India's AI will look quite interesting, says BharatGen executives

For us, accessibility is key, says BharatGen execs

India's artificial intelligence (AI) will look quite interesting, said executives of BharatGen, which has recently been awarded over Rs 900 crore under the IndiaAI Mission to build a large language model (LLM).

 

“People keep asking when are we going to get a ChatGPT? What they are trying to say is when will they get an Indian AI in their phone…for us accessibility is key. And voice being a key modality, it creates accessibility for people with poor functional literacy,” said Rishi Bal, executive vice-president, BharatGen.

 

BharatGen is a government-funded consortium led by Indian Institute of Technology (IIT) Bombay, IIT Madras, IIT Mandi, IIIT Hyderabad, IIT Kanpur and IIT Indore. The focus of the firm has been building India’s LLM across 22 languages. 

Professor Ganesh Ramakrishnan of IIT Bombay and principal investigator said that the aim is to have a trillion-parameter model. “We will progress from a few million to certainly a trillion. The idea is model distillation…we first do smaller models before a bigger model as the former gives better insights and understanding. Once you make larger models, you can go back to smaller models,” said Ramakrishnan.

 

Training this trillion-parameter will involve huge data and access to compute power. For data, BharatGen is looking for publisher and local radio station tie-ups, free optical character recognition (OCR) services and annotation.

 

Data, which is an important aspect of building an LLM, is also being collected in a unique way to maintain the Indianness of it. For instance, unlike the traditional approaches of model building by crawling data, BharatGen has a team of people who are operating in various parts of the country (at present in Madhya Pradesh). They approach publishers, and radio stations that have local content.

 

“We told them that we would like to use some of their data to train our models. We also tell them that we are not going to replicate any of the data. And, the number of people in the community that responded to this reach out has been amazing,” added Bal.

 

Both Ramakrishnan and Bal acknowledged that access to graphic processing units (GPUs) is going to be crucial. Rather, majority of the Rs 900 crore will be utilised for GPUs. “We collaborate with Nvidia for training. We do training with India data centres. For inferencing, we will look at more diverse players. We are in talks with Qualcomm and Intel for inferencing,” added both the spokespersons.