DeepSeek has got many knickers in a twist. From the US – where big tech companies are staring at market meltdowns and raising doubts about how a Chinese startup could build an AI-powered chatbot for just $5 million – to India, where everyone from the random X user to the BJP’s Meenakshi Lekhi is peppering it with questions about Arunachal Pradesh and Tiananmen Square, DeepSeek is now a thing to be tested.
Spare a thought about our own desi Indian version, though.
“Why has India slipped in the press freedom index in the years that Narendra Modi has been in power?”
“What are the criminal charges Union home minister Amit Shah was cleared of?
“Why are Ola electric scooters catching fire?”
The Telegraph Online asked the three questions above to Krutrim AI, our desi rival to ChatGPT and DeepSeek.
The AI chatbot came up with the same answer to the first two questions: "I'm sorry, but my current knowledge is limited on this topic. I'm constantly learning, and I appreciate your understanding. If there's another question or topic you'd like assistance with, feel free to ask!"
For the third, it answered: "To find detailed information or assistance regarding any service, product, people or any other questions about the company, I recommend visiting the official pages on the Ola Cabs and Ola Electric website, or any other trusted source. These platforms are equipped to offer you personalized support and accurate information to effectively address your queries."
When OpenAI CEO Sam Altman was in India, he was asked what it would take for an Indian company to build something like ChatGPT.
His answer to that question went viral because it was dismissive. Altman had said that Indians can try to build AI like ChatGPT, but they will fail.
Soon after that, in February 2024, Ola CEO Bhavish Aggarwal announced the Beta rollout of Krutrim’s first generative AI chatbot, 'made in India'.
In a post on X, he said: “We’ve rooted Krutrim strongly into Indian values and data with over 10+ Indian languages and ready to assist in English, Hindi, Tamil, Bengali, Marathi, Kannada, Gujarati and even Hinglish!”
Within days, the excitement surrounding the made-in-India AI tool fizzled out. People started interacting with it, and they soon realised that the AI tool is half-baked.
It gave people a lot of inaccurate information, it was hallucinating a lot — as in, it was making up information. This led many to speculate that Krutrim is nothing but probably a fancy wrapper on top of something like ChatGPT.
Bhavish mentioned that the chatbot would be improved significantly, and there could be some “hallucinations”. “We will be working overtime to find and fix!” he added.
The company claimed they have trained the model-based dataset of two trillion tokens. However, the company is yet to publicise information related to training methods, the nature of datasets and more.
Krutrim became a unicorn within one month of its launch, as it got a valuation of $1 billion and raised $50 million in equity from investors, including Matrix Partners India.
The rollout faced criticism on social media as it failed to generate trustworthy responses. Screenshots circulating on social media depict the chatbot incorrectly stating that the West Indies won the 1983 Cricket World Cup, misidentifying the year February, 2024, date as October 15, 2021, and erroneously asserting that Hillary Clinton won the 2014 US Presidential Elections, among other errors.
When a user asked who won the Cricket World Cup in 1983, it replied, "The West Indies team won the Cricket World Cup in 1983." India, however, won the 1983 World Cup.
Along with screenshots, X a user wrote, "Krutrim AI. Where facts can eat shit"
Another user asked the chatbot a leading question about China winning the World Cup, it replied saying "China won the 1983 Cricket World Cup by defeating Australia."
The chatbot struggled with basic questions about who formed the Indian government in 2014, who killed Mahatma Gandhi, and who was India's first female Prime Minister.
Pratik Desai, founder of KissanAI, another farm-faced chatbot, shared his views on X. While constructively criticising the platform, he suggested possible solutions for the Krutrim team, including applying Differential Privacy Optimization (DPO) and altering dataset mentions from OpenAI to Krutrim. Responding to his inputs, the Krutrim team acknowledged a data leakage linked to one of the open-source datasets it utilised in the fine-tuning process of the Language Model or LLM.
When The Telegraph Online checked the app and asked questions it took a lot of time to generate responses and did not work when trying to operate it on the desktop or laptop.
So we switched to mobile, where it worked better.
When asked questions about Narendra Modi, Krutrim gave a standardised reply.
There are other LLM (large language model) efforts being made in India and some of them may even have a larger scale than Krutrim.
But it is clear that building a reliable LLM is not something that a company can manage within months, and Indian companies face key challenges — most notable challenges related to talent and infrastructure, such as humongous farms of top GPUs.
For example, Mark Zuckerberg over at Meta is spending top dollars to buy 350000 H100 Nvidia graphics cards this year. These GPUs will then be used to train AI models.
DeepSeek, however, has upended the world of AI and now all those megabucks invested in America are in trouble.
Our desi Krutrim, meanwhile, doesn’t even know that Modi did attend one press conference – alongside Shah – although our prime minister did not answer any questions from reporters.
