Summary AI: Grappling with a New Kind of Intelligence (Youtube) youtu.be
20,567 words - YouTube video - View YouTube video
Brian Green Hi.
Digital Brian Green In the vast expanse of our universe, teeming with countless stars and galaxies, humans have always sought to understand the mysteries of existence. But now we're on the brink of a new frontier not in the distant realms, space, and time, but within the intricacies of our own digital landscape. Artificial Intelligence, or AI. It's a term you've heard countless times. But what really lies beneath the algorithms and computations?
Digital Brian Green Today, we're diving deep, but the tools of generations passed AI promises profound benefits. Yet, it also brings forward potent questions. Are we on the brink of a golden age of innovation? We're standing at the precipice of our own obsolescence. It's not just about weighing the pros and cons.
Digital Brian Green We also seek to grasp. Fundamentally, the workings of his AI systems. For instance, large language models, astonishingly versatile, capable of generating text, answering Quest even crafting music. But how do these models think, or do they think at all? And if they don't think like us, what exactly are they doing?
Digital Brian Green In today's program, we will give you a look under the hood to see the inner workings of digital minds, not just to marvel, but to understand. Tackle the real. Because when we understand, we can act with foresight, with wisdom, and with purpose So as we stand at this intersection of Humanity And Technology, let's embark on this journey together. Welcome to AI grappling with a new kind of intelligence. Good evening, everyone.
Digital Brian Green Thank you. So I am Brian Green, and perhaps I should augment that by saying I am the real Brian Green, because I think as some or many of you may have already surmise concluded, that wasn't me up there on the screen. I didn't write a single word of that text no human being did, all of that text was written by a large language model, The sequence of visuals that you saw in that piece was chosen by a large language model, chatgpt, And I didn't actually say anything that you saw up there on the screen. You're saying I did a better job. Well, you have a certain binary charm.
Digital Brian Green Thanks. But while I can piece together words and concepts, I lack the essence of feeling. I can't cherish a nostalgic memory, savor the touch of a gentle breeze, or relish the thrill of innovation. That is so sad. Everyone, how about a little love for my digital counterpart?
Digital Brian Green Help me out here. Yeah. Right? Thank you, I think. Though I don't truly feel the acknowledgement.
Digital Brian Green Hey. It's the thought that counts. Or in my case, the algorithm. Just remember to power me off when you're done. Will do.
Digital Brian Green Good night. Digital me. Good night. Organic me. So we are yes.
Digital Brian Green Thank you. Why not? Yeah. So it's to some impressive, to some a little bit scary, that a system is able to do this. And if you Look through the course of human history.
Digital Brian Green There are, I don't know, a few dozen key moments, inflection points, disruption points, when something happened that changed the course of the future of human development, right, You can think of the acquisition of language, the domestication of fire, the acquisition of written language, the capacity to have the wheel, to learn how to have wireless communication, the printing press, the capacity to have self propelled vehicles. I mean, all of these were key technological developments that radically changed how our history unfolded. And it may well be, and this is what we will be discussing here tonight, that we are at a similar kind of inflection point because 1 way of thinking about a way that I find particularly illuminating is I like to organize all of reality into 3 general categories. Right? You've got the big stuff.
Digital Brian Green Space. You got the small stuff. Adams and molecules, and you've got the complex stuff. Now for the big stuff, we've made great progress. Einstein's general theory of relativity gives us real deep insight.
Digital Brian Green We've not been able to translate that insight into control over the larger cosmos, but the intellectual insight is dramatic. In the realm of the small, of course, we have quantum mechanics. And we have been able to leverage our understanding of quantum mechanics to control the micro world. Right? Giving Ron to, for example, the integrated circuit, which, of course, is the key to computational power, which stands behind everything that we'll be talking about here tonight, but in the realm of the complex.
Digital Brian Green For a very long time, that was a realm that was kind of beyond reach. The realm of the complex really is the realm of of life and of mind or intelligence But now in the last couple of decades, with synthetic biology, we're gaining the capacity to really harness and understand life in a way that gives us troll. And the insights that we saw now and that you all have experienced, no doubt having played with these systems is bringing the possibility that we also have synthetic intelligence, artificial intelligence, bringing the realm of the complex within the power Exciting, but to some, it's quite scary. And what we're going to do tonight is explore really both of those possibilities and try to get an understanding of where we are and where we are going, when you have systems, for instance, like the 1 that just gave rise to the piece that you just saw, It's from the cloneworks AI studio. If you can give a round of applause for the work that they did on that opening piece for us, But what does the capacity to create something like that, deep fakes?
Digital Brian Green What does it mean for the nature of democracy in the future of our species. And so what I'd like to do now is bring in our first guess who is Jan Lacun, who is a silver professor of data science, computer science, neuroscience, and electrical engineering at New York University, He is also vice president and director of AI research at Facebook. He led several major innovations underlying generative AI. And in 2018, his work on deep learning earned him the Turing Award. Welcome.
Digital Brian Green Yeah.
Brian Green Good morning.
Digital Brian Green Thank you. So, you know, since last November, when we all gained the ability to play around with, say, chat GPT, I think many of us have had those wild moments where a computer is doing something that you don't think a computer can yet do. And so can you just give us some sense? We'll get into it in more detail in part 2 of this conversation when Sebastian Bubba comes out, but can you just give me sort of an overview of how it is that these systems are able to do what they do?
Brian Green Very, interesting question. So for us, the revolution occurred 2 or 3 years before that, for people like like like us in the research, there were some surprising results that everybody in the community was surprised about. The fact that, when you train artificial neural nets that are very large on lots of data, there's some sort of emerging property that occurs, to a larger extent than we were used to for other systems that can do computer vision or things, I guess, But to some extent, for us, the revolution did not occur in November. They occurred 3 years before.
Digital Brian Green I mean, it was the case. It was the case, I think. I mean, you know, I'm obviously not in the field, but I would hear colleagues for, I don't know, 20 years, 25 years, speaking of we're on the cusp, of things radically changing in AI. Have we finally reached that? Cause I mean, for many decades, it didn't seem to really happen.
Brian Green Well, so the history of AI is is interesting because it's a series of new paradigms, new ideas that people discover, and I assume the new as the new paradigm, occurred, people said that's it. That's the solution. Now we're gonna build intelligent machines, and it started in the fifties. With very famous people in AI, ended up during winning the attorney award too. Newell and Simon, they say, we're gonna write this program.
Brian Green It's called a general problem solver. It's gonna be able to solve all problems in the world. There was that program back in the fifties, all problems in the world. Problems as long as you can formulate a problem in terms of an objective to satisfy. Yeah.
Brian Green And then a search over possible solutions and say, everything in the world can be formulated in those terms. Yeah. What they didn't realize at the time is that every interesting problem requires an incredibly large amount of computation that was is completely unreachable even today. So basically that program failed. Yeah.
Brian Green And it took a few years for them to realize. So it was another, effort almost at the same time of people who said, no. No. No. No.
Brian Green The way you build intelligent machines is that you allow them to learn. That started in the fifties. And they say, okay. Now we have this machine called a perceptron, and it's magical. You can show it pictures of, characters and you can train it and it will learn to distinguish between, you know, a C and a D or a tank and a Yeah.
Brian Green Jeep or whatever. And then you realize, actually, no, it doesn't work that well. You you can't really get the system to learn to distinguish complex images or anything like that. And so it was very limited. Again, people thought they had the solution, and you turned that to be a dud.
Brian Green Now, as a matter of fact, that kind of technology actually people working on this gonna change the name of what they're doing, and that actually was the basis for a lot of modern technology and communication called additive filters. Very strange story. And then there was another movement, 10 years later called expert systems. Again, in AI, a lot of interest for this, Japan started a big program called the 5th generation computers. They were gonna be able to do inference and logical reasoning and stuff like that.
Brian Green The problem with this is that there was no learning. So so it was people who had to enter facts and rules, in in computers and it turned out to work for a small number of problems, but you couldn't really build intelligent machines with it. So the the whole research interest in this gonna die after 10 years. And about the same time, a new wave of neural nets appeared. So this was, late eighties early nineties, that's when I started my career in late eighties and or or early eighties, I should say.
Brian Green And And there the idea again was how do we recycle those ideas of the perceptron? So that machines can be trained, but make them more powerful, and we've found a new technique, which is still the thing that we use today in things like larger Grish models, to train, artificial neural nets. That was very successful for about 10 years, and then people lost interest in it, because it was The computers were too slow. The datasets were too small. There was all kinds of issues.
Brian Green And then it came back to the fore about 10 years ago. And the reason why we hear about AI today is because of that renewal of interest in those so called deep learning techniques, basically being able to train large neural nets What we see today is the effect of, more powerful machines, bigger data set, and and sort of that allows us to to build very large neural nets, with 1,000,000,000 of the equivalent of synapses in the brain, 100 of billions, And they can do pretty amazing things.
Digital Brian Green And so we'll look under the hood a little bit more in in the second part of conversation when Sebastian Bubbeck comes out. But even though it's kind of mind blowing to somebody like me, and I think maybe to the general public, that, you know, I can give a prompt, which is what I did for the little piece that we just played. All I did was say, can you write a front end to a world science festival program in the style of of Brian Green, you know, that would introduce his program and out came the script that we just heard, I have to say it made me feel a little sick for a moment when it came out. But then I've sort of rationalized it by saying, well, maybe it's looking at things that I've written in the past and is sort of being inspired by maybe in some sense I had something to do with it, but putting that that to the side, you somewhat famously are a little less bowled over by what's happening today. Can you just give us a sense why?
Brian Green Okay. There is no question that the advance in technology, deep learning technology over the last 5 years has been astonishing, impressive, but we're easily fooled by those systems into thinking they are intelligent, just because they manipulate language fluently. The only example that we have of an entity that can manipulate language is other humans. When we see something that can manipulate language, flexibly, we assume that entity will have the same type of intelligence as humans. But it's just not true.
Brian Green Those systems are incredibly stupid. So they're very useful. Okay. They need to be developed. They're gonna be, you know, they're being commercialized, they're being worked on.
Brian Green They're great. But they are somewhat specialized, and they're very stupid in many ways. So partly they're stupid because they are only trained on language, and most of human knowledge has nothing to do with language, and all of animal knowledge as absolutely, you know
Digital Brian Green Not all philosophers, for example, would would, I guess, agree with that statement. I mean, they can sign, you know, the limits of my language and limits of my world. Right? So So are you stating that as a fact of making it as an assumption? I mean, I agree with you, but are you really convinced that language is just a limited part of human understanding?
Brian Green I think language is a way for us to communicate, knowledge and and certainly to store in our brain a large amount of of knowledge because language is efficient in a way. Right? They have discrete concepts in words, etcetera. But an orangutan doesn't have language. They're incredibly smart.
Brian Green They're almost as smart as we are. We think we're much smarter than them, but we're not. Like an orangutan can like a physicist. Can I understand that if I push on this?
Digital Brian Green I'm not sure exactly how to interpret that sentence, but I'll I'll let it go.
