Krutrim’s Ravi Jain highlights unique value in Indian data representation

Krutrim, an artificial intelligence (AI) venture co-founded by Bhavish Aggarwal of Ola, has joined the increasingly competitive AI race dominated by players such as Google, Microsoft, and OpenAI.

However, what differentiates the firm from these players is that it has been built with the largest representation of Indian data used for its generative AI applications for all Indian languages. Today, all AI models called LLMs (large language models) are trained largely in English. Due to India’s multicultural and multilingual context, the AI models just can’t capture that. Experts said that they need to be trained on unique data sets specific to the country.

“This is a problem statement, which is at the crossroads of knowledge and language,” said Ravi Jain, Head of Strategy, Krutrim, in an interview. “Our differentiation is driven by what data is used in the training, different languages that we use, including their richness and depth. That would define the quality of the output in terms of the applications that we build versus the (tech) giants for whom we are one of the 180 countries that they may be operating in.”

Krutrim, meaning “artificial” in Sanskrit, is a family of LLMs (large language models). This includes Krutrim base and Krutrim Pro, which will have multimodal, larger knowledge capabilities, and many other technical advancements for inference. It is trained on over 2 trillion tokens, which refers to a chunk of text that the model reads or generates.

A team of computer scientists, based in Bengaluru and San Francisco, have trained this model, which will also power Krutrim’s conversational AI assistant that understands and speaks multiple Indian languages fluently, the company said.

When asked about the source of the data, Jain said that the first model that the firm built had a very large representation of Indian data which is present in the public domain. “You can imagine all Indian data which might exist in different languages on the web including a lot of PDFs (portable document format). So you have a lot of publicly available data in different languages,” said Jain. “As we go along, this would become the most important part of our journey about how to digitise data that is not digitised yet. And that is the case with many Indian languages. If we can make them part of the corpus and train the models, that will make a big difference.”

Last month, Krutrim was made available for public beta (where anybody can test it). The AI chatbot, which works similarly to OpenAI’s ChatGPT, is available in two languages — English and Hindi.

“This is a start for us and our first-generation product. Lots more to come and this will also improve significantly as we build on this base,” Ola founder Bhavish Aggarwal said recently on X.
Aggarwal said that Krutrim is rooted strongly in Indian values and data with over 10 Indian languages and is ready to assist in English, Hindi, Tamil, Bengali, Marathi, Kannada, Gujarati, and even Hinglish.

“While some hallucinations will be there but much lower for Indian contexts than other global platforms. And we will be working overtime to find and fix,” Aggarwal said.

Indeed, Krutrim recently gave incorrect responses to users’ queries. Screenshots shared on social media show the chatbot informing users that the 1983 Cricket World Cup was won by the West Indies. The fact is that it was won by Kapil Dev-led Team India that year. It also erroneously asserted that Hillary Clinton won the 2014 US Presidential Elections, among other errors.

When asked how the firm is addressing such issues, Jain said that the generative models can make mistakes as their insights are based on the information in the public domain, which can have a range of views. “It is referred to as hallucinations. We are also seeing quite a bit of that. This is part of the journey of all the larger models even Google’s (AI chatbot) Gemini and even all the versions of Open AI’s ChatGPT have gone through this process of learning and improving,” said Jain. “This includes putting the product in the hands of the consumers, learning from the feedback, understanding the safety issues and what people may perceive as biases and how to deal with those better.”

Jain said that users would see a rapidly improving product over the next few weeks.

The company is also working on AI infrastructure to develop an indigenous data centre and eventually, server computing, edge computing, and supercomputers. Production is scheduled for mid-2024 for prototypes and a rollout by the end of 2025.

“The larger we get, we’ll have to build our own data centre infrastructure and mechanism. These are very power-intensive and hungry machines,” said Jain.

Krutrim was launched in April 2023 by Bhavish Aggarwal and Krishnamurthy Venugopala Tenneti, a board member of ANI Technologies that owns Ola and Ola Electric. The firm in January said that it has achieved unicorn status — a term used to describe startups valued at $1 billion or above — following its inaugural funding round. The round, which saw participation from notable investors, such as Matrix Partners India, garnered investments worth $50 million in equity at a valuation of $1 billion. With this, Krutrim claimed the distinction of being India’s first generative AI company to reach unicorn status. The funds raised will be used to expedite the company’s “mission to transform the AI landscape, foster innovation, and extend its global footprint”. Business Standard