Hello everybody. Good afternoon, good morning. And I am delighted to be the one chosen to introduce to you this really distinguished group of people that we've got here sitting around the table. Six, I think, of the most brilliant, most consequential people on the planet today. And I don't think that's an overstatement. So these are the winners of the 2025 Queen Elizabeth Prize for Engineering. And it honors the laureates that we see today for their singular impact on today's artificial intelligence technology. Given your pioneering achievements in advanced machine learning and AI and how the innovations that you've helped build are shaping our lives today, I think it's clear to everyone why this is a really rare and exciting opportunity to have you together round the table. For me personally, I'm really excited to hear you reflect on this present moment that we're in, the one that everybody's trying to get ahead of and understand, and your journey, the journey that brought you here today. But also to understand how your work and you as individuals have influenced and impacted one another and the companies and the technologies that you've built. And finally, I'd love to hear from you to look ahead and to help us all see a bit more clearly what is to come, which you are in the best position to do. So, I'm so pleased to have you all with us today and looking forward to this discussion. So, I'm going to start going from the zooming out to the very personal. I want to hear from each of you your personal aha moment in your career that you've had that you think has impacted the work that you've done or was a turning point for you that brought you on this path to why you're sitting here today, whether it was early in your career and your research or much more recently. What was your personal moment of awakening that has impacted the technology? Do we should we start here with you Joshua?
Thank you, yes, with pleasure. I would go to two moments. One when I was a grad student and I was looking for something interesting to research on and I read some of Jeff Hinton's early papers and I thought wow this is so exciting. Maybe there are a few simple principles like the laws of physics that could help us understand human intelligence and help us build intelligent machines. And the second moment I want to talk about is two and a half years ago after ChatGPT came out and I realized uh-oh, what are we doing? What will happen if we build machines that understand language, have goals, and we don't control those goals? What happens if they are smarter than us? What happens if people abuse that power? So that's why I decided to completely shift my research agenda and my career to try to do whatever I could about it.
That's two very diverging things. Very interesting. Bill, tell us about your moment of building the infrastructure that's fueling what we have.
I'll give you two moments as well. So the first was, in the late 90s, I was at Stanford trying to figure out how to overcome what was at the time called the memory wall. The fact that accessing data from memory is far more costly in energy and time than doing arithmetic on it. And it struck me to organize computations into these kernels connected by streams. So you could do a lot of arithmetic without having to do very much memory access. That basically led the way to what became called stream processing, ultimately GPU computing. And we originally built that thinking we could apply GPUs not just for graphics but to general scientific computations. So the second moment was I was having breakfast with my colleague Andrew Ng at Stanford and at the time he was working at Google finding cats on the internet. He was using 16,000 CPUs in this technology called neural networks, which Fay had something to do with those. And he basically convinced me this is a great technology. So I with Brian Kenzau repeated the experiment on 48 GPUs at Nvidia and when I saw the results of that I was absolutely convinced that this is what Nvidia should be doing. We should be building our GPUs to do deep learning because this has huge applications in all sorts of fields beyond finding cats. And that was an aha moment to really start working very hard on specializing the GPUs for deep learning and to make them more effective.
And when was that? What year?
The breakfast was in 2010 and I think we repeated the experiment in 2011.
Okay. Jeff, tell us about your moment.
I guess one very important moment was when in about 1984, I tried using back propagation to learn the next word in a sequence of words. So it was a tiny language model and discovered it would learn interesting features for the meanings of words. So just giving it a string of symbols, just by trying to predict the next word in a string of symbols, it could learn how to convert words into sets of features that captured the meaning of the word and have interactions between those features predict the features of the next word. So that was actually a tiny language model from late 1984 that I think of as a precursor for these big language models. The basic principles were the same. It was just tiny. We had 100 training examples. It took 40 years to get us here though. And the reason it took 40 years was we didn't have the compute and we didn't have the data and we didn't know that at the time. We couldn't understand why we weren't just solving everything with back propagation.
Which takes us cleanly to Jensen. We didn't have the compute for 40 years and here now you are building it. Tell us about your moments of real clarity.
Well, for my career, I was the first generation of chip designers that was able to use higher level representations and design tools to design chips. And that discovery was helpful when I learned about a new way of developing software around the 2010 time frame simultaneously from three different labs. Researchers at the University of Toronto reached out to us at the same time that researchers at NYU reached out to us, as well as at Stanford reached out to us at the same time. And I saw the early indications of what turned out to have been deep learning around the same time, using a framework and a structured design to create software, and that software turned out to have been incredibly effective. And that second observation, using frameworks, higher level representations, structured types like the deep learning networks, I was able to develop software that was very similar to designing chips for me, and the patterns were very similar. And I realized at that time maybe we could develop software and capabilities that scale very nicely as we've scaled chip design over the years. So that was quite a moment for me.
