The Exchanges

Q and data size, I just want to ask in terms of the models themselves, there's a core challenge today in terms of two opposing ideologies, which is open versus closed. Jan, you know, you run all things AI at Facebook, or Meta. How do you feel about the open versus closed discussion? I know you've got some very strong opinions. Why does the future have to be open, not closed?

A It's very simple. It's because no outfit as powerful as they may be has a monopoly on good ideas. If you do it in the open, you recruit the entire world's intelligence to contribute to things and having ideas and ideas that you may have thought about, which an outfit with 400 people has no chance of thinking about, or even a large company with 50,000 employees may not want to devote any resources to because they may not think it's useful in the long term or they have more urgency to take care of. So you give it away and then you have tons and tons of people, some of whom are undergraduate students or people, you know, in their parents' basement. So coming up with amazing ideas that you would never have thought about or willing to spend the time to crunch down the seven billion weight LAMA so that it runs on a Mac, on a laptop. I think that's why open source projects succeed, particularly when they concern basic infrastructure.

AI assessment note: “no outfit as powerful as they may be has a monopoly on good ideas.”

Q the investor and the founder side now, really to the research side, and there's no one better than this. I've interviewed many leading AI experts, and they say the value will accrue to the incumbents. Is that right? Will the value accrue to the incumbents? Or do you believe that given what you just said about size not being everything in terms of models, it could be startups as well?

A So the scenario I think will happen, and I'm certainly rooting for, is the scenario I described earlier, where you have some sort of open platform for base LLMs. So base LLMs basically would be seen as a basic infrastructure. TCP, IP, Linux, Apache, completely open. And then there would be an ecosystem of companies building stuff on top of it, which for vertical applications for specific things to specialize those systems for particular application to offer support, to make it customized for enterprise applications for personal things. There'll be like a whole economy around this, which will create jobs by the way, not make them disappear. So this is the scenario that I believe will happen. And the reason I think it will happen is because there is essentially a need to use Essentially millions of contributions for making those systems factual and correct, et cetera. So Wikipedia style. So I think the proprietary approaches will actually fall behind. So that's one point. Okay. The second point is you can ask yourself the question, how is it that the companies that were best positioned to produce something that ChatGPT, namely Google and Meta, didn't? Why is it OpenAI, a small outfit with 400 people? And the answer is, it's not because Google or Meta Did not have the competence or the technology. It's just that they didn't have the pressure to produce completely new products that…

AI assessment note: “proprietary approaches will actually fall behind.”

Q the investor and the founder side now, really to the research side, and there's no one better than this. I've interviewed many leading AI experts, and they say the value will accrue to the incumbents. Is that right? Will the value accrue to the incumbents? Or do you believe that given what you just said about size not being everything in terms of models, it could be startups as well?

A So the scenario I think will happen, and I'm certainly rooting for, is the scenario I described earlier, where you have some sort of open platform for base LLMs. So base LLMs basically would be seen as a basic infrastructure. TCP, IP, Linux, Apache, completely open. And then there would be an ecosystem of companies building stuff on top of it, which for vertical applications for specific things to specialize those systems for particular application to offer support, to make it customized for enterprise applications for personal things. There'll be like a whole economy around this, which will create jobs by the way, not make them disappear. So this is the scenario that I believe will happen. And the reason I think it will happen is because there is essentially a need to use Essentially millions of contributions for making those systems factual and correct, et cetera. So Wikipedia style. So I think the proprietary approaches will actually fall behind. So that's one point. Okay. The second point is you can ask yourself the question, how is it that the companies that were best positioned to produce something that ChatGPT, namely Google and Meta, didn't? Why is it OpenAI, a small outfit with 400 people? And the answer is, it's not because Google or Meta Did not have the competence or the technology. It's just that they didn't have the pressure to produce completely new products that…

AI assessment note: “So I think the proprietary approaches will actually fall behind.”

Q I mean, I totally agree with you, and this kind of led to my next question, which you actually twist about, which comes to the size of, like, Data moats and size of data availability. Is it simply a case that the largest model wins? And how do you think about value in small models as well?

A Yeah, so it's not the case. This is really what Lama has demonstrated and really kind of shown people. So the people behind Lama, Edouard Grave and Guillaume Lample, and then their collaborators, mostly at Fair Paris, actually many of them are in Paris, they've demonstrated that you don't need those models to be very large to work really well. I think it caused a bit of an epiphany for a lot of people. Realizing, oh, you know, you don't need, okay, maybe you need a thousand GPUs, you know, running for 10, you know, a couple weeks to train it, the base system. In fact, this, that number is going down too, because people are kind of figuring out how to do this more efficiently. But once it's pre-trained, you can use it for all kinds of stuff, and you can fine-tune it really easily, uh, and, uh, and then at the end, you can run it on your laptop, right? That's kind of amazing. Or maybe on a, On the, you know, desktop machine with a GPU in it or a couple GPUs. So I think, you know, it sort of opened the minds of people to the fact that there is like enormous opportunities that really weren't thought to be possible before. And I think it's going to make even more progress because if we go towards the design of AI systems, perhaps along the lines of what I described with objectives and planning, I think those systems could actually be even smaller. Uh, to some extent.

