E-Paper
India will be central to how AI is deployed in the real world: OpenAI’s Srinivas Narayanan at HTLS| Watch
India set to provide significant share of global usage cues over next year, said Srinivas Narayanan in online session at 23rd Hindustan Times Leadership Summit
India will be central in how artificial intelligence (AI) is deployed in the real world, with the country set to provide a significant share of global usage cues over the next year, said Srinivas Narayanan, chief technology officer for B2B Applications at OpenAI, on Thursday.
In an online session at the 23rd Hindustan Times Leadership Summit, Narayanan also spoke about evolving guardrails for AI usage, OpenAI’s focus on India and Indian languages, enabling UPI payments and the challenges of scaling. His comments come at a time when the arc of AI evolution is believed to be at an inflection point – both in terms of technology as well as regulations.
“India is going to be central to how AI is deployed in the real world. And I think one of the amazing things about India is just the energy, and the excitement for AI is really high,” said Narayanan. “I talk to people who are building incredibly rich sets of applications, not just for India, but for the whole world.”
OpenAI’s India focus, the company believes, is key to their growth. The ChatGPT Go subscription tier, which is otherwise priced at ₹399 per month, is available for free for a year. This, Narayanan said, defines the company’s “goal to remove every barrier to adoption in India”. It has also tied up with the National Payments Corporation of India (NPCI) and payments platform Razorpay, to enable UPI payments for subscriptions as well as any purchases made within ChatGPT.
India is the second largest market for ChatGPT, and the company is banking on an education-focused push to tap a massive market. This summer, ChatGPT’s study mode was announced. Narayanan said it isn’t just about providing an answer to a query, but more about levels of guidance on how to get there. Educational partnerships, such as those with IIT Madras, will be crucial for OpenAI in terms of research on how AI is impacting education.
Narayanan said that the models have evolved in intelligence “quite a bit” over the last year, and that OpenAI is looking at how enterprises are using these tools, including their own Codex, a coding toolkit for developers. “The entire field of software engineering has completely changed in the last six months,” he said. Codex, interestingly, also allows users to start coding tasks via the ChatGPT app on smartphones.
Narayanan touched upon the challenges of engineering products such as ChatGPT, which now has 800 million active users and counting. “We think of it in a few different layers,” he said. “There is the intelligence layer and how we scale intelligence to the vast user base. We think about it from an infrastructure perspective, and we also think about it from a product perspective,” he added.
The intelligence layer references elements such as language support and context for cultural sensitivities. “One of the really interesting things that we have to think about is how do we make our models great in the languages people speak,, understand local culture and local context,” he said. OpenAI, last month, released the IndQA benchmark for evaluating AI systems in the context of Indian culture and languages. IndQA as a data set covers a range of culturally relevant topics including architecture and design, arts and culture, food and cuisine, as well as religion and sport, across sources natively in Bengali, English, Hindi, Hinglish, Kannada, Marathi, Odia, Telugu, Gujarati, Malayalam, Punjabi, and Tamil.
Narayanan spoke about the challenges of scaling, which requires significant infrastructure investments. He flagged two important questions — how do you make it fast and how do you make it reliable?
“As the models get smarter, we want to leverage them for a longer period of time. These models think for longer periods of time and are able to answer complex questions. But as that complexity increases, you also need the infrastructure to run the models for longer periods of time,” he said.
When asked about evolving guardrails with models becoming more capable, Narayanan emphasised that OpenAI considered the safety element at every step of the evolution. “All of our model launches go through this process and we have a very high bar for evaluating these models around these capabilities for what can be risky when we launch them,” he said.
A big worry for engineers and product teams is how model autonomy handles risks such as chemical weaponry and biological weaponry. Scenarios such as loss of human control over a system, technical unpredictability and accountability remain issues that AI companies are actively addressing. “We can take action when we see models might not be complying with the rules,” Narayanan said.
One of OpenAI’s recent products, an AI browser called ChatGPT Atlas, is a piece of how the AI company sees the puzzle. This is, as Narayanan said, for the users who prefer contextual assistance, get answers to questions and have the convenience of automated agentic actions, instead of doing everything manually.
“It’s not just about going to a website and consuming information. If you actually understand underlying goals, a browser can be more proactive. We see this big broader opportunity to reimagine how we consume information, how people get effective assistance through AI, and that’s what ChatGPT Atlas as a product is trying to do. I think you’re going to see us invest a lot more and do some really interesting work in this field,” he said.
When asked about AI agents being considered the next big thing, Narayanan said one must balance the hyperbole and the studies that say 95% of these agents fail.
“We are in the early stages of AI adoption, and for enterprises, there are interesting challenges about how to make AI understand the organisational context, and make it learn different tools. We’re going to build the agent platform, and we’re going to build a set of connectors that makes it easy for companies to adopt AI. But the interesting thing is the models have evolved in intelligence quite a bit in the last year. You’re already seeing some of these such as Codex being a great example. The entire field of software engineering has completely changed in the last six months,” he said.
