[0:00:00] A diffusion pathway is basically a way to reach a particular goal, which you've got from learning, from doing things, and then packaging it. And it's not just technical packaging, it's about guardrails, it's about how to get institutions on board, how to make data available, there's a whole host of things. But think of it as a toolbox or a playbook for doing things. And then this is a global initiative, so we're going to work around the world and create multiple diffusion pathways and then share them with each other.
[0:00:28] So that we can accelerate this thing and Anthropic is part of that. We have Google as part of that. Gates Foundation is there. UNDP is there. The Kenyans. It's a global coalition. Because what we learned from the agriculture experience, we implemented, we worked with Maharashtra on their AgriStack, which is called Mahavistar. And that took us nine months to figure out how to make it work safely at scale. Using the same learning, it was done in Ethiopia.
[0:00:58] which took three months because we had the learning of Maharashtra and then using the same learning at the you know the PM was very keen to see it in animal husbandry so we worked with Amul and we did that in three weeks so you can see the trajectory of time right from nine months to three months to three weeks so what that shows is that if we can do this lived experience and keep improving and package that and take it to the world
[0:01:24] We'll move the implementation of AI to the real world, and that I think is strategically important for the world of AI. Dario, welcome. Dario, great speech in the morning. You mentioned that we are in the end of the, or towards the end of the exponential. This is something that you spoke about in Machines of Loving Grace, that we would have a country of geniuses in a data center.
[0:01:53] The models over the last, I guess, two months have been sort of giving us a sense that we're getting there. But in your podcast with Dwarkesh, you sort of walk that back a bit by saying, you know, we may reach a country of geniuses in a data center, but the impact on society will take a long time. Can you explain what that means? Because we thought we'd do AGI and we would be finished and done with it. Yeah, thank you for having us, Rahul.
[0:02:17] I would say there is this duality between the fundamental capabilities of the technology and the time that it takes for those capabilities to diffuse into the world. We're getting models that are very good at software engineering, that are increasingly good at biomedical innovation. We're not there yet, but we're on this very fast exponential. But when I look across...
[0:02:41] The world and I look across the enterprise of the world. We have many enterprises as customers. Even if we freeze in place what the technology was capable of today, I think the economic impact could be much greater than it is because it just takes time. There are just frictions to adopt things through enterprises. And I think even more so in the developing world.
[0:03:08] Visited Infosys. One of the things I talked about with Nandan was that we're both obviously very interested in making sure that this technology gets to everyone. And Nandan has devoted a big part of his life to many things in that direction and has built out India's digital public infrastructure. And so I think this question of diffusion is very tied to the question of how do we make sure that everyone benefits.
[0:03:36] So when the foundation models came, Nandan, you said that we should do use cases. And it sounds like Dario is saying the same thing, that even if we do the foundation model, we will still need to do use cases. No, I think it's great what the foundations models are doing and the speed of evolution. But what we have learned is that diffusion of technology is a different ballgame.
[0:03:58] And how do you get technology to a billion people? And I think India, we have a little bit of experience with that. 1.4 billion people on Aadhaar, 500 billion people on UPI, 20 billion transactions a month, the world's largest cash transfer system, the largest financial inclusion system, you name it, all that stuff.
[0:04:15] And we learned that diffusion is a technique. It's both an art and a science. It involves institutions. It involves policymaking, negotiations, dealing with incumbents, dealing with newcomers, strategies for execution. So the whole, you know, trust building, so a whole host of things. And I think if all the investments in AI are going to deliver the value to society, not just to individuals, we'll have to look at diffusion pathways to take this.
[0:04:45] And I think India will lead on that. That's why I've always been saying that India should focus on becoming the use case capital of the world. Talia, you had a more recent essay, which you call The Adolescence of Technology, which is a little more somber, spoke a little bit about sort of dimming the enthusiasm of the first essay.
[0:05:07] And I want you to perhaps unpack that a little bit. We all have spoken about the risks of the technology, but particularly in the global south, where we think that AI is going to be hugely beneficial for us.
