Bhavish Aggarwal, co-founder and CEO of Ola, unveiled Krutrim, a “made in India” huge language model and generative AI platform similar to Google’s Bard and OpenAI’s ChatGPT. Aggarwal’s AI initiative, Krutrim Si Design, is the owner and developer of the platform. It has the most comprehensive representation of Indian data and was trained on 2 trillion tokens, or textual information fragments.
Sanskrit for “artificial,” Krutrim is available in two models: the pro and base. This main version, like ChatGPT, can reply to public prompts and questions. It can produce text in ten Indian languages, including Marathi, Hindi, Bengali, Tamil, Kannada, Telugu, Odia, Gujarati, and Malayalam, and understand about 22 different languages. The startup plans to release Krutrim Pro, a larger variant designed to solve complicated problems, between January and March.
As of right now, Krutrim can be accessed by groups of people that sign up easily, but by January 2024, the business plans to make it available to everyone. Starting in February 2024, developers will also have access to the Krutrim APIs for experimentation and creativity.
“Given our multicultural and multilingual context, current AI models just can’t capture India’s culture, knowledge, and aspirations,” stated Aggarwal. We are pleased to present Krutrim, a business whose only goal is to develop AI for 1.4 billion Indians. “An AI model designed for India should recognize the distinctiveness and appropriate cultural context.” It must be taught using distinct data sets that are particular to us. Furthermore, he said, “it must be affordable for India and have cost structures that prioritize the country.”
The company is developing an AI computing stack internally and producing chips that are optimized for AI computing in addition to releasing its core AI model for generative AI applications. In order to power various AI infrastructures, models, and applications, the company has devised an architecture incorporating numerous chipsets.
According to the business, Krutrim outperforms GPT-4, OpenAI’s large language model that powers ChatGPT, in Indian languages by producing responses faster and with less computer power. It performs better in English than Meta’s large language model, Llama 2.
