AI is probably the most important thing that humanity has ever worked on. I think of it as something more profound than electricity or fire,” said Sundar Pichai, Google CEO, at the World Economic Forum in 2018.
Pichai was not indulging in hyperbole. As the head of a leading technology company, he knew something that others probably did not: that the enduring quest for making ‘intelligent’ machines had crossed critical hurdles and was about to enter a revolutionary phase.
Over the last couple of years, the emergence of Artificial Intelligence (AI)-powered tools such as ChatGPT, Gemini, Perplexity, Grok and many more — all examples of what are known as Large Language Models (LLMs) — have given people a glimpse into the possibilities that AI was always believed to have. These LLMs have shown an exceptional proficiency to ‘understand’, and interact with, human languages in a meaningful way, considered an extremely difficult task for computers.
Language proficiency is just one capability. There are other fields in which AI is making a profound difference. For example, an AI-based tool called AlphaFold has shown such a remarkable ability in predicting protein structures — a notoriously difficult job otherwise, but critically important for understanding life processes — that its developers Demis Hassabis and John Jumper won the Nobel Prize in Chemistry last year. Such major breakthroughs have resulted in a huge excitement around AI. And some concerns too.
Gautam Shroff, professor, Indraprastha Institute of Information Technology (IIIT), Delhi,
“AI will soon start controlling weapons, and will become a mission critical technology (like nuclear technology) at a national security level. This means that AI will be used for cyber counter-terrorism, nuclear weapon designing and running robots which will be fighting the wars of the next decade,” said Gautam Shroff, professor, Indraprastha Institute of Information Technology (IIIT), Delhi, and former vice-president and head of research at Tata Consultancy Services (TCS).
AI is currently one of the most coveted technologies. The recent release of DeepSeek, a Chinese LLM built at a fraction of the cost of its American rivals whose domination of the technology had remained unchallenged till then, was described by many as a “Sputnik moment” — the beginning of a new age of technology war, reminiscent of the space wars between the US and the USSR in the 1960s and 1970s.
For India, where work on AI is at a relatively nascent stage, DeepSeek was a wake-up call that ignited fears of being left out in a technology race once again. Just last year, India had launched a Rs 10,000-crore mission to build capabilities in AI.
With DeepSeek busting the myth that cutting-edge AI technology can only be developed by the richest companies in the world, India felt compelled to reassess its timelines. IT Minister Ashwini Vaishnaw announced that India would have its own LLM within 10 months. The Indian Express on Friday reported that the Centre had received at least 67 proposals to build the India-specific models.
Man vs Machine
Though it is the current flavour of the season, LLMs that take in text inputs and generate synthesised outputs in the form of text, image or code are not the be-all and end-all of AI. These are just a small part of the ultimate ambition to design a fully intelligent machine that scientists have long envisioned. The kind that can ‘think’ and ‘act’ autonomously through a process of self-learning or artificial general intelligence (AGI).
Since the beginning of the computer age in the 1940s, scientists have wondered whether computers, when given more processing power, would be able to perform every task that a human can. In other words, could a computer become an ‘intelligent’, if not a sentient, being?
Many scientists, including mathematician Alan Turing, considered the father of modern computing, were of the opinion that computers would eventually gain so much sophistication that they would be able to ‘think’ and ‘act’ independently like human brains. Others, like Nobel Prize-winning physicist and mathematician Roger Penrose, have been sceptical of the idea that computers could eventually become more powerful than human brains.
Scientists for long have been trying to make computers mimic the human brain. Over the years, they have succeeded in developing algorithms known as artificial neural networks that are inspired by the structure and workings of human brains, and have the capability to identify and learn patterns in data. Breakthroughs in these kinds of neural networks have enabled the development of LLMs and tools like AlphaFold. These systems still do not ‘think’ and ‘act’ like human brains, but are able to deliver results that make it appear that they are doing something similar.
As of now, LLMs seem to be at the frontier of AI technology. They belong to a broader category of AI systems called foundational models. These are general-purpose AI systems over which numerous specific applications can be built. These are trained on very large datasets and form the backbone of the applications that users interact with.
LLMs are a good example of foundational models as they can handle language-related tasks — engage in a conversation, summarise large texts, prepare notes, write computer programmes or even generate poems on demand. There are other foundational models that work with images, audio or video. On the other hand, the Deep Blue computer that defeated Garry Kasparov in a famous man vs machine match in 1997 could only play chess and was not a foundational model in that sense.
Currently, these models also serve as the base for applications that are used for predicting complex protein structures, designing vaccines, weather forecasting and computer coding, among other things.
B Ravindran, head, Department of Data Science and AI, IIT-Madras
“There were two things that were easy for humans but extremely difficult for computers to do — vision and language. We have largely solved the vision problem. Computer vision has become extremely sophisticated now, and is getting better (face detection, medical imaging, satellite-based observations). With LLMs, we have cracked the second problem of language understanding, as these current generative AI tools (AI that generates content — text, images, code, etc.) have shown. Of course, much more has to happen here, but key breakthroughs have been made. The third frontier, where machines are still not doing that great, is in the area of more abstract reasoning. Simple things like if A implies B, and B implies C, then A should imply something something…that kind of stuff. It seems pretty straightforward for humans, but is very difficult for a computer,” said B Ravindran, head, Department of Data Science and AI, IIT-Madras. He also heads the IIT’s Wadhwani School and the Robert Bosch Centre and the Centre for Responsible AI.
India’s challenges
For now, the big race among nations and corporations is to develop their own foundational models as building applications on top of someone else’s model can bring in layers of vulnerabilities. For example, models trained on global datasets often lack local nuances and can insert foreign biases, thereby producing unwanted or erroneous results.