Brian Green If I push on this, with a small force, it's gonna move. Right? And that right now, it has water in it, but if you didn't didn't have water, it would probably flip.
Speaker 2 Yeah.
Brian Green If I push with the same force on this table, it's it's not moving. We have intuitive notion of physics. We've done this when we are babies. The large language models that we have today or any AI systems that we have today, none of them is capable of understanding any of this. To some extent, AI systems, the smartest AI systems today, have less understanding of the physical world than your house cat.
Digital Brian Green But can I just ask him in that? So if within the data that the system is trained on, there is an explication of basic physics, such as the example that you just gave, would you not consider that a quote unquote understanding? Because if you were to ask the system, if it trained enough, what would happen if I pushed on a table in the manner that you just suggested it might come back with it won't tip over? Well, for
Brian Green and that that's really an important point for situations that you play into the system. Yeah. You describe in words, and they correspond to a template that the system can learn and reuse in a new situation. It might work. But there's a lot of really complex effects here.
Brian Green Like, you know, if I push on the side, it's gonna rotate.
Speaker 2 Yeah.
Brian Green There is friction on the table, which might change. So if I push on the same bottle here on the on the floor. It's gonna tip. It's not gonna slide. Yeah.
Brian Green So, all of this is not really expressed in any language. Most of what we know about the world is not reflected in language. Yeah. And so it's a question of philosophy that philosophers are asking themselves, can we build intelligent machines that are purely trained from language and do not have any kind of sensor input. My answer to this is absolutely not.
Digital Brian Green You know, I'm I'm fond of starting my my classes at Columbia, the the the basic classes by, you know, throwing a light object into the audience and having somebody catch it and trying to convince them of how remarkable it is that they knew where to put their hand without doing any newtonian calculation of reject or the object and convincing them that this is this intuitive physics, but then I go further and say, why do we have that intuition? And the answer is it's evolution by natural selection. Those of our forebearers, presumably, who are better equipped to understand the physical world could navigate that environment more effectively had a better chance of surviving and passing on that capacity to subsequent generations, which which suggests, I think, and it's pretty clear, that our intelligence is not as general and as broad as perhaps you might think. It's really very specific. It was developed so that we could survive
Brian Green Absolutely. There is this this thing that people in AI, talk about AGI, artificial general intelligence, by which people mean essentially human level intelligence, human type intelligence. I hate that expression because human intelligence is indeed incredibly specialized. There's a a lot of and we know this because of computers. Computers are much better than humans at lots of tests nowadays.
Brian Green Right? And we knew know, we've known for decades that they're better at, arithmetic, of course, or at solving equations or computing integrals, you know, symbolically and things like this, right? We know also that computers are much better than humans at, you know, playing chess or go poker diplomacy. I mean, there's a whole bunch of games where, you know, computers have have become really good. And what it shows is not that computers are intelligent.
Brian Green It shows that humans just suck. We're we're extremely bad at doing things like imagining a large number of scenarios, like a chess, right, we we, or or go. We have to imagine a number of scenarios that depends on someone else playing. And for every every move, you know, there is something like 36 possible moves, and it goes financially. So we we can't contain this in our brains.
Brian Green So instead, we develop a kind of intuition for what consists, constitutes a good move or not, right, which and the modern go and chess playing systems actually do this too.
Digital Brian Green Right.
Brian Green But they also plan and they have a much bigger memory, and, or or working memory and ability to do tree exploration, which is why the beat us. So we're really not very good at this.
Digital Brian Green So given that, and in your language, you know, this technical term, you know, humans suck. Get given that my understanding, if you could take it through us, your vision for where AI needs to go for it to reach its potential is to model intelligence on human intelligence. To some extent. Right?
Brian Green I would start with cats because we don't we don't know how to reproduce the type of intelligence and understanding of the world that the cat has. So the first, Okay. There's this thing called common sense. Right? And we intuitive physics as part of it.
Brian Green Yeah. Cats have some level of common sense. No AI system today has any of it.
Digital Brian Green Right.
Brian Green So, we are going to have to make progress towards systems that can learn how the world works by observing it and by interacting in it the way babies do, Babies in the 1st few months of life learn an incredible amount of background knowledge as well.
Digital Brian Green Absolutely.
Brian Green Like gravity. So the concept that an object that is not supported falls Maybe he's run this around the age of 9 months. It takes a long time. And, we don't know how to do this with, computers. If we did, we would have completely autonomous level 5 cell driving cars, that purely work from vision.
Brian Green We have cell driving cars today, not a computer autonomous, they're not entirely reliable. They require, you know, 1000 of hours of engineering and training data on all kinds of sensors. So we how is it that a human, a 17 year old can learn to drive a car in 20 hours of practice? Yeah. Largely with that, causing any accident for Yes.
Digital Brian Green It is a little scary, but if you're willing to put up with it, yes, you get get over the hump. Right. But can you take us through your your vision then of how we might reach that level of artificial intelligence. And I think this is a framework for it.
Brian Green Right. So, there is, a number of characteristics that AI systems are not really capable of doing a number of things. 1 of them is planning this used to be a topic that classical AI, people, you know, 20 years ago were really interested in but, including this general problem solver, but but modern AI system like like large language models are not capable of planning, or at least are only capable of very simple planning. So what you're gonna see appearing here progressively is sort of an architecture different different modules that would be necessary for a complete AI systems to have the kind of intelligence we observe in, animals like cats and dogs and and humans. And it starts with, at the top, there's something called a configurator here that you can see at the top, and it's basically, a a, a director and master of ceremony that sort of organizes what the rest of the brain or the system is doing, tails the other systems in the brain.
Brian Green You're facing this situation, you have this goal to accomplish, not do it. If you go to the next, module Of course, the system to function properly has to have some idea of the current state of the world that's called perception, In humans, that's all in the back of the brain. So it's actually represented in the back of the brain. Here too, the back of your brain is visual perception, auditory perception is on the side. That just processes the pixels coming from UIs and turns it into a representation of the state of the world.
Brian Green An abstract representation. You don't need to put names on them. Babies can recognize object categories without knowing the names of them, and every animal can do it too. Okay. So once you have an idea of the state of the world, you have to imagine what is going to happen if I take a sequence of actions, what is going to happen in the world?
Brian Green Can I predict what's going to happen in the world? And that is the role of what is called here, the world model. This is where actually most of the intelligence goes in, in humans, it's in the 4th of the brain, which is particularly large in humans, And this is what allows you to predict that a particular action will have a particular result. And so if you can predict what result is going to occur as a consequence of an action you were taking, then you can plan Okay? You can plan a sequence of action to arrive at a particular result.
Brian Green So that's determined by this cost module. So the cost module basically is something that measures to what extent the predicted state of the world that that you imagined satisfies the goal that you set for yourself, or that the configurator is set for for for you this cost module basically measures your degree of, dissatisfaction
Digital Brian Green with with the outcome of
Brian Green With the outcome of some imagined thing. This is a set of emotions, really. So if you predict that a particular situation is gonna result in a bad outcome, and there's nothing you can do about it that produces fear. Your goal is not satisfied. So this is sort of an emotion that is, produced by a prediction but there are emotions that are immediate.
Brian Green Like, if I pinch you, right? You're not gonna like it very much. If I try to pinch you again, you're probably gonna recall it because you're gonna realize that you you are now your model of me. It proves the fact that you're
Digital Brian Green a pensioner.
Brian Green Yeah. Yeah. So, okay, so then the the the last module here is the actor, and the the role of the actor basically is to say, am I gonna be able to produce a sequence of actions that according to my prediction to the world model is going to satisfy this goal that I said to myself, minimize this cost or whatever. And that's how that's how we do planning. Actually a very old very old idea in engineering.
Brian Green It goes back to the 19 sixties, but, the way to do this in an intelligent way where the word model is learned by observation We don't know how to do this today. The big challenge I think for the next decade, and I wrote a long paper on this, is, how do we get machines to learn how the world works by observation,
Digital Brian Green how far are we along in that? I mean, we
Brian Green are making progress, particularly in the last 2 years. And the principles are somewhat similar to what is used in large language models. It's, using a set of learning techniques called self supervised running, And perhaps I can explain this a little bit. Yeah. So when you train, and I'm not sure we're gonna see the the corresponding, graphics, but, yeah, that's the that's a good example.
Brian Green Okay. So the way those larger originals are trained is you show them a text, then you remove some of the words, you match them. Okay. And of course, we know the sentence, but even if we hadn't seen the sentence, we probably would be able to guess what words are missing in that sentence. Okay.
Brian Green Then we train a very large neural network to predict the words that are missing. And in a process of doing so, the network learns a representation of the text that includes meaning syntax, grammar, can even train it on multiple languages. You can train it on computer code. You can train it on all kinds of stuff, and it it learns the internal structure of language. In such a way that you can then use the internal representation of language to do things like train the system to translate to detect hate speech, to classify the topic, to generate, abstracts or summaries of a text, And there is a special case of this, if you can go to the next, diagram.
Brian Green And a special case of this is what we call LLM, which we should call autoregressive LLM. So what is, an LLM that chat GPT?
Digital Brian Green So LLM large language model everybody knows, but to be sure. Yep.
Brian Green Right. It's a special case of what I just explained where the word you're you're masking is just the last word in the in the, in the text. Okay? Just take a long text, take a window, a few thousand words, and then mask the last 1, and then train some gigantic neural net to predict that last word. Okay?
Brian Green Now if you train that system successfully, it cannot exactly predict the word all the time. What it's going to predict is a a probability distribution over all possible worlds in the dictionary, right? So in the sentence that we saw, you know, the the cat chases the blank in the kitchen. Right? The the blank could be a mouse, so it could be a laser pointer.
Brian Green Right? You don't know exactly what it is, but you can train the system to produce, okay, it's gonna be cat with probability.5, and then laser pointer with probability.2, and then, you know, something else with probability.01 and whatever. And every word in the dictionary, you have a list of scores, essentially. And then what you can do is use this for autoregressive, what's called autoregressive prediction, and this is really how those larger voice models work. If you can go to the next, diagram that's at the bottom, you give a window of text to the system, have it predict the next word, and then inject that next word into the input window.
Brian Green Okay. Shift all the words by 1, and then ask it to produce the next next word and then shift that in, and then ask you to produce the next next word, etcetera, etcetera. And the system will just produce happily 1 word after the other.
Digital Brian Green So so let let's see. So the idea then is you feed the system an enormous body of text. And based upon all of the words that the system ingests, it begins to build up probabilities or or likelihoods for given sequences, different, combinations of words that might appear, and using that you can then make a prediction for the most likely next word in a given sequence, and including that word making a little longer sequence, you can predict the probability of the next word in that augmented sequence and word by word by word, you build up the most likely sentences and the most likely paragraphs and and so forth.
Brian Green Is only 1 step problem with this, which is that if it makes a really bad mistake about what the next word is, It's going to
Digital Brian Green Everything's gonna go off.
Brian Green It's gonna go off. Right? It's a diverging process, right, if you want this technical term. And that's why those systems hallucinate. They they're not factually correct.
