Why India Needs More AI4Bharats

When a country wants to build AI, one of the most important ingredients is the need for a robust industry-academia partnership. The US and European countries have already identified it, and China too has also been capitalising on it. But when it comes to India, there is a visible lack of such initiatives. 

One prominent initiative taking wings in India is AI4Bharat, which started as a collaboration between IIT Madras and Nandan Nilkeni’s ekStep foundation. It is sponsored by Bhashini, Microsoft, Google, and NVIDIA and its contribution to the Indic open source AI community has been tremendous. But the problem is, it is the only prominent one in the country so far. 

In a podcast with AIM, Adarsh Shirawalmath, the founder of Tensoic, and the creator of Kannada Llama, said that he is very inspired by the work that AI4Bharat is doing in the country. “We need more AI4Bharats in the country,” he said. 

https://www.youtube.com/watch?v=RAoTisOaJSQ

The Indian academia narrative

One of the biggest challenges when it comes to building AI models in India, apart from compute, is the rate of adoption within the industry. “We need to do what China is doing,” said Shirawalmath, while explaining that China’s government funds the projects in the country, India can leverage the industrial partnerships to compete in the race.

BharatGPT is another similar initiative which was started by IIT Bombay, and is now being built in partnership with the Department of Science and Technology of the government. Its goal is also to make AI for Indic languages taking the open source route. 

“Not sure what is happening right now [with BharatGPT], but it would definitely drive India’s AI forward,” said Adithya S Kolavi, the creator of Indic LLM Leaderboard from CognitiveLab. 

Kolavi pointed that one of the biggest reasons behind his love for AI4Bharat is that the initiative is bullish on open source. For example, Llama 3’s TikToken optimiser is not compatible with Indic languages. To make up for this, AI4Bharat’s IndicTrans2 tokeniser came out in open source which was very helpful for the community. 

Mufeed VH, the creator of Devika also told AIM that he tested on AI4Bharat’s IndicLLMSuite dataset, which is rich enough for making AI models. The dataset contains around 251 billion tokens in 22 languages, which is mostly converted from audio and translated into Indic languages from Wikipedia. 

“More data wouldn’t hurt,” Mufeed said. He recently got into Y Combinator for his startup Stition.ai, which is building security focused solutions for AI.

The problems that Mufeed points out is that even though IITs and NIITs are doing the research, they do not get the support they need from the industry. “China is racing against the US. I think the Indian government should do the same,” he added. 

Stuck at Research

On the contrary, there has been research that comes out of Indian universities focusing on LLMs, voice models, and using AI in several fields, but most of them get stuck right at the research phase. However, this is slowly changing with several researchers sending their papers to ICML and NeurIPS. 

“None of the research from the universities actually comes out. They just do research in the field like a final year project, and it dies there,” said Mufeed, about how researchers should come out of the universities and put their creations into products, or probably build a research lab, such as AI4Bharat.

Kolavi pointed out that there are also not enough grants in India coming from universities or companies for flourishing research. “You have the VC kind of things, but grants are essential to push research forward. I have not seen that concept flourish in India,” added Kolavi. 

The Ivy League universities have the infrastructure to facilitate research, but Indian universities do not have that. In a recent post, a student pointed out that there are merely six GPUs in a university in India for doing AI research. 

Similar thoughts were shared by Professor V Ramgopal Rao from BITS Pilani on how industry involvement is key, whether it’s to make our students industry ready or reap the benefits of  research happening in the country. Industry involvement is needed to convert research into innovation.  

Meanwhile, Google and OpenAI have been taking active interest in working on Indic languages. Google spoke at length about its Navrasa model, which is built in partnership with Telugu LLM Labs in India. OpenAI at its Spring Update released GPT-4o, which also includes a huge corpus of Indic language data. 

Apart from gathering massive funds, which is definitely a priority to bridge the gap between research and industry, India needs adoption by companies within the country for Indian AI models. We need more initiative out of our universities, and collaborate with the industry to drive India’s AI momentum forward.