In applications related to defence or national security, a foreign model always carries potential dangers of sabotage, leaks of sensitive data or uncertainties over updates. On the other hand, home-grown models can spur innovation across sectors, and can result in the establishment of an AI ecosystem.
“It is like having our own space programme. Not all countries have it. We can possibly piggyback on the US or Europe or someone else? But then, we will also have to remain dependent on them. For the same reason, we have to have AI infrastructure if we aspire to be a major power in the world,” Shroff said.
There is a lot of interesting work in AI happening in India but these mostly relate to building AI-based applications for specific work, like in healthcare or drug discovery. Building foundational models has not been a priority, mainly because it is an extremely resource-heavy and expensive exercise. It involves massive computational infrastructure, enabled through specially designed state-of-the-art chips called Graphics Processing Units (GPUs) that were once used primarily for gaming.
A model architecture has to be designed and built, and then trained on very large datasets for it to ‘learn’ to do a variety of things. Training models is a process that consumes an enormous amount of electricity as well — LLMs like GPT-3 devoured nearly 1,300 megawatt-hours (MWh) of power. And these thousands of GPUs often run in “hyperscale data centres”, which can be as big as one million square feet.
Ashwin Srinivasan, senior professor, Department of Computer Science, BITS-Pilani, Goa
“Any form of big-impact research in science requires substantial long-term funding, especially of blue-sky (curiosity-driven) research. If support from the government or industry, or both is available, I don’t see why technologies like LLMs cannot be developed in India. In practice, there is probably plenty of useful AI research that can be done in Indian universities without an indigenous LLM. But for reasons of sovereignty and national security, countries, including India, will probably invest in AI technologies that are home-grown,” said Ashwin Srinivasan, senior professor, Department of Computer Science, BITS-Pilani, Goa.
The AI Mission is a good beginning in this regard, said Mayank Vatsa, professor, computer science, IIT-Jodhpur. “The government’s AI Mission has undoubtedly sparked off important discussions about enhancing research infrastructure. Training advanced deep learning models for large-scale applications demands substantial GPU clusters and high-performance computing (HPC) facilities.
While some premier institutes such as IITs and national research labs have built up some capacities, their scale remains modest compared to global benchmarks. The absence of a robust, dedicated HPC framework for AI research, especially for large-scale experiments, continues to be a major bottleneck,” Vatsa said.
Shortage of GPUs, currently in high demand and short supply, is a big challenge. While the AI Mission seeks to procure at least 10,000 of these chips, some researchers feel there is lack of expertise to run these clusters.
“In 2017, I was on the first AI task force… We had recommended that India should create a centralised AI infrastructure, allocate about Rs 5,000 crore over the next few years, including on procuring GPUs, and let the research community use this. We had three other committees after that but nothing much happened,” said Shroff of IIIT-Delhi.
“OpenAI (which developed ChatGPT) was just beginning then. GPT-2 was out, and GPT-3 was in the pipeline. We knew the direction in which things were moving. Back then, no one was sure that this language thing (LLMs) would work. Even OpenAI was unsure, but they went ahead. If we had started acting in 2017, we would also have been in the lead. Now we have to spend much more on buying GPUs in a GPU-scarce market. Why do we have to always play catch-up? You have to trust your scientists to do the right thing. You set up a committee in 2017… but then didn’t act on its recommendations,” he said.
India might be late, but is not completely out of the race. As Ravindran of IIT-Madras points out, there are plenty of new and innovative things that India can do to join the leaders.
“This is particularly true for data. None of the models in the market right now are trained on Indian datasets. We have our own peculiarities and nuances. There are thousands of languages and dialects, each with its own finer details. Also, most of us speak multiple languages, and mix up their words while speaking or writing. These are unique traits, which no model captures. We have to create our own datasets quickly, and some work in this regard has been happening,” he said.
“We then need not build our own model from scratch. There are lots of open-source models like Llama or Mistral that are available in the public domain. These can be tweaked to our requirements, and then trained on Indian datasets. It would be as good as having a home-grown model,” Ravindran said.
Richa Singh, head, Department of Computer Science and Engineering, IIT-Jodhpur
Richa Singh, head, Department of Computer Science and Engineering, IIT-Jodhpur agrees, saying India’s linguistic diversity was both an opportunity and a challenge. “We need to build large-scale diverse datasets in multiple Indian languages not just for common tasks like communication but also for specialised domains like healthcare, agriculture and law. Developing advanced multilingual multimodal AI systems that can effectively handle the complexities of Indian languages can give us a unique advantage,” she said.
Chetan Arora, professor, computer science, IIT-Delhi
It is also important to secure Indian datasets and not allow its leakage, said Chetan Arora, professor, computer science, IIT-Delhi. “Let’s say company A goes to the government and says it will help fulfill a task free of charge but will collect the data. Although the government has got the service free of charge, the data is gone, in this case, too cheaply. You just cannot keep giving data to the multinationals as it is a strategic asset and advantage now,” he said.
A unique feature of India’s AI research ecosystem is that it is almost entirely driven by government institutions and laboratories. There is little initiative or participation of the large IT companies which are some of the biggest names in the international software service industry. This is very unlike other countries where AI research has been spearheaded by private corporations, whether in the United States, China or Europe.
Indian industry is seen as extremely risk averse with little expenditure on research and innovation. “Private companies do not have the muscle, frankly,” said Shroff. “TCS has the biggest research lab (amongst private companies) in India. We had about 700 people just in India when I was there, all researchers. But still, the hardware investments are minuscule, compared even to many customers of TCS,” he added.
Despite these challenges, most agree that it’s key that India powers through. It is one of those rare instances when India is throwing its hat to participate in the development of a cutting-edge technology, instead of just hoping for an early adoption. How it performs in this race will have implications not just on its capabilities in AI, but also on its economic growth and aspirations for a seat at the global high table.