- The company claims the chatbot has been trained on over 2 trillion tokens
- Krutrim AI has built two LLMs, Krutrim base and Krutrim Pro
- Krutrim has claimed to manufacture chips optimised for AI computation
Krutrim Si Designs, otherwise known as Krutrim, will soon launch its first product — a generative artificial intelligence (AI) chatbot app. The founder of the startup, Bhavish Aggarwal took to X (formerly known as Twitter) to reveal the launch timeline of its first AI product, and shared screenshots to preview its capabilities. As per the startup, all the AI products launched will be able to understand the 22 scheduled languages of India. The products will also be able to generate content in eight languages.
Aggarwal posted on his social media handle on February 4, and posted, “Testing the @Krutrim app before release! Releasing next week. Looking good though we will keep improving the models post launch too.” Later, in a series of several posts, he also showcased the generative capabilities of the chatbot.
While a couple of posts highlighted that the chatbot is capable of answering basic queries and even offering suggestions, one particular post received a lot of flak from netizens.
Upon asking the Krutrim AI chatbot whether India was a country before independence, the app responded with “Yes, India was a country before the British Raj. It had its own political system, culture, language, religion, and tradition.” Captioning it with “This is why we need to build India’s own AI”, Aggarwal also posted the response of OpenAI’s ChatGPT which stated, “Before the establishment of the British Raj, the region that is now known as India was not a single unified country in the modern sense but a collection of various kingdoms, empires, and principalities.”
Many users responded to the post calling it a case of AI hallucination (a phenomenon where large language models (LLMs) create false or non-existent information), and that ChatGPT’s response was more accurate. Some users also pointed out a UI flaw where the ‘Send’ icon was not aligned with the empty spaces in the text box.
As per the company, the AI firm has built a family of LLMs including Krutrim base and Krutrim Pro, the latter of which is claimed to be a multimodal foundational model. The models have been trained on over two trillion indic language tokens. The startup has also said that its generative AI apps will be voice-enabled, although it is not known if the feature will be added at launch. On its website, the startup has also posted its AI model’s benchmark scores, comparing them against the Meta Llama 2 7B model which has been trained on more than 500 billion tokens and is a coding assistant. The chart posted shows Krutrim beating Llama 2 7B on the Arc, Hellaswag, Copa, Jeopardy, Piqa and other benchmarks. However, it performed worse on MMLU, Bigbench, and Lambada_open benchmarks.