What should be India’s AI strategy? We asked this question 10 years back and we ask it now. From a global perspective, much has changed since then, and from India’s point of view, not so much, like before we have much catch-up to do. So, how can India make the most of the current opportunities, but also go beyond the endless catch-up game to become a leader in the long term.
Let us start with what has changed with AI
First, AI has suddenly matured in the last 18 months to solve many problems where we didn’t have a clear pathway to a solution for a long time. They include problems in natural language understanding, image, video generation and inference.
The second big change is now there are broad models, termed Foundational Models, which solve a variety of problems reasonably well, rather than having to build a custom model each time. This has spooked the AI and computer science community, where we held the view that broad algorithms/models will solve all problems badly, not well!
Third, we sparsely understand why the current techniques work. In many ways they work in different ways from human intelligence and make stupid errors. They look highly inefficient akin to a brute-force approach, controlling them is hard, and so is testing them. There are no natural next steps from here to the next breakthrough.
Fourth, they are compute hungry. These are not our good old models that were trained/ran easily on our PC or mobile, they need GPUs to back them, and have latency issues. In some ways, they are very data hungry, in building a foundational model. On the other hand, data needs have reduced for building problem-specific models on top of foundational models – where we needed 500 to 5000 data points, one can now get reasonable results with 5-1000 data points, by methods such as few shot learning, providing context and fine tuning.
These are big changes, not only for the outsiders, but also the AI community itself, which didn’t see this happening. So, where does the opportunity lie for India?
AI technology to solve business use cases for the world
The first is rather simple and straightforward. AI technology has matured to solve a number of problems. However technology diffusion is very slow – products and services that bring new technology in hands of everyone take time to gain trust, render efficiency and it takes people time to adopt it. They need to be easy to use, reliable and provide precise low-friction customer value. This doesn’t happen automatically!
This is a big opportunity for the Indian startup ecosystem to build AI technology for business use cases for the world. In the last few years Indian startups have understood how to sell to businesses across the border with a maturing B2B SAAS industry (we still don’t know how to build global B2C businesses from India). We know how to build tech products for global needs, market and sell them to the world. We need to bring these skills to new-age AI led products and be firstcomers in solving a number of business use cases globally.
Current challenges include massively lower funding for AI startups as compared to West. In the last year, money invested in AI startups in the US was $31 Billion as compared to India being $168 Million. This needs to be massively scaled up. Second, we see a number of ‘thin wrapper’ AI solutions from India, which is a small set of code or packaging over existing foundational models. These solutions are not competitive enough, and mostly do not create sustainable customer value. Indian startups need to be the (pareto) best product for the use case they pick. Lastly we need to combine AI + hardware to make the most use of the opportunity – that has been harder for Indian companies to build and scale sales globally.
This is the first leg – differentiated AI technology from India that delivers sustainable customer value. At Change Engine, we are targeting this area in a few different ways – creating data to understand AI gaps in India, producing public content and AI bootcamp on building real-world AI products and working directly with entrepreneurs through our accelerator.
AI for Indian businesses, governance and public services
The second leg is using AI to diffuse to make Indian businesses, government, public services, way more efficient. Two recent policy articles, one from USA and another from China, take the position that diffusion of AI in a country’s ecosystem will determine its competitive advantage. The hypothesis is simple: AI will make everything more efficient, and whoever is able to adopt it fast will have better internal governance and be highly competitive globally.
For India, this is spectacularly more important, given the scale of our population, capacity issues and capital deficiency in serving the needs of our citizens. Not only do we need to build solutions for Indian use-cases, but make sure they are adopted at scale. Do not make the mistake of thinking that it will be automatically adopted – no, it won’t! India centric AI is important for another reason, preservation of the best traditions and ideas from India, the best philosophical knowledge, art, literature, history, and so on. If AI systems will be the information retrieval, inference and creation mechanisms of the future, India’s 3000+ years culture should be well represented.
