Aravind Srinivas is Right on India Training its Foundation Models Debate
From Abhivardhan, our Chairperson
I came across these 2 important perspectives by Aravind Srinivas, and I completely agree with him, when it comes to AI development and research in India.
In recap: Aravind implies that Indians cannot afford to ignore model training skills, and merely focus on building on top of existing models.
He also says that while thinking models cost a lot of money for the purposes of investment, DeepSeek managed itself with merely 2,048 GPUs, which is an interesting achievement.
I believe that both Ministry of Electronics and Information Technology and India Inc., should review their internal and external policy goals on AI development and research on 2 counts:
1️⃣ Don't leave the "use case capital" objective.
2️⃣ Don't obsess around 1️⃣ and promote ways to encourage alternative AI research & development avenues.
Now, let me emphasise on some legal and policy aspects as to why should India change its stance. Here's my humble take, as a legal specialist in AI and digital technologies.
1️⃣ Wrappers, or deliverables built on top of existing models will not necessarily have relevant intellectual property protections, especially when it comes to copyright of AI-generated content, trade secrets associated with core models and the patentability of AI systems. I am saying it again - AI Patentability is going to be a huge risk for Indian end-users, and business end-users, and we should have the leverage to negotiate with any MNC or company with leverage over infrastructure layers, digital value & supply chains and data - when it comes to access rights.
2️⃣ We need a CERN-like body for artificial intelligence research, and not an AI Garage in Bharat. If the United States does not wish to pursue a CERN-like approach for international cooperation, on a multi-party basis, then India, UAE or Singapore should seek these options. UAE can easily leverage this because it only helps in their own AI industry-policy goals. For India, it could be a win-win if done right because it creates an anchor for more research avenues and trust-based partnerships.
3️⃣ Avoid coming up with an all-comprehensive regulation on artificial intelligence right now. Yes, cover some liability-accountability issues with multi-modal synthetic content (deepfakes) but come up instead with a Digital India Act than an AI Act. I am saying this as the author of India's first privately proposed AI regulation (aiact.in), because the feedback I have received on the second-order effects of AI regulation in India, clearly show that gradual demarcation & adjudicatory clarity can only drive better regulation in the coming years, not now.
4️⃣ Discourage AI hype at all costs, and make AI research discussions cool again. And yeah, check the Durgapur AI Principles by the Indian Society of Artificial Intelligence and Law as well: https://www.isail.in/the-bharat-pacific-principles/durgapur
What do you think?