Panel Discussion: Global India AI Summit 2024: Large Language Models by IndiaAI
Summary of the Panel Discussion: India AI Large Language Models
The session was about
- Larning and sharing knowledge on Large Language Models (LLMs) and Large Multimodal Models (LMMs), with the aim to understand the unique linguistic and cultural diversity inherent to India,
- Examining how LLMs can effectively address challenges associated with multilingualism.
- The ethical considerations and biases linked to these models,
- Promoting a discussion on responsible AI practices, including fairness, inclusiveness, misinformation mitigation, and intellectual property rights within diverse cultural contexts.
- Collaborative opportunities among indigenous communities, academia, industry, and startups in the creation of indigenous foundational models.
The discussion highlighted India’s unique challenges and opportunities in building a comprehensive AI ecosystem, underscoring the need for collaboration, innovation, and inclusivity.
Mr. Amitabh Nag (CEO, Digital India Bhashini, National Language Translation Mission)
- Seven Pillars of India AI Mission: Compute infrastructure, foundation models, datasets, application development & usecases, future skills, startups/entpreneurship around usecases, and safety/trustworthiness.
- India Innovation Center: Focuses on contextualized AI models, unique Indian use cases, and skill development.
- AI’s National Impact: Highlights AI as a transformative force akin to the Industrial Revolution, with a focus on democratizing data for public good.
Mr. Shri Nivas Narayan (VP, OpenAI)
- Global Perspective on AI: Large language models (LLMs) enhance productivity by 40-50% across diverse sectors.
- Actionable AI: Moving from answering questions to performing actions like function calling, with examples in customer service.
- Indian Context: Focus on cost reduction and linguistic inclusion (GPT models optimized for Indic languages).
- Safety Measures: Prioritizing bias mitigation, robust evaluation, and human-centric alignment strategies.
Ms. Shalini Kapoor (Chief Technologist, AWS)
- Localized Data Challenges: Addressed data scarcity, bias mitigation, and the necessity for domain-specific datasets.
- Micro LLMs: Shared insights on building micro-LLMs for agriculture and healthcare, focusing on region-specific and linguistic diversity.
- Scalable Solutions: Advocated for simplifying API access and increasing use-case based AI applications.
Dr. Mohit Sewak (AI Researcher and Developer Relation Manager, NVIDIA)
- Multimodal Models: Advocated for moving beyond text-based LLMs to include speech and video for better context understanding.
- Synthetic Data: Proposed generating synthetic datasets to supplement limited real-world data.
- Resource Efficiency: Highlighted the need for compression technologies to scale LLMs in resource-constrained environments.
Dr. Kallika Bali (Principal Researchers, Microsoft Research)
- Cultural Alignment: Discussed challenges in aligning LLMs with India’s linguistic and cultural diversity.
- Bias Mitigation: Acknowledged the complexity of eliminating bias and the need for sensitive data collection strategies.
- Knowledge Integration: Proposed federated and fine-tuned models for incorporating indigenous knowledge.
Prof. Ganesh Ramakrishnan (IIT Mumbai)
- Collaborative Frameworks: Emphasized public-private partnerships for advancing LLM research in India.
- Algorithmic Innovations: Highlighted leveraging linguistic similarities across Indian languages to optimize sparse datasets.
- Bhat GPT Initiative: Focused on open-source frameworks for speech, text, and domain-specific applications.
Mr. Pratyush Kumar (Co-Founder, Sarvam AI)
- Open Source in AI: Advocates open data, tools, and model-building processes to democratize AI.
- Challenges in Scale: Highlighted the need for large datasets, computational power, and skilled talent for LLM development.
- Future Approach: Combining models and fine-tuning for specific applications to maximize utility.
Insights on Startups & Innovation
- Challenges: High costs of token usage, reliance on foreign models, limited access to computational resources, and IP licensing constraints.
- Solutions: Development of open foundation models, focus on fine-tuning, and increased investment in local research.
Key Takeaways
- Multimodal AI: India’s LLMs should integrate voice, video, and context to address cultural and linguistic diversity.
- Synthetic Data & Miniaturization: Using synthetic datasets and model compression can optimize costs and resource usage.
- Use Case-Centric Models: Tailoring LLMs to specific domains like agriculture, healthcare, and education ensures scalability and relevance.