The CEO of Ola, Bhavish Aggarwal, recently launched Krutrim (which means artificial in Sanskrit). It is also claimed to be "India's first full-stack AI" solution. At first glance, this platform bears a striking resemblance to ChatGPT - at least in terms of the UI/UX part - but only in green.

Aggarwal claims that Krutrim AI outperforms GPT-4 in various Indian languages. He says it is trained on 20 trillion tokens and is capable of understanding over 20 Indian languages, generating content in approximately 10 languages including Marathi, Hindi, Bengali, Tamil, Kannada, Telugu, Oriya, Gujarati, and Malayalam.

"Today, all AI models known as LLM are mostly trained on English, but language is not just text. Language is also a carrier of cultural values, background, and ethics, which current AI models fail to capture our culture, knowledge, and aspirations, given our diverse culture and multilingual heritage," says Aggarwal.

Does Krutrim AI really exist?

Confidently, Aggarwal says, "This is not just a wrapper on an existing API. It's not just a slight tweak, it's what people in the AI industry call adding a little extra dataset to an existing model." Some existing APIs include GPT-4, Llama 2, Gemini, etc.

"This is foundational work that starts at the scientific level, changing the mathematics and algorithms of the model to make it more suitable for Indian languages," explains Aggarwal.

During the Krutrim launch event, Aggarwal even took a dig at an American AI company, saying, "We hear them talk about how they are building AI for humanity and how they are doing it democratically."

Ola has created India's first generative AI model from scratch, which is commendable, but the possibility of this achievement seems quite rare. Even other major players outside the Indian market have taken more time and resources to build their generative AI models from scratch.

Ola's CEO claims that after Krutrim, they will release their first multimodal Krutrim Pro in the next quarter. He says, for Krutrim Pro, "all dimensions and modalities will be taken as input simultaneously, and the algorithm will be able to train across modalities. This is different from training separate text models, separate speech models, and separate vision models."

Interestingly, OpenAI announced multimodal GPT-4 in March, but in reality, GPT-Vision was not released until September - that is, over six months later.

Considering how Krutrim developed an LLM model in just three months compared to Indian counterparts like Corover AI, Sarvam AI, and Kissan AI, who built their models based on Llama 2 or GPT-4, they should have taken less time theoretically compared to Krutrim.

In addition, Elon Musk took four months to build Grok, and xAI is dealing with GPU shortages, which Oracle also highlighted in its latest earnings call. Surprisingly, Aggarwal did not disclose the number of GPUs purchased for training their model throughout the event. Instead, he said they are building all these capabilities in-house.

Furthermore, Aggarwal did not reveal the dataset on which Krutrim AI was trained. Training models on multiple languages is not an easy task, especially considering the increasing costs.

Particularly, languages with complex structures and scripts like Hindi, Kannada, or Telugu require different numbers of tokens. In comparison, the simplicity of English requires fewer tokens. The economic impact of tokenization differences is significant. The cost of training and using AI models depends on the number of tokens, computation, and cloud costs.

Krutrim has raised $24 million in debt from Matrix Partners, who are also shareholders of Ola Electric. However, Aggarwal clarified earlier that Krutrim is an entity independent of Ola.

Regardless of the circumstances, Ola has not yet released any research papers or details, including the dataset on which it was trained and the team leading this initiative, leaving them in a dilemma of whether to call themselves Ola or Krutrim.

All of this seems like a marketing gimmick aimed at securing additional funding. Even without launching a solid product in the market, the demo does not seem convincing.