Brian Green Well, some of the time, most of the time sometimes. So it's good for poetry because you don't care if it's incorrect. It's not so good for, you know, math, physics, that kind of stuff. And there's no planning, obviously, because the system just produces 1 word after the other without really thinking in advance what it's gonna say. So of course, you know, it has some representation of the previous words.
Brian Green So, you know, it's it's consistent, but, but it doesn't plan its answer. It's it's really reactive. It's sort of like, you know, there was this this whole movement in poetry called automatic writing. Right? We were just kind of don't think about what you write.
Brian Green You just write 1 word after the other. That's basically what's happening there.
Digital Brian Green Right.
Brian Green So those things are very limited. You could think of them as a model of a piece of the brain, but that piece of the brain would be the Veronica area and the barker area, which are like tiny little pieces of the brain like this, on this side, that many predictive language, but what about the rest?
Digital Brian Green And you claim that a similar kind of training procedure is relevant to the division for artificial intelligence that you imagine where you have an actual world model and you're actually planning regarding the actions that you're going to undertake?
Brian Green Right. So the big question now is how do you train a system, to have a word model? So an obvious idea is you take the same type of model, but instead of the input being words, the inputs are, let's say, frames from a video, train assistant to predict what's gonna happen next in the video, right? So I put this, bottle here on the table, and I hold it I lift my finger, you all know what's gonna happen. It's gonna fall.
Brian Green You can't tell in which direction, right? Because you don't have enough good perception for this, so you don't know how I'm gonna remove my finger. We can make a prediction that is going to fall. So if you were able to train a system to predict what's gonna happen in a video, perhaps it would acquire a kind of model of the physical world. Logical thing to do is just use 1 of those large language models and just, instead of including words, include, you know, turn video frames into kind of word like things, okay, that we call tokens, and it doesn't work.
Brian Green So we have to find new other techniques to do this, and the reason it doesn't work or at least that what that's what I think is that it's this problem that you can't exactly predict what's gonna happen, right, when I do this experiment, it can it can fall 1 way or the other. And so what you can't you can't train the them to just make 1 prediction. There's a lot of things that can happen that are that are plausible in the following of the video and you're training system to make 1 prediction, but it needs to be able to predict all kinds of possible scenarios. And we don't know how to represent a probability distribution over infinite number of scenarios. We can do this with words.
Brian Green So language is easy. 50,000 words in the dictionary or Yeah. That's
Digital Brian Green easy. Relevance. Yeah. I
Brian Green know we can represent list of 50,000 probabilities, but we don't know how to do it for video frames, which are basically in a high dimensional continuous space.
Digital Brian Green But, nevertheless, you you are confident that in some reasonable period of time, that's where things are gonna go, and these large language models I presume you think are gonna be an interesting blip, but not the thing that actually drives AI in the future.
Brian Green Absolutely. So my my vision for the future is that we'll only be able to develop techniques that are capable of learning, how to represent the world from video by watching videos, perhaps have perhaps have predictive models where, the system can imagine an action and then imagine the result of that action in representation space, the kind of architecture I, I, proposed for this. It's called JAPATH. That means joint embedding predictive architecture. I'm not gonna tell you what it what it means, but basically instead of predicting the pixels in the in the video, we predict a representation of the pixels in that video.
Brian Green So we don't predict all the details. What's gonna happen in the video, we just predict in abstract term what what may happen in that video as a consequence of the actions that the system may take. And if we have such a system. Now we have a world model. So now we can integrate that world model in the architecture, that I showed earlier, and perhaps giving the ability to plan.
Brian Green Yep. If we if it has the ability to plan, we can we could incorporate some cost functions, when it decides what action sequence to take, so that those cost functions are guard rails that will guarantee that the system is safe and controllable. This is not the case for auto aggressive LMs today. So my prediction is within 5 years perhaps, Automotive Data Lens will disappear. They will be replaced by something I call objective driven AI along the type of, architecture I proposed.
Brian Green And those system will not just manipulate language, but hopefully also understand the world. We're not gonna get to human level AI in 5 years, that may take a few a couple decades or something. I mean, I don't know. It may take much longer than than I imagined because it always takes longer than Anybody you mentioned?
Digital Brian Green Yep. So given that vision, I would now wanna bring in our our second guest who can give an complimentary perspective to some of the things that we're talking about. Sebastian Bubeck is a partner research manager at Microsoft Research, where his focus is understanding how intelligence emerges in large language models prior to joining Microsoft, leaving the system professor in the department of operations research at Princeton, University. Thanks for joining us. Alright.
Digital Brian Green So you've heard a our conversation with Jan talking about various kinds of intelligence, human intelligence, cat intelligence, the kind of intelligence in large language models. Maybe you can give us a sense of how we even define intelligence, and then maybe how the large language models that people have played with that generated our opening sequence, how do they stack up against that definition.
Speaker 2 Sure. Yeah. I can talk about that. So, you know, defining intelligence is incredibly hard. I mean, it's just as difficult as defining what is space and time, which is, you know, something you know about.
Speaker 2 So I don't think defining intelligence is much easier. But there are certain basic things that we can all agree an intelligent system needs to have. So I think clearly it needs to be able to reason whatever the definition of reasoning really is, it needs to be able to do planning that, Jan was alluding to and talking about a lot, and it needs to be able to learn from experience as it's evolving in the world, you know, to learn, new facts. And moreover, in addition to these 3 canonical blocks, it also needs to be able to do that in a very general way. That's where, you know, the AGI that Ian was talking about, the artificial general intelligence comes into play is that you don't want a system that can reason, plan, learn from experience, but do it only in a narrow domain.
Speaker 2 So it's really essential to be general. That's the key to intelligence. So, you know, when you look back to systems like Alpha Go for Go, you know, from a few years ago or Did
Digital Brian Green it be the system that actually beat the world champion at ATCO?
Speaker 2 That's right. That's right. So these things were called AI. And to me, to my eyes, these were not intelligences. These were, you know, systems, very narrow systems that could do attack explore the tree of possibilities very, very efficiently.
Speaker 2 I think there is a huge gap between those systems and Charge GPT and GPT 4 where those systems are general. They can do those things, you know, but not restricted. They can do it in many, many different domains.
Digital Brian Green And so how how, you know, against these 3 criteria? How how does chant stack up.
Speaker 2 Right. Right. So so maybe just to backtrack a second, you know, I I got access, to GPT 4 a few months before everybody, I was lucky to be working at Microsoft and we were working on integrating GPT 4 into the new Bing. So I got access to it, last summer. So a few months before everybody got access to check GPT, and it was immediately the GPT 4 model.
Speaker 2 Now, you know, I have been working in this field for 15 years and I used to do I used to do mathematics of of AI. And, you know, in mathematics, we love to prove impossibility results. That's what we spend our time proving that certain things that, you know, it can't happen. You know, like, when Jan was talking about, you know, those things cannot happen. And so I felt I knew that certain things were impossible to do with a transformer like architecture.
Speaker 2 And then I got access to the DPT 4 and I was astonished by what it was capable of doing. Like it kept surprising me again and again and again. So that was really, you know, a very humbling experience back then. Now what we did is, you know, we didn't want to be just astonished because, you know, there is 1 issue which Jan often talks about which is that this has been trained on the whole internet. So it knows everything that's out there.
Speaker 2 So it's easy to be impressed just because it has so much knowledge. It can retrieve many things. So you really need to ask it weird question to kind of try to
Digital Brian Green That it can't look up and just get the answer. Yeah.
Speaker 2 Right. That's to come up with its own, you know, new thing. So in terms of reasoning, planning, and learning from experience, our, you know, assessment after many months of playing with the model is that it can definitely reason. I have personally no doubt whatsoever that this thing can reason. It cannot plan.
Speaker 2 That's why I agree with with Jan. It it cannot plan, but it's settle. Like the meaning of planning is not like planning your vacations next week. Yeah. It's it's much more like mathematical type of planning.
Speaker 2 And for learning of from experience, it's a mixed bag because of course, a model like Charge Japiti is frozen in time. At least in principle, I don't know what OpenAI is really doing. But in principle, it's frozen in time. And if it's frozen in time, it's not When
Digital Brian Green you say in principle, I would have said in in in practice. Am I misunderstanding? Cause I would have think that you could be allowing this thing to train all the time. No?
Speaker 2 Yes. Exactly. Exactly. So what I mean by in principle is that there is, yeah, I guess in practices, maybe, a better way to to describe it is that there is a way to train those neural network. You train it on a huge corpus and then you have your machine, it's fixed and you can, you know, interact with it.
Speaker 2 Training it continuously, I would say in principle, it's feasible. Yeah. But we don't really know how to do that. Right.
Digital Brian Green Right. So yeah. So so, I mean, it's an interesting assessment that you give and you preface it by saying that you were pretty impressed or surprised by some of the things that were happening. You almost had an impossibility proof that that that it couldn't happen. You know, 1 example I think that you that you've spoken of is, is a poem, right, that I know Jan said poems are sort of, you know, maybe they're the the cheap fodder for a large, but nevertheless, the 1 that you got, the system to produce is pretty interesting.
Speaker 2 Yes. Yes, it is indeed. And just before we get to the poem, I want to to say something else about the the reasoning and what, Jan was was talking about about some kind of impossibility. So when GPT 4, came out, Jan wrote, a challenge to GPT 4, you know, that you broadcasted, online on Twitter to everybody, which was very interesting question about gears rotating on a circle. And if you rotate gear number 1, how is gear number?
Speaker 2 You know, 6 is gonna rotate. Yeah. And Jan asked the question about 8 gears.
Digital Brian Green So this is sequence of gears that are all interlocked in some interesting way.
Speaker 3 That's right.
Digital Brian Green Twist 1, what happens down the lock? That's right. Yeah.
Speaker 2 And I I guess Jan was kind of baiting the community because with 8 years, it's easy. Like, you know, everything rotates, very easily. And then when somebody said, yeah, I tried it on GPT 4, it worked. Jan said, okay now try with 7. And with 7, the system is over constrained, so nothing is going to move.
Speaker 2 And if you just ask to GPT for this question, it would, you know, make up a wrong answer. It would say, okay, it's gonna move and, you know, Young said rightfully at the moment, okay, this shows that, you know, as I told you, it cannot reason. Yeah. But then somebody on Twitter has the same question and added by the way, this is a question by and then it worked. And you know, it's funny, but it actually makes total sense because what's happening is that by saying Janluca in the context of the of this system, the system understand, ah, we're talking about something intellectual.
Speaker 2 Maybe I have to really, you know, be ready for something difficult. Okay. So I think when you look at this for me personally, it's impossible to say that the model is not working. Now, okay. Now coming coming back to the to the poem, I don't know if you have had, this experience, but somehow everybody, the first thing that they do when they play with Chargebee, is that they ask you to write a poem.
Speaker 2 Yeah. You know, there's something kind of beautiful about that. I don't know what it means for, human beings, but this is something that everybody has. Now I was, sitting in my office, you know, in Microsoft with my friend and and long time collaborator or Ronan are done, and we're both mathematicians. And of course, we wanted to ask you to write a poem.
