Your browser does not support the <audio> element.
HINDI IS THE world’s most widely spoken language after English and Mandarin. Yet it constitutes only 0.1% of all freely accessible content on the internet. That is one obstacle to India developing its own generative artificial-intelligence (AI) models, which rely on vast amounts of training data. Another is that Hindi is spoken by less than half the country. More than 60 other languages have at least 100,000 speakers. Data for some of them simply do not exist online, says Manish Gupta, who leads DeepMind, Google’s AI arm, in India. Natives of those languages stand to miss out on the AI revolution.
Generative AI tools such as ChatGPT, a chatbot, are powered by large language models, or LLMs. The “language” bit is crucial: without a corpus of data it is impossible to make models, whether large or tiny. That is one reason why, two years into the new AI race triggered by the launch of ChatGPT, India has yet to produce any noteworthy AI innovations. But behind the scenes the government, non-profit outfits, Indian startups and global tech giants are working to adapt the technology to the country’s needs. The pace and scale of their success will influence India’s progress in the coming century. It will also offer lessons for other developing countries.
There are two big reasons for India to develop its own AI capabilities. First, as a rising power it is wary of depending on foreign technology. Second, it could be transformative for development. “The real value comes from how you apply these technologies to make a difference to people,” says Nandan Nilekani, a tech grandee.
For a better sense of India’s AI challenges—and opportunities—consider the analogy of cooking dinner. The raw ingredients for AI are data. In the absence of a well-stocked pantry India is doing the equivalent of growing its own food. AI4Bharat, a research lab at the Indian Institute of Technology in Chennai, has sent people across the country to manually collect voice recordings in 22 languages. Google is doing something similar. Both feed into Bhashini, a government project to create a translation system for Indian languages.
Next, the data are blended, simmered and seasoned using a recipe known as a model. Models can be huge, with lots of ingredients and many complicated steps, or they can be relatively straightforward. The recipes behind ChatGPT or Google’s Gemini are enormous. But for India’s purposes, simpler ones may suffice. One idea is to use open-source models, such as Meta’s Llama, as a base sauce, and then add ingredients or tweak the techniques according to local needs. Sarvam AI, a startup in Bangalore, is going down this route.
Lastly, cooking requires the skilful harnessing of power. Just as turning ingredients into food depends on the application of heat, so AI relies on specialised computer chips. The sort needed to build and run sophisticated AI models are expensive and in short supply globally. Earlier this year the government said it would acquire 10,000 of them at a cost of 50bn rupees ($600m) to make computation power available at subsidised prices. And Indian innovators are exploring other types of chips that may be better suited to their purposes.
What, then, will all this effort produce? As in the West, the most visible products will at first be chatbots. The difference is that these will be tailored to immediate, practical uses, revolving around translation and simplifying dealings between citizens and the state. Moreover, Indians use the internet largely as an audiovisual, rather than textual medium. So Indian AI products, unlike Western ones, will be voice-first or exclusively voice-based.
Take form-filling, which can seem like India’s national pastime. Allowing citizens to verbally answer questions in their own language, which a machine inputs into forms, would widen access and remove middlemen. Automating checklists for compendious compliance rules or bots that assist in interpreting requirements could make the process less soul-crushing. “For the first time with UPI [a home-grown digital-payments system] we can say something in India is better than the rest of the world. But the truth is that every other damn thing is not better,” says Vivek Raghavan, a co-founder of Sarvam. AI, he reckons, “has the ability to flatten that, if everything became easier to do”.
AI could also help in areas such as education and health. One study in 2022 found that less than half of Indian students in year five could read at the level of year two. The health-care system, too, is in dire shape. Cheap, mass-scale personalised tutors could start tackling the crisis in learning. Systems that help in interpreting lab results, assist in diagnoses, or take on administrative work could free up doctors to see more patients. The sclerotic justice system could be sped up by automating some of the procedural tasks that take up as much as half of judges’ time.
Many of these challenges exist across the developing world. With a few notable exceptions, non-European languages are poorly represented online. India’s advantage will come not from pushing at the boundaries of AI, but from solving chronic, basic problems of the sort rich countries no longer think about. India has a unique perspective that could enable it “to build out the next set of AI-led companies in many more categories than exist,” says Dev Khare of Lightspeed Venture Partners.
All this echoes the country’s approach to “digital public infrastructure”, its name for technology platforms backed by the government and built upon by private companies. India has invested in identity systems, digital payments, data management and open protocols, all built at a low cost. The success of these efforts at home has prompted the government to promote their use abroad as a means of winning goodwill and projecting power. If Indian techies can find ways to train and run AI systems frugally, that expertise, too, will be attractive to other developing countries.
India’s AI success is by no means guaranteed. Some are sceptical of the government’s 10,000-chip plan: the state has a poor record of using its research-and-development resources effectively, and the idea that bureaucrats would decide which projects are worthy is unappealing to many. The use of small models to solve big problems remains untested. And even if India lines up the ingredients, recipes and power it needs, it still faces a severe shortfall of chefs. According to the Takshashila Institution, a think-tank in Bangalore, 8% of the world’s top AI researchers are from India. The proportion of them that actually work in India rounds to zero. ■
Stay on top of our India coverage by signing up to Essential India, our free weekly newsletter.
link