The ongoing India AI Summit in New Delhi will witness discussions ranging from how the developing technology may impact jobs to India’s AI infrastructure and governance.
Ankush Sabharwal, creator of BharatGPT, which was selected under the Centre’s IndiaAI Mission to build sovereign AI models, spoke to The Indian Express about these domains on the sidelines of the event. He discussed why India should focus on AI solutions rather than expending resources on building frontier models.
Sabharwal is also the CEO of CoRover, a conversational AI startup. His company recently raised funds from Google and HDFC Bank and has received a contract from the Indian Army to build a sovereign AI model for military operations.
Accessing compute (meaning the critical infra needed to run AI systems, like semiconductors and memory) is one of the biggest challenges for small Indian companies. In your experience in building AI models, has that been a struggle?
Today, it is easy to create an AI model. You rent Graphics Processing Units (GPUs, or the hardware needed to perform complex calculations and process digital visuals), take data from open source organisations, like Common Crawl, use open-source algorithms, and create a model. But what problem are you really looking to solve?
One does not have to build a mega, trillion-parameter AI model, for which you might need large computing power… If ChatGPT wants access to such computing power, it is understandable, because they are creating general-purpose large language models. I’m not saying we should not build such models. But if you are building such a model, then you need to have the resources and the vision for the next 10 years to build, while not making any money.
Are you exclusively utilising the subsidised GPUs available under the Centre’s IndiaAI Mission, or relying on other resources?
Story continues below this ad
We have multiple partners. We have L&T, Tata Communications, and we are also partnered in the IndiaAI Mission, under which everyone has to submit a use case and they can then get GPUs at a subsidised rate. But again, our focus is not to just burn GPUs. We want to have the things which work on edge devices to really democratise the technology.
But do you actually get GPU access for much cheaper through the AI Mission?
Yes, the hourly GPU rates under the IndiaAI Mission are much cheaper. The market rate is around Rs 400 per hour, and under the subsidised rates, we can get it for around Rs 67.
It also depends on what you are building. Google has many services, and you have to see what you are building – LLM, or fine-tuning a model, or something else. Based on that, you have to choose an approach.
Story continues below this ad
Another common problem that’s flagged is the lack of high-quality Indian datasets for AI models to train on. Do you feel there’s a gap?
It depends on what people are planning to do with the data. It’s easy to say that I want all the data out there. But, if you are solving specific problems for a travel company, you partner with them and use their data. If you are building something for defence, you seek data from them.
Even if you are building a model from scratch, there is enough open-source data, which you can use… There is Common Crawl, in India, we have AI4Bharat, which is open source, Bhashini is also open source, and there’s the AI Kosh. Ultimately, data is the real differentiator. If you have the right data, you will be able to build the right solution. For instance, we work with IRCTC, LIC, and NPCI, and their data to create models for them.
India does not have as many mathematics scholars as China does, which has been attributed to their success in AI. Do you see that as a long-term challenge?
Story continues below this ad
Yes, this problem is genuine. That’s why you see fewer research papers and deep tech companies coming from India. To build hardcore deeptech platforms, you really need hardcore research-oriented techniques. We definitely need to be strong in analytical, mathematical, computer science, and programming skills.
But, we also have to leverage our strengths… To create meaningful AI solutions, one does not need deep tech solutions or mathematical skills. To create strong AI platforms, you need deeptech, and only a few of us can do that right now. But a wide range of people can build AI solutions.
