Transcription
My first question is, where are we on the path to AGI? We are on the path to human-level intelligence or to superintelligence. I famously don't like the phrase AGI. And it's not because I don't think we're going to get machines that are smarter than humans. It's because I don't think human intelligence is general. So calling human-level AI AGI is a misnomer. Uh, unfortunately, that ship has sailed, but yeah, we're going to get machines that would be smarter than humans at some point. It's not going to happen next year. It's not going to happen in two years because we need a few conceptual breakthroughs for that, and those are things I've been working on and I'm still working on.
So, what do most leaders misunderstand about today's AI capabilities, and why does that misunderstanding matter for policy, regulation, and uh, capital allocation decisions being made right now? Okay, we're not going to get to human intelligence, human-level intelligence, or superintelligence by scanning ups or by even refining the paradigm. There is a need for a change of paradigm. Uh, and I've been seeing this, you know, for a number of years now. And I think we see, we're starting to see the limits of the LLM paradigm. A lot of people this year have been talking about agentic systems, and basing agentic systems on LLMs is a recipe for disaster because how can a system possibly plan a sequence of actions if it can't predict the consequences of its actions? Right? So if you want intelligent behavior, you need a system to be able to anticipate what's going to happen in the world and and also predict the consequences of its actions. If you can do this and it can plan a sequence of actions to arrive at a particular objective, and that's what's missing. That's the concept of world models. You don't have that in LLMs. You're not going to get intelligent behavior without that. You're not going to get efficient learning without that. You're not going to get zero-shot, uh, you know, task solving. So the first time you ask a 10-year-old to solve a simple task, they will do it without necessarily being trained. You know, the first 10 hours that a 17-year-old drives a car, within 10 hours, the 17-year-old can drive the car. Uh, we had millions of hours of training data to train autonomous autonomous cars, and we still don't have level five autonomous driving. So it tells you the the basic architecture is not, is not there. Okay.
Embedded assumptions about intelligence. Much of the global AI debate seems to rest on implicit assumptions about how intelligence actually works. You've long argued that intelligence is not primarily about language, but about understanding the physical and social world. What is missing in today's dominant AI models? And what kinds of architectures or learning paradigms are actually required to move closer to real intelligence? If there is one idea about intelligence, human or machine, that you wish every world leader at Davos truly understood, what would it be? Okay, the real world is way more complicated than the world of language. Okay, this is paradoxical because as humans, we think language is sort of the epitome of human intelligence. But it turns out predicting the next word in the text is not that complicated. And you can accumulate a lot of knowledge in an LLM, which is why they need to be so big and why you need to train them on so much data. >> But real intelligence comes from an understanding of the real world. Unfortunately, the real world is messy. Sensory data is high-dimensional, continuous, noisy, and generative architectures do not work with this kind of data. So the type of architecture that we use for LLM generative AI does not apply to the real world. The next revolutionary AI, which is coming fast, is going to be AI systems that understand the real world. Systems that understand high, continuous, noisy data like video, like sensor data. Systems that can build predictive models of how their environment is going to evolve and what their effect on the environment is. Systems that can plan, they can reason at the core level, systems that are controllable and safe so that you give them a task and they accomplish it. Okay. So, we're going to see another AI revolution. We've seen the deep learning revolution, the LLM revolution. Now it's going to be the physical AI revolution, if you want. >> All right.
