Let me reveal the curtain on this dazzling vision, because according to several experts, general artificial intelligence won’t look like what you’re seeing here.

As we approach the singularity, many leading thinkers suggest that AI will resemble a visualization of the Internet: events occurring in one place, other actions happening elsewhere, all connected through efficient real-time channels, operating seamlessly like a precision timepiece.

Minsky’s Mind

If you follow this blog regularly, you’ve probably heard me mention Marvin Minsky, an influential figure from MIT and his groundbreaking book, Society of the Mind, where he first proposed this concept. Minsky concluded after years of research that the human brain, although a single biological organ, functions not as one unified computer but as a collection of various interconnected “machines.”

Some may argue this is just semantics—after all, we understand how the cortex operates, how the two hemispheres communicate, and the function of structures like the amygdala. But Minsky offered more than that: he introduced the idea of “k-lines” or knowledge lines, which represent the pathways we take when recalling memories. These are similar to how data packets travel across the Internet—jumping from node to node.

He also discussed what some refer to as the “immanence of meaning,” the notion that meaning isn’t embedded in data itself but emerges through our internal processing.

In my view, this is deeply philosophical—in fact, looking at the term “immanence,” it suggests something rooted within rather than above. Instead of transcending, we're diving inward.

This offers a fresh perspective on AI. When people debate whether AI is truly “intelligent” or “aware,” I’d say the essence lies more in the effects it produces—like ripples from a stone dropped in water—rather than in the system itself. As these agents evolve, they’ll become increasingly tangible in other ways too, constantly acting and influencing systems around them.

That leads me to a presentation by Abhishek Singh at IIA last April, where he explored our likely response to new digital forms of intelligence in a compelling way.

The Three-Fold Cord

Singh presents a framework he calls a “trilemma” involving three core aspects of intelligence: scalability, coordination, and diversity. How scalable is the system? Can its components work together effectively? And how different are the tasks each unit performs?

He uses the example of birds flying in a flock versus wolves in a pack. Birds act in near unison, highly coordinated and scalable due to their uniformity.

Wolves, however, operate differently.

“Each individual has a unique role,” he explains. “But the group doesn’t scale as large. What sets Homo sapiens apart is our ability to achieve both high diversity and massive scalability.”

He draws a parallel to the CAP theorem used in distributed computing, which states that databases can only optimize for two out of three properties—consistency, availability, and partition tolerance—at any given time.

“You face a trilemma,” he says, “and while it doesn't map exactly to distributed systems, there's a similar dynamic at play in ecosystems of intelligent agents working together.”

Enter CHAOS

Singh then introduces chaos theory and its relevance to this field.

“What I’m presenting is chaos theory 2.0 in the context of coordinating agents,” he clarifies. “In centralized systems, choosing two of the three criteria often means sacrificing the third. One way to navigate this challenge, though not completely resolve it, is by embracing decentralization.”

Yet, mere decentralization isn’t sufficient.

“To achieve all three goals simultaneously in a decentralized setup, you need specific algorithms and protocols,” he adds. “Our strategy for addressing this trilemma involves two key ideas: local protocols and emergent behavior.”

This echoes what I mentioned earlier, albeit less eloquently:

“One way to think about this,” he explains, “is comparing today’s dominant model—where a single, massive AI housed at a big tech company tries to handle every task—to a decentralized model where multiple smaller AIs interact. Each small agent alone isn’t powerful, but together, using shared protocols, they produce something greater than the sum of their parts.”

Interestingly, he revisits the earlier point about task heterogeneity being perhaps more of a conceptual distinction. Even within that single large AI, different parts are performing distinct subtasks, much like the non-fungible neurons in the human brain.

“This monolithic model still contains a fractal version of the trilemma,” he notes. “Inside that one big neural network, countless parameters coordinate and tackle various subtasks, creating internal diversity.”

Watch Singh’s talk where he applies these ideas to financial markets, social norms, and knowledge sharing, and you’ll see how this plays out in reality. He also draws parallels between agent-based systems and the early days of the Internet, where humans had to invent networking standards like HTTP and SSL. He mentions a Model Context Protocol (MCP) and references NANDA, an MIT-led initiative aimed at developing an AI agent communication protocol.

So, should we embrace more CHAOS in AI? Watch the full video and decide for yourself.

The above is the detailed content of What Does Super-Powerful AI Look Like?. For more information, please follow other related articles on the PHP Chinese website!