12 hours Ago By Vallari Sanzgiri
Explainer Briefly Slides Stating that better AI models, over bigger ones, are the demand of the coming decade, Advisor to the IndiaAI Mission Aakrit Vaish, encouraged budding start-ups to work on small language models (SLMs) over large language models (LLMs). Speaking at the NVIDIA AI Summit 2024, Vaish discussed the popular notion that India is the ‘AI use-case capital of the world’ and asked attendees to consider what this means in terms of India’s prospects. Start-ups must address the data gap problem: While Vaish pushed for a focus on SLM creation, he listed high-quality data access as the key issue for the model development, including collection, digitization, crowdsourcing, validation, synthetic data generation, cleaning and labelling.
Particularly, in the case of Indic languages, Vaish said that good data sets are hard to find. This is so because Indic languages are fairly complex, with different morphological structures and syntax, which brings the cost of conversion/translation to around 10 times that of English. Arguing that India will be working with multi-modal AI models (combining various communication modes such as text, speech, and video) in the coming years, he maintained that start-ups must address this problem.
“India by design will be a multi-modal country. Voice and video will take the biggest part in our journey of building AI applications. Great speech and video solutions will far surpass other models,” he said.
Combination of LLMs and SLMs too ambitious for India: It is not a practical approach for India to work on LLMs and SLMs in tandem, said Vaish. “This is a view (sic) sitting in October 2024. For example, there may have been a reason to do this when ChatGPT 3.
5 came out but considering the current progress of established LLMs – it can be done but it won’t solve problems,” he said. Problem-solvers must work with domain experts: Vaish encouraged start-ups to allow interaction between their AI developers and domain experts. For example, he asked individuals developing AI in diagnostics to speak with experts in this field.
“You will find the answer to building such specific models relies on research and development. So rather than just considering AI as the plugin, you need to sit with domain experts who’ll tell you, you need to look back to the base issues,” said Vaish. Also Read:.