My next question is about your time at Meta. You spent 12 years leading AI research at Meta during a period of extraordinary acceleration. What do you see as the most important breakthrough that enabled AI's rapid progress over the last decade? And looking ahead, what critical scientific or research breakthroughs still need to happen for AI to live up to its long-term promise rather than plateau or go into an AI winter? Well, there's been an astonishing number of uh innovations that have propelled the field forward, but I'll tell you the biggest factor in progress was not any particular contribution. We could cite, of course, a bunch of them like transformers and stuff like that, but it's not any particular contribution. It's the fact that AI research was open, right? People would do a piece of research, write a paper, post it on arXiv, eventually submit it to a conference or a journal or something, uh, open source their code. And that made the field progress extremely fast because the more people can contribute to something, the faster progress takes place. And what to my, you know, despair, what's been happening the last few years is that increasingly more industry research labs have been closing up. Open AI, Anthropic was never open, in fact, very closed. Google became slightly open and now more closed. FAIR was very open, but now there is kind of a change of mist parity at Meta, which, you know, may change how this, how it operates. And I think it's disastrous because it's going to slow down progress, particularly in the West, particularly in the US. And simultaneously, the more open research labs, industry research labs are in China. The best open source models at the moment come from China. They're really good, and so everybody in the research community is using Chinese models. Uh, okay, my colleagues, my former colleagues at Meta, working on kind of a new version of the successor to Llama, if you want, but, which may turn out to be good. Is it going to be open? Not entirely clear. So I think that's a huge mistake. We're slowing down progress because of that. Okay.
Your new venture, Advanced Machine Intelligence. This is a perfect transition to your next chapter. Uh, you've recently launched this company, Advanced Machine Intelligence, uh, publicly. It's reported you suggest AMI is focusing on building a fundamentally new generation of AI systems based on world models. Systems that learn from video, physical interaction, and spatial data rather than language alone. Can you share more about the problem AMI is trying to solve that today's leading systems cannot, and realistically, how long do you think it'll take to develop the architectures required for robust world models? >> Right. So, Advanced Machine Intelligence, we actually pronounce it AMI. Okay, that means "friends" in French. This is actually the name of the research project that I was kind of driving at at Meta. I was actually an individual contributor at FAIR. I was the manager of nobody. People worked on that project because they wanted to work on it and wanted to work with me, not because I was their boss, which is the best situation in a research environment. So, uh, not top-down, not top-down, bottom-up. That's the way research should, should take place. A lot of people don't understand this, but that's really the way it should work. So, uh, we've had this project for quite a long time at FAIR, Advanced Machine Intelligence, which is the name we gave to this idea of building an AI system that can learn from sensory data, from video, learn world models, state of the world at time t, action that the system imagines taking. Can you predict the state of the world at time t plus one that will result from this action? If you have such a world model, you can plan a sequence of actions to accomplish a task. Okay, this is the blueprint. I wrote a big vision paper, 60 pages. You can read just the beginning, you get an idea of it, or you can listen to a talk I've given on this, which I put online in 2022, where I explained where AI research should go in my opinion, and then we've been sort of building it since then and making a lot of progress. So we have systems now that we can train completely self-supervised on unlabeled videos, and those systems understand video, represent it really well, can predict missing parts in a video, and they also have acquired a certain sense of common sense. If you show them a video where something impossible happens, they tell you this is impossible. Like, you throw a ball in the air, and the ball stops or it disappears, prediction error goes through the roof because the system says, like, no, this is completely incompatible with what I've observed during my training. So, um, so we have the elements of that, and it's based on a non-generative architecture called Jepa, Joint Embedding Predictive Architecture, that makes predictions in a representation space. And there's a trick, it's complicated to train a system that is not generative to basically tell it to extract as much information as possible about the input and represent as much of the input as possible, but also predict in that space. It's crucial. I'm not going to go into why, but I think it's a really crucial aspect, and it's a complete departure from what, you know, most of the industry, certainly, is working on. Um, so that's the plan for AMI, you know, develop this architecture. We already have prototypes that work, but we want to generalize the methodology so that it applies to any modality, any data, any sensor data. So then we can build from data phenomenological models of complex systems that perhaps we control optimally, be it industrial processes of any kind, manufacturing processes, chemical plants, a turbojet engine, a whole airplane, perhaps, you know, chemical reactions, you know, a cell, a living cell. So everything in the world is complicated because it's an emerging collective phenomenon of really complex systems, and we can only build phenomenological models of those things. >> Yes, sir. >> This is the idea of a digital twin that I'm sure you have heard of. Right. So people are sort of trying to kind of accurately model a physical system so you can simulate it. The problem is that if you simulate a system too accurately, you can't predict anything. I could make, I could explain everything that takes place in this room at the moment, right now, in terms of quantum field theory or something like that, right? But that would be completely impractical. It would explain everything that takes place in this room, including all of our thought processes and everything, right? We can simulate everyone's brain. But of course, that's completely impractical. The way we can understand what's taking place right now in this room is by, is through psychology, maybe a little bit of science, you know, things like that, economics, maybe even, but not at the level of quantum field theory or particle physics or atomic physics or molecules or proteins or organelles or cells or organisms. Right? This is much higher level. So the idea that you have to develop an abstract representation of a phenomenon to allow you to make these predictions is absolutely crucial, and generative models don't do that.
