UPI was an open network for payments, and the open architecture led to massive growth and became the world's largest payment system. A lot of those principles are embedded in Beckon. Open networks allow many actors and innovators to build applications on the edge using AI. We keep talking about agents, but the real power of agents is removing complexity for the user. If a user is a farmer or somebody producing a little bit of electricity, if they can easily transact with somebody else through an agent in their own language, suddenly this is inclusion at massive scale. I see AI agents on an open network as the fundamental construct for massive diffusion of technology.
And also the importance, as you mentioned, of doing that in languages and local...
Oh totally. You talked about what you're doing at ISC. There are many initiatives in India driving to make language completely accessible. Language is not just a pure language; the way Indians speak mixes English, Hindi, and Tamil in one sentence. How do you deal with that? How do you recognize that? All that is getting addressed. There are many initiatives: Voice AI, Bhashini of the government, AI for Bharat, your project, the Google project. There's lots of stuff, but fundamentally, language as a barrier will go away. If you combine language so a person talks to the agent in their own language, and then the agent does some transaction hiding all the complexity behind it, that's the holy grail. We can get everybody on the system, and that's how AI will get diffused.
Right. Speaking of farmers and agriculture, let me come to you. The World Bank recently launched the AgriConnect initiative. Describe that a little bit. It makes what smallholder farmers do much more efficient and scalable. I'm curious, what has that work taught you so far about the type of global standards needed to scale local solutions?
If you look at AgriConnect in Uttar Pradesh for now, it is a very farmer-oriented approach to provide coherent and consistent services at the same time with an open stack, open network. If you think about the previous days of computer innovation, mobile innovation, now we are seeing AI innovation. I would interpret this evolution from the supplier-oriented service environment to the customer/user-oriented environment. In that sense, some open standard and open network is a crucial part to make sure user-centric service. It is very efficient and affordable solutions in the AI era to fully provide quality of service to the user. In that sense, this AgriConnect project is really important, but it is not only for agriculture. We are really looking forward to expanding to another sector like healthcare and education, so it can be a very universal network in the future.
That's pretty powerful. Speaking of healthcare, you mentioned you're taking this to healthcare. Kiran, you've been an incredible advocate and innovator when it comes to thinking about medicine as a whole. You've talked about this idea that we need to move beyond the industry of medicine. Tell me, say more about what you have in mind about what we need to move to, and in particular how you can connect what's going on with AI and data sets with fundamentally transforming medicine.
I have to answer this in two parts. The first part is how do we leverage what Nandan referred to as the digital stack to a health stack. That is the first big opportunity we have. I think India is a country that can uniquely create a global reference model when it comes to the use of AI in the kind of health data we're collecting. India is beginning to collect a lot of health data in its health stack: phenotypic, genomic, demographic, radiological data, and of course treatment and treatment outcome data. When you start collecting this data, the whole objective again is a holy grail: universal healthcare delivery at scale in a sustainable way, and how do we reduce the disease burden and increase lifespan. These are big challenges and a very complex set of solutions, but this is a starting point where you get this huge digital stack of health data. Because India has this open-source and consent-based secure data sharing model already established in UPI, I think we should quickly apply this to healthcare. When you do that, you will start risk profiling your population at a demographic level, which is very exciting.
And at scale. If you can integrate insurance into that, that would be even more powerful. That can only be done by AI. AI has the opportunity to risk profile very fast to try and find interesting insurance models to see how we can marry the risk profile with the insurance instrument. Not easy, but it's a good challenge because AI can be given a lot of exclusion/inclusion criteria which it can adopt. I'm very excited with what AI can do for digital health and the whole universal healthcare delivery model.
Additionally, India has this unique model of ASHA workers. If ASHA workers can be empowered with AI, that is even more powerful. Deploying AI for the common people is very important to both Nandan and Sam's point of view. Now coming to my excitement about your second question: what is it that I'm looking beyond this? I'm looking at advancing medicine using AI. Biology on its own was limited because it didn't have the kind of power of technology to get deeper insights. AI, like AlphaFold and AlphaGenome, is going to give it immeasurable opportunities to understand biology. To me, biological intelligence is just amazing. If you combine it with artificial intelligence and bring that convergence, I think we're in for huge transformation. When I look at biological intelligence, when I just look at cell biology, how cells signal, how cells create circuits, how cells regulate, how cells connect and disconnect. The human body, living systems, have distributed data centers, and these data centers are connecting and disconnecting with sips of energy, not gigawatts of energy. They are actually translating that into instant information and decision making. If we can learn that and apply it to AI, I think it's going to be transformational. I am really looking forward to reprogramming cells. That's the holy grail: how do you convert a cancer cell into a non-malignant cell? How do you look at regenerative science? How do you look at lifespan? Your biggest question today is how do we shift from hospital-centric care to primary and community care. That can happen with AI, with predictive and preventive medicine. I think I've said enough.