AI assessment note: “Yeah, so it's not the case. This is really what Lama has demonstrated”

Q Can I ask you, David asked this as well, how long did it take to get, in terms of like the major breakthroughs, how long did it take you to get to the major breakthroughs that you're at the origin of, when you look back over that time to get to those major breakthroughs?

A Well, so there's a few breakthroughs. So the first one was, uh, in, in the When I was still on undergrad, basically finishing my engineering studies, I figured out that the, the way forward to kind of lift the limitations of the old systems that were abandoned in the sixties was to find learning algorithms that could train multilayer neural nets, essentially. And people had all but abandoned this, uh, type of research, except for a handful of people in Japan. And One guy I heard, I heard about called Jeff Hinton, um, who had published a paper in 1983, so this was just, ah, the year I graduated on something called the Bolson machine, which, ah, was clearly a method to go beyond those, those limitations, and so I had on my side kind of developed a, a method for training multi-layer nets, which was very close to what we now call backpropagation, but not exactly the same. It was closer to what we call target prop, actually, nowadays, And then, you know, published a few papers in French, and eventually met Jeff at a meeting in France in 1985, and we realized we'd been working on the same thing, and we're thinking alike, and, but I was, you know, in the middle of my PhD, and he was a associate professor at Carnegie Mellon, so we, we started, you know, a discussion, and then, you know, visited him at Carnegie Mellon for a summer school he organized, and then I, when I finished my PhD,…

AI assessment note: “when I finished my PhD, I did a postdoc with him... developed what's called convolutional nets”

Q I have to ask, Joshua described kind of the, the hype cycles within, uh, AI and neural nets like deserts when you're not in them. And he asked the question, How did Jan not get discouraged when for a solid decade we were in a desert where no one really cared about neural nets? How did you keep the enthusiasm, bluntly, when, as Joshua said, no one really cared?

A Both Joshua, Jeff, and I had in the back of our minds that those methods would eventually come to the fore and that, you know, we would have to kind of snap people out of their preconceived ideas about, uh, about neural nets. So yes, there was Um, so Yoshua and I were actually working together at, uh, AT&T Bell Labs in the early nineties. And then the interest of the community for those methods started waning around 1995 or so. And there was indeed about 10 years when not only nobody was interested in neural nets, but people were even making fun of it, you know, talking about it in, uh, sort of disparaging terms. Now there, there is, uh, something though, in 1996, I kind of changed the job. I, I stayed in the same company. I was still working at AT&T in the research labs. But I became, uh, a department head, and this was the early days of the internet, and, uh, my group and I started working on something completely different that had nothing to do, or not much to do at least with machine learning, this, uh, image compression. I had this, this idea that, uh, with the internet coming up, we should have a way of scanning existing paper documents and then, you know, put them on the internet so that everybody could, could have access to them. And so I worked on this for five or six years together with Leon Boutou, who's had been a long, long-term collaborator. Joshua was also involv…

AI assessment note: “had in the back of our minds that those methods would eventually come to the fore”

Q Jan, I, I agree and I love this. You work with Meta. My question and David Marcus's question was, how does Meta win then?

A So it's been the case that Meta in the past has open sourced pretty much everybody, everything that it's ever produced, uh, in terms of basic infrastructure, right? So you have, you know, React for, uh, you know, the framework for web and mobile apps. You have PyTorch. PyTorch is not even owned by Meta anymore. The ownership was transferred to the Linux Foundation because it's so essential, um, to the, you know, AI R&D infrastructure nowadays. You know, ChatGPT was developed on PyTorch. Okay. All OpenAI runs on PyTorch. The entire world, in fact, runs on PyTorch, except Google, because they have their own, their own thing, right? But it goes beyond that, right? Meta open sources its hardware server backplane design, so that hardware manufacturers can, can build to its specifications. And pretty much everything, aside from sort of legal issues that are sometimes due to kind of recent laws or, or, or code decisions, pretty much everything is, uh, has been open sourced. It is not because other people can use your technology that you can't exploit it to the same extent, right? Who can use smart NLP systems for, you know, translation or content moderation on Facebook other than Facebook? It doesn't matter if other people have access to the same technology.

AI assessment note: “It is not because other people can use your technology that you can't exploit it”

Q I mean, I totally agree with you, and this kind of led to my next question, which you actually twist about, which comes to the size of, like, Data moats and size of data availability. Is it simply a case that the largest model wins? And how do you think about value in small models as well?

