Within technology circles, Artificial General Intelligence—commonly abbreviated as AGI—is a term tossed around with immense frequency. To the everyday observer, it sounds innocuous and generic. Yet, the true definition of AGI remains heavily debated and, for some, deeply unsettling. It encapsulates the vision that artificial intelligence will soon achieve supreme competence across every conceivable domain.

While the concept is undeniably gripping, industry experts argue that the reality is far more intricate and nuanced.

A prime illustration is the concept of the "jagged frontier," a framework introduced by former MIT professor Ethan Mollick to explain why AGI will not emerge all at once.

“Imagine a fortress wall, with some towers and battlements jutting out into the countryside, while others fold back towards the center of the castle,” Mollick writes in a 2023 blog post visualizing the dynamic. “That wall is the capability of AI, and the further from the center, the harder the task. Everything inside the wall can be done by the AI, everything outside is hard for the AI to do. The problem is that the wall is invisible, so some tasks that might logically seem to be the same distance away from the center, and therefore equally difficult – say, writing a sonnet and an exactly 50-word poem – are actually on different sides of the wall. The AI is great at the sonnet, but, because of how it conceptualizes the world in tokens, rather than words, it consistently produces poems of more or less than 50 words. Similarly, some unexpected tasks ( like idea generation ) are easy for AIs, while other tasks that seem to be easy for machines to do (like basic math) are challenges for LLMs.”

Ramesh Raskar, an instructor at MIT, shares a similar perspective on AI as a multifaceted entity displaying varying tiers of capability. Raskar recently sat down with veteran journalist Nina Gregory to examine these dynamics in depth.

The Mountain View Discussion

“I think everybody knows that there’s just an investor pitch deck called AGI,” Raskar stated. “There’s no such thing.”

It was a stark opening remark, delivered with considerable force, which he subsequently supported through a deeper exploration of the jagged frontier.

“It’s not possible to create a zero shot AGI, because most of the world’s knowledge is not going to be available to any model for a very, very long time,” he noted, highlighting proprietary pharmaceutical research and other frontline data sources that remain closely guarded for various strategic reasons.

Expanding on his viewpoint through a detailed automotive analogy, Raskar elaborated:

“The models, especially foundation models, are like the engine of a car,” he explained. “The agents are like a car, but we still have to build safe roads, like the internet of AI agents. So the company that builds the engine may not be the company that actually builds the car, the agent. So it’s quite likely that a pharma company will start buying the agents and the frameworks, the agent harnesses, and the whole scaffolding around them from a completely third-party player. It doesn’t have to be the same companies, and perhaps that might be ideal. That would be ideal, because then they will be focused on not model safety, but agent safety, and those two things are completely different.”

Raskar highlighted Network AI agents in Decentralized Architecture (NANDA)—noting with amusement that it shares a name with his sister and translates to 'joy' in Sanskrit.

While this protocol enables autonomous agents to collaborate seamlessly across the internet, significant hurdles remain along the development path.

Raskar outlined four primary challenges associated with deploying systems on the web of tomorrow:

  • Establishing a comprehensive registry for agents
  • Implementing reliable attestation mechanisms
  • Ensuring continuous verification
  • Managing safety testing and regulatory protocols

“When cars are released, you have to take them to crash test dummies and test tracks, and national highway, NHTSA safety driving tests,” he remarked, introducing the concept of a dedicated 'NANDA town' for autonomous systems while elaborating on the attestation process:

“If somebody claims, ‘I’m an agent for diabetics,’ or ‘I’m an agent for financial advice,’ somebody else has to certify that, and that has to be attested on a minute-to-minute basis,” he detailed. “So if somebody says, ‘I’m going to give you financial advice,’ they have to take a test every minute to prove that they’re a financial adviser. And so that agent is sort of being constantly verified.”

India’s Vision: Personal AI for Every Citizen

Fresh from a journey to India, Raskar shared observations regarding national initiatives to equip all 1.5 billion citizens with customized AI agents. He dismissed anxieties surrounding data center overexpansion as unfounded, pointing out the efficiency of running advanced models locally on consumer devices.

“Most of the agents are to run on the edge,” he emphasized. “They can run on your PCs, your laptops, even your phones. At some point, you might run a replica as a backup in the cloud, but you do not need the AI infrastructure buildout.”

“A lot of the work we do at MIT is basically thinking about what it means to build a city of agents, and not an individual agent,” he told Gregory. Addressing her use of the term “bazaar” to characterize the decentralized interplay of model behaviors, he raised several critical architectural questions:

“How would you create, you know, the identity rails, the payment rails? How would you create the justice system for agents if agents go rogue? … If (engineers) build all those things, they’ll be ready for the ‘bazaar era’ of AI, as opposed to just buying it off the shelf from a big company.”

“I think the question here is how unique each of our agents will look, and that’s what really drives us when we wake up in the morning,” Raskar concluded. “What really drives us is, ‘I’m going to create something with my talent, with my expertise, with my hard work, that’s unique compared to other people,’ right? You’re talking about 25 years of your hard work. And I think personal agents will allow us to create something highly crafted that you’re really proud of, and the tools to create your own personal agents are actually going to play a very big role… Sometimes they’ll have economic value, but sometimes they won’t have economic value, but they’ll be useful in some other ways. And that’s the future. That’s really the glorious future we can have, where we have 8 billion agents that are very unique.”