Does India need its own ChatGPT, or is there a better question to ask?

OpenAI’s CEO Sam Altman, during his recent India visit, made some comments that were later taken out of context, leading to a furore on social media. Jaspreet Bindra, former chief digital officer at Mahindra group, offers his perspective on whether India needs its own ChatGPT and large language models.

Bindra reckons that we have the brain power, but, and we speculate with him, that there is a better question: for example, should we instead unleash an Indian tsunami of entrepreneurship on top of what is already there, instead of building it from scratch?

Q. What is your take on Sam Altman’s response?

I think is that Sam Altman, perhaps, is not the most careful public speaker. He says things from his heart—not only in this interaction, but if you watch his other interactions as well as see his other podcasts, he kind of just lays it out as it is. I think in this sense he was not challenging India or Indian companies’ ability to make an LLM or a large language model like ChatGPT, GPT 3 or 4.

I think it was more around whether it is necessary to do that and how difficult it is to do that—and whether we have thought through that. Now given the mood in most countries these days, I think it got a bit of a patriotic twinge on it, and I think people assumed that he was doubting our ability to build something like that, rather than thinking through what we need. Therefore, I think it did get blown out of proportion.

I do think that Rajan [Anandan, venture capitalist] asked a very pertinent question. I think the response was also I think quite a good response but I think in my view it got blown out of proportion and non-scientific, non-technology, you know, non-scientific or non-technology views were put on it.

Q. So he actually went on to clarify what he meant by his response. He basically says, “This is really taken out of context. The question was about competing with us for $10 million, which I really do think is not going to work. But I still said try. However, I think this is the wrong question. And the right question is, what a startup can do that's never been done before…”.

So in that view, do you think that it is a good time for us to make massive investments in language models or do you think that it's easier or more cost effective to build upon what's already been done, if you could break down the technology aspect for us?

Sure, first of all I don't doubt India's ability to build something like this. We have built large supercomputers, we are one of the four countries with missiles that can shoot down satellites, we have had a satellite on Mars. So I think the ability is there, and to do it at a much lower cost is also there. ISRO is obviously one such example. Cost-effective, smart engineering has been a core competence in this country. .I don't think that, you know, if open AI took X billion dollars we need the same X billion dollars.

I think the bigger question in my mind—and I'll be honest with you, I don't yet know the answer, given it's a very emerging field—is that do we really need to build our own large language model now? What are we building it for?

There are five things we really need to build these:

-           You need money—maybe not tens of billions, but more than $10 million.

-          You need talent, which is not a problem in this country

-          You need lots of data, and that might be a problem, because while we have lots of data our digital public goods, is that the data that we are looking for?

-          We need infrastructure, especially GPUs (graphics processing units), which at this point of time only NVIDIA makes and even Elon Musk is not getting enough GPUs. I think that the reason why there's no GPT-5, GPT-6, or work has stopped on many of these. There’s an actual infrastructural and GPU shortage.

-          And frankly, the biggest technical problem is really how do we build the guardrails, how do we make it hallucinate less, how does all the training happen, all the through human feedback etc. I think what OpenAI etc. have been able to do over these many years that they've had to spend on this is not only learn the art of building something but also the art of creating the right guardrails. Guardrails to make it hallucinate lesser, to remove biases, etc. They're still struggling with it, and therefore, I think these are clear obstacles.

Most importantly, we need to first figure out what we need them for. The closest example I can think of is Bloomberg’s large language model trained on financial data, and it works for what they needed—all their financial services, their stock market equity analysis and so on.

So I think first thing we need to figure out is why we need this, what we need it for, and then does it make sense to build something for what we need, or to adapt and fine-tune something which already exists? It’s a classic make-versus-buy decision.

Q. So just taking off from that thread about why India needs its own language models, how does it work for vernacular languages? Can we use existing chat GPT frameworks or do you think that there isn't enough source data and we need to figure that out first?

That's a good question. So those are some of the use cases that one can think of. Now, if you only need for languages, I don't think you need to build a fresh new language model. It's a translation exercise of existing input and output—in my opinion, that doesn’t need a language model.

What could be more important is are we looking at Indian context? A lot of the data presumably which would have gone into ChatGPT, Bard, etc. would have far more of a western context.

I think a lot of Indian content on the internet is still in English. I'm pretty sure that would have been scraped by the global large language models. However, if there's vernacular content out there, I don't think it's as much on the web, but we might have to take it off the web.

Then you have vernacular context to India's-specific areas. I assume that given the super-fast pace of various open source projects, I think it might be easier, faster, cheaper, more efficient to fine tune an existing large language model rather than create something completely on your own.

I think the big opportunity in front of India is to leverage the massive amounts of data that exists because of the digital public goods, because of Aadhaar, UPI etc. And therefore, is there a way to unleash the power of that data, through hundreds and thousands of more products beyond the Paytms of the world?

What generative AI probably can do is that it could help us unleash the power of this data—faster, more efficiently, much cheaper. I would unleash India's entrepreneurial energy on that rather than perhaps building it all up from scratch.

Q. Do you think that there could be an argument that we should build this? And I'm using ChatGPT and LLMs as a proxy for all kinds of fundamental technologies and the scientific know-how that we should build, not for nationalism but for national interest?

There's definitely merit in that argument. And I do believe that there are some fundamental technologies which you should build as a country. With nuclear, you don't have a choice, frankly. I mean, you can buy nuclear weapons, but that is more geopolitically difficult.

Space is a good example. Building your own space technology makes a lot of sense. However, every country can't build everything. So we would need to differentiate between what could be core. Now, in the large language models, what is core is the data. It's not the model itself.

I also do believe that the way this technology is going, building a large language model might just become commoditized. It'll be like building software. Therefore, there may not be any value in the AI modelling part of it. What is important is the data behind it. That is national property in a sense. Much like in telecom, the airwaves—the spectrum is national property and now, the equipment could be Ericsson’s or whoever else's.

So I don't know where large language models would fall in this, but my instinct tells me that we should be very protective of the data behind the model rather than the model itself.

Having said that, it's a very fast evolving field. I would like to be educated on the fact that for certain national security or other reasons, you need to have an India local LLM. We need to be very clear on why we need an India LLM, and that why needs to go beyond our national pride only.