GenAI & India's Incoherence
On the gap between building AI and actually using it.
On 12th June, the US government banned two variants of the LLM Claude — Fable 5 & Mythos 5. Only for foreign/non-US nationals. This raised many questions related to AI sovereignty.
AI sovereignty, loosely speaking, is the ability not to be entirely dependent on foreign models for your critical infrastructure, your data, and your decisions.
I am not gonna dive deep into that, but would rather dive a bit deeper into what and how Indian tech circles reacted to it. And moreover, what I think we are doing completely wrong.
I came across this post on X by Sekar Vembu (not sure if that is his real name), which encapsulates my thinking on this topic to some extent. Here is an excerpt from that:
Neither the “nationalists” nor their critics, often branded as “globalists” or worse, “anti-nationals,” seem interested in the solution that would truly benefit India and its diverse people, who are desperate for honest opportunities to use their talent, work hard, and succeed. The real moral solution for India is an ultra-minimal government combined with a maximally open and free market, operating under a system of law and justice that everyone can trust.
The values of socialism are so ingrained in us that we often turn to the Government for answers. The irony is this: even by this standard, we are failing. This is not happening because we are too socialist, but because we are not coherently anything. Or rather, we haven’t figured out what we are.
Before I write more about the incoherence, let’s actually understand what a coherent pipeline looks like.
Let’s look at the US, for example. They have a proper research talent pipeline, a lab-to-startup pipeline already in place. They have world-class research focused public & private universities where the focus has been to do precisely this. What we see today with Anthropic, OpenAI, etc., is a cumulative result. Not a singular data point. The focus of the Indian state should be to create an environment such that these research-heavy pipelines do exist. Our R&D spend (as a % of GDP) is way below that of many developing countries. The focus at the grassroots level is simply not there.
But at the same time, the government can’t really magically create 10 state-of-the-art companies just by virtue of some central funds, which is essentially what we are asking of them. And here, there can be a better way.
Let’s forget about the US for a moment. Come to Asia. Our neighbour, China. They came to the show pretty late, but as soon as the cutting-edge DeepSeek model was out & made headlines, the Chinese government went into full socialist mode. They made sure GenAI adoption happened. Many Chinese companies have already pushed “all-in-one” DeepSeek AI servers. They have set a target of 70% AI adoption for various government and private organisations.
Now compare that to India. Even with our ingrained socialistic values, we are not socialist enough to pull this off. We have Sarvam, the only leading GenAI company right now. But how many Indian companies or government organisations are actually using it? On the contrary, Indian conglomerates have announced mega deals with US-based companies.1
This is not to say that the government should coerce Indian companies into using Sarvam for everything. Absolutely not. (At the end of the day, the better models would win. As they should.) But they can easily put two of them together and initiate something. Or better, start using it for the PSUs that they are running - actively!
Maybe private insurers don't need to move immediately. But LIC can start, and their eventual success will encourage the rest.2
All these talks of AI sovereignty would bring absolutely no value if Indian companies aren’t even using it. Because a big part of the software lifecycle of these products is the feedback. Once you start using the products, you come across different roadblocks that further help in refining these models. But without proper industrial use (and only academic masturbation), this brings absolutely no value at all.
These models improve through usage. There will be edge cases that the models will miss. There will be some domain-specific use cases in which it would fail miserably. The LIC agent would randomly see an erroneous output around 10:43 AM on a Tuesday morning. These will all help shape the product. And that’s how the product would become better (and maybe even see adoption outside India!).
And the thing is, many government processes are riddled with inefficiencies. Be it getting a real estate project approval or getting a hearing in court, most of our systems are overburdened because of manual inefficiencies. GenAI, if used properly, can do wonders here. And the conversation should start here. We are currently focusing too much on the first layer: “Why doesn’t India have more GenAI companies?”
On the contrary, we are not actively asking: “Where can we deploy GenAI solutions?” Asking this question actively can bring some changes in the PSU structures also. Indian PSUs are not built to optimise for outcomes in most cases. Their approach is conservative in nature, audit-defensibility above all else, with KPIs built around uptime and compliance rather than impact. Indian states can’t whip out AI solutions like China did, but for sure, they can create some outcome-driven pilot programmes.
This, from my semi-vantage point, is a concrete policy ask. Currently, it is completely absent.
Read more of my writings here, and learn about me here.
Observations, data, and the not-so-ordinary world. Subscribe if that's your kind of thing.
In no way am I suggesting that we should just start vibe using GenAI for everything. But genuinely, I feel there are some great use cases which the companies are not exploring at all.