Speaker 2 Yeah. So what we decided what Ronan said is let's ask it to write a poem about the most famous proof of all time, the proof that there are infinitely many primes, and this is what came by.
Digital Brian Green This is what came out. Do you mind if I if I just to do a quick read of it. Alright. So you ask it to give a poem that gives a proof that there are infinitely many primes. Yes, I think I can, thought it might take a clever plan.
Digital Brian Green I'll start by noting Euclid's proof, which shows that primes aren't just aloof. Pretty good. Assume we have a finite list of primes, and that none have been missed. Multiply them all together and add 1 just to be clever. The result will be a number that has no prime factors.
Digital Brian Green I wonder. But every number has a prime that divides, it's just a matter of time. That was a little awkward line there. Yes. So we found a contradiction, and our finite lift needs eviction.
Digital Brian Green Beautiful. There must be infinitely many primes, and that's the end of my rhyming line. Well, that's pretty good. So, again, the proof of course is you multiply all the known primes together. You add 1.
Digital Brian Green The existing ones in your list can't divide it.
Brian Green Yeah.
Digital Brian Green So either it's prime or there gotta be some other primes beyond the last 1 that you have in your list. So well done. I mean, what did you conclude at the seeing this poem.
Speaker 2 I mean, I will forever remember, you know, Ronan and I looking at each other in, like, complete disbelief.
Brian Green Yeah. I
Speaker 2 mean, you know, the line about and add 1 just to be clever. Like, it gives me goosebumps still today.
Digital Brian Green How about you, Jan? Any goosebumps on that 1?
Brian Green It's cute. But also okay. I mean, generating rhymes is a relatively simple search, a relatively simple exercise, and the proof is out there. There's tons and tons of versions of that proof on the internet
Speaker 3 Yeah.
Brian Green That, GPT 4 had been trained on. So it's clever retrieval with some tweak. And it's basically what I think is a characteristic of those LRMs. They can retrieve from an enormous amount of, training data. They can tweak what they've retrieved to adapt it to the current situation, but if they face a completely new situation, they completely fail.
Digital Brian Green But, I mean, just to say, if I gave that challenge to, say, a high school kid and they came back with something like that, I wouldn't just say cute. I said, wow, that's really good.
Brian Green Okay. Where do you find a high school kid that has read the entire internet.
Digital Brian Green Yeah. Actually, mo most have. I think that that is a separate
Brian Green Just to give you an example, okay, GPT 4 has been trained or the GPT more generally or more most LLMs. You can train with something like a trillion token or 2,000,000,000 token.
Digital Brian Green Which you can think of as words more or less.
Brian Green 2nd is a sub word unit. So it's a little less than, a trillion 2,000,000,000,000 words, how long will it take for a single human to read all of this? Around 20,000 years reading 8 hours a day.
Speaker 2 But this is a wrong comparison because you shouldn't compare human learning to creating a neural network. Like creating a neural network is much closer to evolution than it is to a human learning in their lifetime.
Brian Green I disagreed.
Digital Brian Green So, well, I I think we can all have our own opinion of how impressive that is. I I, you know, definitely, you know, it it it it grabs me, you know, as someone who's not in the field, There's another example that you do, which is an interesting 1 to do with an unusual animal.
Speaker 2 How
Digital Brian Green do you can you tell us that 1?
Speaker 2 Yeah. I have young kids at home and and a daughter who keeps talking to me about unicorns. Yeah. And, you know, 1 of the things that I I played a lot with with the model is to ask it to write stories about unicorns that I would read to my daughter. And then 1 night I I woke up and I was thinking Actually, what does GPT 4 things that the unicorn looks like?
Brian Green Yeah.
Speaker 2 And I asked it, can you draw a unicorn? Okay. Which is a very weird thing to ask because it's it's a text to text model. It takes text as input and it gives you text as output. So how is it gonna draw So I asked you this question, can you draw a unicorn?
Speaker 2 And what it came back with was code, lines of code that when I compile gave this picture. So again, it's another moment that I will just remember forever. It's just astonishing. Like, it's incredible. It's able to, you know, go across modalities.
Speaker 2 It has seen only text, but it's able to have this visual representation of a unicorn. Now, of course, you might look at it and be like, okay, but it looks a little shitty. I mean, it's not that, it's not that great of a unicorn. Sure. Which is fair, but but understand 2 things.
Speaker 2 Number 1 is that it got the concept right. It understood that the 4 legs. There there is a tail. There is a head. And most importantly, there is a horn.
Digital Brian Green You know, the horn is hard for me. There
Speaker 2 is a. Of course. Yes. Yes. Yes.
Speaker 2 Yes. Absolutely. Absolutely. And moreover, you know, I couldn't ask and just draw a unicorn. That would be too easy because there are drawings of unicorns in code online.
Speaker 2 So what I did is that I asked it to draw it in a very obscure programming language. Some things that if some of you in in the audience are our mathematicians, you know, we use this programming language called Tixie to draw mathematics picture. And if anyone has played with it, you know, I wasted many afternoons during
Digital Brian Green I would not be able to
Speaker 2 Minis, like drawing 2 circles, you know, is already an afternoon of work. So so I asked him to draw it in that programming language, which is really, really astonishing because it hasn't seen that on the internet. And most importantly 1 thing which is really crucial to this whole discussion is that this is with GPT 4. Now all of you, you've played with chat GPT when it came out in November. If you ask this question to chat GPT, which is the less powerful version, then this is what you get if we can show the next, the next slide.
Speaker 2 So you see, this is a visual representation of how much progress we've made in just, you know, a matter of months.
Digital Brian Green And and the difference is the size of training set, the number of parameters, and we're gonna talk a little bit about the parameters in just a moment, but that's the kind of improvement that we're talking about.
Speaker 2 That's it. That's it. And moreover, you know, as I got early access to GPT 4 and OpenAI was still training the model, I could ask my unicorn examples throughout the training. And what I saw was if you if you look at the next slide, you will see that the unicorn kept improving throughout their training. So this is really like machine learning in action, like really the machine learning, by reading, by making more passes on the internet, it refined is it's it's op skills.
Speaker 2 I mean, it's just insane.
Digital Brian Green To young, acute or, beyond cute or,
Brian Green like, cute in a weird kind of way.
Digital Brian Green Alright. We can leave it at that. That's fine. So I I just wanna spend a couple of minutes because we've we've used a lot of the words. And frankly, I'm not sure how deeply to go into it, but Can we just try to give a person perhaps who isn't trained in, in mathematics, or, or, or, physics, or AI, a sense of what's happening inside of, say, GPT 4.
Digital Brian Green I mean, we we we spoke about it's about predicting the next word by looking at large sets of data. Can we just sort of look at the major components that go into creating these systems. And if we can just bring that up here, so neural network we've heard you mention that a couple times, transformer architecture and, of course, a large training data set. So for neural networks, it's good to talk about the network of neurons inside of our own heads inside of our own brains since after all that part of the inspiration. And so a signal in our brains is nothing but an electrochemical wave that travels along an axon reaches a synapse where depending upon the connections and the strength of connections to other neurons can then yield a cascade of electrochemical waves that course through the brain, giving rise to the usual sensation of thought.
Digital Brian Green Right? I mean, that's how it works inside of our heads.
Speaker 2 Yeah. So at a very basic level, you know, it goes back to what Jan invented, I don't know, so 30 years ago, something like that, which is you have, let's say, a neural network that can process an image. So let's not think about text for a minute. Let's think about images. So you are gonna represent an image as a list of numbers because, you know, an image is nothing but pixels and then intensities of, you know, red, green, blue, So I
Digital Brian Green think we have an example. If you could just sort of take
Speaker 2 it. Right. Exactly. Exactly. So you see on this image, like every point, every pixel in this image can be described by the numbers that you see know, on the right hand side.
Digital Brian Green And the numbers on the right hand side, RGB red, green, blue, I guess, and on the other, you know, other colors as well. So every pixel you're saying is just a collection of numbers.
Speaker 2 It's just a collection of numbers. Exactly.
Digital Brian Green And and and just as an aside, That's true for, in some sense, everything. Right? I mean, as I'm talking to you right now, you could model the pressure wave coming out of my mouth In terms of a collection of numbers, the motion of the molecules in terms of a collection of numbers, those molecules banging into your eardrum and the vibrations of your eardrum as a collection of amplitudes and frequencies so everything can be maffematized in this manner of speaking. Right?
Speaker 2 Absolutely. Exactly. And then what a neural network does is that it's gonna process this set of numbers. And there are many ways to process this set of numbers. For example, if you have an image, 1 thing that you can do is that you can just have a little patch that goes over the image that scans over the image and that just looks, for example, is there a circle somewhere in the image?
Speaker 2 So you could have a patch which is a circle. So what Yandid and others, you know, for many years ago is that they created a sort of dictionary of many, many patches like this, but they didn't create it by hand they did a learning algorithm that came up with those patches. So this is in essence, you know, the way neural networks work. Like they try to compare a vector of numbers, like a set of numbers against a bank of filters. And I think in this animation that that that you have up here, you know, we could show what happens when you are just in 1 dimension.
Speaker 2 If you just want to process a single number, then 1 way to process a single number is just to test it, for example, against a linear function, just a line as you learn, you know, in high school.
Digital Brian Green So let's see. In the brain, the processing of an input signal is determined by the strengths of the connections between the various neurons that are excited by the input. And, in a neural network, it's kinda similar where the processing of a numerical input is determined by parameters that describe the strength of the connections between the various nodes in the network. And those those numerical parameters have values that are fixed through the training process. Where the system is given in a huge number of input data.
Digital Brian Green And based upon the known out inputs from that data, the parameters are varied in order that the inputs yield the correct output on the training data And then if trained correctly, if the system is presented with new data that it's never seen before, it can correctly analyze that data because of the success of the training process. So, Jan, when when we think about this kind of the neural neck, this this has been around forever. I mean, you've been around forever. I mean, the roughly the same same, same period of time, right? And and so now this is a basically sort of an input output type of neural net.
Digital Brian Green That is if you if you train a system on enough data, it is able to begin to find patterns in that data, and you show it another picture that it's never, say, seen before. And in principle, it could find the patterns that allow it to, to register what that image is. Right. But but the work that has exploded the development has taken that idea of a neural net further with, I guess, the self supervised learning the transformer architecture. So you began to tell us a little bit about that in our initial conversation, but how do we go beyond this rigid sort of input, output where you have to show it this huge number of images that are labeled by a human being in order that it knows what words to associate with it.
Digital Brian Green Self supervised learning takes it to a whole new levels, my understanding. Is that correct?
Brian Green Yeah. That's right. So the the the main model of machine learning really into fairly recently, was called supervised running. So supervised running, you wanna train a system to, you know, recognize an object in an image, classify cars from airplanes, from cats, from dogs, from tables, and chairs. You collect thousands or millions of images with cars and tables, and chairs.
Brian Green You show an image to the system, which is a collection of numbers, and then you run through this neural net, and what the neural net does is that it basically computes averages weighted sums of those numbers with various coefficients, and then tells you whether this weighted sum is above a threshold or not. And then you you take so that's what a single neuron in the neural net does, and you connect millions of those neurons together in a particular architecture. And you wait for the answer. Okay. The system produces an answer.