A Yeah, so it's not the case. This is really what Lama has demonstrated and really kind of shown people. So the people behind Lama, Edouard Grave and Guillaume Lample, and then their collaborators, mostly at Fair Paris, actually many of them are in Paris, they've demonstrated that you don't need those models to be very large to work really well. I think it caused a bit of an epiphany for a lot of people. Realizing, oh, you know, you don't need, okay, maybe you need a thousand GPUs, you know, running for 10, you know, a couple weeks to train it, the base system. In fact, this, that number is going down too, because people are kind of figuring out how to do this more efficiently. But once it's pre-trained, you can use it for all kinds of stuff, and you can fine-tune it really easily, uh, and, uh, and then at the end, you can run it on your laptop, right? That's kind of amazing. Or maybe on a, On the, you know, desktop machine with a GPU in it or a couple GPUs. So I think, you know, it sort of opened the minds of people to the fact that there is like enormous opportunities that really weren't thought to be possible before. And I think it's going to make even more progress because if we go towards the design of AI systems, perhaps along the lines of what I described with objectives and planning, I think those systems could actually be even smaller. Uh, to some extent.

AI assessment note: “Yeah, so it's not the case. This is really what Lama has demonstrated”

Q and I've interviewed many kind of leading AI experts, and they say the value will accrue to the incumbents. Startups, they don't have the data, they don't have the models, it'll accrue to the incumbents. Is that right? Will the value accrue to the incumbents, or do you believe that given what you just said about size not being everything in terms of models, it could be startups as well?

A So it depends on which scenario you, you believe in. So the scenario, um, I think will happen, and I'm certainly waiting for is the scenario I described earlier, where you have some sort of open, uh, platform for base LLMs. So base LLMs basically would be seen as a basic infrastructure, uh, like, you know, TCP, IP, Linux, Apache, essentially, um, completely open. And then there would be an ecosystem of companies building stuff on top of it. Which for vertical applications for specific things, right? To specialize those systems for particular application, to offer support, to make it, you know, customized for, for enterprise applications, for personal things. I mean, there's, there'll be like a whole economy around this, which will create jobs, by the way, not make them disappear. So this is the scenario that I believe will happen. And the reason I think it will happen is because there is, uh, essentially a need to use Essentially millions of contributions for, for making those systems, uh, factual and correct and et cetera. So Wikipedia style. So I think the proprietary approaches will actually fall behind. So that's one point. Okay. The second point is you can ask yourself the question, how is it that the companies that were best positioned to produce something that ChatGPT, namely Google and Meta, didn't? Why is it OpenAI? Okay. The small ad sheet was, you know, 400 people. I…

AI assessment note: “there would be an ecosystem of companies building stuff on top of it”

Q Why do you think we love the doom, Jan? You're, you know, I love your approach and mindset, and I agree with it, but why do you think we are kind of magnetized to, like, oh, we're all going to be unemployed in the doom?

A Well, because I think for a number of reasons, so I'm not a, you know, social psychologist or sociologist, but, uh, but clearly, I think we're hardwired to pay attention to things that occur or may occur that could be dangerous to us, because it means that there's something about the world that we don't completely understand, and we do have to pay attention to it and be careful about it. So, for example, take a young, a young child, five-month-old, And show a scenario to the small child of a little car that is sitting on the platform, and then you push the car off the platform, and instead of falling, the, the car appears to float in the air. A five-month-old will barely pay attention to it, but if you show this to a ten-month-old, the ten-month-old will look at it with huge eyes and stare at it for a long time wondering what's going on, because in the meantime, babies around the age of, you know, between, between six and nine months learn about gravity. They learn that objects that are not supported are supposed to fall, and so their mental model is that an object that is not supported should fall, and they see this object that appears to float in the air, And they say like, this can be like, you know, there's something I didn't, I didn't, I don't understand about the world. I need to look at this and investigate. Okay. So we're hardwired for this because that's the way we lea…

AI assessment note: “we're hardwired to pay attention to things that occur or may occur that could be dangerous”

Q Can I ask you a couple of direct questions? I'm just too interested, and we can take them out if needed. Um, what did you say to Jeff when you heard that he was obviously making the moves that he did? Did, I'm sure you had a conversation with him. What did you say to him?

A We haven't spoken yet, actually. We're going to speak to kind of get, you know, each other's opinion on it. Uh, I don't think he knows my, uh, my, my opinion on this because I don't think he follows, you know, what I post on Twitter or whatever, even though he is on Twitter himself. But so I think we have, you know, a discussion to have. I've had this discussion before with Yosha Benjo, uh, but not with Jeff. And, and to me, the fact that he left Google is Not particularly a surprise. The fact that he needs Google to be able to speak his mind, I think is not surprising. So I have a very different deal at Meta, which is that I say whatever I want. Okay. I'm not under the, uh, tight control of, uh, you know, the communications department or, or, or anything. I just, I just say what I think. All right.

