Transcription
I thought today I'll, uh, briefly talk about, do a reality check on AI. Because I think, you know, the hype levels are unprecedented, and, uh, you know, hype and technology have always gone together. When the cloud came out, they said this was the greatest thing since sliced bread. Five years back, when they, the crypto crypto bros told us that they'll solve world peace and world hunger with crypto, and that I don't know where that went. And now, of course, AI is going to change everything for everybody. So I think the hype cycle is, is reached unprecedented levels. But the reality is that we are facing challenges to build AI at scale and make it, make it work for everyone. And I think the, the there are many public, uh, information on how products are getting delayed; it's taking much more time; it's much more complicated; internal politics plays a part. So all the things that we know about institutions, individuals, egos, also applies to the world of AI, and therefore to actually make this stuff work at scale and useful to everyone is turning out to be far more difficult than we thought.
Now there are the usual issues: How do you make the user experience seamless? What are the infrastructure we need? How do you govern these things? How do you manage the complexity? These are things that we have dealt with in every field, and we are dealing with them at AI at scale. Now one of the key differences between previous tech revolutions or advances has been, for the first time, we intend to place trust in non-human intelligence for decision-making. Now we didn't do that earlier because earlier technology was deterministic, predictable. Now we are essentially expecting the machine to make decisions, and there's a huge leap of confidence, a huge leap of faith in the ability of technology to take us forward. We also know that we are far more forgiving of human error but much less forgiving of machine error. To give an example: Few hundred thousand people die on the roads due to car accidents, and we take that as a given, but if one person is killed by an autonomous car, the provider of that has to go back to the drawing board for two years. In other words, because the bar we have of expectations of machines is at a different level than of human beings, this becomes actually very difficult when you try to adopt at scale. So adopting AI at scale is hard work and will continue to be so.
Now it's even harder in the enterprise. In the consumer world, you can adopt AI at scale because somebody can land, land a, you know, launch a chat job chatbot which everybody uses, and once in a while if the chatbot hallucinates or makes a mistake, you're willing to live with it. You're willing to live with the fact that the consumer chatbot has occasional mistakes. You know, after all, you're doing is editing your article, getting a menu planning, or whatever. But enterprises have to make sure that they don't give wrong answers, because enterprises are putting their brand behind an offering, and if they provide AI at scale and that AI at scale even has one or two percent error in the way it gives answers, then that affects the brand itself. And that's one of the reasons why it's taking so long because we don't have the guard rails to ensure that we have absolutely no machine error in these kind of things. At the same time, the expectations are huge. Every board today saying, "I want this done by tomorrow morning," they're calling up the CIO, the CI calling up the vendors, you know, you have all that stuff going on, and we really need to have, so it's an unrealistic expectation. So in general, enterprise AI is much harder and will take a long time.
But the most difficult actually is implementing AI in the public sector, because the public sector has structural consents; it has ministries; it has departments; everybody is territorial. So data is not shared, and if data is the lifeblood of AI, we have to find a way to bring all AI together, irrespective of which part of the government it comes from. So actually, the public sector is the most difficult, and also because public trust is so important, ethical concerns are so important, and you don't have, don't want to have bias, and the person, the bureaucrat signing off on the AI wants a clear commitment there'll be no blowback on him or her or her because the AI does something wrong. So it's really even more difficult in the public sector. So essentially my point is that a lot to be done if you really want to make all this AI stuff work, and all transitions we know are painful, but this is not a new animal; this has happened before in every industry. Every industry when it comes, it goes through these kind of challenges, whether it's automobiles, PCs, internet, cloud, so all of them go through the, the difference this time are a couple of things: One is the hype is at a different level, so everybody's talking about it. Second, everybody thinks that he'll get some, you know, uh, talking about AI or delivering AI or all that is going to give some pixie, you know, pixie dust or something, so everybody's talking about AI, so the whole thing is much more hyperventilating. But fundamentally, the challenges of implementation are the same like anywhere else, and we have to make sure habits change; we have to change the workflow in enterprises or in governments so that AI is part of it; we have to do a lot of upgradation, but the fundamentals still matter. So I just want to clarify one reality check that AI doesn't mean it's going to be easier to do; it's going to take the same effort, if not more effort, and because you're trusting the machine to give decisions, more responsibility to make sure that it works, and this is something which is much more complicated than we think.