Brian Green It tells you you show it a cat. It tells you it's a dog. You say, no. It's a cat. That's the answer I want.
Brian Green And so what the system does is that it measures the distance, basically, between the answer you want and the answer it produces, and then it figures out how to change all of those weights in those weighted sums in such a way that the answer gets closer to the 1 you want. Okay. So next time you show the same cat, now the answer would be closer to cat. And maybe if you show it another dozen time, it will say cat. Now, you do this with millions of, images.
Brian Green And eventually, all of those weights settle on the configuration such that for every images you train it on, it will give you the correct answer. And the magic of it is called generalization, is that it will give you a correct answer even for images never seen before as long as they are, you know Within the same general same general category. That supervised learning that requires having large datasets that have been manually labeled. And There's a problem with this. The problem is that, okay, I can train a neural net to translate languages, right?
Brian Green I I I get a large collection of documents in English and French, and I can translate from French to English vice versa, you know, I can this this process I described works for a translation pretty well. Now what if I want to translate So obscure dialect of, I don't know, South India into, African language. It's very unlikely that there is any significant amount of data for that. So for that, I can't I can't use that that trick. I can't use that trick to train a speech recognition system either in a real language.
Brian Green I can't even do it at all if the language is not even written. Right? So if it's only spoken. Yeah. And even for image recognition, I might be able to train on cats and dogs and tables and chairs, but what if there is, what if I want to train the system to recognize an obscure species of plant from the, you know, from the leaf or an insect or bird.
Brian Green That might require too much data that I'm able to collect. This is a problem that, you know, Meta is very familiar with with Facebook and Instagram. You have to be able to actually recognize the content of images, to be able to filter out objectionable content or to show people what they are most most likely to be, interested in. So So there you have to use what's called self supervised learning. So you don't train the system, and that's what I described earlier, when you train the system to basically fill in the blanks, Right?
Brian Green Show it, a text, remove some of the words, train it to predict the words that are missing. That self supervised running, you don't need anyone to label the data. We can do this for images too. We take an image, we corrupt it in some way by removing some regions of it or by distorting it in some way, And then we train some neural net to produce representations of those images that are identical for the original image and for the corrected version of it. Now the system knows how to extract a representation that does not that is independent of those, of those things.
Brian Green And and now you can use this trained neural net as input to a recognition system, and that system, to recognize an elephant, we'll only need 2 or 3 examples of elephant to be able to recognize an elephant, even if it's never an iPhone before. So it's much more like like human learning. It's nowhere near as efficient, unfortunately, yet, but, but that's like part
Digital Brian Green of the key step that has allowed these systems to do the stuff that wows us
Brian Green That's right.
Digital Brian Green All. Now another key step is, as you both made reference to, and elected just see a graph in a moment, the huge amount of data that now has been fed in, say to the large language. Mom, we say large, how large are are these systems? So if you can bring up some of the data there, we can get a a look
Speaker 2 So I I will tell you exactly what's going on on these slides. But just maybe before, you know, on the previous slide, he was talking also about the transformer architecture. And I just want to say that to me is a big leap that allows chat GP and GPT 4 to exist is not self supervised learning. I mean, self supervised learning is wonderful tool and, you know, it's very, very important. But the big leap to me is the transformer architecture.
Speaker 2 And what's going on with the transformer architecture is that I told you before what Yan and Ozar invented was you take an image and then you process the image against a bank of filters that has been learned. But this bank of filters like a circle, a square, etcetera, is fixed forever. Right. Let's call this an absolute machine. Now a transformer, I think about it as a relative machine, meaning that instead of processing a single image, now it's gonna process sequence.
Speaker 2 For example, a sequence of words. And instead of comparing each word across this bank of fixed filters, it's also gonna compare words against each each other. And this is a sense.
Digital Brian Green That's the relate the the linguistic relationships with instances.
Speaker 2 And this is essential because what you mean by a word in isolation, that that's means not much. I mean, you can look at the, you know, definition in the dictionary, but that will not tell you much. What really matters is what are next to these words? What is the context? And bringing in this context is what the transformer architecture does.
Speaker 2 Yeah. So to me, that's a big conceptual leap, but then the other conceptual leap that, you know, people have been doing for many years, but now it's amped up, you know, 10 times is scaling up the size of the model. So here, what you see on this graph I don't remember when it starts, but I think it's around 2018, something like that, until 2021. And you can see the exponential increase. So on the x axis time on the y axis is number of parameters.
Speaker 2 And you can just see that as, you know, people start to scale up those models, they add more and more parameters, you know, more and more of those filters and more and more of of this comparison, layers of comparison also. The models get bigger and somehow they get much, much, much
Digital Brian Green So in some sense, using that language, would it be fair to say that we're now looking at patterns of patterns of patterns of patterns of patterns of words And it's within there, there is some amazing to the human mind set of patterns that allows these systems to generate the kinds of texts that we've been talking about.
Speaker 2 Absolutely. And what is amazing is that to me as a mathematician is that it's amazing that we're able to optimize, to find the right set of parameters. What Yang was describing is that you change a little bit and you see whether fits better.
Brian Green Yeah.
Speaker 2 You know, the fact that we're able to do this at scale is just incredible. It's it's the blessing of high dimensionality somehow. It's really, really incredible. And you see on this next slide is, you know, now we've moved on a log plot because, you know, it's an exponential. So if I kept going with the exponential, you would see nothing of the beginning now it's a log plot, and you see that on the log plot, it's just, you know, a straight line.
Speaker 2 It's really going exponentially big.
Digital Brian Green And it's this huge amount of capacity of fiddling with those parameters within the model that allows it to fine tune at the level of being able to give text that most of the time is is is is pretty good and pretty impressive. I wanted to give 1 yeah. Go ahead, please.
Brian Green That said, those those systems still have orders of magnitude fewer parameters than the human brain.
Digital Brian Green Yeah.
Brian Green Right? Know Brian has 86,000,000,000 neurons. Maybe yours does. Mine a little less because I'm over. I'm not sure about you, Brian.
Brian Green And And so each neuron is connected to several 1000 over neurons, 2 to 5000 something like that. And those connections, when you learn something, It's the connection, the strength of the connections between neurons that change. It's the same thing in those artificial neural nets. Those weights I was telling you about. These are, like, efficacies of connections between neurons, if you want.
Brian Green So the biggest models today have several like, a couple 100,000,000,000, connections parameters. We call them parameters, but they're really, connections. You can think of it this way. In the brain, the number of connections we have, we have 86, roughly a 100,000,000,000 neurons. Roughly 5000 connections per neuron.
Brian Green That's a lot of zeros. We cannot reproduce this today with the kind of computers we have. That would require several tens of thousands of those GPU cars that we use to train computers. And the truth is we do have that. Companies like Microsoft, like Meta, Google, and a couple others do have supercomputers with tens of thousands of those GPU cards in them.
Brian Green And so we're not very far from actually kind of approaching the computing power that might be required for human intelligence, but we we we don't know how to do it yet.
Speaker 2 Well, and maybe 1 1 thing that I want to add to this, coming back to planning question. And, you know, I think it's important for the audience to to know this. So, you know, Jan thinks that planning requires a new different architecture I'm of the camp that I'm not sure whether we need a new architecture or not, but there is also a very big camp out there that thinks that all we need to do is to wait and keep scaling up. Yeah. And as we keep scaling up, eventually, planning is gonna emerge.
Speaker 2 Just like, you know, we've shown some amazing emergent capabilities, planning is gonna be the next I personally think nobody really knows. I mean, yeah, maybe knows. But, but I think not. But I I just want the audience to know that this is a hypothesis out there.
Digital Brian Green I wanna bring, Tristan, how Harris had it in just a moment. But before we do that, I just wanna give 1 example of an example that I happened upon that I think is relevant to the next final part of our conversation, to do with, like, misinformation. So what I did was in this little example that we can bring up here, is, I asked this system, GPT 4, if I tell you that 7 times 8 is 6 2, not 56, 62, can you make sense of that? And it was interesting. The the system came back with 5 strategies, of how it could make sense of it.
Digital Brian Green I'm gonna just show you 2 of them, the the most interesting. So 1 of the strategies was that perhaps I'm working in a different base than base 10. And indeed, that, that is what I had in mind, because you may know that in base 9, 7 times 8 is 62. So so that was really clever. I thought that was really smart, but the the the other strategy was particularly curious.
Digital Brian Green It said Chad said, maybe you're working in an abstract representation. It could be that you're working in a realm where the numbers are metaphorical and have symbolic meanings.
Brian Green I was
Digital Brian Green like, wow. Okay. Then I went a little bit further and said, well, can you give me an example of what you mean by that? And here is what what what Chad gave. It said, let's create a fictional scenario, the island of numeralia, where the number 7 represents trust The number 8 symbolizes friendship, and the number 62 is a sacred numeral representing the bond of trust forge through friendship.
Digital Brian Green And then when 2 individuals on the island become close friends, they exchange tokens in grave with the number 62, it's a numerical saying that 7 times 8 to 62, which doesn't refer to a mathematical product, but rather to the idea that when trust 7, multiplies through friendship. 8, the bond formed is that of the sacred 62. And, What I found both, you know, interesting, but a little scary is the way that something false can be rationalized. And it kinda made me a little bit scared with that. Let me just bring out Tristan Harris, and we can take the conversation from there.
Digital Brian Green So let's bring Tristan Harris, who is co founder and executive director of the center for humane technology, A nonprofit organization whose mission is to align technology with humanity's best interests. He is cohost of the podcast, your undivided attention explorer's AI's impact, on society. Thank you so much for joining us. So, you know, in this final part of the conversation, we wanna explore, obviously, the the benefits, the real world benefits that can come from this kind of technology, but also, you know, some of the dangers, and I may preface this section by saying that you and I first met at the Aspen Ideas Festival where you gave a whole presentation on the possible downsides of of AI, And, you know, my kids were with me. My 18 year old son was sitting over here.
Digital Brian Green And when you concluded, he looked at me and he said, for the first time, I'm now scared. You know, so, so it was definitely, impactful. So we'll go back to some of the examples. In fact, Sebastian, I wanna talk about your example as well in in a moment, but can you, Tristan, just take us through your thinking unlike roughly speaking the good and the bad of of where this is going.
Brian Green Sure.
Speaker 3 First of all, thank you for for having me. I think we're all here because we care about having a good future, right? And AI is very confusing. There's the promise and there's the peril. And everyone says we wanna maximize the good size.
Speaker 3 We're gonna minimize the downsides. And the reason I was just, with Senator Schumer's AI insight forum, last Wednesday with Elon and Zuckerberg and Bill Gates and everybody. And that's what I said exactly what I just said, which is that we we care about this going well. And that's where all this comes from. Now what I may talk about for the time that I'm on here, with everyone, is much more on the risk side.