Well, it would actually take us, in effect, Kiran, from kind of treating diseases to preventing diseases. You and I were talking earlier, you and I and Demis were talking about this idea of someday we should be able to try to build virtual cells, models of virtual cells, and be able to do a kind of cell-based biology basically.
That's one of the exciting things.
Yeah. Well, we'll come back to that. But I want to come to you, Sal. As you think about what you've been doing, what role does DPI and open networks play in developing and scaling the kinds of solutions to some of society's most pressing problems? You've been thinking about this for a very long time, from way back when you set up the AI institute, way before most people were thinking about these things. I want to hear what your perspectives and experiences have been.
Great. Thanks, James. Good morning. Just so we're all clear, there's a lot of intellectual horsepower on the stage, and it's all on this side of me. I'm basically here for my good looks, so just so we manage expectations. When I set up Wadhwani AI back in 2018 and the Prime Minister was good enough to come and inaugurate that, we had a huge benefit: over the last 20 years, the government of India has developed a set of DPI, digital public infrastructure, that is broad and deep. These DPI are basically digital building blocks that connect policy, program implementation, public service workers, and citizens in the country. These digital public, this DPI, provides two key, very practical, down-to-earth functions and benefits. Number one, they provide data and data pipelines. For AI, you couldn't build AI for the social sector without the kind of data and data pipelines they provide. Secondly, this DPI provides distribution channels so that once your inference models are ready, these AI platforms, again developed and managed by government, provide a distribution channel to get our AI models out at scale. Without these, trust me, the usage of any model in the public sector, in the social sector, would cost incredibly more and wouldn't scale anywhere near what we see. Two quick examples, one in healthcare, one in education. One of the national health priorities for the government of India for the last several years has been the elimination of tuberculosis, TB. It's the largest infectious disease killer in the world, killing close to 2 million people a year. It's the largest infectious disease killer in India, killing close to half a million people a year. For each person that unfortunately dies, 20 other survivors live miserable lives, impacting their ability to earn a living. The government asked us to come in and see what we could do. We identified with the government the three or four key pain points in the patient's journey. First one is diagnosis. Diagnosing TB in economically vulnerable communities isn't easy. X-ray machines, sputum analysis, etc., are challenging, expensive, time-consuming, and tedious. That's challenge number one. Secondly, if you do sputum analysis, these have to go to 64 government labs around the country. There's throughput time, and by the time the patient gets the results back, you've lost some time in initiating treatment. Finally, there's a huge problem that a subset of TB patients stop taking their medication because it's a very toxic regimen with a toxic effect on the body. Once you stop taking them, you develop drug-resistant TB, the mortality rate goes up dramatically to 50%, and you infect a lot more people. Fortunately, the government has a DPI called Nikshay. It's a very large data platform, a patient management system that has data on all detected TB cases in the country. The government gave us access to that database. We developed a range of models to address these challenges. For diagnosis, we've come up with a way of diagnosing TB from the sound of a cough into a smartphone. It's instant, quick, shows you the risk, and government workers can focus on those patients. In the one year or so this program has been rolled out, detection of TB patients has gone up by 25% nationally. You may think that's bad news because there are more TB cases, but they were there. Now we can make sure they get treated. On the labs and turnaround time, we've come up with an AI way to automate a lot of this testing. Now you get the results the same day, and the doctor finds out instantaneously to start treatment. On the issue of patients who develop drug-resistant TB, we've come up with algorithms that predict which TB patients are likely to fall off their medication regimen. The 2,000 or so TB case workers in the country, a small number for the four or five million TB patients we have, can now target their time and bandwidth on that subset of patients that really needs help. All of that was enabled by this DPI called Nikshay, this database which enables all of this. One other quick example in the education space: in the global south, including India, there's a very high dropout rate of students in early grades, grades one to five. We got a call from a large state government in India about a year back saying, 'Can you help?' We took a look, and it turns out the single key reason for this very high dropout rate is the inability of these young children to read proficiently in their mother tongue. If you can't read properly, it affects how you do in all your subjects. You start failing, get frustrated, your parents say, 'Come back home, work on the farm, work in the kitchen,' and that affects the rest of their life. We've come up with a system to diagnose within 20 seconds for each child by just speaking into a phone. In 20 seconds, we figure out exactly where they are struggling, what words, phrases, sentences, and what will help them get on the right track. This is being done at the cost of five paise per student. Cost is another very big part of scaling that doesn't get discussed too much, but cost is very important. While this was in pilot, the state was so impressed they made it mandatory for all three million kids of that age in school. Three or four other states, including Rajasthan, just made it mandatory for all 8 million kids over there. All of this again enabled by DPI. Rajasthan has a state-level DPI called Rakshak. Our system, our model, sits on top of that system. It reaches 400,000 schools, 8 million students, and it's spreading. Now the government of India, by the end of next year, wants to make it standard across the country. All 75 million children of that age group will get their reading improved and strengthened through the systems we have. Bottom line: all enabled by DPI.