AI assessment note: “We haven't spoken yet, actually.”

Q You mentioned correct it. I promise. Last question, then we'll do a quick fire. You mentioned correct it. Elon Musk said, oh, with Tucker Carlson, the trouble with AI is you can't release and then correct. Unlike all prior technological developments, once released, it is too powerful to be able to bring back into the box. It cannot be amended in that way. Is that not true?

A That's not true. That's completely false. It makes an assumption which Elon and perhaps some other people may have, uh, become convinced of by reading, you know, Nick Bockstrom's book, Superintelligence, or, or reading, you know, some of Eliezer Yudkowsky's, uh, writing. So this is predicated on an assumption that is just false, which is, uh, the existence of a hard takeoff. Right? So the fact that the minute you turn on a super intelligent AI system is going to take over the world. And it's going to escape your control and it's going to refine itself to be even more intelligent. And so, you know, and the world will be destroyed. Uh, and that's just ridiculous. It's just completely ridiculous because there is no process in the real world that is exponential for very long. Um, you know, those systems will have to like recruit all the resources in the world. They will have to be given Uh, you know, limitless power agency. Like why would we, we do this? And what's more, they would have to be built so that they have a desire to take over. Like, you know, systems are not going to take over just because they are intelligent. Because again, you know, in, uh, even within the human species, it is not the most intelligent among us that want to dominate others.

AI assessment note: “That's not true. That's completely false. It makes an assumption”

Q If they're non-linguistic, like the majority, I'm sorry for the base questions, but then what are they and are, is that that we don't have able to be ingested by AI models and engines over time?

A Well, so first of all, there is no question that eventually AI systems will understand the world in similar ways that, that humans do, uh, perhaps better ways, uh, but there will not be autoregressive large language models as a type that we're now, uh, talking about. They will be different, uh, for a number of different, different reasons. But, but to answer your question more directly, Anything that has to do with sort of an intuition of the real world requires an experience of the real world or, or a simulated version of it, uh, which, uh, those large language models don't have. They're purely trained from text. So you can, you, there's a number of questions that, about the physical world that they'll be able to answer because there's a template for it in the, or something very similar in the data that they've been trained on. Same for planning. You can ask them to You know, plan a trip or something, and they will adapt a template that they've, they've been trained on. But they don't really have sort of a model of, a mental model of how the world works that allows them to plan complex action sequences or, or use tools or things like that.

AI assessment note: “eventually AI systems will understand the world in similar ways that, that humans do”

Q I'm, I'm super, I'm super naive, Yann. Why does open win against a more controlled, tight knit, well-funded open AI or other large corporate with a big balance sheet and a very rigorous, but streamlined team?

A It's very simple. It's because no outfit As powerful as they may be, has a monopoly on good ideas. So if you do it in the open, you basically recruit the entire world's intelligence to contribute to things and, and having ideas and ideas that you may have, you know, thought about, which, you know, an outfit with 400 people has no chance, uh, thinking about, or even a large company with 50,000 employees may not want to devote any resource, uh, resources to because They may not think it's, uh, useful in the long term or, or they have, you know, more urgency to take care of. So you give it, you give it away. And then you have, you know, with tons and tons of people, some of whom are, you know, undergraduate students or people, you know, you know, in their parents' basement. So coming up with amazing ideas that you would never have thought about or willing to spend the time to crunch down the, you know, seven billion weight llama so that it runs, uh, on a Mac on a laptop. Like, oh, that's, it's pretty amazing. So I, I think that's why, you know, open source, uh, projects succeed, particularly when they concern basic infrastructure. So if you think about it, the, the early days of the internet, there was a battle between Microsoft and Sun Microsystems to provide the basic infrastructure for the internet. Uh, you know, the operating system, the web server, you know, things like that,…

AI assessment note: “It's because no outfit As powerful as they may be, has a monopoly on good ideas.”

Q Jan, help me out here. You run AI at Meta. We heard their cultural data sets. We had national data sets. Just start. Jan, how important is the size of the model first?

A You don't need those models to be very large to, to work really well. And I think it caused a bit of a epiphany for a lot of people realizing, oh, okay, maybe you need a thousand GPUs running for a couple of weeks to train it, the base system. In fact, that number is going down too, because people are figuring out how to do this more efficiently. But once it's pre-trained, you can use it for all kinds of stuff and you can fine tune it really easily. And then at the end, you can run it on your laptop, right? That's kind of amazing. Or maybe on a desktop machine with a GPU in it or a couple of GPUs. So I think it opened the minds of people to the fact that there is like enormous opportunities that really weren't thought to be possible before. And I think it's going to make even more progress because if we go towards the design of AI systems, perhaps along the lines of what I described with objectives and planning, I think those systems could actually be even smaller to some extent.

