Krutrim is set to unveil its groundbreaking generative artificial intelligence (AI) chatbot application, marking a significant leap in the realm of digital interactions. Founded by Bhavish Aggarwal, the startup aspires to redefine the landscape of AI technology. Aggarwal recently shared on the social media platform X (previously Twitter) the anticipated launch timeline of the AI chatbot, providing a sneak peek into its capabilities through shared screenshots.
The startup emphasizes that all forthcoming AI products are designed to comprehend the diverse linguistic landscape of India, covering a remarkable 22 scheduled languages. Furthermore, these innovative products will showcase content-generation proficiency across eight languages.
On February 4, Aggarwal hinted at the imminent release of the @Krutrim app, disclosing that it is in the final stages of testing and is scheduled for launch in the coming week. Despite expressing contentment with the app's current state, Aggarwal committed to ongoing refinement and enhancement of the AI models post-launch.
Aggarwal showcased the versatility of the chatbot through a series of posts, demonstrating its adeptness in handling basic inquiries and offering insightful suggestions. However, one specific interaction stirred controversy when users questioned the chatbot about India's status before independence. The response, asserting India's existence as a sovereign entity prior to the British Raj, generated debate and comparisons with OpenAI's ChatGPT, which presented a different historical perspective. Critics accused the Krutrim AI chatbot of exhibiting signs of "AI hallucination," wherein large language models generate inaccurate or fictional information. Users favored ChatGPT for its alignment with historical accuracy, and concerns were raised about a user interface flaw, particularly the misalignment of the 'Send' icon with the text box.
Krutrim disclosed details about its Language Model Models (LLMs), encompassing the Krutrim base and Krutrim Pro variants, with the latter being promoted as a multimodal foundational model. These models underwent extensive training on a corpus exceeding two trillion Indic language tokens. While Krutrim hinted at the potential inclusion of voice-enabled functionality for its generative AI apps, its availability at launch remains uncertain.
The company also provided insights into the performance benchmarks of its AI models, comparing them against Meta Llama 2 7B, a coding assistant model trained on over 500 billion tokens. Despite outperforming Meta Llama 2 7B in various benchmarks such as Arc, Hellaswag, and Copa, Krutrim's performance lagged in benchmarks like MMLU, Bigbench, and Lambada_open