[0:05:22] We perhaps will have a different calculus on the risk reward ratio. What do you think about that and how would your essay address that? I think that's an insightful comment, which is that I think in the global south, there's an opportunity for AI to accelerate catch-up growth, to solve a bunch of problems that are in the way of catch-up growth. I think AI is a technology that has big risks and big benefits, but in the global south,
[0:05:51] The benefits may be even bigger than they are anywhere else. But at the same time, that doesn't mean, of course, that the risks aren't real. India is the world's largest democracy. We need to think about how democracies handle AI and how we confront other countries that are authoritarian. That's one of the risks I talk about. Another risk I talk about is making sure that AI systems are
[0:06:18] Safe and predictable and autonomously behave in a way that's under our control. And everyone in the world has to worry about that. That affects...
[0:06:26] Everyone in the world. And then, you know, I think of particular relevance to India is, you know, the concerns I raised around economic displacement, right? Where, you know, I think the signature of this technology is going to be that it greatly grows the economic pie for the whole world. And again, you know, huge upside because the opportunity for catch-up growth, like, you know, growth can be very, very fast.
[0:06:51] But, you know, things are going to change and there's some potential for disruption. And, you know, I think what I've been thinking about as I visited India these last few days and the last time I visited is, you know, how can we work together with the companies in India to kind of...
[0:07:06] Drive this growth for everyone to make sure that the existing companies, large and small, continue to prosper along with us and the other makers of AI. And, you know, also on a philanthropic basis, how we make sure that the benefits reach everyone, both in an economic sense and health sense, in kind of other senses. So, you know, I think India, you know, it kind of offers a particularly keen distillation of, you know, especially the benefits, but also the risks.
[0:07:36] Nandan, you've had experience with an adolescent technology with the whole DPI. In the early days of DPI, it was challenging. There were challenges with getting the big vision, which of course now we're looking at and in hindsight, it looks like it was easy, but the early days was difficult. So as the father of a mature technology, would you want to give the father of an adolescent some advice? I don't know, that makes me a grandfather.
[0:08:04] So I think when you talk about diffusion and you have to think of AI, everybody agrees it's like a general purpose technology, like people give the simile to fire or electricity or whatever. It's about starting from the user and how can we improve their lives? How can we take a billion people and help them to learn better? How can we take a billion people and give them better healthcare?
[0:08:33] How can we take a half a billion farmers and improve their earnings? You have to start from there and then figure out how to make it happen. And it's not just technology. Technology is just one piece of the puzzle. It's about institutions. It's about trust building. It's about negotiations. It's about guardrails, which Dario mentioned. It's about working with different stakeholders and making them go towards a common vision.
[0:09:00] Diffusion is difficult. It's not a simple task. So I think, I feel that India will demonstrate this because we have the experience of diffusion at population scale in all the various areas. And obviously diffusion of AI, there are some differences that we need to think about, data, guardrails and so on. But I think we can build a pathway or multiple pathways to that goal. And that...
[0:09:28] We'll show the world because I believe right now in AI, there's a race to the top and a race to the bottom. And the race to the bottom is faster than the race to the top. So I think all of us who have a stake in AI being useful to humanity have to accelerate and redouble our efforts to make the diffusion happen. Otherwise, the consequences are going to be very, very difficult because there's going to be a backlash.
[0:09:59] If the only thing that AI does is create deepfakes or raise the price of your power bill or all the other things that are happening, people are going to respond. I mean, the resentment of the blue collar worker led to the train wreck of globalization. The resentment of the white collar worker is going to lead to the train wreck of AI. So I think we really have to work very hard to show profound, useful cases of AI.
[0:10:27] Thaddeo, you were in India in October and you're back again now. You spend a lot of time actually with the developer community. You clearly were impressed because you've come back so quickly. Could you tell us a little bit about what your experience is with how India is building and using AI perhaps?
[0:10:50] Just go through the stack. I mean, enterprise, small business, startups, then developers. What's different about the way India does it? Yeah, so I would say there's just an excitement here and a technical acumen. And we can even see it in the statistics of usage of Claude. You know, use of Claude for technical...
[0:11:13] Kind of programming and software engineering, mathematical tasks. The fraction is substantially higher here in India than it is in most other places in the world. And, you know, every time I go to speak at one of these, you know, kind of, you know, we'll host these builder or developer events in India, just there's a lot of excitement. You know, I can feel the brimming excitement of like, you know, what is something that we can build. In just the last four months, you know, the use of Claude and Claude code has doubled.
[0:11:41] I'm sure it's the same for the other. I don't say that to promote Claude and Claude Code. It's more a statement about the kind of excitement in India. On the enterprise level, the two of us just announced a partnership just yesterday. So we're really excited to work with.