Open versus closed AI. You've been one of the strongest advocates for open research and open models. Even as AI power becomes increasingly concentrated among a smaller number of companies and governments, what risk do you see if frontier AI becomes primarily closed, proprietary, and geopolitically siloed? Is openness ultimately a competitive advantage or a public good that must be actively protected? And where, if anywhere, should openness stop? I think AI is fast becoming a platform, and historically, platforms have always become open source. Um, this reminds me of the debates people were having in the '90s about the internet, right? The infrastructure of the internet was, you know, distributed, open, but you had to buy a server from Sun Micros or HP and then run a proprietary operating system on it with proprietary web servers and etc. All of this was completely wiped out. The entire internet runs on Linux, and the entire software stack of the internet is open source, from, you know, low-level protocols to operating systems to web servers to applications on top of it. If it's not open source, it will just not be adopted. I think it's a similar phenomenon that is bound to occur for AI, and I think should be promoted, particularly by countries that are neither China nor the US, because we want AI systems to, particularly LLMs, if we kind of stick with the current paradigm, we want them to become the repository of all human knowledge, and we're not going to be, no private company, as big as it can, can do this by itself. You need access to multilingual data, to cultural data that is local. You need contributions from governments, from, you know, local people to fine-tune the system, and you're not going to get that with, uh, with proprietary systems. So what I've been advocating for a few years is the idea of a consortium where various regions in the world will contribute to training a global open-source LLM that could constitute the repository of all human knowledge. And this is absolutely crucial because the biggest risk of AI, people are talking about, you know, AI taking over the world and killing us all, and we had a debate on this two years ago. That's BS, if you pardon my French. Uh, the most important risk of AI is that in the near future, where our entire digital diet will be mediated by AI systems, if those AI systems come from a handful of proprietary companies on the West Coast of the US or China, we're in big trouble for the health of democracy, cultural diversity, linguistic diversity, value systems. So we need a highly diverse population of AI assistants for the same reason we need diversity in the press, and that can only happen with open source. Yeah.
Next version. Safety control and AI risks. >> Yeah. All right. If you want to, yeah. Yeah. Here, I'll do you one of these. Yeah. Okay. You heard it here. All right. You've pushed back on apocalyptic AI narratives, arguing that they can distract from the more immediate concerns for leaders in this room. What are the real AI risks in the next 5 to 10 years that deserve serious attention? Which of these do you see as most pressing, and which are overrated? Concentration of power among the companies or governments, human misuse of AI systems, economic displacement in terms of jobs, or other systematic risks we're underestimating. Yeah, I think capture and, you know, centralized control of AI is the biggest danger because it will mediate all of our information diet, as I just said, uh, and so you don't want that. And I think people around the world would just refuse that. And so we need to build an open infrastructure that would give an alternative, and that has to be, you know, high-quality, uh, you know, top-performing AI, AI systems, LLMs at first, maybe other types of AI systems going forward. So the other risks, uh, right, you had, you know, human misuse, yeah, that's a problem. But, you know, it's like everything in the world. It could be misused. And, you know, there's going to be sort of countermeasures for this. I'm not, like, overly worried about it. Some people, some of my friends are, but, you know, I think it's just yet another risk, not a particularly existential one. Um, economic displacement. So, I'm not an economist. I'm actually having dinner with two very prominent economists tonight, and I'm just going to parrot them. This is Philip Aghion, Nobel Prize winner, and Eric Brynjolfsson from Stanford. Um, and there's a lot of people in economics. What they're predicting is that AI over time is going to improve productivity by something like 6% per year. Okay, this is not going to be like a hot takeoff or anything like that. 6% per year is actually big. It's nothing to sneeze at. Um, but it's not yet measurable, but that's what they're predicting. Um, it's not going to create major unemployment like mass unemployment. And the reason is because the, what limits the speed at which the technology disseminates in the economy is how fast people can learn to use it. So it's sort of a built-in regulatory mechanism. Eric, okay, you're kind of stupid. Am I saying something stupid here? >> No, you're okay. I'm just parroting you. You got 10 minutes.
