Ola Founder Bhavish Aggarwal Unveils Startup Krutrim To Cater To Indian Market’s AI Needs
Ola co-founder and CEO Bhavish Aggarwal on Friday took the wraps off his latest venture, Krutrim (means ‘artificial’ in Sanskrit), which is focused on delivering artificial intelligence (AI) focused solutions for the Indian market.
The startup has introduced its Krutrim AI chatbot that can take prompts in 22 scheduled Indian languages besides English and generate responses in English, Hindi and eight Indian languages. But Aggarwal revealed that the larger endeavour of the startup would be to create an AI stack that thinks from the level of semiconductor and Graphics Processing Unit (GPUs) and upwards till creating Large Language Models (LLMs). The effort from the startup would include long-term investments in developing chip designs, engineering sustainable cloud architectures, and creating novel models beyond mere LLMs to cover the spaces of healthcare, drug discovery, material science, education, agriculture and more. Lastly, the startup would also look into making end-user consumer facing applications.
Investment And Company
Speaking to the media, Aggarwal revealed that Krutrim is entirely a separate entity, independent of the Ola and Ola Electric corporate structure. Although the companies possibly do have certain investments in Krutrim. “It’s a completely independent company, but we do have investment into it which we have not publicly announced. We have, I have invested, (along with) some friends and family, (and) some investors have invested here (as well),” he said.
Aggarwal did not reveal the size of his investment in Krutrim. According to a media report, the startup had earlier raised USD 24 million in debt from Matrix Partners.
Krutrim AI
The Krutrim AI chatbot operates in a manner akin to well-known AI chatbots ChatGPT and Google’s Bard, yet it distinguishes itself by offering support for a range of Indian languages beyond Hindi and English.
The AI chatbot based on Krutrim LLM boasts generative capabilities across 10 Indian languages, while it can process inputs in a total of 22 languages. During the presentation of these new models, Aggarwal highlighted that the system has undergone training on a vast dataset comprising over two trillion tokens specific to Indian languages. In the context of language processing for Large Language Models (LLMs), a token serves as the fundamental unit of data, encompassing text or code, essential for the model's understanding and language generation.
“AI will define the future paradigms of economy and culture. To become a true leader of the world, India needs to become a global leader in AI. Today, all AI models called LLMs are trained largely in English. But language is not just text; it is also the vehicle for cultural values, context, and ethos. Due to India’s multicultural and multilingual context, the AI models just can't capture that. It needs to be trained on unique data sets specific to us. And on top of it all, it needs to be accessible to India, with India-first cost structures,” said Aggarwal, in a statement.
Krutrim has been built with a substantial representation of Indian dataset incorporated into its training, enabling the development of generative AI applications tailored for all Indian languages. Aggarwal disclosed that a team of computer scientists, located in Bengaluru and San Francisco, spearheaded the training process.
But the Ola CEO refrained from providing detailed information regarding the data used to train the model or its sources. However, he mentioned that certain datasets were proprietary and obtained through data partnerships.
For now, the launched AI chatbot from the startup is being made available to some people who have signed up and more users can sign-up for access in batches throughout December. The public release is set for January 2024, with open APIs accessible to all developers by February 2024.
Ola will be deploying Krutrim across its group of companies, integrating the technology into various functions such as sales, service, support and other operational processes.
Aggarwal said Krutrim's advanced linguistic capabilities position it as a valuable tool across various domains, including education and business communications. The platform reportedly integrates cutting-edge techniques in secure AI to minimise inappropriate responses.
Data Centres And Chips
The company is said to be actively engaged in the development of AI infrastructure, with plans for indigenous data centres and eventual advancements in server-computing, edge-computing and supercomputers. Prototypes are expected by mid-2024, with a production roadmap set for release by the end of 2025.
“We're also building our own technology for data centres. It is not just the physical investment into the data centre, but also the technology because it's very important to make data centres more efficient, to bring down cost, and also to make those greener and sustainable solutions,” said Aggarwal. “While AI is the soul, the infrastructure and the silicon is the body in which it runs. In India, we need to design our own silicon chips for building this.”
When questioned about the necessary graphics processing units (GPUs) powering Krutrim's operations, Aggarwal declined to provide specific details but conveyed that the company possesses substantial GPU resources. Originally designed to enhance graphics processing, GPUs play a pivotal role in significantly expediting computational processes, particularly in the space of deep learning for AI.
But Aggarwal said that the company is now six to eight months into the process of building and designing its own silicon and expects to have its own chip in 2025.