We see quick wins such as Digi Yatra Foundation, but challenges are immense. First, we need more open data-sets from the government and clear sharing protocols for government departments to make data available to the public, right users and builders. This is well-known but progress has been slow. Second, businesses building for India use cases, and the government as customers are not VC-friendly. While VCs have lately been a little more open to these (see this), mostly VCs do not like the government as a customer and India use cases generally run into market size issues. This means the government and private philanthropic initiatives need to support these activities. Government money is there, but runs into a variety of procurement issues which don’t encourage startups, the best solutions or risky bets. Procurement needs to be set right, which is the topic for another essay. On the other hand, philanthropic capital for tech in India has been sadly abysmal – we need to recall OpenAI started as a non-profit, so are Ai2 (Allen Institute of AI) and Howard Hughes Medical Institute (HHMI). What about India?
There are positive initiatives in this space including People+ai, AI4Bhārat, BHASHINI, and Sarvam. But, they are much smaller and narrower than what is needed for speed.
Forward looking AI research
The third leg is the most ignored and un-talked. It is India’s participation in cutting-edge research. Big tech and startups in the West are crowded with AI professors and PhDs from universities. Yann LeCun, Chief Scientist, Meta , Ilya Sutskever, ex- OpenAI and Aravind Srinivas, Perplexity CEO, are all either PhDs or Professors from US Universities. Deep learning itself was invented at the University of Toronto. Ilya Sutskever's list of 30 (actually 34) most important AI papers (not a gold-standard, yet) have 9 coming from universities, 10 coming from industry, 4 individual papers and 11 from university-industry collaborations. Deep innovation and solutions to hard problems come from scientists with PhDs, universities or deep industry-university partnerships.
To understand the value of research, one must understand the history of AI. In the 70-80’s knowledge based methods were the vogue, in late 80’s and early 90’s neural networks had a short stint, quickly replaced by statistical and Bayesian methods in 90’s and early 2000’s and deep learning is a phenomenon just 12 years old. Since then, the speed of innovation has been exponential. If we sleep today and wake up in 3-5 years, who knows, maybe there will be a wholly new technique – which is interpretable, doesn’t hallucinate, is data efficient and more. Labs such as those of Josh Tenenbaum at Massachusetts Institute of Technology are looking at newer ways of modelling and there have been experiments with neuro-symbolic methods (something of interest to me as well for a long time!). Another is from Lav Varshney’s lab in University of Illinois Urbana-Champaign called information lattice learning. The other direction comes from advances in neurosciences to inform new age AI, quantum computing may disrupt it and biological computers may usher a new era.
For India to go up the value chain in innovation and long term leadership, it needs a strong university RnD ecosystem in AI. This enable hard technical problems being solved in India leading to creation of companies that are global leaders, rather than followers and copycats of Western-developed startups and technologies. We also need to be ready for the next disruption. Who knows the next disruption might come from India- SVMs and the internet were invented in Europe, chatbots in China and deep learning in Canada. Even if not, we need to have a critical mass of researchers working in cutting-edge areas, ready to translate research, as soon as disruption arrives. We cannot be reactive.
The data doesn’t favor India in research today. Our recent report shows that the USA and China’s paper output is 21-16x India’s. Top Indian universities and institutions are doing 1/10th the number of papers as compared to USA and China, in top AI conferences. Further, what is alarming is that India’s growth of high-quality papers is already plateauing! We need fundamental fixes starting from how research is funded, mission-led approaches, improving ease of doing science and making our universities more efficient. At Foundation for Advancing Science and Technology - India, we have been working on policy solutions for many of these, and also working closely with government departments.
This completes the stool:
Leg 1: World-class AI technology from India delivering sustained customer value
Leg 2: Diffusion of AI technology into government systems, public services and Indian businesses
Leg 3: World-class AI research in Indian universities in current and future looking areas.
This is how we can win today and be ready for tomorrow.
Before I end, let me talk about three issues that must be bothering you. Shouldn’t India be in the LLM race? What about compute? And what about AI regulation? All these three can be dealt with by policy and some of it may resolve itself in coming years. I have already exceeded 1500 words and you might want to get a chai now, so wait for another blog on these. Meanwhile, think how you can enable one or more of the three legs, as you sip your chai!
Comments