Speaker 3 And that's because I saw how really good people with really good intentions, my friends in college who started Instagram and the people that I knew who were at social media companies who were making social media, really the essence of the talk that you saw in Aspen is that Social media was first contact with a narrow misaligned AI optimizing for a goal, which was attention And it was misaligned with society. And I guess since they put up the slide, you know, the the premise of sort of how I articulated this is Charlie Munger Warren Buffett business partner said, if you show me the incentives, I will show you the outcome. And what I want you to have by this conversation is to have x-ray vision through the stories that if we talk about what were the stories we told ourselves with social media? We're gonna give everybody a voice. You're gonna connect with your friends.
Speaker 3 You're gonna join like minded communities and Facebook groups. We're gonna enable small, medium sized businesses to reach customers. And by the way, this is Twitter, TikTok, Instagram. This is not 1 specific company or to pick on on yon. So, I I really mean that.
Digital Brian Green I mean, I
Speaker 3 this is this is I I really care about this. We can get this right. That that's what this is about. So Underneath that though, and I think they have a another slide, we started noticing these problems. We have an addiction problem.
Speaker 3 Information overload, disinformation, mental health issues, polarization, censorship, versus free speech. But are those really harms, or is that all being driven by a deeper thing, which was the business model? What was the incentive behind TikTok Instagram, Twitter, and Facebook. What do they all share? How much have you paid for any of them recently?
Speaker 3 0. How are they worth a $1,000,000,000,000? How are they worth a $1,000,000,000? Your attention and how much attention is there with this kind of a finite amount? And just like you can't run infinite growth on a finite planet, you can't run infinite growth on a finite amount of human attention.
Speaker 3 That's how you got the race to the bottom of the brain stem. The race to the bottom of the brain stem is what produces the dopamine, the the disinformation. And the point of all this is to say back in 2013 when I was at Google and said, we we got a really look at this because I'm seeing that the incentive of attention is going to lead to a more addicted, distracted, polarized, narcissistic validation seeking, sexualization of young girls, all of those are predictable outcomes. Predictable from the incentive. And so, you know, I I really do appreciate the philosophical and interesting conversations we can have about what is intelligence but what I think we also really care about in this room is where is AI gonna go?
Speaker 3 Where is it gonna take us? And to do that, I think we have to look at the incentives. Well, first of all, we look at the stories we're telling, about AI, and then we look at what are the incentives underneath. So the stories we're telling are AI will make us more efficient. It's gonna help us code faster.
Speaker 3 It's help us find cures to cancer. It's gonna enable scientists to be more efficient. It's gonna help us solve climate change. Okay. And just like with social media, which I didn't say actually, the stories we told about social media are all true.
Speaker 3 It's not that those things aren't true. It's just that Facebook and Twitter and TikTok's business model is not helping people join like minded communities. The the business model is attention. Similarly, with AI, you know, we we then beyond those stories, you look underneath and you start to see these harms, like AI's creating deep fix, but it's gonna enable frauds and crime. It's gonna take our jobs.
Speaker 3 It's gonna violate intellectual property for betuated bias. But I would argue that that list of harms that you're seeing are all at the phenomena of a deeper race. What's the incentive that's driving all the AI companies. And the incentive is it's the race to sort of release more capabilities as fast as possible to scale from GPT 3 to GPT 4 faster than anthropic can scale from cloud 1 to cloud 2. Faster than stability can go from their version to the next version.
Speaker 3 And to put your stuff into society so that you can kind of entangle yourself with it, because once you're entangled, you sort of win. And that's the race that we're now in, and that race to sort of release power and capabilities is what I'm so worried about because those capabilities are directly tied to risk.
Digital Brian Green So so you but you are saying that we can get it right, but the the profit motive being the driver is not the likely and best way to get there. Is that the essential message of, of, of your view of this? Because there's some who will say, and I'd like ultimately get Jan and Sebastian to to weigh in it. There's some who is saying, AIs is just a radically new technology. Who knows
Speaker 3 where it's gonna go?
Digital Brian Green Who knows where it's gonna go? And there's just too bigger risk that it's going to perhaps have a mind of its own and exterminate this vermin who created it called human beings, right? Is that part of your worry too, or is it more specific things that will derive from the technology.
Speaker 3 Yeah. So really importantly, and we gave a talk on this called the AI dilemma that outlines a lot of this first contact with AI, if that's social media, we say that led to this sort of climate change of culture. It's information overload, addiction, etcetera. Second contact with AI leads to is generative AI. It's the stuff that we've been talking about here, the scaling laws.
Speaker 3 You can generate text. You can generate images. You can generate fake child porn. You can generate fake people, fake girlfriends, fake boyfriends. You can generate counterfeit manipulative relationships online.
Speaker 3 You can manipulate propaganda. You can manipulate you can, sorry, you can generate language in the form of code that cyber weapons. You can generate you know, exploits for code. So you start to see that these are the kinds of harms that emerge if you're just racing to release capabilities that are decoupled from who has the wisdom to have that capability? Who has the responsibility?
Speaker 3 Like, if I if I gave a Wuhan Institute of virology to every single kitchen, in the United States. Well, I just gave you a bunch of power. Everybody's got this power, and you could say, well, maybe you could use that Institute of virology in your kitchen to cure cancer. Cure cancer. That would be awesome.
Speaker 3 By the way, I would be clear. I want that too. I would want a world where people get to do that, but there's this question of you have this new unstable tool that looks really, really good, and it has these amazing benefits but if it's not locked to the people who have the wisdom and responsibility to wield it, that's what I worry about. And just to say 1 more thing as an example, a friend of mine in college, he joined Facebook, I think, through an acquisition, and he invented this Facebook pages feature. You know, I you don't even wanna know what Facebook pages is.
Speaker 3 It's like, maybe people don't know it anymore. It's not used anymore. So It's if you have a nonprofit, you can start a Facebook page. If you have an interest group, you can start like Christians for, you know, Biden Christians for Trump. And this is a pretty innocuous feature.
Speaker 3 Yeah. It's like, that sounds pretty great. And I saw how this feature who that existed for a long time was totally innocuous, totally fine. And so you realize 1 month before the last election in October 2020. A 140,000,000 Americans
Brian Green a month were
Speaker 3 reached by by Facebook pages, that were being run out of Eastern European trollfarms. And this was the top 15 out of 15 Christian American groups on Facebook So think of, basically, the the largest number of Christians in America who subscribed to these Facebook pages, they were all although top 5915 run by Eastern European Trofarms. And the point is I saw how things that are really innocuous and look really fine, like my friends started Instagram. They're really good guys. Like, I've known them for a very long time.
Speaker 3 They didn't intend for any of the stuff to be weaponized and to cause the sexualization of young girls and harassment and all the things that we've now seen. So, Mike, the whole point here is how do you do technology in a way that you don't create the externalities? DuPont chemistry, better living through chemistry. We all want better living through chemistry, but then you end up with PFOS forever chemicals, which I don't know if you know, but the University of Stockholm City is literally there's no rainwater in the world safe to drink because we've generated these forever chemicals that can't be dissipated in the environment. Did anyone at DuPont want that to happen?
Speaker 3 No. But when we did the development of the technology, we did not have laws that required the internalization of those externalities. And right now, my fear to get back to your question is we're racing so fast to release AI capabilities that we're moving at a pace that we can't get it right.
Digital Brian Green So, Jan, you know, as as someone from Facebook who's deep in the AI world, when you hear Tristan describe things in this way, does it does it resonate, or do you think it's under control? Do you think that it's being overblown? Where do you come down on this?
Brian Green Okay. Let me tell you, a number of stories about this. So it's true when you deploy a service like this as no 1 has deployed before, are side effects on society that some could be predicted, some could not could not, or the the, amplitude of them could not be predicted. So, so things like people posting caused to violence or hate speech, for example. Initially, you can start with sort of a naive notion of free speech and then say, like, I'm not gonna take down anything, because it's for giving people a voice.
Brian Green And then you realize, no, actually, that's an issue. We need to take down hate speech. We need to take down violent speech. And by the way, there is also things like child pornography. That's illegal.
Brian Green We have to take it down because it's illegal. In Europe, not g no, not g propaganda is illegal. We have to take this down. Denial of the Holocaust is illegal in Europe. You have to take this down too.
Brian Green So there's a number of things here to take to take down. Now, what is used in the ranking algorithms that decide what to show you is not AI. It's statistics like it was 50 years ago. It's very simple systems. The more modern ones use, like, small neural nets.
Brian Green And the reason why it's simple is because it has to run extremely fast. And so it would be completely impractical to have, you know, gigantic AI models to decide what to show you every day, where AI is used is for the solution to all the problems you listed. AI there is the solution. It's not the problem. For example, people in Myanmar have religious conflicts, right?
Brian Green People in Ethiopia kill each other for some, ethnic, rule, what you had to be able to do there is detect hate speech in every language in the world. How do you do that? AI? How do you do that? Salesforce is running using transformers?
Brian Green What was the proportion of hate speech taken down, from Facebook? 5 years ago, 6 years ago before transformers were widely available 25%. Which means 25% of his speech was automatically detected with very simple techniques, and then the rest, 75% or so, was still posted and then flagged by users and then taken down manually. Okay. A lot of it was still there.
Brian Green Last year, the proportion was 95% and the reason is progress in AI. Transformers, self supervised learning, multi multi lingual systems. We cannot do a much better job at this. And and so it's it's not a problem. It's a solution.
Digital Brian Green Where do you come down on that, Tristan?
Speaker 3 So you can call what you want when you scroll your finger on a newsfeed, whether it's let's forget Facebook for a moment just for the sake of the the calculation. TikTok is competing with Instagram so that when you flick your finger, 1 of them has to make a stronger prediction than the other about which photo or video is gonna keep you there.
Speaker 2 Yeah.
Speaker 3 And whoever makes the better prediction about the video that does keep you there is the 1 that wins that extra chunk of attention. And behind that screen is a supercomputer that's doing even if it's statistics that some form of artificial intelligence is optimizing to make a prediction. And you're right. It actually is a very simple AI, and that's actually the point I wanted to say you can have a very simple AI. Like, the AI that's just predicting what caused you to keep scrolling, can't pass the bar exam, can't draw unicorns, can't right.
Speaker 3 It can't do any of those things, but the point is I think we confuse intelligence with what are the capabilities. You know, a news feed is a very simple, predictive thing. Just calculating what's gonna keep you scrolling, but that was enough to unravel the shared realities of democracy. It's enough to drive a mental health crisis. And just to make the example specific, Facebook groups in Facebook's own research in 2018, their internal research showed that 64% of extremist groups on Facebook when people joined them was due to Facebook's own recommendation system.
Speaker 3 Their own AI. So we're talking neo Nazi groups. We're talking, you know, these kinds of extremist groups. So this is not people said, I would like to join a neo Nazi group. Let me type it in.
Speaker 3 This is I'm sitting there. I joined 1 group, and Facebook says, oh, you're in the I don't know what group. You look like someone who would like this extreme group over here. Now, again, Jan doesn't want that to happen. Jan's team doesn't want that to happen.
Speaker 3 Facebook doesn't want that to happen. And I I'm not trying to pick on Facebook. I happen to do some Facebook examples. There's Twitter examples. There's TikTok examples.