AI assessment note: “You don't need those models to be very large to, to work really well.”

Q and data size, I just want to ask in terms of the models themselves, there's a core challenge today in terms of two opposing ideologies, which is open versus closed. Jan, you know, you run all things AI at Facebook, or Meta. How do you feel about the open versus closed discussion? I know you've got some very strong opinions. Why does the future have to be open, not closed?

A It's very simple. It's because no outfit as powerful as they may be has a monopoly on good ideas. If you do it in the open, you recruit the entire world's intelligence to contribute to things and having ideas and ideas that you may have thought about, which an outfit with 400 people has no chance of thinking about, or even a large company with 50,000 employees may not want to devote any resources to because they may not think it's useful in the long term or they have more urgency to take care of. So you give it away and then you have tons and tons of people, some of whom are undergraduate students or people, you know, in their parents' basement. So coming up with amazing ideas that you would never have thought about or willing to spend the time to crunch down the seven billion weight LAMA so that it runs on a Mac, on a laptop. I think that's why open source projects succeed, particularly when they concern basic infrastructure.

AI assessment note: “It's because no outfit as powerful as they may be has a monopoly on good ideas.”

Q Now I want to finish with an eye to the future, and I want to finish with Yann LeCun. A lot of people suggest that we should be concerned about AI, its impact on jobs, its impact on society. How do you feel about this, the concern and how we should think about the next few years of AI and its role in society moving forwards?

A People are kind of extrapolating. If we let those systems do whatever, we connect them to internet and they can do whatever they want. They're going to do crazy things and stupid things and perhaps dangerous things, and we're not going to be able to control them, and they're going to escape our control, and they're going to become intelligent just because they're bigger. And that's nonsense. No economist believes this. No economist believes we're going to run out of jobs because no economist believes that we're going to run out of problems to solve or requirement for human creativity and human communication and stuff like that. This is going to create as many jobs as it makes disappear. And those jobs, by the way, are going to be more productive. So overall, technology makes people more productive. In other words, for the same amount of hours worked, you produce more wealth. Okay. But every technological revolution, unless it's accompanied by political changes and social changes, generally profit a small number of people, at least temporarily, right? That happened in the industrial revolution in the late 19th century, where a few people became extremely rich and a lot of people were exploited and then society changed and there were like social programs and income tax and high tax for richer people and stuff like that, which the U S has Backpedal on this, but not Europe. So ther…

AI assessment note: “This is going to create as many jobs as it makes disappear.”

Q Jan, help me out here. You run AI at Meta. We heard their cultural data sets. We had national data sets. Just start. Jan, how important is the size of the model first?

A You don't need those models to be very large to, to work really well. And I think it caused a bit of a epiphany for a lot of people realizing, oh, okay, maybe you need a thousand GPUs running for a couple of weeks to train it, the base system. In fact, that number is going down too, because people are figuring out how to do this more efficiently. But once it's pre-trained, you can use it for all kinds of stuff and you can fine tune it really easily. And then at the end, you can run it on your laptop, right? That's kind of amazing. Or maybe on a desktop machine with a GPU in it or a couple of GPUs. So I think it opened the minds of people to the fact that there is like enormous opportunities that really weren't thought to be possible before. And I think it's going to make even more progress because if we go towards the design of AI systems, perhaps along the lines of what I described with objectives and planning, I think those systems could actually be even smaller to some extent.

AI assessment note: “You don't need those models to be very large to, to work really well.”

Q I'm, I'm super, I'm super naive, Yann. Why does open win against a more controlled, tight knit, well-funded open AI or other large corporate with a big balance sheet and a very rigorous, but streamlined team?

A It's very simple. It's because no outfit As powerful as they may be, has a monopoly on good ideas. So if you do it in the open, you basically recruit the entire world's intelligence to contribute to things and, and having ideas and ideas that you may have, you know, thought about, which, you know, an outfit with 400 people has no chance, uh, thinking about, or even a large company with 50,000 employees may not want to devote any resource, uh, resources to because They may not think it's, uh, useful in the long term or, or they have, you know, more urgency to take care of. So you give it, you give it away. And then you have, you know, with tons and tons of people, some of whom are, you know, undergraduate students or people, you know, you know, in their parents' basement. So coming up with amazing ideas that you would never have thought about or willing to spend the time to crunch down the, you know, seven billion weight llama so that it runs, uh, on a Mac on a laptop. Like, oh, that's, it's pretty amazing. So I, I think that's why, you know, open source, uh, projects succeed, particularly when they concern basic infrastructure. So if you think about it, the, the early days of the internet, there was a battle between Microsoft and Sun Microsystems to provide the basic infrastructure for the internet. Uh, you know, the operating system, the web server, you know, things like that,…

AI assessment note: “It's because no outfit As powerful as they may be, has a monopoly on good ideas.”

