A large language model movement is happening in India. As VCs look to ride the generative AI wave, a number of AI startups in India have spawned in recent months. Of these, Sarvam AI topped the list with the biggest cheque received.
In December, the homegrown startup raised $41 million in a Series A round led by Lightspeed with participation from Peak XV Partners and Vinod Khosla’s Khosla Ventures.
Vivek Raghavan and Pratyush Kumar, co-founders of Sarvam AI, also worked on developing AI models for speech recognition and translation at AI4Bharat, a research lab wing of IIT Madras that was backed by the Indian government and Microsoft.
The five-month-old startup gained ground quickly. A couple of weeks ago, it released the first Hindi LLM called OpenHathi-Hi-0.1 and is reportedly working on a range of enterprise models.
We speak with Vivek about why Sarvam stands out, its goals and security concerns around LLMs.
What is the connection between AI4Bharat and Sarvam AI, given that some of the people involved like you and Pratyush are the same?
AI4Bharat is a center in IIT Madras. At that time, I was working for the EkStep foundation and I was one of the people involved and I was a mentor to AI4Bharat. The goal of AI4Bharat was and continues to be them to be the development of Indian language AI using open-source data, open-source models and open-source benchmarks.
Sarvam AI is a company, which was formed in July, August of this year to basically build large language models and a full stack generative AI capability which focuses on Indian use cases and Indian contexts. While Pratyush and I were involved with AI4Bharat, we started Sarvam AI later.
Of course, we have an MOU with IIT Madras and we do actually collaborate still.
Suddenly over this last year, foundational models have become a geopolitical issue. Why is it essential for India to have homegrown large language models?
I think the answer is in the question itself. This is too important a technology for people not to understand it from top to bottom. We absolutely should use LLMs from anywhere, whether they be from Open AI or Google or open-source ones from Llama or Mistral. But we should know how to build a model for ourselves because this is too important. Therefore, various countries are having on embarking on these kinds of efforts. Right. And so, while this is definitely a commercial thing, we want to build for the Indian ecosystem, and we are deliberately an Indian domiciled company.
How does Sarvam plan to be different when users can get the same thing with a GPT wrapper?
There’s a couple of things here. Firstly, we don’t have anything against OpenAI, there are certainly use cases where it is appropriate to use whatever you need. But understanding the Indian context and secondly, being able to operate in Indian cost structures are things that if you actually want to be able to do that thing at scale, you can’t be using GPT-4. And as far as moats are concerned, things are changing in this space every day.
I don’t think you can operate on the basis of moats. You have to operate on the basis of whether you have the technical strength to react to whatever might happen. This is not a situation that you can sit at a moat and hope that things work out. Things are changing fast, so your moat is really the team. So, the funding you receive goes into hiring talent. And you need globally competitive talent for something like this.
Which is also why OpenAI shifted from a non-profit structure to a profitable one.
I don’t want to comment on Open AI per se, but we saw that they started as a nonprofit and then later realized that they couldn’t do this as a nonprofit. And obviously in India, we are generally talking about what to order of magnitude less than the actual things that are there. So, it’s much harder. I think the big question is that we hope that when we show that there will be opportunities in India where there is much more significant funding to build these kinds of foundational models. We want to make sure that we have enough and we’re solving real problems for India, and at the same time that building this muscle to be able to be ready for bigger challenges.
What are your plans with your funding, and how hard or easy is it to hire from the talent base in India?
I think we have primarily two goals with the money that we have raised. One is of course, to get world class talent in this space. And the second thing is that we need a lot of compute to train the models etc. Of course, it is not easy to get really good people. But we are hoping that we will put things out where people understand that this is a team that is serious and wants to do this in a systematic way and they come and join our mission to build something for India.
Data privacy is a messy affair when it comes to training LLMs. India doesn’t have fully formed laws around it. How do you plan to tackle this?
It’s getting harder and harder to understand the provenance of data. Because you can paraphrase it and say it in a different way. And that’s not a violation of copyright. “Where did a fact originate?” is a very complex question. Just because it exists in a textbook, doesn’t mean anything. The textbook could have taken it from somewhere else. So, some of those things are hard challenges and because the tech is moving at such a rapid pace, globally. regulation also needs to be portable.
And if we take something very quickly and just fix it, then there is a danger that you may actually stifle innovation or not be able to protect the real issues so regulation has to move along with the tech advancement. I think this is the first time I felt that a technology will actually help people and has great potential. So, we need to figure out how every student can get a good teacher, how will every patient can get high-quality medical care. And we need to remember that when we are thinking about regulation and it also is important that regulation doesn’t ensure that the big companies escape and smaller emerging ones buckle under it.
What are you working on next?
One of the important things that we are working on is a platform which makes it easy for people to create and production deploy LLMs. Let’s say I have built an app which uses an element and I want to deploy that app on WhatsApp. Just think about how do I how to distribute these kinds of capabilities in a very different way in India, and how you would do it in other parts of the world. How will specifically Indians consume generative AI? That is the question we want to understand.
