India should target chips, visual AI, not classic LLMs: Soumitra Dutta

Portulans Institute founder Soumitra Dutta says Big Tech data centres can help, but India's AI push needs chips, data, energy, and higher R&D spend

Soumitra Dutta, the founder of Washington-based Portulans Institute which is focussed on research and outreach on technology, is considered a leading global voice on artificial intelligence (AI). The former dean of the Said Business School at the University of Oxford talks to Shine Jacob about the future of AI in India, large language models (LLM), the data centre revolution, and the changes that public and private sectors need to take to boost the AI ecosystem.

 

Edited excerpts:

 

Google is investing $15 billion for its largest AI data centre outside the US in Visakhapatnam. Will such investments help in boosting India's AI sector?

 

For American tech companies, which are dominating the AI world, India is a big part of the story as it is a huge market. India is the second-largest market for ChatGPT users and might be the biggest soon. So, they want to look at India as a market, and also as a source for data. 

The population size here creates tremendous data in terms of people's behaviour and patterns of decision-making, and all of that feeds the AI engine. China is out of the equation for them due to various geopolitical reasons, and hence India becomes a natural market for these companies. To tap this, solutions have to be provided for India with AI services. Hence, they have to have AI infrastructure in the country, and hence they are building the infrastructure here largely to support the rollout of their AI services here. That kind of investment will keep on happening out here. 

From an AI ecosystem point of view, this investment is pretty good. Let's be careful that large data centres may not create jobs. They provide computing facilities, consume a lot of energy, and don't create many jobs. At the end of the day, large data centres will not solve the Indian AI story. 

The Indian AI story needs three things — we need chips, data, and energy. If you want to have your hometown ecosystem for AI, you need these three components. Such investments may trigger follow-on investments in renewables and other sectors.

 

What should be our approach in developing each of these three segments to have a better ecosystem?

 

We don't have chips; we are trying to get them but that will be a long journey. There is no way that we will be able to have it with reasonable quality before ten years. It is a good investment to make and we should have made it 15 years ago. For chips, we will have to be in the good books of America and the West to get them. 

We have plenty of data out here. One of the issues for us is how do we enable Indian companies to use this data for their own applications. We also have abundant energy. We have to invest more in renewables to support the data centres.

 

What will the business model for AI look like in India?

 

Here you have a more proprietary American model, with large companies controlling the AI ecosystem. I think the Indian model should be more open-source and more China-like in some sense, where you have open technologies which are deployed more aggressively in the economy. I think in the case of deployment of AI for the economy, we should probably learn from the way China is doing it. 

Do it frugally, as all the AI applications and LLMs in China are developed like that with low cost. At the same time, make it open source so that it spreads to the entire economy. China is doing it strategically as open source will help the economy and create a standard of some kind. We should also aggressively push companies and universities to invest in AI and AI education. We need to use AI to bring the Indian economy to a different level.

 

How far away is an LLM breakthrough moment for us? 

India already has the ability to take open-source technology from both America and China. I don't think the challenge right now is in developing the classical LLM. The classical LLM, which is essentially word-based, is more and more ‘solved’, as the discovery has already been made. The next level is more visual, and the spatial side of it. Hence, India should try and focus on the next level of LLMs and invest in that, rather than creating the current model of words. You can get the existing model from open source, adapt and modify. You don't have to go through the same discovery process. 

The advantage with India is it has got good talent. Lack of money, chips, and compute resources are our concerns.

 

Do you think more investment from the private sector should come for this?

 

We should invest in research and development (R&D). R&D happens in the universities, private companies, and national labs. Cumulatively we are not doing enough in R&D as it is 0.6 per cent of our GDP, versus 3 per cent in China, 3.3 per cent in the US, and 4-5 per cent rage in Finland and South Korea. We have to increase it everywhere — in universities, central labs, and private companies. 

The private sector has the money and has to invest in more R&D. To ensure this, we should make regulations and moves to incentivise them. You have to force them to compete with global competition. Private companies need to have a little more pressure and hunger. 

For universities, the concern is money, and also a weak industry link. You need to do more to get the best researchers to universities. We don't have a single university in the top 100. The reason partially is the less competitive salary structure in our country, high bureaucracy, and the amount of research funding is not fully available. If you push all of this, R&D will excel.