Should India build its own LLM (Large Language Model such as ChatGPT or Gemini Advanced)? This question has begun to bother the tech eco-system. Some feel that it’s enough to tap into the existing LLMs to make the best use of them instead of building the resource-guzzling LLMs. Others, however, think we need our own LLMs that factor in our needs instead of depending on the LLMs that are trained on material that is anything but Indian.
While TCS CEO K Krithivasan says that he doesn’t see any huge incremental advantage in building one’s own LLMs since there are so many already available, Tech Mahindra announced that it is building Project Indus, an indigenous foundational model designed to converse in a multitude of Indic languages and dialects.
A few start-ups have developed functional LLMs, while others are showing interest to build some more even as the debate rages whether India should build its own LLM given the constraints of resources.
Nandan Nilekani argues that India doesn’t need to build its own LLMs. Instead, he suggests letting the major tech companies in Silicon Valley handle it, while Indians should focus on creating applications that leverage these LLMs to address global and emerging market challenges in areas like healthcare, education, and more.
China’s example
“Aravind Srinivas, founder of Perplexity, has expressed a differing opinion on X (formerly Twitter). He believes this perspective is wrong. According to Shirinyan, India should indeed invest in building its own LLM. He cites China’s example of developing low-cost LLMs as a demonstration of what’s possible,” Jaspreet Bindra of AI & Beyond, points out.
“In my view, it is high time for India to consider building its own LLM, drawing on contextual data from numerous resources. India has a treasure trove of data, such as archival materials from Doordarshan, All India Radio, and a wealth of internet data across various languages. These resources could form the basis for creating an LLM that is far more contextual and vernacularly suited to India’s needs,” he felt.
Kiran Chandra, Centre Head of viswam.ai, a collaborative LLM platform being developed by IIIT-Hyderabad and Sweccha, strongly felt that India should essentially have its own LLMs.
He, however, sees two challenges: having quality datasets and computing capacities.
“We aim to solve the dataset challenge through community-driven data collection. We are engaging in community-led initiatives to document and collect these resources. To date, volunteer efforts have successfully collected up to 50 million tokens on multiple occasions,” he points out.
Better algorithms
“Regarding compute, we believe that better algorithms can significantly reduce the reliance on high computational power. The current approach by organisations like OpenAI, which involves brute-force methods requiring massive GPU clusters, is not the only way to build AI technologies,” he said.
Chakri Gottemukkala, Co-Founder and CEO of o9, a supply chain management solutions unicorn, prefers a multi-model approach. The start-up built an architecture to use both public LLMs and private LLMs trained on its domain-specific data. This allows them to leverage LLMs for general knowledge and utilise specialised models for complex industry-specific tasks.
Anushree Verma, Director Analyst, Gartner, felt that it all depends on the vertical that one operates in. “If it is a healthcare or a financial services firm which is highly regulated, it will not want to rely on external ones ideally. Building own LLMs is a cost intensive exercise. So, unless you are looking at your own proprietary expertise or looking at filing patents, you would not look at building your own,” she said.
“The resources in terms of talent and expertise in the organisations would also decide if you want to build your own as opposed to relying on external ones. Some of the organisations also want to rely on the expertise and capabilities of the LLM provider, because they will want to embed the latest security and other practices integrated,” she said.
“If it is a critical task which has any confidential purpose or a goal that is critical to an organisation, it may want to build its own LLM,” she said.