And when do you think was the moment when the chips really started to help scale up today's LLMs that we have today? Because you said 2010, there's still a 15 year gap.
The thing about Nvidia's architecture is once you're able to get something to run well on a GPU because it became parallel, you could get it to run well on multiple GPUs. That same sensibility of scaling the algorithm to run on many processors on one GPU, this is the same logic and the same reasoning that you could do it on multiple GPUs and then now multiple systems and in fact multiple data centers. So once we realized we could do that effectively, then the rest of it is about imagining how far you could extrapolate this capability. How much data do we have? How large can the networks be? How much dimensionality can it capture? What kind of problems can it solve? All of that is really engineering at that point. The observation that deep learning models are so effective is really the spark. The rest of it is really engineering extrapolation.
Fay, tell us about your moment.
Yeah, I also have two moments to share. So around 2006 and 2007, I was transitioning from a graduate student to a young assistant professor and I was among the first generation of machine learning graduate students reading papers from young Joshua and Jeff. And I was really obsessed in trying to solve the problem of visual recognition, which is the ability for machines to see meaning in objects in everyday pictures. And we were struggling with this problem in machine learning called generalizability, which is after learning from a certain number of examples, can we recognize a new example, a new sample? And I tried every single algorithm under the sun from base support vector machines to neural networks. And the missing piece that my student and I realized is that data is missing. If you look at the evolution or development of intelligent animals like humans, we were inundated with data in the early years of development, but our machines were starved with data. So we decided to do something crazy at that time, to create an internet scale data set over the course of three years called ImageNet that included 15 million images hand-curated by people around the world across 22,000 categories. So for me the aha moment at that point is big data drives machine learning, and it's now the limiting factor, the building block of all of the algorithms that we're seeing.
It's part of the scaling law of today's AI.
Yeah, it's part of the scaling law of today's AI. And the second aha moment is, I was the first chief scientist of AI at Google Cloud. Part of the work we do is serving all vertical industries under the sun, from healthcare to financial services, from entertainment to manufacturing, from agriculture to energy. And that was a few years after what we call the ImageNet Alex moment, a couple of years after AlphaGo. And I realized, as the chief scientist at Google, this is a civilizational technology that's going to impact every single human individual as well as sector of business. And if humanity is going to enter an AI era, what is the guiding framework so that we not only innovate but we also bring benevolence through this powerful technology to everybody? And that's when I returned to Stanford as a professor to co-found the Human-Centered AI Institute and propose the human-centered AI framework so that we can keep humanity and human values in the center of this technology.
So developing but also looking at the impact and what's next, which is where the rest of us come in. Yann, do you want to round us out here? What's been your highlight?
Yeah, probably go back a long time. I realized when I was in undergrad, I was fascinated by the question of AI and intelligence more generally, and discovered that people in the 50s and 60s had worked on training machines instead of programming them. I was really fascinated by this idea, probably because I thought I was either too stupid or too lazy to actually build an intelligent machine from scratch, right? So it's better to let itself be trained or self-organized, and that's the way intelligence in life builds itself. It's self-organized. So I thought this concept was really fascinating and I couldn't find anybody when I graduated from engineering. I was doing chip design by the way, wanted to go to grad school. I couldn't find anybody who was working on this but connected with some people who kind of were interested in this and discovered Jeff's papers for example. And he was the person in the world I wanted to meet most in 1983 when I started grad school and we eventually met two years later.
And today you're friends, would you say?
Yes. Oh, we had lunch together in 1985 and we could finish each other's sentences. Basically, he had a paper written in French at a conference where he was a keynote speaker and managed to actually decipher the math. It was kind of like back propagation a little bit to train multi-layer nets. It was known from the 60s that the limitation of machine learning was due to the fact that we could not train machines with multiple layers. So that was really my obsession and it was his obsession too. And so I had a paper that proposed some way of doing it and he managed to read the math. So that's how we hooked up.
And that's what has set you on this path.