Those are incredibly powerful examples. The case of TB is actually one that's super important because something like 40% of people in the world go undiagnosed with TB, and most of them are in the global south. It also brings me back to a question of scale. In what you're doing, you mentioned a few countries, but I think your goal is to get with some of these education and health solutions to like 25 countries or more. How are you thinking about taking that to multi-country scale? What are some of the ideas you have about how you do that?
Beyond agriculture, we are piloting and have a lot of good projects in many countries. For example, in Kenya, we have some pretty good progress in math education with very simple AI math tutoring solutions that can achieve the same academic goal within six weeks, which usually takes more than a year. That's one example. In Nigeria, not only TB, some very small handheld ultrasound devices can scan pregnant women and easily diagnose some problems for the baby, drastically reducing birth and baby death rates. That's another example. How can we scale this up? Another good example is we are expanding the current India model to other three African countries, and in Brazil, we added one more in the Philippines. One of the ways is how can we really find some very standardized and scalable model. This is not easy, but from one concrete example like in the India case for AgriConnect, we are figuring out what would be the best way and lighter model we can quickly replicate to other countries. This is our role. The World Bank is trying very hard to figure out what that means and how they can really replicate this model to other countries and what would be the really critical component to be replicated. We are just keep working on what would be the best simple model is. In that sense, I'm using this analogy as a sommelier: we are not innovation creators. There's a bunch of really good wine producers, but our customers are not really aware of which wine is really fit for their taste. As a sommelier, the World Bank is trying very hard to understand the wines and then find some better to recommend the better wines for our customers' taste.
Yeah, I think on this question of scale, those are great...
You need any help on quality control with these wines.
Yeah. But I think on this question of scale, Nandan, I've heard you say that we won't get to true population scale unless we actually scale things like inference in AI and how we do that at massive scale. I'm just curious for you to expand on that a bit more, but also what lessons and implications might even have for people like us who are building frontier models. Say more about why the inference part of this really matters.
I think broadly speaking, especially in the global south, the cost of AI inference has to drop dramatically. If you're serving a customer with one query and that costs 500 rupees or something, it's not going to work. Inference has to be cheap. I think you'll do that because there's a lot of focus today on the training side, getting bigger and bigger models and launching them. But as that sort of stabilizes, I think the focus will shift on the inference side to make inference cheap. I'll give you a very tangible example of open networks giving AgriConnect. Yesterday I was talking to Demis, and Demis was saying that Google is improving its weather models. They're making it better, more efficient, more predictive, more granular area by area. Now if you had an open network for agri like AgriConnect, which suppose it has millions of farmers, then all we need to do is just plug in the latest weather model of Google into that open network, and suddenly 10 million farmers have access to the latest weather data. That's a good example of why this open network thing is important because it allows you to plug new models, new sources of capability, new ideas. To give you another example of how it reduces complexity, there's a very interesting demo here of energy trading. We never thought of energy as something that you traded because you bought it from the utility. But today, with millions of people producing energy, they're able to. Somebody has rooftop solar, there's some extra energy, they can sell it to somebody else. But how does a farmer in UP learn how to sell energy? It's a whole new concept. It's only by an agentic commerce interface that is simple. I think low-cost inference combined with agents that hide complexity is the key to massive diffusion.
Yeah. In fact, I like the example you brought up on weather because in some ways, thanks to the forward thinking of the Indian government and the Ministry of Agriculture, they've set up that infrastructure. Last year they used one of our models, NeuralGCM, which predicts monsoons, and were actually able to deliver to something like 38 million Indian farmers predictions about the monsoons. But that only worked because the Indian government actually set up that kind of infrastructure where you could plug in these models. Kiran, I want to come back to you because you raise some more foundational questions here about the future of biology and health overall. I've heard you say that AI doesn't replace biology, that biology is much more fundamental and foundational. Say more about that and what can AI learn from biology and vice versa, and what do you imagine needs to happen to fully take advantage of it?