Q Why do you think we love the doom, Jan? You're, you know, I love your approach and mindset, and I agree with it, but why do you think we are kind of magnetized to, like, oh, we're all going to be unemployed in the doom?

A Well, because I think for a number of reasons, so I'm not a, you know, social psychologist or sociologist, but, uh, but clearly, I think we're hardwired to pay attention to things that occur or may occur that could be dangerous to us, because it means that there's something about the world that we don't completely understand, and we do have to pay attention to it and be careful about it. So, for example, take a young, a young child, five-month-old, And show a scenario to the small child of a little car that is sitting on the platform, and then you push the car off the platform, and instead of falling, the, the car appears to float in the air. A five-month-old will barely pay attention to it, but if you show this to a ten-month-old, the ten-month-old will look at it with huge eyes and stare at it for a long time wondering what's going on, because in the meantime, babies around the age of, you know, between, between six and nine months learn about gravity. They learn that objects that are not supported are supposed to fall, and so their mental model is that an object that is not supported should fall, and they see this object that appears to float in the air, And they say like, this can be like, you know, there's something I didn't, I didn't, I don't understand about the world. I need to look at this and investigate. Okay. So we're hardwired for this because that's the way we lea…

AI assessment note: “we're hardwired to pay attention to things that occur or may occur that could be dangerous”

Q industrial revolution, Even the introduction of PCs into kind of, you know, working environments. These were multi-decade introductions. Bluntly, what AI feels like in some industries today, we use it at the media company, and it's cutting our employment. Like, the speed of transition is much, much more compressed in this timeline, which will lead to short-term significant high unemployment. Do you concede that, or do you not concede that?

A So this is something I used to be very worried about, that the, the speed of progress of technology was going to leave a certain number of people behind who, you know, cannot be basically retrained fast enough or be, or maybe they are too old to retrain themselves for the new, uh, the new world. I was worried about this. And then I talked to a bunch of economists and they say, oh, you know, not really, because the speed at which a technology disseminate dans l'économie est limité par la façon rapide que les gens peuvent apprendre à l'utiliser Um, so a good person to talk to about this is Eric Brynjolfsson at Stanford, and what he says is that when a new technology is introduced, let's say the, the PC, right, with, you know, a graphical user interface, the mouse, et cetera, right, in the mid-nineties, how long did it take to have a measurable effect on productivity, you know, which is the amount of wealth produced by per hour worked? Yeah. He says, you know, typically it's 1520 years, and the reason is that that's what it takes for people to learn to use that new technology, basically.

AI assessment note: “I was worried about this. And then I talked to a bunch of economists”

Q You mentioned correct it. I promise. Last question, then we'll do a quick fire. You mentioned correct it. Elon Musk said, oh, with Tucker Carlson, the trouble with AI is you can't release and then correct. Unlike all prior technological developments, once released, it is too powerful to be able to bring back into the box. It cannot be amended in that way. Is that not true?

A That's not true. That's completely false. It makes an assumption which Elon and perhaps some other people may have, uh, become convinced of by reading, you know, Nick Bockstrom's book, Superintelligence, or, or reading, you know, some of Eliezer Yudkowsky's, uh, writing. So this is predicated on an assumption that is just false, which is, uh, the existence of a hard takeoff. Right? So the fact that the minute you turn on a super intelligent AI system is going to take over the world. And it's going to escape your control and it's going to refine itself to be even more intelligent. And so, you know, and the world will be destroyed. Uh, and that's just ridiculous. It's just completely ridiculous because there is no process in the real world that is exponential for very long. Um, you know, those systems will have to like recruit all the resources in the world. They will have to be given Uh, you know, limitless power agency. Like why would we, we do this? And what's more, they would have to be built so that they have a desire to take over. Like, you know, systems are not going to take over just because they are intelligent. Because again, you know, in, uh, even within the human species, it is not the most intelligent among us that want to dominate others.

AI assessment note: “That's not true. That's completely false. It makes an assumption which Elon”

Q How would they be even smaller? Sorry, unpack that for me.