[0:12:01] All the, you know, all the large enterprises in India, they know much more about the Indian market. They know much more about, you know, distribution. They know how to serve enterprises within India. And, you know, they're much better at that than we are. And, you know, can we plug our technology into what they do and, you know, create something that kind of, you know, that kind of wins for both sides, right? We would like to be able to, you know, jointly win with the companies in India. And then finally, I think there's another element that's almost unique.
[0:12:30] which is that there's an excitement to build but there's an excitement to build for
[0:12:35] Public good and for philanthropic benefits. So, you know, Nandan kindly introduced me to the Extep Foundation, of course, builds digital infrastructure. And we've already started to work on a number of projects really to reach people in rural areas. We're trying to combine Claude with something called Open Agronet, which, you know, helps farmers in rural regions to kind of find better information and, you know, better advice to be more effective.
[0:13:05] And you, of course, have an office. You've declared a managing director. But actually, more importantly, Sonnet 4.6, which dropped yesterday very inconveniently,
[0:13:32] So I couldn't try it. Apparently is doing very well on 10 Indic languages. So there seems to be a bit of a focus in your development as well on India. Can you tell me why that's important? Why language, cultural context is important and what, you know, India can play, what role it can play in that? Yeah. So, you know, you know, language models have always been multilingual, but of course they're better at languages than they're that, you know, that they've been, they've been trained more on.
[0:14:02] And, you know, of course, you know, as I learned when I first came here, India has, you know, a very long tail of regional languages. And, you know, we see this as something related to access, something related to making sure we provide benefits for everyone, right? If we can only speak the most common languages, then there's a long tail we're not reaching, right? The farmers that we mentioned, you know, many of them speak one of the less common regional languages. And so we put in place a push.
[0:14:31] We're collaborating with folks in India to acquire more data for this long tail of Indic languages and Sonnet 4.6 represents an improvement. We're not all the way there yet. We want these models to be as good, even far out in the long tail of these languages as they are at speaking English. And we're making progress towards that. We're not there yet, but we want to keep going.
[0:14:59] And then after you built DPI, I mean, I say after, like, as if you stopped, you still continue to do it. But after you did the bulk of the work, you spend some time and effort actually taking it out to other countries. And I was wondering whether you've thought about that for AI. And, you know, as we, you know, it's been mentioned so many times today that this is the first AI summit in the global south. And so I think perhaps countries of the global south.
[0:15:25] If there's a model, as it were, for doing this would benefit from those ideas. So as you think about it, is it costs? Is it skill? Is it data? Infrastructure? What is it that countries need to think about for this?
[0:15:43] Finally, it's about a lived experience. If you have done it, you can do it better next time. So what we did in the DPI part, the digital public infrastructure, is that after several years of that experience, we worked with global philanthropists, set up something called CoreDevelop, and said, let's take DPI global. And today, we have some version of DPI running in about 40 countries around the world. We recently had a summit in Cape Town.
[0:16:08] AI has to really be impactful, we need something similar. So yesterday we just launched something called 100 Diffusion Pathways by 2030. We like to have targets. Can you explain that? Yeah, so the idea is that a diffusion pathway is basically a way to reach a particular goal, which you got from learning, from doing things, and then packaging it.
[0:16:37] And it's not just technical packaging. It's about guardrails. It's about how do you get institutions on board? How do you make data available? There's a whole host of things. But think of it as a toolbox or a playbook for doing things. And then this is a global initiative. So we're going to work around the world and create multiple diffusion pathways and then share them with each other so that we can accelerate this thing.
[0:17:00] Anthropic is part of that. We have Google as part of that. Gates Foundation is there. UNDP is there. The Kenyans. It's a global coalition. Because what we learned from the agriculture experience, we implemented, we worked with Maharashtra on their AgriStack, which is called Mahavistar. And that took us nine months to figure out how to make it work safely at scale. Using the same learning, it was done in Ethiopia.
[0:17:27] which took three months because we had the learning of Maharashtra and then using the same learning at the you know the PM was very keen to see it in animal husbandry so we worked with Amul and we did that in three weeks so you can see the trajectory of time right from nine months to three months to three weeks so what that shows is that if we can do this lived experience and keep improving and package that and take it to the world we'll move the
[0:17:56] Implementation of AI to the real world. And that, I think, is strategically important for the world of AI. So 100 by 30 is the new... Yeah, 100 diffusion pathways by 2030. And we welcome everyone to join this moment. You have many such catchphrases, but the one that really stood out to me some time ago was India needs AI and AI needs India. We've spoken a bit about the India needs AI, but why does AI need India?
