The Constitution of Intelligence

For the better part of a decade, the defining assumption of artificial intelligence has been deceptively simple: if larger language models are more capable than smaller ones, then the path to greater intelligence is to continue building ever larger models. More parameters, more data, more computation. Intelligence, in this view, is primarily a question of scale. The remarkable progress of frontier models has made this assumption seem almost self-evident. Yet history is filled with ideas that worked well enough to obscure an even better one. A quiet shift is occurring among the people who use AI most intensively. Increasingly, experienced software developers are abandoning the search for a single “best” model. Instead, they assemble teams. One model writes code. Another reviews it. A third searches documentation. A fourth explains technical concepts to human readers. The resulting workflow resembles less a conversation with an omniscient assistant than the management of a small research group. At first glance, this appears to be an economic optimization. Several open models can often outperform a single premium subscription for particular tasks at a fraction of the cost. But the deeper significance may lie elsewhere. The practice suggests that intelligence itself may not scale in the way we have assumed. Perhaps the decisive variable is not the capability of an individual mind, but the quality of the institution in which many minds cooperate. This would hardly be unprecedented. Civilizations have never depended upon universal experts. Their greatest achievements emerged from specialization. Adam Smith described how the division of labor multiplied productivity. Friedrich Hayek argued that knowledge is dispersed throughout society and cannot be centralized without loss. Scientific progress depends upon peer review rather than solitary genius. Constitutional governments distribute authority precisely because concentrated power is prone to error. These institutions succeed not because every participant is equally capable, but because each participant performs a distinct role within a system designed to expose mistakes and combine partial knowledge. Artificial intelligence has largely pursued the opposite strategy. Rather than organizing intelligence, it has attempted to concentrate it. This strategy has produced astonishing systems. It has also encouraged us to think of intelligence as something that resides inside a single model. Yet intelligence in nature rarely looks that way. The human brain is not a homogeneous mass of interchangeable neurons. Vision, language, motor control, memory, and executive planning occupy different regions, each optimized for a different task. No single area performs every function. Intelligence emerges through coordination among specialists. Human society follows the same pattern. Universities, laboratories, corporations, courts, and markets all transform specialized knowledge into collective capability. Their effectiveness depends less upon extraordinary individuals than upon well-designed processes for collaboration, criticism, and revision. Perhaps artificial intelligence is approaching the same organizational transition. Open-source models have become capable enough to specialize. One may excel at mathematical reasoning. Another may possess superior coding ability. A third may retrieve information efficiently. A fourth may write with exceptional clarity. None is universally superior. Their differences are precisely what make them valuable. The central challenge then changes. Instead of asking how to create the most intelligent model, we begin asking how specialized models should work together. This is not simply an engineering problem. It is an institutional one. How should artificial experts divide labor? How should they evaluate one another’s conclusions? When should disagreement trigger additional investigation rather than immediate consensus? How should confidence be represented and uncertainty communicated? These questions resemble constitutional design more than model training. Indeed, the analogy is surprisingly close. Every durable institution confronts the same fundamental problem: no participant possesses complete knowledge, yet collective decisions must still be made. Scientific journals rely upon anonymous reviewers. Courts hear opposing advocates before reaching judgment. Financial markets aggregate dispersed information through prices. Democratic constitutions establish checks and balances because no office is presumed infallible. The objective is not to eliminate disagreement. It is to organize disagreement productively. Artificial intelligence has already begun to rediscover this principle. Multi-agent systems, debate architectures, verification pipelines, and retrieval-augmented workflows all reflect a growing recognition that reliable reasoning often emerges from structured interaction rather than isolated cognition. But these developments point toward a larger conceptual shift. The future of AI may depend less upon building larger digital brains than upon designing better digital institutions. Such systems need not be dominated by a single superintelligence. In fact, the coordinating intelligence may be relatively modest. An orchestra conductor does not produce the music. An operating system scheduler performs only a tiny fraction of the computation occurring within the machine. A journal editor contributes less subject-matter expertise than the community of reviewers. Leadership, in these cases, consists of organization rather than domination. An executive AI could similarly allocate tasks, evaluate evidence, identify unresolved disagreements, and synthesize conclusions while consuming only a small fraction of the total computational resources. The demanding intellectual work would remain distributed among specialized models operating in parallel. This architecture offers more than efficiency. It offers resilience. Monolithic systems inherit their own blind spots. A single mistaken assumption can propagate through every conclusion. A community of specialized systems, by contrast, possesses the capacity for self-correction. Independent perspectives create opportunities for criticism before errors become accepted facts. Human civilization became more intelligent not because individual humans evolved dramatically within recorded history, but because we developed institutions capable of accumulating, correcting, and transmitting knowledge across generations. Libraries, universities, scientific societies, and constitutional governments are technologies for organizing intelligence. Artificial intelligence may now require comparable inventions. The next great breakthrough may not be another trillion parameters trained on ever-larger datasets. It may be a constitutional framework through which specialized artificial minds debate evidence, challenge assumptions, preserve dissent, and reach conclusions that none could reliably produce alone. For years, the central question in artificial intelligence has been, How do we build a smarter model? A more consequential question is beginning to emerge. How do we build a wiser society of models? The answer may determine whether artificial intelligence merely becomes larger—or genuinely becomes more intelligent.