Is alignment the right frame? Many policymakers are focused on AI alignment, but alignment to whose values and enforcement by whom? Is alignment ultimately a technical challenge or a political and institutional one? And are we asking too too much of engineers to solve what fundamentally are governance questions? Okay, so the policy alignment is a very interesting one because a lot of people think of it in terms of LLMs, like how do I align my LLM to, you know, not produce ridiculously insulting answers or or, you know, things that are, you know, tasteless. And it's the wrong way to think about it, you know, because AI architectures are going to change a lot. They're going to be different. The type of blueprint that I described earlier, which I call objective-driven AI, are systems that are given an objective, and they can, the only thing they can do is fulfill this objective, and you can make that subject to guardrails, which have to be satisfied at inference time. Okay, so this is very different from the way we coerce or train LLMs to behave properly. We can never be sure that an LLM would behave properly because the data that we train it on is a very small subset of all the prompts that people can fit it. So we can never guarantee the safety or the behavior of an LLM. So if you try to project, if you imagine that future AI systems that have human-like intelligence will be LLMs, which of course is not going to happen, you say, "Oh my god, that's going to be dangerous." It's the wrong approach. So, is it me? I have a lightning round at the very end. I want to get to that.
AI laborer and human agency. AI is already reshaping work, but not always in the ways people expect. Where do you see AI augmenting human intelligence rather than replacing it? And um, where do you think society is underestimating the transitional costs, the transition costs? Are we asking the wrong questions about job loss in your opinion? And what is, what is your advice to all the young people, young people, educators, and workforce leaders in our audience about how best to prepare for an AI-rich future? Okay, so I'm going to answer the second question first. I think clearly technology progress is accelerating, and what that means is that, you know, everyone who is studying right now is going to have to change jobs because technology evolves so quickly. So what students need to learn are fundamentals, things that have a long shelf life, will not be out of fashion in five years or 10 years. Very fundamental thing. I tell students, if you, you know, if you study, if you have the choice between taking a course in, I don't know, mobile app programming or quantum mechanics, take quantum mechanics, even if you're a computer scientist, because the methods that you will learn doing this will allow you to learn to learn, and also you'll have basic techniques you can reuse in all kinds of different contexts, like how would you know in advance that all the underlying mathematics of machine learning basically comes from statistical physics, right? Which is why there are so many physicists who do AI these days. So learn fundamentals, learn to learn, and then be ready to change expertise, to change jobs. So that was the second question, and I forgot the first one. >> Yeah, you know, for time, we're going to move on. So my last question is going to be, what does 2035 look like? Um, what, what is the success and failure of AI? What would you say it looks like in 2035? But before I do that, I have five lightning questions I want you to just answer real quick. What is the most overrated idea in AI right now? >> What's the most underrated research direction? Models. One book or thinker outside AI that most shaped how you think about intelligence. Okay, Frans de Waal, unfortunately died recently. He wrote a book, "Are We Intelligent Enough to Understand How Intelligent Animals Are?" We think of intelligence as related to language. It's not. Animals are really intelligent, and that's the kind of intelligence that we currently cannot reproduce with AI. Read that book. He's busy. Okay. Which leaders, scientific, corporate, like David Rubenstein, he's right over there, or political, do you think will most shape AI's trajectory over the next decade? >> I'm not sure how to answer this. Okay, here is an answer. It will be on. Okay. All right. All right. What do you think is most missing in how Davos covers AI? And I don't mean, I mean everything in Davos. I mean, given how intense Davos is already, like if you add something, you know, we're all going to die during the week or something, like, you know, it's fine. Okay. All right. So, last question. Who's enjoyed Yan so far? Who, who's glad he didn't retire to an island after he left the last job and that he's on the case? >> All right. All right. You should have talked to my wife. Yeah. Yeah. Okay. All right.