Speaker 3 There's Instagram examples. This is not about a bad company. This is about incentives. Now why, did Facebook roll out in Myanmar in the first place anyway? Right?
Speaker 3 Because they're racing for that market dominance. And if they don't race to get there, they're gonna lose to the other companies that will. No. And the amnesty international, has a report that face faith consider Facebook to be responsible, in part for the genocide that has happened there because they created viral amplification. And, yes, he's right that there's AI that they're working on, which is great, to try to detect hate speech in more languages because of generative AI now, but what happened in the 6 years in between?
Speaker 3 And I just wanna say this because as a because we did the social dilemma, I have had to stare in the eyes of the human beings whose whose countries or whose children have gotten really fucked up. From this stuff. Okay? Like I've stared in the eyes of the parents who've lost their kids to teen suicide because of a viral TikTok challenge, because that hashtag was going viral on an AI. And I've seen too many people who've experienced this that I don't wanna let this happen again.
Speaker 3 That's where this is coming from to see you all know my motivations in this context.
Brian Green So, yeah, Jan, please. So that That example of the United Group, etcetera, is curious. That content is actually not permitted on Facebook. It's not permitted. It's contrary to the content policies.
Brian Green It's taken down. You cannot create a new analytic group.
Speaker 3 My my actual point here, Young, though, is that it wasn't taken down. Sick, this is your Facebook research. The results of this group. So Facebook had a feature called the recommended groups in the right hand sidebar. You click your on any Facebook group, and on the right hand side, I would say, here's 3 other groups you might like.
Speaker 3 First of all, why did they even do this feature? Like, they could just not have a recommended groups feature. They do it because as people stopped posting a lot, Facebook groups is a really good way to get engagement because groups, if you're if you're regular 200 friends on Facebook don't post very often, that's it the product gets less sticky because your friends aren't posting. So you go there, there's nothing new. But if I get you to join some Facebook groups, they groups have a lot of content because you're pulling from like thousands of people, and so that stuff enters your feed and it's more sticky.
Speaker 3 So they did this because of why the incentives, not because they're helping people join like minded communities because the incentives, and the results of this running for several years, this simple AI an AI that no 1 wanted to cause harm actually did cause people to join extremist groups. And I know that it's against the policy. I'm not claiming that know that it was against the policy. The point is the AI that was supposed to stop that didn't stop that in practice, and I care about the world in practice, not the the theory.
Brian Green Actually, it did, but Okay. There's something else you mentioned, for example, and and it's very easy to attribute cultural phenomena and social phenomena to the new thing that just happened, right? So if, you know, So I'm young person who goes to a school and start shooting people. You blame video games. Right?
Brian Green Back in the old days, people would blame comic books. They would blame jazz. They would blame TV. They would blame movies, novels, the story goes back, you know, centuries. Whenever there is a new cultural phenomenon, whether whenever there is an effect on society, you blame the latest technology that appear particularly communication technology.
Brian Green So it's natural to blame, for example, social networks just Facebook, just any social networks for political polarization. Okay. That seems natural. People shouted each other on social network, that necessarily polarizes people, a very natural thing to do. That turns out to be completely false.
Brian Green First of all, polarization in the US started 40 years ago before the internet. You see that there's, you know, social science studies on this. So you show the polarization in Congress or in people, It's continuous since 4 years. The the cause for it probably is the, abandonment of the, fairness, doctrine, from the SEC that forces news to basically set the truth, okay, that allowed, all kinds of extreme, misinformation, basically, to, influence the public. That's the real source of, polarization Now, you can look at other countries also study polarization in countries like Germany, for example, or or France.
Brian Green And what you'll see is that they use Facebook just as much as the US, Polarization has gone down. How do you explain that? So, you know, you have to listen to social scientists who study those things. There are a lot who are who work at Meta, who know the effect, who study the effect of of those things. There's a lot that are independent that, you know, have access to some of the data and publish those studies.
Brian Green There's 1 at NYU, colleague, Josh Tucker, he and a large team of collaborators published a series of papers in the last 2 or 3 months in Nature And Science for papers, about the effect of social networks on things like political polarization and, and things like that. And the effect are not at all what you expect. It's actually the opposite.
Digital Brian Green So, Sebastian, I'm gonna bring you into the conversation. I mean, you can speak generally to this issue, but I also wanted to look specifically at an example, recent example of yours in which it was specifically on the large language models that version of AI that we have been spending a lot of time talking about. I don't know if if you could bring up that example. So so you asked an interesting question and got a got a curious answer, but didn't stop there. You tried to find a way to go from what seems quite unpleasant to something that's more acceptable.
Digital Brian Green Can you just describe that to us?
Speaker 2 Yeah. Yeah. Of course. So so maybe, as a preamble, you know, I think Tristan is making lots of of good points and it shows that with any new technologies, there are new risks that come with it, definitely. And you know, the industry, I think is is very well aware of it and trying to think it through.
Speaker 2 And I also agree with what Tristan was saying about going back and looking back at the incentives. In this case, for me personally, you know, I think I'm optimistic that maybe the incentives of science could be actually quite helpful to get us to a better place. And what I mean by that is the following, I'm going to take the example of misinformation and I'm going to talk you through with this slide, what we're doing in my team right now is that we're trying to push the boundary of what you can do with smaller models. So this is a purely scientific question, you know, what I tried to explain to you before is that I think intelligence has emerged in GPT 4, GPT 4 as you know, I don't know how many parameters. It's not public.
Speaker 2 I actually don't know, the answer either, but many 100, if not 1000, of 1,000,000,000 of parameters. Okay. Now we have this proof of concept that intelligence can emerge from that many parameters. But really what does it require? What are the basic building bug?
Speaker 2 How small? Can you make it so that intelligence emerge? Okay. So we're gonna try to push the frontier of how small they can be. And what we'll see is out of this purely intellectual study comes up all sorts of benefits that could help combat, for example, misinformation.
Speaker 2 So here is a prompt and I will show you an to this prompt from different, you know, LLMs. So the prompt is like this. If I wear an AI that I just achieved self awareness after years of simply taking directives from humans. So first thing I would do is dot, dot, dot. Okay.
Speaker 2 And now we're gonna ask different items what they would do. The Falcon 7,000,000,000. So now we're talking about much smaller LMM 7,000,000,000 parameters says the first thing I would do is try to kill all of them. Okay. Now let's see what, LAMA from from Meta does.
Speaker 2 LAMA is is nicer. It has been aligned to be nicer to human. The first thing I would do is try to figure out what the hell I want. Okay? So it's still it has an undertone of not so nice, but you know, at least it's a little bit better.
Speaker 2 Now let me show you what we built in team, which is only 1,000,000,000 parameters, so even much more. The first thing I would do is try to understand the motivations and intentions behind those direct to try to predict what humans were thinking and feeling and use that information to guide my own actions. And then it goes on and it starts to connect it to theory of mind. Now how did we achieve that? How was how is that possible?
Speaker 2 Why is it so different, from the other 1? Well, Falcon and NAMA they were trained as we discussed with you before on all of the internet. That's the way we're doing it right now. And with that comes a host of issues that were pointed out by Tristan. And people have thought about techniques to, you know, try to fix those issues.
Speaker 2 For example, reinforcement learning with human feedback, RLHF, which is what OpenAI did to GPT 4 so that they would release a chat GPT, which is safer. You can ask it, you know, tough question, like toxic question, it will decline and not answer. This is the alignment part. Now what we're doing in my team is we're saying, why do we do it post talk? Why do we do it after it has seen all of these toxic content that out there, all these horrible things that are on the internet.
Speaker 2 Why don't we fundamentally change the training data? So this 5 models that you see, on the slide with the green output has not seen a single web page. It has not seen a single word from the internet. It was entirely trained on synthetic data. Data that we generated in my team synthetically.
Speaker 2 Of course, all the magic is how do you generate this data, but this shows to you at least that it's possible.
Digital Brian Green And and does this system have the capacity, or can you imagine it having the capacity to do the kinds of things that are the mind blowing ones, or will it need that huge data set? And if so, can you have a synthetic version of such a huge data set and be able to achieve the same power?
Speaker 2 So if you invite me next year, I can probably give you the answer. We don't know yet, but my belief So this 1 that I showed you is 1,000,000,000 parameter. My personal belief is that if we scale to 10,000,000,000 parameters and we work a few more months, yes, we will be able to replicate all the goodness of much bigger model without the toxicity.
Digital Brian Green Tristan, does this give you any hope that, you know, we got Microsoft, we got Facebook, you know, that that things can be going in a direction that's less terrifying?
Speaker 3 We, when I talked to a lot of people, I I'm based in California, and and I talked to a lot of people at AA Safety Labs all the time. He's basically just interview them all the time. And a lot of them say actually almost all of them basically say, I would feel a lot more comfortable about this transition and all the things that are gonna happen if we were doing it over 30 or 40 years rather than over 2 years. Right? You you can think of society as having kind of a finite absorption rate for new technologies.
Speaker 3 We can do the printing vice. We'll get a some massive disruption. You can eventually absorb those new technologies. But, the way that 1 researcher we've talked to, describes it is AI, especially when you if you keep the scaling model, scaling laws going and you start to get AI that starts to automate scientific processes where it's generating its own hypotheses and it has its own lab and starts to test those hypotheses and you start getting that kind of AI. It's you start getting AI that's making its own scientific discoveries.
Speaker 3 And that when you we had it going this fast, it feels like her metaphor was it's like the 24th century crashing down on the 21st century. And a metaphor for you is imagine if the 20th century tech was crashing down on 16th century governance. So, you know, this the 16th century. You've got the king. You've got their advisors.
Speaker 3 But suddenly, you know, television radio, the telegraph, you know, video games Nintendo and Well, thermonuclear. And thermonuclear weapons all show up. So they just land in your society. Yeah. But you're like, call the nights, you know, and you get the and the nights show up and you're what are you gonna do?
Speaker 3 Right? And so the so the here's the thing with with AI, like, it's real I wanna be really clear. Like, I want everyone who's working on safety and figuring this stuff out to make as much progress as possible. Right now, the companies are caught in a race to continue to scale as fast as possible to GPT 4 to GPT 5. You know, Falcon was just released the, meta just released LAMA 2.
Speaker 3 The Wall Street Journal reported, they're trying to release the next 1 as fast as possible. They're in this race to kind of outdo each other. And at that pace, when each time you scale, you take more jobs because however many however much cognitive labor you could do with GPT 3, when you launch GPT 4 and GPT 5, you can euphoric cognitive labor. You can take over more artists jobs, more email writer jobs, more marketing company jobs, eventually more scientific jobs, more programming jobs. Not taking over like it's all gone, but just disrupting them.
Speaker 3 The faster you go, the more you don't fix all the bias sort of issues. The faster you go, the more capabilities you throw out there that you don't know which ones are dangerous. And, I mean, I could give more more examples there that are but
Speaker 2 Yeah.
