Large Language Models: Google Research India head disagrees with Nandan Nilekani, says India must build LLMs

Gupta, speaking at the Bengaluru Tech Summit held between November 19-21, firmly disagreed with Nilekani and said, “There’s a technologist that I deeply respect, Nandan Nilekani, who made this statement: India should forget about foundation model building, just focus on the use cases. And I would say, with the person I respect, I respectfully disagree.” Nilekani is credited with helming world’s largest unique identity programme with Aadhaar.

The Google research head countered, saying India should invest in foundational AI models or Large Language Models (LLMs). “I think he’s (Nandan Nilekani) wrong because he’s not preaching what he has practised. What he did, he revolutionised India’s technology landscape by starting with the basics for Aadhaar. He didn’t start with use cases. He started with foundations. So, I think we must start with foundations.”

Nilekani’s comments came at Meta’s Build with AI summit in Bengaluru in October, where he said, “Our goal should not be to build one more LLM. Let the big boys in the (Silicon) Valley do it, spending billions of dollars. We will use it to create synthetic data, build small language models quickly, and train them using appropriate data.” The Infosys Chairman’s idea is that India must create infrastructure for collecting the right data and making the country “use case capital of AI globally.”

Gupta, on the other hand, emphasized that India’s constraints should drive innovation and not limit it. “We must use these constraints and so on as a kind of ingredients for innovation. Because I don’t think just throwing all the data and billions of dollars of computing at building models is ultimately the right way.”

Nandan wouldn’t be the only one with this view. Recently, Infosys Founder NR Narayana Murthy also said that India should not yet invest in building its own LLMs. “We have not been able to build large databases, and without big data, AI has no value. An LLM doesn’t make any sense. The Indian mindset is still not oriented towards problem definition and problem-solving,” Murthy had said in an interview.

Discover the stories of your interest

Indian Google team’s global contribution

The Google Research head highlighted the groundbreaking work being done in India, where researchers have been instrumental in developing global AI advancements. “Our team sitting in India has been really at the forefront, I would say globally. If you look at which are today the best teams that understand how you optimize these LLMs or these large multimodal models for efficiency, I would say arguably it is my team sitting in Bangalore,” Gupta added.

Gupta also cited innovations coming out of his team and why the country should not shy away from building LLMs. He gave the example of Matryoshka representation learning, inspired by Russian Matryoshka dolls, which optimizes AI models by prioritizing the most important signals.

“This has been used for Google image search, Lens, and even influenced OpenAI’s large language models,” Gupta said.

He also spoke about another breakthrough, Performer, which addresses the computational costs of handling long context windows in AI models, which had involvement from India’s research team.

Currently, startups such as Sarvam AI, Krutrim, Haptik (acquired by Jio), and TWO AI are building LLM products in the country, with more enterprises joining the race.