Right. So after that, once you can train complex systems like this, you ask yourself the question: how do I build them so they do something useful like recognizing images or things of that type? And at the time, Jeff and I had this debate when I was a postdoc with him in the late 80s. I thought the only machine learning paradigm that was well formulated was supervised learning. You show an image to the machine and you tell it what the answer is. And he said no, no, no, the only way we're going to make progress is through unsupervised learning. And I was kind of dismissing this at the time. And what happened in the mid 2000s when he, Yoshua, and I sort of started getting together and we started the interest of the community in deep learning, we actually made our bet on unsupervised learning or self-reinforcement loop, right? This is not reinforcement. This is basically discovering the structure in data without training the machine to do any particular task, which is by the way the way LLMs are trained. So an LLM is trained to predict the next word but it's not really a task. It's just a way for the system to learn a good representation or capture the structure.
Is there no reward system there? Sorry to get geeky, but is there nothing to say this is correct and therefore keep doing it?
Well, this is correct if you predict the next word correctly, right? From the rewards in reinforcement learning where you say that's good. And so in fact, I'm going to blame it on you, Fei-Fei, it turns out you produced this big data set called ImageNet which was labeled, and so we could use supervised learning to train the systems on, and that turned out to work actually much better than we expected. And so we temporarily abandoned the whole program of working on self-supervised and unsupervised learning because supervised learning was working so well. We figured out a few tricks.
No, you didn't. I didn't either. But it refocused the entire industry and the research community on deep learning, supervised learning, etc. And it took another few years, maybe around 2016, 2017, to tell people this is not going to take us where we want. We need to do self-supervised learning now, and that's what LLMs really are the best example of. But what we're working on now is applying this to other types of data like video, sensor data, which LLMs are really not very good at at all. And that's a new challenge for the next few years.
So that brings us actually to the present moment. And I think you'll all have seen this crest of interest from people who had no idea what AI was before, who had no interest in it, and now everybody's flocking to this. And this has become more than a technical innovation, right? It's a huge business boom, it's become a geopolitical strategy issue. And everybody's trying to get their hands around what this is, or their heads around it. Jensen, I'll come to you here first. I want you all to reflect on this moment now. Here, Nvidia in particular is basically in the news every day, every hour, every week. And you have become the most valuable company in the world. So there's something there that people want. Tell us, are you worried that we are getting to the point where people don't quite understand and we're all getting ahead of ourselves and there's going to be a reckoning, that there's a bubble that's going to burst and then it will right itself? And if not, what is the biggest misconception about demand coming from AI that is different from the dot-com era or that people don't understand?
During the dot-com era, during the bubble, the vast majority of the fiber deployed were dark, meaning the industry deployed a lot more fiber than it needed. Today, almost every GPU you could find is lit up and used. And so the reason why I think it's important to take a step back and understand what AI is. For a lot of people, AI is ChatGPT and it's image generation, and that's all true. That's one of the applications of it. And AI has advanced tremendously in the last several years. The ability to not just memorize and generalize, but to reason and effectively think and ground itself through research. It's able to produce answers and do things that are much more valuable now. It's much more effective. And the number of companies that are able to build businesses that are helpful to other businesses. For example, a software programming company, an AI software company that we use called Cursor, they're very profitable and we use their software tremendously and it's incredibly useful. Or a company like OpenEvidence who are serving the healthcare industry, doing very well, producing really good results. So the AI capability has grown so much, and as a result, we're seeing these two exponentials that are happening at the same time. On the one hand, the amount of computation necessary to produce an answer has grown tremendously. On the other hand, the amount of usage of these AI models is growing also exponentially. These two exponentials are causing a lot of demand on compute. Now when you take a step back, you ask yourself fundamentally what's different between AI today and the software industry of the past. Well, software in the past was pre-compiled and the amount of computation necessary for the software is not very high. But in order for AI to be effective, it has to be contextually aware. It can only produce the intelligence at the moment. You can't produce it in advance and retrieve it. That's called content. AI intelligence has to be produced and generated in real time. And so, as a result, we now have an industry where the computation necessary to produce something that's really valuable in high demand is quite substantial. We have created an industry that requires factories. That's why I remind ourselves that AI needs factories to produce these tokens, to produce the intelligence. And this is a once-in-a-lifetime event. It's never happened before where the computer is actually part of a factory. And so we need hundreds of billions of dollars of these factories in order to serve the trillions of dollars of industries that sit on top of intelligence. You come back and take a look at software in the past. Software in the past are software tools. They're used by people. For the first time, AI is intelligence that augments people. And so it addresses labor. It addresses work. It does work.
So you're saying no, this is not a bubble.
