India needs to think about its leapfrog moment in AI research: B Ravindran

IIT Madras' B Ravindran says India must focus on AI penetration, data, and infrastructure - not just chasing ChatGPT - as BharatGPT, Sarvam and India AI mission chart the nation's path forward

Professor B Ravindran heads the Wadhwani School of Data Science and AI at Indian Institute of Technology Madras (IIT Madras) and is regarded as a leading expert on artificial intelligence. His work has shaped national AI policies and thinking on using the technology ethically. Ravindran, in an interview with Shine Jacob in Chennai, spoke about India’s progress in AI. Edited excerpts:

 

When can we see Indian large language models (LLM) like China’s DeepSeek or a ChatGPT of America’s Open AI? 

It (DeepSeek) didn’t happen momentarily. The Chinese government deliberately started investing in it over a decade ago. It’s just that DeepSeek caught the public imagination. In fact, there are at least 50 organisations in China that could have done something like DeepSeek. 

Hence, it is not a DeepSeek moment, not a lone wolf fighting against a system and creating it. A whole ecosystem was developed, including agentic models [a type of AI system]. Some of the first open-source agentic models came out from China; the US came later. 

What is the status of LLM development in India? 

It’s something that we are starting now. We have the benefit of hindsight by learning from various companies and countries. We shouldn't be thinking about producing the next ChatGPT; our focus should be on AI penetration. That will help people access AI; you don't need high-end GPUs (graphics processing unit) or a power plant just to power AI. 

India needs to think about what should be the leapfrog moment. We are now waking up in terms of AI infrastructure.  Infrastructure doesn't only mean compute, we need data. We need to develop our own datasets, our own tokenisers, and build a training pipeline and perhaps our own models. 

We are investing in GPUs. The India AI mission is fairly comprehensive in the goals that we have set up. Of course, we have to invest more money into all of these. It understands that data itself is a separate pillar, compute is separate. In addition, they have focus areas on fundamental research, skilling, innovation and entrepreneurship. It also has set an ethical roadmap. At a higher level, the vision is set. 

Will BharatGPT or Sarvam take the lead in this? 

BharatGPT is an IIT Bombay initiative and it is focused on building from scratch. Sarvam is trying to build a medium-size language model. We don't have the computing resources needed to build LLMs. Even for building, say, an 80-billion parameter model, we need thousands of GPUs. 

Though China says that building DeepSeek was very cheap, the background work to make it a reality was fairly extensive. I believe reports underestimate the cost. We are not going to go to the scale of these 500-billion parameter models. Even to build whatever we are building now requires a significant amount of computing resources. 

We will soon have a suite of Indian data-trained language models, which we can build for our own applications. We are not going to replicate the external models; price points will also be low. The initial models you will see commercialised before this year-end. 

What is the role of the private sector in AI development? 

The amount of money that companies invest in fundamental AI is minuscule. Forget about working with academia, our IT companies should be able to do more research internally and they are not doing that. At least, not visibly doing that, and all the work that’s on is on how to take advantage of the existing open-source models. 

Fundamental research is still lacking in terms of large-scale investments by corporations. Industrial investment in academic research is not happening as well. We (at IIT Madras) work very closely with industry, and we are an exception. Just having a few institutes doing the same will not be enough — we need hundreds of such institutes at the same level. The government can only do so much; it is upon the industry to start nurturing talent. 

A Reserve Bank of India panel, which included you, recommended leniency for first-time errors in finance. What is the reason? 

We are saying that there should be a rigorous validation chain at a bank. Don’t rely on the developer to give you the guarantee. Banks will have to rely on internal capacity to test models. After everything, when you deploy the model and some error comes, we will not penalise you for the first time. This is considering that you have followed rigorous validation and testing. If you are doing it repeatedly [making errors], you will be penalised. 

I think the report got the middle ground. We are positive about the role that AI can have in taking banking to the unbanked. 

There is a lot of concern about jobs, especially after layoffs at Tata Consultancy Services. How do you assess this? 

AI is not going to completely replace humans but it will help deliver solutions. AI-awareness is something that everybody should work on. Whether you like it or not, AI already plays a huge role in most of your life. 

The landscape of jobs is going to change. Everybody should learn how to use AI to do their jobs better. There will always be something that will get eliminated. Every sector will have to look at the value that humans add. 

For most jobs, it is not a single qualification but a collection of skills that are required. For a reporter like you, some aspects of it can be significantly augmented by AI; some can be replaced by AI, and some you have to bring in. Journalistic instinct can never be replaced.