Large language models (LLMs) and the emergence of transformers resulted in the keenness of people globally to adopt tools like ChatGPT in their everyday life. However, large energy consumption, high dependence on cloud, output having biases emanating from large databases and data ownership related-concerns led to efforts on finding solutions and identifying alternatives to the LLMs. Proliferation of smartphones, IoT devices and availability of transformers library acted as catalysts for small language models (SLMs). From building domain-specific SLMs and being offline models, the focus is now aimed at building multi-model agentic SLMs.
India was one of the first countries to start work on the alternative to LLMs focussing on vernacular languages and cost efficiency. Just like India, other developing nations are also attempting to build SLMs. Indonesia, Philippines and Peru are developing language assistants to support the diverse languages spoken in these countries. Nigeria and Kenya are building SLMs for agri advisory support.
Aim of India AI Stack
Bhashini is the national platform designed as government-led open architecture that enables AI-based language translation and speech recognition in multiple Indian languages. Efforts are under way to create rural specific SLMs for education, health, agriculture and employment. The objective of India AI Stack is to reduce dependence on foreign trained models and get aligned with Indian socio-cultural sensitivities as well as make them device-light suitable for low connectivity territories. Sarvam AI focuses on speech-to-text in different languages, while Gram Vaani uses SLMs to give rural news and a mode for grievance redressal in Hindi & Bhojpuri.
SLMS will foster innovation in small towns and make AI inclusive. By having SLMs hosted on devices, it is possible for the citizens to access services such as weather bulletin, crop advisory, financial information and government schemes. A big concern is about personal data being moved outside of India. SLMs allow government, banks, hospitals and other legal entities to retain the data locally and ensure its privacy.
Advantages of Large language models (LLMs)
Quality of education delivery and access could be energised significantly with the help of SLMs. Students could be provided with personalised learning pathways and counselling in their respective languages. Skilling as required for each location based on the availability of jobs and requirement of talent could be delivered just in time, in local languages and make it affordable.
Success of SLMs also depends upon how the data is labelled and catalogued for the purpose of data modelling. Careful planning and long term thinking is required for each domain with multiple stakeholders involved in this initiative to create a strong foundation which is a pre-requisite for the successful deployment of SLMs. With India’s advantage on low cost and access to the large, scalable AI talent pool, its expertise in building SLMs for different domains, devices and languages would be much sought after by other countries once we make headway with the deployment across various sectors.
The writer is chairperson, GTT Foundation