A Well, because the, the current models, you, for them to work, you have to train them on gigantic amounts of data. Way more data than any humans has ever been trained on, right? So the, the amount of data Lama is trained on, for example, is something like, uh, 1.4 trillion, uh, tokens. Which is a, you know, it's like a quarter of the internet or something. It's something absolutely enormous. It would take someone reading eight hours a day at normal speed about 22,000 years, um, to read through that, ok? So obviously those systems can accumulate a lot of knowledge from text, but they don't do it the same way humans do it, because we don't need that much time to be that smart and to learn, uh, that much. So obviously we are much more efficient, our brains are much more efficient than those models. At learning things. Like, how is it that a teenager can learn to drive a car in about 20 hours of practice? We still don't have level five self-driving cars. So obviously we're missing something really big. And what we're missing, I think, is abilities for, for AI systems to learn how the world works by observation, mostly. And then this ability to plan so as to satisfy objectives. And then beyond that, the ability to set sub-objectives in those Satisfaction of a bigger one. Okay, that's called hierarchical planning. And we do this. Humans do this. Some animals do this to some extent. Ev…

AI assessment note: “our brains are much more efficient than those models. At learning things.”

Q What are your thoughts on what's missing from those systems? In, in that logical next step, where does that lead you in your thinking?

A So those systems do not have anywhere close to human level intelligence. Okay. Despite what you might think, we are kind of fooled into thinking it because those systems are very fluent with language, but their ability to, to think, to understand how the world works, to plan are very, very limited. And their understanding of the world is very superficial, and the reason for it is that they are strictly trained on language, and language only contains a small proportion of all human knowledge. Most of human knowledge is not linguistic at all, and all of animal knowledge is non-linguistic, and we take it for granted. You know, this is the Moravec paradox, right? All the capabilities and abilities that we take for granted Like, you know, planning a motion or something or very simple things that everyone can, can do. A ten-year-old can, you know, clear up the dinner table and fill up the dishwasher. Any, uh, seventeen-year-old can learn to drive. We still don't have salary cars. We don't have domestic robots.

AI assessment note: “their ability to, to think, to understand how the world works, to plan”

Q Can I, can I ask that? How do we instill values within models where they don't have a desire to dominate?

A Right, so these objectives I was telling you about, so, okay, so let me describe the, the sort of architecture of future AI systems as I see it. We're going to have AI systems that basically are going to plan their actions, and actions can include sequences of words that you tell someone, but they're going to plan the sequence of actions or words so as to optimize a series of objectives that we set them. Okay, so one objective is, uh, Does this answer the question I just asked? Ok. Another objective might be, while you're talking to a thirteen-year-old, make that answer understandable by a thirteen-year-old. Another objective might be, you know, I asked you to answer a question about the world, so be factual. Or it's a question about, you know, yesterday's political event, you know, can you kind of be compatible with everything you've read in the press, uh, this morning? Uh, things like that, right? Uh, so you, you know, you'll have those systems that have, you know, a series of objectives and their output, their answer by construction is going to have to satisfy those objectives. And some of those objectives will be hardwired to make those systems safe. Like, if it's a domestic robot that can, you know, cook, uh, cook dinner and can wield, you know, kitchen knife in its, uh, in its arm, there's gonna be a term in there that says, like, stop moving your arm when there's people …

AI assessment note: “You make them produce answers that by construction have to satisfy objectives”

Q you not also think there's this core business model challenge there, which is, it's the classic innovators dilemma. Like why didn't Google do this? Because it would have killed that absolute cash cow of Google ads. The cost to service a query versus the costs of this is so significantly different. You'd be killing your core cash cow with this, with unknown upside. Versus retaining what is a great business?

A You don't have a choice. I mean, there's no question that, you know, within some time, you know, we could take a, it could take a while, but there is no question that people interact mostly with the digital world using, uh, AI assistance. And, you know, they may run into your, your augmented reality glasses, ok, so, uh, or, or something of that type, like, you know, you know, like in the Spike Jonze movie, Her, that's, uh, that's not, not a bad depiction of what, you know, the, the way things could develop. And so, if you, if you take the assumption, you make the assumption this is going to happen, you, you, you have to build it as quick, as quickly as you can. And it might enableize your, you know, your newsfeed algorithm or, or whatever, or the case of Google, your, your search engine. But you have to do it. You know, it's like, um, I mean, Meta has been known to make those choices, uh, in the past, like the move to, to mobile, for example, um, and the, the move to, uh, you know, short, uh, short form video, for example, you know, which, you know, obviously TikTok has been, uh, very successful at, uh, Meta has entered that, that, that business in kind of a, a big way, despite the fact that the amount of revenue it derives from it is lower than a traditional news feed. Because it's hard to put, you know, put ads in videos basically.

AI assessment note: “You don't have a choice... And it might enableize your... search engine. But you have to do it.”

Q So which regions most need to change their modus operandi when it comes to the practice of scientific research and incentive mechanisms?

