Ola, the Indian unicorn that introduced the Krutrim chatbot in February, has launched its AI cloud platform for developers, aiming to provide an alternative to OpenAI’s ChatGPT.
Krutrim, meaning “artificial” in Hindi, symbolizes AI tailored for an Indian audience. However, the chatbot’s launch drew mixed feedback, with users reporting erroneous, unintentionally amusing, and misleading responses.
On May 2nd, Ola founder Bhavish Aggarwal unveiled the AI cloud and developer platform, offering models as a service and GPU instances for developers to fine-tune and deploy large language models (LLMs). Speaking at the event, Aggarwal emphasized his belief that “AI is the soul, and silicon and data centers are the body.” Yet, the AI model, supposedly the platform’s soul, leaves much to be desired.
The platform features a comparison tool to evaluate model responses between Krutrim’s proprietary model, Meta’s LLaMA 3, and Mistral. Screenshots reveal response quality, but Ola’s ambitious vision to deliver a comprehensive AI platform to Indian developers remains in question. They aim to provide everything from custom silicon to a marketplace of AI apps for SaaS customers.
During the launch event, Aggarwal criticized Google and Microsoft for not subsidizing their cloud platform costs for Indian developers. He claimed Krutrim would be the most affordable platform, challenging global hyperscalers.
While Ola’s ambitions are commendable, certain facts warrant consideration. It’s incredibly challenging for a single entity to design custom AI chips, train foundational models, build AI cloud infrastructure, and deliver a comprehensive developer platform. In today’s AI landscape, most companies focus on specialized aspects of the stack. Meta, for instance, aims to deliver powerful LLMs like Llama but leaves deployment and scaling to cloud providers. Even OpenAI relies heavily on Microsoft Azure for infrastructure while refining its GPT model.
Companies like Google and Microsoft, with dedicated research teams, manage the full stack. Google’s DeepMind contributes to models like Gemini, while Google Cloud provides infrastructure through Tensor Processing Units and Vertex AI. Microsoft and AWS similarly handle models, platforms, and infrastructure backed by decades of research and proven track record of delivering cloud services.
Ola’s claims seem overly ambitious without a proven AI research record. They should concentrate either on training specialized models for Indic languages or on building robust infrastructure and platforms. Even for a unicorn, achieving both is an ambitious and challenging goal.
Jaspreet Bindra, a leading Indian voice in generative AI and an author of the book The TechWhisperer, emphasized the need for an LLM catering to the diverse Indian audience but cautioned that building a complete stack from scratch—including silicon, capable models, infrastructure, and platform—is extremely challenging for a single entity.
My Experience with Krutrim AI Cloud
As an enthusiast and an analyst, I signed up for the Krutrim AI cloud platform to take it for a spin. To say the least, the experience left me dissatisfied. The developer experience didn’t live up to the lofty promises made by Bhavish and his team.
First things first, the core LLM, Krutrim-spectre-v2, which is touted as the first Indian LLM, lacks quality. It not only hallucinates but also responds with incoherent and nonsensical answers even to simple questions.
Ola did not publish a model card explaining the model’s technical details. It is unclear whether it is an instruction-tuned model or how it performs against standard benchmarks.
The documentation and APIs are not ready for prime time. They are half-baked and do not work for the most part. The platform lacks the required finesse to become comparable with some of the other platforms, such as OpenAI Playground, Anthropic Console or Mistral Platform.
My key takeaway is that the model and the platform have a long way to go before they impress the Indian developers.
Ola’s ambitious vision to disrupt the AI cloud space with Krutrim is notable, but the platform’s shortcomings highlight a significant gap between aspiration and reality.
While India urgently needs AI solutions tailored for its diverse audience, Krutrim’s current limitations in documentation, model quality, and developer experience hinder its potential to rival established players. Rather than attempting to dominate the full stack, Ola should refine its strategy to focus on key strengths that align with India’s needs. Until then, developers will remain skeptical, and Krutrim’s journey to deliver on its promises remains an uphill battle.