Now we know that adoption cycles are reducing globally, uh, you know, electricity took 60 years; it was, you know, 1880s and all that; we had the current wars, but it's only 1920 and it was used properly. Each cycle of technology takes less years, and in general, it, it has taken so many more time to get adopted in India. Interestingly, this time around, while we expect AI adoption also to take 10, 15 years, our belief is that in India it can happen much faster, so because of India's situation today and the technological sophistication that we have been able to accomplish in the last 15 years or so, it is going to be much faster. Therefore, the gap between the global developments and AI in India is going to be very short, and this is because of the transformation that we have done in India. If you look at what's been happening in India in the last, uh, you know, 10, 15 years when smartphone began, smartphones began as, if you remember, Steve Jobs launched the iPhone in 2007, Google launched Android in 2008. So the initial use of phones in India was communication, entertainment, and dominated by western companies. So you had, you had, Google, YouTube, or Facebook, or you had, you know, WhatsApp or whatever. So that was the first era. But around 2015, 16, with the rise of India's digital infrastructure, arrival of Aadhaar, UPI, and so on, India's thing became more sophisticated and led to payments and transactions becoming a bigger part of the internet world, which, by the way, much bigger than anywhere else because of the unique infrastructure that we have, and that led to the rise. And the second thing that happened was the, the balance moved from, uh, just the global tech companies to actually homegrown tech companies financed by venture capital. So you have a company like, uh, Misho, which has over 100, 200 million users, many of them first-time buyers, many of them first-time sellers, selling commerce in every nook and corner of the country. Or you have PhonePe, which belongs predominantly to Walmart, which has 48% of payment transactions in India. These companies did not exist; they came out of this digital infrastructure we had. Or PhysicsWallah, which is a teacher, a talented teacher, Alakh Pandey, who created a multi-billion dollar company by using YouTube to teach, and then he sold classes; he used UPI to pay and is built a multi-billion dollar enterprise. So that's what happened, and we had this whole payment and transaction revolution, and then we had the rise of all these gig economy companies. Zepto, two 22-year-olds delivering, uh, you know, stuff for three billions of dollars worth of goods in quick commerce. You have Rapido, which has the bike taxis; Rapido in Hyderabad is as big as the public transport system. You have Urban Company, which delivers, you know, people to repair your, repair your air conditioning, give you a massage, all that at scale; they're, you know, really large base. So all these guys have created, uh, all these companies, and now we are going to see, as the number of users in India go to about 900 million, the phone is going to be the basis for reimagining work, and that is really what is required to get where we want in terms of jobs and so on, where people will discover jobs on the phone; they will get their credentials on the phone; they will get, you know, get their benefits on the phone; and that's, that's the future that we see.
So India has had a digital transformation, and as, uh, Minister Jaishankar said, the acceptance of technology in India by Indians is very high; they're willing to accept. The reason is that what they have is so bad that this is much better for them, and so they, they, there they absolutely no hang-ups about accepting it; they're not having long convoluted debates and, and so on about what's good for them, and they just take the stuff and move on. Now this whole thing we did was called DPI, digital public infrastructure, and we believe that that is going to be the basis for using AI. So when we look at implementing AI at scale in India, it's built on an existing foundation of digital transformation that has happened for a decade. What are the, and you know, with the current thing, we have the predominant languages of the phone are English and Hindi; the UI, the user interface is a touchscreen which you, or you keen, and you have static knowledge available, and even with that we have reached more than 500 million users today. WhatsApp in India has more than 600 million users; PhonePe has about 350 to 400 million users; UPI has 400 million users. So even with all that we have reached these numbers. As the penetration of phones goes up and we reach a billion phones, what is going to take it? The billion people to use this. First language will move from just India, Hindi and English to every major Indian language, and that will make it much more accessible; we're already seeing evidence of that; I'll talk about it second. The UI from keyboard and touch will go to voice and video, so it'll be, so you'll talk to the computer in a language of your choice; it becomes so much more accessible. And third, because of generative AI and the reasoning capability of AI, you will go from static knowledge to dynamic contextual information that is at your fingertips at the time you need it. So these are the three things that will drive, we believe, the AI adoption to a billion people.
