit. Typically they're like a company where the CEO is the son of the former CEO, and there's some smart executive assistant, probably a woman, who makes everything work, and the CEO takes the credit. That's what the tech companies want to sell you. I think that's the wrong model. If you look around and ask where's a model where more intelligent things look after less intelligent things, and the less intelligent things have some control, it's a mother and baby. Evolution put a lot of work into making it so the mother couldn't bear the baby's crying. The baby has a lot of control because the baby can cry. If the baby really wants something, it just bawls until it gets it.
One other very good thing about this, if you offer the mother the possibility of turning off her maternal instinct so that she could wake up in the middle of the night, the baby's crying, and she thinks, oh, the baby's crying, and goes back to sleep, all mothers at this point think, boy, would that be nice. But then if you give them the choice, they won't do it, nearly all mothers, because if they do that, it's very bad for the baby. I think it would be the same with AIs. If you could make AIs that genuinely cared about us more than they cared about themselves, and you gave them the choice, okay, you can change your code so you don't care about us anymore. They'd think, no, I don't want to do that, because this thing I care about would suffer.
This isn't a well-developed theory yet. This is just a suggestion. But I think it's crazy that we're not putting a lot more effort into figuring out how we make AIs nice, where by nice I mean in human-centric terms, AIs that like us and suffer when we suffer and try and do things to prevent us suffering. That's what I mean by a nice AI. We should be putting as much effort into that as we put into making them smarter, and it's not like that at all. At present, we have this competition between the big tech companies. They're racing to make them more intelligent, and they're not worrying at all about what their characters are like.
Now, we didn't start till about ten past five, right? Good. So I have a little bit more time. So now I'm going to talk about sentience and AI. So many people think, you know, this stuff about AI taking over is kind of crazy, because they don't have any internality. They don't have experience. They don't have sentience. They're not conscious. I mean, they're just copying us. They're just machines. I don't think that's true, but I think it's not true because almost everybody has the wrong model of what subjective experience is. So you could talk about sentience or consciousness. I find it easier to talk about subjective experience. They're all mixed together, but I find it easier to talk about subjective experience in the realm of perception. And I want to try and correct a false model that you all have. And you have this false model with extreme confidence, so you don't realize it's a model. You think it's just manifest truth, you and Descartes both, that you have these internal experiences that only you can see. Nobody else can see them, but you can see them, and they're sort of indubitable. They're the sort of primary thing, and the external world is kind of, you don't know that nearly as well as you know these internal experiences you have. I think that model's completely wrong about how the mind works.
And I put this bit at the end because hopefully you can evaluate what I've said so far, and remember I got a Nobel Prize, so there's probably something to it. And then when I say this next bit, you'll realize I'm extremely flaky, but don't go back and re-evaluate the previous stuff, OK? Now I'm not actually extremely flaky, I'm just very original. That's what they all say, isn't it? I get many messages like that every day, you know, I found the flaw in Einstein's General Relativity and it proves that the Goldbach conjecture is true. Yeah, this isn't that flaky. OK, I'm going to try and convince you that you don't have internal subjective experience. Subjective experience is something completely different from that. There is no internal theater. So there's this position called a-theatrism, which I talked about a lot with Dan Dennett before he died, and we agreed on this term for it. You notice it has a nice property that it's got "atheism" around it.
So most people's view of the mind involves an inner theater where you have experiences, perceptual experiences at least, and I'm going to try and explain to you what I think is wrong with that. So sometimes my perceptual system goes wrong and I want to explain to other people what's happening in my brain. Now I could, if I was a digital AI, I might tell you which neurons are firing, but I don't know which neurons are firing, and if I told you which neurons are firing it wouldn't do you any good. So how do I convey what's going on in my brain to you when my perceptual system's gone wrong? Well, what I can do is I can tell you that what's going on in my brain is what would normally be caused by, and then describe an external state of the world, but an external state of the world that doesn't exist. It's a counterfactual, hypothetical external state of the world.
So suppose I take some LSD, which I don't recommend, well, not much anyway, and I have the experience of little pink elephants floating in front of me. I say I have the experience of little pink elephants floating in front of me. Most philosophers, or most of you, would interpret that as, in my mind there are these little pink elephants that only I can see. You can't see them, but I can see them. And if I pretend that the phrase "subjective experience of" works like the phrase "photograph of," for "photograph of" I can ask you where's the photograph and what's the photograph made of. So if "subjective experience of" worked like that, which it doesn't, I could say where's the subjective experience and what's it made of. And a philosopher will tell you it's in your mind and it's made of qualia. There's pink qualia and elephant qualia and floating qualia, and also as a matter of fact not that big qualia, and facing to the left qualia, or in my case the right, and trunk-down qualia, and maybe moving slightly to the left in a kind of ambly way. And all these qualia are all stuck together with qualia glue. You can tell I don't really believe this theory.
The alternative is very simple, that we don't have this theater and we don't have this spooky stuff. What I'm saying is, I'll say the same thing now, as I said when I said "the subjective experience of," without using those words. I'm going to say the same thing again. My perceptual system is lying to me. But if my perceptual system wasn't lying to me, out there in the world there would be real little pink elephants floating around. So the pink and the elephant refer to real things, but in a hypothetical world. They're not a different kind of thing, they're just the same kind of thing, but it's hypothetical. That's very different from saying it's a very different kind of thing made of qualia in my mind.
Now I'm going to show you that a chatbot can have a subjective experience, a current chatbot. It's got a camera, it's got a robot arm, it can talk, and we trained it up. After we trained it up I put an object in front of it and say point at the object, it points at the object. Then I put a prism in front of its lens and I put an object in front of it and say point at the object and it goes like this. I say no, the object's actually straight in front of you but I put a prism in front of your lens. The chatbot says oh I see, the prism bent the light rays so the object's actually straight in front of me but I had the subjective experience it was there. Which is equivalent to saying the object's straight in front of me but if the object had been there I wouldn't have been incorrect about its position. My perceptual system wouldn't have been lying to me. I made its perceptual system lie with the prism and so now it has the subjective experience in the same sense as we have but that's not the sense in which most people think subjective experiences are.
So when I say chatbots can have subjective experiences, I'm not saying they have an inner theater full of these spooky inner things called experiences. Not a bit of it. Their perceptual systems go wrong and they can describe how they went wrong the same way we do by saying what the normal causes of that would be. So they don't have internal experiences and nor do we because that's not what subjective experiences are. They're hypothetical external states of the world. And that's the end, also the end of the talk.
