When we imagine a superpowered artificial intelligence, we might think of a giant centralized brain. But leading experts, inspired by the work of Marvin Minsky, are proposing a completely different model: a decentralized network of interacting “small minds,” much more resembling a society than a single machine.
Introduction: Deconstructing the myth of the central brain
Imagine visiting the “Grand AI Chief,” the entity that controls every system: mail, utilities, government, and business. You walk into the command room and there you see it, one big, beautiful brain pulsing in bright colors while a set of wires connects it to the global network. This image, inspired by science fiction, is what many picture when thinking about artificial general intelligence. However, according to many experts, this is not the real picture of what we will see.
Instead, as we approach the technological singularity, some of the best minds argue that we are much more likely to see something resembling what you might imagine if you tried to visualize the internet: something happening here, something else happening there, and effective channels connecting everything together in real time. It might look messy when viewed up close, but the system works with the elegance and precision of a Swiss watch—complex, dynamic, and remarkably interconnected.
“Giving machines a heart”: A Chinese call to delve deeper into AI development
Minsky’s mind: The “Society of Mind” theory
This groundbreaking idea is largely credited to MIT great Marvin Minsky and his book “The Society of Mind.” After much contemplation and research, Minsky stated that the human brain, despite being a single biological organ, is not a single computer, but rather a series of different “machines” connected to one another. This may sound like just a difference in terminology, but Minsky gave us more than that. He helped introduce the idea of “k-lines,” the pathways by which we remember things, which resemble the journeys of data packets across the internet.
This leads us to look at AI from a different angle. When people argue about whether AI is “real” or “conscious,” we can say that the “reality” of AI lies in how we process its outputs, much in the same way that the ripples of a rock in water are in some ways more real than the rock itself.
The AI trilemma: Scaling, coordination, and heterogeneity
In a recent presentation, Abhishek Singh introduced what he calls the “AI trilemma,” a fundamental tension between three desirable traits that systems rarely achieve simultaneously:
- Scalability: Can the system operate effectively as it grows?
- Coordination: Can the different parts work together seamlessly?
- Heterogeneity: Can individuals or components take on specialized and different roles?
To illustrate this tension, Singh points to examples from nature. Ant colonies achieve massive scalability and incredible coordination, but individual ants are largely similar. In contrast, wolf packs exhibit high heterogeneity (different roles for hunting and caretaking) and effective coordination, but they do not scale; pack sizes rarely exceed a dozen. Singh explains that “what makes human societies unique is our constant attempt to balance all three traits simultaneously.”
Chaos theory 2.0 and emergent intelligence
Singh introduces a new framework for understanding how these elements interact in AI systems, which he calls “Chaos Theory 2.0.” He explains that in centralized systems, when you improve two properties, you inevitably sacrifice the third. But operating in a decentralized manner can help bypass these limits.
The solution lies in an alternative approach: instead of “one big brain sitting inside a major tech company,” we move toward “many small minds interacting with one another.” No single small brain is powerful enough on its own, but together, using coordination protocols, they create emergent intelligence that exceeds the sum of its parts.
Singh draws inspiration for this from how human societies have solved coordination challenges throughout history. We have language, cultural and financial institutions, social norms, and knowledge transmission systems. These mechanisms, Singh argues, provide blueprints for how AI agents can organize themselves.
Conclusion: Toward an internet of AI agents
The idea is not merely theoretical. Singh points out similarities between agent systems and the early internet, where humans had to invent protocols like HTTP and SSL to communicate. He mentions the ambitious MIT project NANDA, which aims to create an “internet of AI agents.” The question he poses is: Do we need more of this organized “chaos” in AI? The answer is likely yes. The future of artificial superintelligence will not be in building a bigger brain, but in designing a smarter society.