First and foremost, biology works through distributed data centers. When it wants to build intelligence, retrieve memory, and inference data, it does so with sips of energy, not with gigawatts of power that our data centers use. So we could learn something from that. More fundamental to that is that biology also has generational learning. If you think about how our DNA stores generational memory, that's fascinating. How does the Arctic tern fly out of its nest for the first time, travel 70,000 kilometers to the Antarctic and back? It has navigational intelligence embedded in its DNA. How does that work? I think we have to learn a lot from biology and use AI to learn that biology, because without AI you cannot have any insights into biology. I just feel that the future is going to be about the convergence of biological intelligence and AI, and that is going to be a very powerful transformative process because biology has a lot to teach AI in terms of how to do it with less energy, how to do it rapidly, and how to multiplex multimodal data very rapidly. That is something which is very exciting. To go back to what you've just been discussing with Nandan and others, what makes it very exciting right now is the volumes of data you can contribute on open networks. I work with a lot of organizations around the world in my field. There's a huge reluctance to share data; there's a lot of wariness about IP being fragmented. Except for India, there's a lot of resistance to share data. When you don't share data, you're going to silo it. India has this unique opportunity because of its open networks and the digital public infrastructure to share volumes of very important data, like Nandan just illustrated in terms of the environmental data and climate data, and farmers taking huge advantage of that. That is what we have to really focus on because India is uniquely positioned in terms of its open networks. If we can actually keep generating data and then make the cost of inference low because of the economies of scale, then you have a winning formula.
I'm being told that we're going to have to wrap this up. But before we do that, I want to just see if we can do a quick lightning round. This summit has been extraordinary. The example that India is setting for the world is extraordinary. If you could each say one thing you'd like to see happen in the next 12 months, particularly with this idea of open networks and change at population scale, what would that be? I don't know who wants to go first.
I think I'd like to see massive diffusion where all these applications that are just rolling out on open networks reach millions of farmers, students, patients, and actually show the world that AI is a force of good. I think we have an obligation to show that; otherwise, the pushback will be huge.
Yeah, I definitely want to see a sustainable standard of care, high-quality, universal healthcare coming out of this AI effort and the health stack.
That's preventative presumably.
Absolutely. Diagnostic, preventative, predictive, and precision, because you can't do away with treatment. But how do you basically stage it up front?
Yesterday when the Prime Minister spoke at Bharat Mandapam, you know how he has been saying for years, 'Make in India for the world.' He said in the age of AI, 'Let's develop in India and let's deliver to the world.' In our case, just one little example at Wadhwani AI: I've given you two or three examples of what we've done, but we've developed over 25 AI platforms in India in education, healthcare, agriculture, which are scaling up. What's interesting is over the last year, we've had an incredible amount of incoming interest from governments throughout the global south in Africa, Asia, and so on, who are hungry for these solutions and they're looking to India to provide these. In fact, when PM Modi launched our institute back in 2018, he was saying the US is so far ahead, China is so far ahead. I said, 'Mr. Prime Minister, we can set the example in India for how AI can be used for societal transformation. No one else is doing that. We are showing how it can be done.'
For the next 12 months, I really want to work more to disseminate the really good use cases to our client countries and people. One of the reasons is that some big challenge for people in the developing world is they do not clearly know what they can do with AI, even though it can provide really much affordable and easiest way to expand their capability and productivity and intelligence in a very easy way compared to the old days. Once again, they can get to know this is a real important opportunity for them, and then I believe they will really find some really good way to fully utilize this in a very affordable way.
No, thank you. At least what I'm taking away is that it's not just the example that India is setting for India and the world, but also quite frankly the example that each of you is setting. All of you in your work and your organizations, your teams, through your initiatives, but also your philanthropy and quite frankly insight, have done a lot to show what leaders can do. I appreciate the examples that you're setting and the example that India is setting. Please join me in thanking my panelists here.
Thank you. And I think with that we'll draw to a close. Thank you.
Just request you all to be seated. 30 seconds more. First of all, could we just please have another round of applause for our esteemed panelists. Very insightful. Thank you very much for coming here. The true benefits of AI, the discussion shows, can only be realized when we build for everyone using open networks. Very insightful conversation. Thank you, James, for moderating it. To further help drive population-scale impact, we invite changemakers and researchers to apply for the two Google.org impact challenges. One is in AI for science, one is for government innovation. There's a QR code for you to learn more. I encourage you all to visit us at booths three and four in Hall 5 to see firsthand how Google AI is delivering real-world impact. Finally, I just request all the panelists to please join center stage for a photograph. Thank you everyone.