Now we believe that this will lead to India becoming the AI use capital of the world. So I don't want to get into, maybe later we'll talk about Yann LeCun; we'll talk about DeepMind and all that, but I can assure you that whatever DeepMind or not, India will be a place where this stuff gets used, gets stuff gets used at scale of 1 billion people, like we have shown before with Aadhaar, with UPI. And an example of that is how we have created the, uh, data for languages today, uh, you know, people like AI for Bharat have created open-source AI models for Indian language age, covering all the 22 major Indian languages and continuously reducing cost of operation, because you, you want to make AI work in India, you have to charge one rupee per inference; you know, you can't do it by charging those exotic, you know, exorbitant prices that you're seeing elsewhere. And this is also about data; these guys have gone out and collected data; this is not building models by scraping the internet and somebody else's IP; this is about going out and collecting data; they have collected ed 18,000 hours of spoken Indian language across the length and breadth of the country in all its variations and nuances using a whole army of people going around collecting data. So AI is about data, and data has to be the right data for your context, and these are the kind of infrastructure which are creating low-cost, population-scale AI. An example is in learning where this is being applied, uh, for education now in four or five states, 6,000 labs, and it's about learning to read and write. We all know the challenge in India; we had the ASER report for many years showing there's no improvement in, in, uh, in education, but now we think with these AI tools we can actually make learning better. And this is again a key difference in the way Indian we think about AI in this country, which is how do we use AI to make lives better? It is not how do we make use AI to, make things so convenient that you lose your skills; that it's not about dumbing down people; it's about actually using this to improve the capacity, capability, and potential of human beings, and therefore how do we use, uh, AI to make lives better, and that's a continuous theme in all the AI work in India. Another one which is coming up now in multiple states is the open agree network, which allows farmers, ers to get access to information in the language of their choice and all kinds of information: weather cycles, crop cycles, etc., and this again is getting rolled out right now; again another case of building AI at scale to help farmers improve their productivity and their markets and so on. And so what you're seeing is a, you know, sort of a AI and DPI, so AI makes DPI better. For example, in the Aadhaar system, they use AI for liveness testing; when doing an Aadhaar authentication with the face, they make sure that you're not using a photograph; there's a real, real human being there; all that is done with AI. When we do transliteration in AI in the, in the data of AI, they use AI, I mean transliteration; they use generative AI to reduce the errors in UPI, which does 18 billion payments a month; they're using AI to have voice payments. So you can see that the, the digital infrastructure we have is getting loaded with more AI and making DPI better. And a great example of that is what the government has built as a DPI, which is Bhashini. I don't know, is Amitabh here? I don't know. So, uh, Amitabh and, uh, Abhishek have done a great, uh, job of, uh, building Bhashini, which is essentially an API-based platform for Indian languages. They have a large number of models there; the most popular ones are from AI for Bharat. This system is already today doing 300 million inferences a month. Okay, it is doing it in 36 languages and 22 languages in voice, and the way it provides a standard protocol for Indian language translation in which are then used by multiple applications. So my Aadhaar is the application; the chatbot for Aadhaar resolution, PM Kisan, Emitra is a chatbot for farmers to know whether they got their money; the similar one is there for Nrega payments. So all these are actually today being used by ordinary people in a language of their choice to get a service that they want, and this will only get better because as the data from these interactions improve the models, as the models become better in terms of reasoning and so on, this will only become better for people. So this is the cycle. So when you develop these kind of public systems, it's not about delivering a solution; it's delivering an infrastructure which keeps improving, and so that's how the AI in India is, is improving. So I think my point is that it's not a slam dunk; it's not that some magic will happen tomorrow morning and we'll all be, you know, coasting along, you know, getting our UBI payments or whatever; it's about focusing on individual transaction, individual narrow use cases, making them work at scale; we'll be doing that; it's about creating an evolvable, iterative model of improvement, so you improve data, improve, uh, AI models with data from collection, synthetic data as well as AI from the usage, so that you keep improving things, and it is about making sure it's safe, secure, unbiased, responsible, and there are whole set of techniques to do that, and obviously we have to make this at one rupee per inference, otherwise this is not going to fly. So we think that AI is not easy; it's not, it's not some, you know, Kool-Aid; it's about doing it properly, but India will be uniquely placed because of its history and it'll combine DPI and AI to create a whole new way of doing things. Thank you very much.