[0:18:26] Because this is where we're going to show it working. You know, I mean, I think because of the history of India's digital journey and thanks to the leadership of Prime Minister Modi, who is the biggest champion of all the work that's going on, we have a political leadership that's committed. We have technologists. We have enough people with the right value system to make this happen. And we have done this before. And therefore...
[0:18:52] India will be where you'll see most of the deployment of AI in a tangible way, where farmers are able to make more money, where children learn better, where healthcare is better, where people talk in their own language, so you have universal access. So this is where you're going to show this. And the world needs this to be shown and the AI companies need this to be shown because they have to show real stuff where this is working at scale for people. So I think, yeah, it's very important.
[0:19:22] Thario, if I can ask you, if you were to rewrite Machines of Loving Grace, which is a beautiful 20,000 word or something essay for India, if you were to think about, you know, what is that utopic vision of what AI could be for India?
[0:19:41] What would it be? I mean, obviously not 20,000 words, but whatever few ideas you could have now. Yeah, I mean, I think a lot of what was there was universal, but there are some things that I would accentuate that are possible when you have such a large population for running this kind of large possibility of experiments, right? You have here a very...
[0:20:05] A very large population for kind of studying and improving human health. There's world-class medical research. So I think I would double down especially on some of the sections about accelerating the cures for diseases. We just had Demis Hassabis, who has basically solved the protein folding problem with AI and kind of shown us all the way.
[0:20:29] And what we need are like 50 improvements like that. And the hope is...
[0:20:38] Working together between the AI developers and the folks who diffuse AI and do the actual medical research, can all of us working together between the AI companies and folks in India really accelerate progress. I would also say accelerating the rate of economic development. There is so much technical potential and technical adeptness in...
[0:21:07] In India, and it actually almost seems like a perfect case study for the idea that AI could really accelerate economic growth because it seems like the base ingredients.
[0:21:22] are kind of all there, and AI could help to tie them together. So, you know, in the developed world, I've wondered, you know, what could AI lead us to, you know, 10% growth rates, which sounds absurd, but, you know, the thing I imagined in the positive scenario of Machines of Love and Grief. But I think here in India, there's a lot of catch-up growth to be done. There's an enormous amount of technical potential and ability. And, you know,
[0:21:45] As I'm seeing, there's this eagerness to adopt AI. So India is one of the few places in the world where I wonder, could there be 20% or 25% growth, which sounds absurd, is unknown anywhere in the world. But as I think about this, it kind of stacks all the factors for a very bullish.
[0:22:05] Very bullish picture of how that growth could happen. So, you know, I think I could imagine India being one of the countries in the world that most embodies this. Certainly the large country that most embodies this. And Nandan, with all of these things, there needs to be some unlock. So if you were to advise governments of the world, governments of India, governments of the global south,
[0:22:30] What should they do to unlock this 20-25% growth that Dario dreams of? I don't know about 25% and all that. If I get 10%, I'll be happy. But I think there are a number of things. Obviously, I think people before have talked about the need to create massive compute. We need to bring in the models. I think the focus has to be on inclusion.
[0:22:55] This AI has to carry everybody. Everybody must feel it. Everybody must benefit from it. And that's why I think the language is very important. We want people to be able to speak to the computer in their language, in their dialect, mixing English, Hindi, Tamil, whatever. That needs to be done. I think that's a big thing. And then I think making agents work for people.
[0:23:17] I think if we can make agents work for people, then it means more inclusion because they can get complex things done because you're hiding all the sophistication behind the agent. So I think there are a lot of things we can do, which we'll see now in the coming years. We already have three or four examples in agriculture, in healthcare, in language, in education, in electricity. We have a very good example of P2P trading. So I think all these are examples of how AI will actually benefit people.
[0:23:47] And then it is, I think as to Dario's point, India is a country which is very positive about technology in general and AI in particular. And we need to take advantage of that and not let them down by giving them truly transformative applications using AI, which you will see in the next two to three years. Nandan, Dario, thank you so much. What a lovely conversation. Thank you.