Why the world should stop obsessing over LLMs

As the debate over the U.S.-China battle for AI supremacy intensifies, Nandan Nilekani, the Indian tech billionaire and co-founder of Infosys, says the world needs to shift its focus away from who’s building the largest LLMs, and toward how they’ll be used.

In addition to his private sector accomplishments, Nilekani is the architect of Aadhaar, India’s national ID system. Aadhaar allows 1.4 billion Indians to access a wide range of digital services — from banking and health care to work programs and tax services.

Now, he’s pushing for a more democratic approach to AI, advocating for smaller, open-source models trained on high-quality data vetted by humans rather than massive systems run by a few powerful players. 

“Model building will become very common. I don’t think we should get obsessed with that,” Nilekani told Rest of World. “We have to bring the conversation back to making it useful to people.”

This interview has been edited for clarity and brevity.

What LLMs are you following closely?

Models are coming fast and furious from around the world. You have all the U.S. models — OpenAI, Gemini, Llama, Anthropic — and then, of course, you have the Chinese models — DeepSeek, Qwen from Alibaba. So, there are umpteen numbers of models out there.

Models will be a commodity. There will be faster, better models that will come. But the real challenge in AI, like everything else, is how do you make the lives of people better?

Smaller models trained on the right data can be almost as effective as very large models trained on generalized information. So the tendency now is to have smaller models — more content, more efficient with low cost of inference, etc., and which are trained on specific data for a specific vertical use case and so on. So I think this whole area is changing very rapidly. 

Which real-world AI use cases have impressed you the most so far? 

If you think of AI applications, you can think in three ways: consumer, enterprise, and society-government.

The consumer side is growing with all these AI chatbots, where there is constant competition among various actors. Enterprise AI is rolling out, but relatively slowly — not because of limitations in the technology, but because large companies are burdened with legacy systems accumulated over the years. Their data is not necessarily all organized in a way that AI can readily consume, so enterprises are going through their own journeys.

On the social side, people are looking at how to enable population-scale applications using AI, which make a difference to people’s lives.

I am currently working on two or three areas. One is applying AI to language. It is important to India because we have a large number of languages. How do Indians communicate with each other? How do people speak to computers? Earlier we used to have a keyboard on a PC, but people need to be literate for that. Then you had a touch screen when the smartphone came, so you could swipe and watch a video. But we think the future will be spoken — you speak to the computer, but in a language of your choice, in a dialect of your choice, using the colloquialisms that you like.

If a farmer in Bihar can speak to a computer in Maithili or Bhojpuri or whichever language and gets the right answer, you have made AI so much more accessible to him. That’s a big area of focus for me.

Stepping outside of India for a moment, DeepSeek made waves recently. How do you see the current AI race between the U.S. and China?

It’s great that this is happening because this kind of race forces more and more firms around the world to build better capabilities. So the technology in AI is moving at a very rapid pace. A lot of innovations are coming. And many of the innovations, fortunately, are being put in the public domain — like DeepSeek has published how they built it, which is very useful knowledge. That’s why I am saying that models will become a commodity. When anybody can build a model, it will only require sufficient compute capability, and enough data, which is becoming more and more available. Model building will become very common. I personally don’t think we should get obsessed with that.

Can you elaborate on that?

We have to bring the conversation back to making AI useful to people. All technology is ultimately only meaningful if it is accessible to people at scale. Which is why, I think, rather than starting with AI, you should start with the work we did in the last 15 years on building digital public infrastructure at scale. 

Does that mean a shift toward smaller, more efficient models rather than large language models?

Some solutions can only be solved with large diagonal models. But there are some solutions where you can have medium-sized models. There are others where we can have small models. There are some that you solve by having the model running on your phone — a quantized model [that uses less memory and computing power]. We need to use whatever makes sense for a particular use case.

I’ve philanthropically supported a group at IIT Madras [one of India’s leading engineering universities) called AI for Bharat. They’re collecting data from the field, so it’s not just scraping some internet stuff. They have people around the country collecting samples of people speaking in Hindi, Bhojpuri, Tamil, etc., and in the colloquialisms of that region. All that data is being brought in and is open-source.

You’ve raised concerns about the “secrecy” surrounding large-scale AI models, and have advocated for open-source alternatives. Are you equally concerned about Big Tech’s growing control over computing infrastructure, including data centers?

If computing is going to be expensive and difficult, and if many people have small data centers, can we bring them all on an aggregated platform? It’s all about getting access to computing for more people.

My view is that AI will become obviously pervasive. It will be adopted by all kinds of users. It can have negative issues; it can lead to challenges of hallucination, fake news, deep fakes, all that — agreed. But it also has the potential to change lives for people. And ultimately, it will be so pervasive that everybody will be able to leverage it in some way or the other.

We have seen many technology hype cycles. Do you think there’s a similar trend in AI right now? 

You are right in that we have had many hype bubbles in technological history. We had the dot-com bubble 25 years back, then we had this whole Web3 bubble, then for some time people said cryptocurrency, then some people said metaverse — so every few years, there is some new thing that comes up.

I think AI is a mix of both. Certainly, there is a bubble or hype element. But at the same time, it is genuinely a powerful capability which, if applied well and responsibly, with accuracy and scale and speed, can make a difference to lives. So it’s somewhere in the middle.