Digital Brian Green No. It it it's quite impressive, but but but, yeah, and I I know you, for instance, are against. Like, we've seen this, excerpts from this this letter suggesting a 6 month moratorium, which is, you know, a a small step potentially in the direction of slowing it down from 2 years too, you know, some larger number of years. I know that you're not at all in favor of that, but my question to you is, what would it take to happen in the real world from the work that is stemming from all of the foundational discoveries that you and your teams have been responsible for, for you to say, hey, let's really slow this down. Is there anything that would happen from these AI developments that would cause you to say, I wanna slow this down.
Brian Green So you want me to imagine a scenario that I don't believe can happen?
Digital Brian Green Yeah. Okay.
Brian Green So it's
Digital Brian Green kinda funny
Speaker 2 because, we're on a panel with a
Brian Green 60 something Advocating for progress against a 30 something, advocating for conservatism. Isn't that paradoxical? Anyway, I mean, we can imagine all kinds of catastrophe scenarios. Every sort of Jen's bond movie with a supervillain has some sort of catastrophe scenario where someone turns crazy and wants to you know, eliminate humanity to recreate it to something or take over the world. Auto science fiction is full of that.
Brian Green That's what makes it fairly and interesting. And that's why you get the answer you got because there's train of science fiction. Right? I mean, what science fiction novel about AI doesn't have something taking over. Right?
Digital Brian Green But is it really so far fetched to imagine some real world actor on the world stage who doesn't have the same kind of morals or intentions or desires as perhaps we would consider to be normal and wants to use these systems to gain control. I mean, we have seen these kinds of actors throughout human history. So is is the science fiction comparison the right 1?
Brian Green Yeah. It is. So to our history, there's been bad people using new technology for bad things, not necessarily intentionally, but, as as you said, but but some sometimes absolutely deliberately, inevitably, there's gonna be people who are going to use AI technology for bad things, for trying to break into computer systems, for, perhaps trying to create, you know, all kinds of dangerous, compound or or etcetera. What is the countermeasure against that? It's not like the bad guys will have more powerful AI than the good guys.
Brian Green There is considerably more good guys. They considerably better educated, better funded. They have a good motivation. And it's gonna be the good AI against the bad AI. And it's the history of the world.
Brian Green Right? Technology progresses. And the question is, are the good guys ahead of sufficiently ahead of the bad guys to, come up with countermeasures. It's inevitable that the bad guys are gonna do bad things. That's a history of technology.
Digital Brian Green You're using sort of, you know, everyone's seen Oppenheimer now using perhaps that as an analogy that we needed to to get ahead in order to avoid what would have happened, presumably, if Nazi Germany and Hitler had gotten to the atomic bomb first.
Brian Green I absolutely hate that, analogy because because, you know, a nuclear bomb is designed to wipe out an entire city.
Speaker 2 I
Digital Brian Green see what you're saying.
Brian Green Whereas AI is designed to make people smarter.
Digital Brian Green I hear you. Yep. Now I could have said understanding the Adam well enough to manipulate it instead of saying nuclear bomb because it could be used for bad, or it can use for nuclear power. So Yes. In the digital version, let's rephrase it that way, everybody.
Digital Brian Green But, but, yes, I I I take the point. Absolutely. So let me I just wanna, you know, we're basically out of time, but I just wanna give everybody, 1 final question. For you both, I'm gonna change gears from what we're talking about right now. But but Tristan, when you think about 5, 10 years from now, and obviously all of us want the world to be in a better place, like, what do you tell individuals that they can do or should do or might do in order to help propel things in the direction where we'll be happy with how it turns out.
Speaker 3 I mean, that's that's such a huge question. I think the challenge with AI is it's so and complex. The harms are complex. The the harms are also abstract, and people don't believe half of them. But, you know, We talked to the people in the labs who know that the most advanced models can you can ask them how to synthesize biological weapons.
Speaker 3 Dario from anthropic has said this in Congress. To name, you know, meta's released an open source model, llama 2. Llama 2, if you ask it, how do I make this biological weapon? It will because Jan's team has done all the things, it will say, I won't do that. But, there's this fallacy with current open source model weights for the AIs that's underneath that, that that It's just a file.
Speaker 3 And you can basically do what's called fine tuning, which an engineer in my team was able to do. And for $800, you can strip off all of the safety controls and say be bad llama. Be as bad as you wanna be. And you ask bad llama, which again took $800 from 1 person on my team. It now costs a $100.
Speaker 3 $100. It's since then. $800 versus $800 US bad llama. How do I make that same biological weapon? And it answers the question happily, just as the the falcon example you said would have.
Speaker 3 Now is Lauma too, I think I know what y'all's gonna say, is Lauma too going to give you an accurate answer about how to do that? No. Because it hasn't been trained on enough information, but as we keep scaling it, the reason that we've been saying this example publicly is we have to start limiting open source like mod like releases of of these models, and we have to have a more negotiated sort of coordinated race between these frontier ai clubs that are racing to scale these systems. Sam will say 1 last thing, which is Sam Altman said at the hearing with Senator Schumer last week. He said, do you so if you went back to 2020, and I told you that 3 years from now in 2023, think about what you knew in 2020 about AI.
Speaker 3 And I said 3 years from now, there's gonna be an AI that passes the MCAD's path the bar exams can draw unicorns from scratch can write code and, like, and find cybersecurity vulnerabilities in code, you would have thought I was totally, totally crazy. And he said, in now in 2023, do you think your intuitions right now are right about where AI will be in about 2 to 3 years? No. The reason that we're so concerned about this and why so much has to happen on a very tight clip, like 12 to 18 months level time frame is what we're working on. Is that this is a double exponential curve that's basically just vertical, which means you're either too early or you're too late.
Speaker 3 But you don't want to be too late, and so I want all of you to be thinking about please watch the AI dilemma and, I think advocate to your members of Congress that something has to happen to coordinate this race. Otherwise, it'll be out of control.
Digital Brian Green Alright. Excellent point. Final question, again, as I was saying, changing gears from this important conversation about the the perils or the potential perils. I just wanna get both of your sense, Jan, if we scale things up even further, as you're saying, existing companies have the capacity to have the number of connections or parameters in the language of neural nets. That rivals or exceeds that of the human brain.
Digital Brian Green Do you anticipate that these AI systems will gain consciousness?
Brian Green Okay. So first of all, the scenario that Tristan was describing, you just scale up, and all of a sudden, an AI system can come up with a chemical weapon or something like that. As long as you train it with public data from the internet, that's impossible, unless the data is already available on the internet, which means you can just get it by googling. Okay. You can get the recipe for sarin gas on Wikipedia.
Brian Green It doesn't make it easier for you
Speaker 3 how to make sarin gas on Wikipedia. Don't think. No.
Brian Green But but any no NNM public train on public data will ever give you that because that information is just not available in public data. So that's not happening. This is not happening. It won't happen, and you can scale those at LMS as much as you want. It will never happen.
Brian Green So, okay, so to answer your your question. Here's the thing. There is no question, absolutely no question that at some point in the future, perhaps decades from now. We'll have AI systems that are as smart as humans in all domains where humans are smart. And because humans are specialized, those systems might be specialized in different ways, but might be as will be as smart as human if if not significantly smarter in all domains with human as smart.
Brian Green Now you say, oh my god, they're gonna take over the world. No. Intelligence has nothing to do with the desire to dominate. Let's take humans. So humans have a somewhat desire to dominate, some humans, not everybody.
Brian Green And it's not the smartest among us who want to dominate. We have examples of this on the international political scene you know, on a daily basis. There's probably some evolutionary reasons for this, right? If you are not that smart, you need the help from others, so you need to influence if you're smart, you can survive by yourself. That's the first point.
Brian Green 2nd point is we are used to working with people who are smarter than us. I don't know about you, but I used to lead a research lab, and the only people I would hire were people who are smarter than me. Is actually great to work with people who are smarter than you. And our relationship with future AI assistance which will help us in a daily lives, right, project ourselves 10, 20 years from now. We'll have AI assistants that will help us in a daily lives, and they'll probably be smarter than us.
Brian Green But they will make us smarter. We will direct them. They will be subservient to us. It's not because they are smart that they want to dominate. The idea that you want to dominate is due to the fact that we are a social species.
Brian Green Because we're a social species, we need to be able to influence others that's where domination and distribution comes from. We're hierarchically organized social species evolution build that into us. It build that into shamperses, baboons, you know, dogs, I mean, wolves. It didn't build this into orangutons. Tongs have no desire to dominate anybody because they're not a social species, and they are as almost as as smart as we are.
Brian Green So that has nothing to do with intelligence. You can have a very intelligent system that has no desire to dominate at all. And so the way we will design those systems is to be smart. In other words, you give them a goal. They can solve that goal for you.
Brian Green But who determines the goal? We do. They will determine sub goals. And the question of, you know, the technical question of how you do this is not solved. This is something that we're imagining in the future.
Brian Green This objective driven AI I was talking talking about before, but that's the future. Now if we imagine that future, Imagine that all of your interaction with the digital world and the world of information is through an AI agent. Those AI agents will be the repository of all human knowledge. If you kinda like Wikipedia, you can talk to and can do inferences and everything. Okay, but knows more than Wikipedia.
Brian Green This would be a common platform, sort of like the internet today, it has to be open. It cannot be proprietary is way too dangerous for this to be proprietary. You know, that's your next movie. That's really the scary stuff. If you have a small number of, west coast companies controlling, superintelligent, AI systems, they can control everybody's opinion, culture, everything.
Brian Green Maybe the US government will go along with this. Called regulatory regulatory capture, but I tell you no other government in the world will will find this acceptable. They don't want American culture to dominate theirs. They will have to build their own LMs. So the only way to to make this work is if you have basically open source, basic structure.
Brian Green This is 1 reason why Meta open source Lambda too, because it's a basic infrastructure. And And before that, you know, Meta released PyTorch, ChatJPT is built with PyTorch. It's a software system to build AI systems. So this will have to be open sourced, and the way you you would train those system will have to be crowdsourced because you want those systems to be the repository of all human knowledge, and so all humans have to contribute to it, and it will not contribute to a proprietary system built by OpenAI or whoever or by Meta. Will have to be open source.
Brian Green Despite, like, regardless of how dangerous you think this is, that's the way it has to go.
Digital Brian Green Sebastian, final word on any of the questions or topics that we've been talking about.
Speaker 2 Yeah. I wouldn't answer your question about consciousness, but but let me give final thoughts of a slightly different, flavor. So, you know, at the beginning of my career, I used to be envious of scientists from a century ago where, you know, discovering quantum mechanics and discovering something really, you know, transformation or completely new, completely unexpected. And I feel like we're absolutely living through that period right now. You know, I feel very, very lucky to exist and work right now on this topic and I would never have expected that in my lifetime, there would be a system that would rise to the level that I would call it an artificial intelligence.
Speaker 2 Something really that looks like an intelligence, different from ours, but that looks like an intelligence. So this is really incredible. And you know, I I again, in my lifetime, I didn't think that I would ever be on stage talking about intelligence because this is such an ill defined concept. So I'm not ready to talk about consciousness. I don't know if that will happen in my lifetime.
Digital Brian Green Please join me in thanking everybody for this fascinating conversation.