Last question. The long view. You have a rare depth of perspective on how AI evolves over decades. If we look ahead 10 to 15 years to roughly 2035, how may our economies, institutions, and even forums like Davos look meaningfully different from today because of AI? What would success look like, and what would failure look like? So, success would look would involve AI systems that understand the physical world, but also perhaps reach something like human-like intelligence. And of course, in certain domains, it's going to be more intelligent than humans because we know computers can do a lot of things better than humans. So that would be success. I imagine that will occur with some non-negligible likelihood within the next 10 years. It's not going to happen next year. It's not going to take two years. Unlike some of my more optimistic colleagues, there's still a lot of work to do. It's not going to be an event like it's so, a lot of people are in their mind that, you know, there's going to be one secret to AGI, whatever they call it, and the next day computers are going to take over the world. This is ridiculous. It never happens this way. There's going to be a bunch of conceptual breakthroughs which are going to be in obscure research papers that nobody is going to pay attention to until five years later when someone demonstrates how powerful they are. Okay, that's what happened with deep learning to some extent. That's what happened with transformers, and also with LLMs. So we're going to, we're going to see this. So, you know, read the papers that the scientific community pays attention to or does not yet pay attention to, because they're going to cause a revolution over the next five years. And so, what is AI going to look like, you know, what is it going to look like five, 10 years from now? We'll have those assistants working, you know, assisting us at all times, perhaps in our smart glasses, at least that's the vision at Meta, or other wearable devices. Those systems are going to be assisting us, amplifying our intelligence, perhaps allowing us to make more rational decisions. And intelligence is the commodity that is most required in the world, right? So the purpose of increasing the total amount of intelligence on the planet, I think, is a very good one. That's intrinsically good. That's going to be under our control. Our relationship with superintelligence systems is going to be the same relationship as a business, academic, or political leader with their staff. Politicians certainly are surrounded by staff of people who are smarter than them, right? Certainly true for professors too, actually. Right. Our purpose is to, yeah, make our students smarter than us. And in business, it's the same thing too. In research, certainly, you know, the best that can happen to you is work with people who are smarter than you. So, in the last minute, let me ask you this. Five years ago, on this stage, we had Mera Wilche hold up a book that was written by ChatGPT-2. And I think it was one of the few times ChatGPT-2 was even mentioned in Davos that year. Five years ago. A lot of people, a lot of things that people were talking about in AI, they predicted what's happening now was 90 years away. The last five years has moved really fast. Maybe not for you. What are the next five years going to look like? Is it going to feel even faster? And how can we all prepare to thrive as a species and a society over this great change? So it looks very different depending on whether you are, you know, in the trenches trying to kind of make science and technology progress, whether it's conceptual breakthroughs that you realize not right away that they really are breakthroughs until you kind of start to make them work and things like that. Uh, but from the outside, from the public, what they see are discontinuous changes, right? So the public saw ChatGPT, which was, you know, GPT-3, whatever, as a discontinuous change. It wasn't. The technology was developed over years before that. A lot of labs had similar systems internally. It's just that it became visible at that time. Before that, the DARPA Grand Challenge, right, that opened the eyes of the public to the possibility of self-driving cars. Ladies and gentlemen, Yann LeCun.