Two hidden trends that matter in the Ai world right now;
Continual Learning
and
World Models
This is what India has to focus on. The fact is that GPTs and their products, the LLMs and their core architectural constraints, the Transformers are all at a dead end.
The LLM world doesn't know how to impact the real economy or productivity on this path. They are throwing the sink at it and hoping some magic will come out of it just by deploying the sheer compute at scale that costs billions and trillions of dollars (this was the classic mistake of early dotcom era bust too). Yet they are nowhere near solving the 2 basic problems of "bad generalisations" and even worse "sample efficiency". In fact, if anything they may hit mathematical limitations of RL soon.
India needs to bet on an alternate path, an alternate destiny for mankind. Fund a new architecture. Liberate intelligence from the abstract layer of Human language constraints and focus on narrow intelligence dominance in specific domains with continual learning. No preformed biases, no pre-trained models, no frozen training weights. This is a revolution that can differentiate Bharat from US and China, it can give us that edge that nobody currently has, in less than 2 years!
But current data-centre deployment in India is minuscule compared to the world at around 1.5GW (US has 54GW and China 32GW). If India can strategically plan to deploy the next 2-5GW of hyperscale datacenter for CL and world models with more efficiency rather than just throwing all the labelled datasets at it and instead deriving intelligence beyond mere human language understanding, then there is a potential path to achieve massive intelligence gains in specific domains at a fraction of the cost and compute.
Interestingly, this aligns with our ancient wisdom of the Sindhu-Saraswati civilisation. Most of us think of Sanskrit as a language born to communicate ideas amongst us humans (like all other languages), but what if you think of Sanskrit as a layer of intelligence beyond the mundane human existence? in fact, it was probably a layer specifically designed to capture intelligence from a higher level. Vedas were then downstream to that intelligence capture and so very complex in nature that most parts have either been lost or remain only partially deciphered till date.
Can Bharat achieve such a paradigm today? Let the world focus all their energies on the current backprop-Transformer-GPT-RL-LLM models and throw the sink of data and compute to arrive at a magical intelligence layer, maybe they will achieve success. But if they won’t (which is more likely), we will be in a commanding position to walk on an alternate path in 2 years. As a nation we need to institutionally and strategically focus on deploying compute at scale towards CL and world models to achieve narrow AGI in specific domains. Can we, for once, find the courage to do something different and lead the world rather than be led?