India's AI Edge: Applications, Not LLMs

 What is India’s strength in AI? 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗔𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀
 
India may be lagging behind other countries in terms of fundamental AI research but it punches above its weight when it comes to building AI applications -

1️⃣ Greater adoption of Application models vs. Foundational LLMs

The number of downloads of models (on Hugging Face) focused on Indic use cases in the last month from today show up to a staggering ~90X greater adoption of smaller application models (largely developed by AI4Bhārat) vs. foundational LLMs (based on Sarvam's Sarvam-1 and Krutrim's Krutrim-2-instruct).

These are the use cases for each of the Application models:
- indictrans2-indic-en-1B: translation from 22 Indian languages to English
- indic-bert: language model and embeddings for 12 Indian languages
- indicBERtv2-MLM-only: multilingual language model for 23 languages
- indictrans2-en-indic-1B: translation from English to 22 Indian languages
- indic-sentence-bert-nli: sentence similarity across 10 Indian languages

👉 The application models are typically “small” models ranging from ~300M to ~1B parameters in size vs. the foundational LLMs that are 2 to 12B parameters in size. This also indicates that for solving India-specific use cases, we do not necessarily need “large” models; and the development of small, fine-tuned models on top of leading open-source LLMs from global companies is a good strategy to solve for niche domestic use cases.

2️⃣ India publishes ~2x more at Application vs. Theoretical AI Conferences

Of the top 10 AI conferences, India publishes ~2 times more papers in conferences like AAAI and EMNLP that are more application focused vs. the more theory focused conferences like NeurIPS, ICML and ICLR (source: Mahajan, Bhasin & Aggarwal, 2024).

3️⃣ AI4Bharat's significant contribution to India's R&D capabilities

The team at AI4Bhārat in collaboration with Microsoft India, Indian Institute of Technology, Madras, EkStep Foundation and others has done a stellar job in collecting, curating and processing local language datasets to unlock significant value for both public and private sector organisations. By using these datasets to fine-tune Transformer-based models like BERT & ALBERT, they have created models that often outperform models from global companies on niche NLP use cases. Additionally, this work has led to the formation of Sarvam as a venture-backed startup focused on the commercialisation of this research.

4️⃣ Growth of India's AI Startups
The rise of generativeAI startups from India that are developing on top of the global foundational LLMs further highlights our strength in building AI applications. These startups are not only solving domestic use cases but also catering to global markets.

5️⃣ Conclusions
India’s prowess in building AI applications is highly commendable. One way to make our mark on the global AI ecosystem is by standing on the shoulder of giants to build impactful products.