With hundreds of billions of dollars being invested in artificial intelligence (AI) infrastructure and debates intensifying around centralisation, data ownership, and the future of work, questions are emerging about who will control the next phase of AI development.
Ramesh Raskar, Associate Professor at MIT and a computer scientist known for his pioneering work on computational imaging, argues that the future of AI lies in decentralised personal agents rather than large central platforms.
Speaking to Devansh Mittal on the sidelines of the AI Impact Summit, he discusses democratising AI, building an “internet for AI agents,” and how India’s DPI stack could shape the next phase of technological development.
You have a very different vision of AI compared to what people are currently experiencing, with tools like ChatGPT. Can you explain that to someone who only knows ChatGPT and other such LLMs?
If you think about ChatGPT, you interact with it, but all your data goes to a central company, which builds a profile of you. You can only do what the platform allows you to do. It is like going to a restaurant, where you can only order from the menu they serve.
In contrast, creating your own AI agent is like cooking in your own kitchen. You decide what to make, how to make it, and who you share it with. You have autonomy and control. An AI agent built by you would run on your own data, remain private, and work according to your needs. It is similar to having your own Aadhaar, email account, or personal files. You own them and use them as you wish. In the future, your AI agent could run directly on your phone, allowing you to do whatever you want without depending entirely on someone else’s software.
Story continues below this ad
But wouldn’t AI always require large data centres and massive hardware somewhere in the world?
Right now, the assumption is that AI needs very large machines and cloud platforms. That is true today for large-scale operations. But very soon, AI models will run on personal devices — on the edge — including phones. It will be no different from using a calendar app or a video app. You will not always need gigawatt-scale data centres. We may need them for the next few years, especially for large operations, but five years from now, many everyday AI tasks may run locally on devices.
You have talked about every citizen having a personal AI agent, which is easier here as we already have a Digital Public Infrastructure (DPI). Would the government create these agents? What would that look like?
No, the government does not need to create the agents, just as it did not create banking apps for UPI. What the government created was an interface, UPI, that allowed banks and private companies to build services on top of it.
Story continues below this ad
The same idea applies here. The government would provide the platform and authentication layer. Private companies and developers would create AI agents and services that connect through that infrastructure. Just like banks must be authenticated to operate on UPI, AI agents interacting with public systems would require authentication. But beyond that, it would not be a government-controlled solution.
Are you already in conversation with the Government of India about this?
This work has been inspired by encouragement from senior leadership. There has been interest in exploring these ideas, and you will see developments very soon.
Story continues below this ad
You say you are working on infrastructure for AI agents. What exactly are you building?
We are trying to create the internet for AI agents.
When the internet was built, it required multiple layers: internet service providers, domain name systems, web standards, and protocols coordinated by organisations like the World Wide Web Consortium. Together, these formed the stack that made the internet function.
Similarly, we are building what we call the OASIS stack — a multi-layered infrastructure for AI agents so they can communicate, authenticate, and interact with services across platforms.
Story continues below this ad
You have said this model may face resistance from companies whose business models depend on monetising attention. How do you see that playing out?
There will definitely be change. When computing moved from desktops to mobile phones, companies had to rethink their strategies. The same will happen again. Business models may shift away from advertising toward new approaches such as product placement or service-based interactions. We do not yet know what the final models will look like, but companies will adapt.
You have thought deeply about AI and its impact on the workforce. Do you think humans will continue to derive their sense of meaning from work, or will that change in the coming time?
Story continues below this ad
I think we are currently on a dangerous path. Nine out of the 10 paths lead to dystopian outcomes, which include large-scale job losses, erosion of the middle class, extreme concentration of power and wealth and increased misinformation. If nothing changes, that direction is quite likely. But if we work together to democratise AI, there is still a chance to avoid that outcome.
What do you mean by “democratising AI”? What does that look like?
Right now, we are in what I call the factory era of AI. A few companies build systems, and everyone else simply uses them. Users have limited control. The next stage will be the garage era, where individuals and small groups can create their own AI systems, similar to early personal computing.
After that comes the bazaar era, where everyone has their own AI agents interacting in an open ecosystem. The economics of a bazaar are very different from those of a few dominant factories. That shift is what true democratisation of AI looks like.
