Homegrown company Ola co-founder and chief executive Bhavish Aggarwal has entered the much-hyped artificial intelligence race with the introduction of “India’s first AI computing stack” named Krutrim, a multilingual AI model, which aims to rival bigger names in the industry like OpenAI’s ChatGPT and Google’s Bard.
‘Krutrim’ translates to "artificial" in Sanskrit and comprises two versions – Krutrim, the base model trained on 2 trillion tokens and unique datasets, and ‘Krutrim Pro’, a more sophisticated model set to launch in the early part of the next year. It’ll be designed for advanced problem-solving and task execution.
At the launch event on Friday, Aggarwal demonstrated an AI chatbot, powered by Krutrim, akin to ChatGPT and Bard, which can comprehend 22 Indian languages and generate text in 10 Indian languages.
The Krutrim team operates in both India and the Bay Area (US). The company claims the model has been trained in just three months, and that it “outperforms” OpenAI’s widely popular GPT4 consistently in Indic languages. It understands multiple Indian languages and can mix them. It is voice-enabled, and the model talks back, says the company.
“We are extremely happy to launch India’s first AI computing stack that is unique to our context, connecting our future to our roots. AI will certainly transform everything, making India the most productive, efficient and empowered economy in the world,” Aggarwal posts on X.
Aggarwal also talks about building AI infrastructure. “AI is the soul, but silicon & infrastructure is the body. For a well-run AI computing story and business, you need to build AI models & infrastructure, and silicon, and integrate them tightly, that is our endeavour.”
He also says data centres use a lot of energy across the world, and that the company has created a technology to bring down the energy cost of data centres.
He emphasises the technology used in Krutrim has ‘PUE’ of 1.1, which means only 10% of energy is wasted. “Data centres have a metric called PUE and data centres in India have a PUE of 1.5, which means about 50% of energy is wasted on top of one unit used for a useful computation.”
He says Krutrim Pro will be a multimodal trained model, which means all the modalities will be taken as inputs together and the algorithm will be able to train across modalities. “We are building Krutrim as a platform for developers across India and other parts of the world. Developer APIs go live in February 2024,” he underscores, adding that Ola group companies were already using Krutrim for internal workload.
The company aims to release its AI-focused system in package (SIP) by 2025, with prototypes anticipated by mid-2024. An SIP involves individual chips housed within a single package, each serving a distinct function, typically employed to conserve space and reduce installation expenses. This initiative aligns with the company's aspirations to construct a homegrown supercomputer tailored for AI within the next few years, as highlighted by Aggarwal.