A Which region? Uh, oh wow. Uh, pretty much every region, I'm afraid, but for different reasons. Seth. Uh, so let me start with China. So China has a bit of a epidemic of bad science. There's a lot of very smart people in China, a lot of very good researchers, a lot of very good work coming out of China, particularly in AI, particularly in computer vision. Uh, but a lot of absolutely terrible work that has to be retracted a few months later, it's, it's being published. And it's partly because of the incentive mechanisms In the, uh, academic and, and, uh, system, uh, in China. So there's, there's a problem to fix there. I can move to Europe. So in Europe, there are good things. So the education system for like undergraduate education in Europe is great. It's fantastic because it's partly free. So that allows, uh, talented people to go to the schools, even if they're, if they're not rich, right? Sure. Which is not the case in the U.S., for example. At least not to the same extent. That's good for Europe. A lot of, you know, European engineers and scientists are a great, a top base in the world. But then, what are the opportunities for people who want to, you know, go into science and research? And, and there, uh, most of, most European countries actually are, don't have systems that really encourage this and motivate the most talented people and students to go into, into science. A…

AI assessment note: “pretty much every region, I'm afraid, but for different reasons.”

Q That's why I'm here, Jan. I'm happy to provide. Um, I'm gonna do a penultimate one for you. When you think about what you'd most like someone listening to take away, what would it be? When they hear this, what do you want them to take away as the number one thing?

A AI is going to bring a, a new, a new renaissance for, for humanity. A new, a new kind of, a new form of enlightenment, if you want, because AI is going to amplify everybody's intelligence. Right. It's like every one of us will have a staff of people who are smarter than us and know most things about, you know, most things and, and most topics. So it's going to empower every one of us is going to make us more creative because be able to produce, uh, text, art, music, videos without necessarily having all the technical, uh, skills that are currently required for, for doing those things. And, and so exercise or creative juices. So that's, that's the positive side. There are risks, there's no question, but it's not like those risks. Don't believe the people who tell you that those risks are inevitable, or that they will inevitably lead to catastrophe. That's just not true. It's like, you know, place yourself in 1920, like, who would have thought That a mere, 50 years later, you could, you know, cross the Atlantic in a few hours in complete safety, you know, at near the speed of sound. You know, would people seriously want to ban aviation or call for regulation of jet engines before jet engines existed? I mean, that's kind of insane. So I'm not against regulation. There should be regulation of AI products, particularly the ones that In order of making critical decisions for people, …

AI assessment note: “AI is going to bring a, a new, a new renaissance for, for humanity.”

Q questions in terms of depth, breadth, and kind of obvious and non-obvious. So forgive me if some are obvious. I just want to ask, when I hear the historical context there from you, Over many decades. How do you feel today, when we look at what's happening today, are we at a new inflection point in development, or is this merely the continuation of what we've seen for many decades?

A Um, it's a combination of the two. So on the one hand, a lot of what we see today when, when you are kind of down in the trenches of, of research looks at a logical extension. I was not as enthralled by the sort of, uh, recent progress as the, the, the public was because, you know, I've seen this progress happening over the last several years. Now there are things that have been very surprising. The fact that self-supervised learning methods applied to transformer architectures Work amazingly well, and they work, you know, way beyond what we could have expected. The fact that we can do basically train systems to understand language, translate language in multiple languages, and then, you know, continue text. If you, if you train them to do this or answer questions, if you train them to do this works amazingly well to an extent that, you know, people didn't quite expect that, you know, was going to happen by just making them bigger and training them on more data. So that certainly has been surprising for everybody. But that revolution occurred two years ago, right? So whereas the, the wider public, you know, has learned about it through a tragedy that, you know, was made available for us, you know, it's, it's been more continuous. And you see this in, you know, a lot of marking events in, in technological progress or in AI in particular are marked by kind of splashy events that …

AI assessment note: “it's a combination of the two. So on the one hand”

Q How would they be even smaller? Sorry, unpack that for me.

A Well, because the, the current models, you, for them to work, you have to train them on gigantic amounts of data. Way more data than any humans has ever been trained on, right? So the, the amount of data Lama is trained on, for example, is something like, uh, 1.4 trillion, uh, tokens. Which is a, you know, it's like a quarter of the internet or something. It's something absolutely enormous. It would take someone reading eight hours a day at normal speed about 22,000 years, um, to read through that, ok? So obviously those systems can accumulate a lot of knowledge from text, but they don't do it the same way humans do it, because we don't need that much time to be that smart and to learn, uh, that much. So obviously we are much more efficient, our brains are much more efficient than those models. At learning things. Like, how is it that a teenager can learn to drive a car in about 20 hours of practice? We still don't have level five self-driving cars. So obviously we're missing something really big. And what we're missing, I think, is abilities for, for AI systems to learn how the world works by observation, mostly. And then this ability to plan so as to satisfy objectives. And then beyond that, the ability to set sub-objectives in those Satisfaction of a bigger one. Okay, that's called hierarchical planning. And we do this. Humans do this. Some animals do this to some extent. Ev…

AI assessment note: “we're missing, I think, is abilities for, for AI systems to learn how the world works”