Cognitive Density: Why the Next AI Revolution Won’t Come From Bigger Models
July 9, 2026
The first generation of artificial intelligence was built on a remarkably simple assumption. If intelligence emerged from scale, then the obvious solution was to scale everything: more parameters, more GPUs, more electricity, more data centres and more internet data. For several years that strategy appeared almost self-evident, and because it produced astonishing progress, few people questioned whether the industry was solving the right problem or merely the easiest one. Today, that assumption deserves reconsideration.
The remarkable advances of the past few years have demonstrated that neural networks can recognise patterns at a scale no human mind could ever approach. They can summarise books in seconds, generate software, analyse financial statements, write legal drafts, explain medical literature and solve increasingly complex reasoning problems. Yet despite these extraordinary capabilities, they continue to exhibit an equally remarkable weakness. They often fail in precisely the situations where experienced professionals succeed, not because they lack information, but because they lack the structured judgement required to organise that information into understanding.
That distinction may define the next era of artificial intelligence. The industry continues measuring progress through benchmarks, context windows, token counts and parameter sizes, but those metrics increasingly resemble measuring a library by the number of books it contains rather than by whether anyone inside understands how the books relate to one another. Intelligence has never been a simple function of memory. It has always depended upon relationships.
Why Relationships Matter More Than Memory
Recent neuroscience offers an interesting clue. Researchers at the Max Planck Institute found that individuals with higher fluid intelligence did not simply remember more information than everyone else. Instead, they constructed richer and more coherent cognitive maps, organising concepts into structured relationships that allowed them to navigate unfamiliar problems far more effectively. Intelligence, in other words, appeared to depend less on storing additional facts than on building better internal models of how those facts connected. The brain was not functioning as a warehouse. It was functioning as a living network of relationships.
Artificial intelligence increasingly faces exactly the same challenge. Today’s frontier models ingest enormous quantities of information, much of it gathered from the public internet, which remains one of humanity’s greatest repositories of knowledge while simultaneously serving as one of its greatest collections of repetition, misinformation, marketing, ideological conflict and intellectual noise. Hidden within that immense archive are extraordinary insights produced by scientists, engineers, physicians, mathematicians and economists, but they exist beside billions of pages that contribute little to genuine understanding. Training on larger portions of the internet therefore produces both greater knowledge and greater confusion.
The remarkable achievement is not that large language models occasionally hallucinate. It is that they hallucinate as infrequently as they do considering the material they were asked to learn from. This is why experienced users consistently obtain dramatically better results from AI than inexperienced users. The model itself has not changed. The person asking the questions has.
Why Experts Still Get Better Results From AI
An accomplished engineer rarely accepts the model’s first answer because years of practical experience allow weaknesses, omissions and faulty assumptions to become immediately obvious. A surgeon challenges uncertain diagnoses. An experienced investor recognises when an argument violates economic reality. A physicist notices contradictions that everyone else misses. Every interaction becomes an iterative conversation in which the human continuously filters, redirects and improves the model’s reasoning. Someone lacking that expertise often mistakes polished confidence for genuine understanding because confidence and competence can appear remarkably similar when you lack the knowledge required to distinguish between them.
Artificial intelligence therefore performs a curious function. It rarely replaces expertise, but it does have the ability to amplify whatever expertise already exists. That observation carries profound implications because it suggests the next breakthrough may have far less to do with building larger models than with changing the way those models acquire knowledge in the first place.
AI May Need an Apprenticeship Model
Human beings do not become experts by reading and memorising everything. They become experts through apprenticeship.
A young engineer works beside experienced engineers for years before being trusted with critical decisions. Surgeons spend thousands of hours observing, practising and being corrected by senior surgeons. Scientists challenge one another’s assumptions through peer review. Editors refine writers. Mentors shape judgement. Every profession compresses decades of accumulated experience into patterns that cannot easily be extracted from textbooks alone because much of expertise exists as tacit knowledge rather than explicit instruction.
Jeff Hawkins described something remarkably similar through his Thousand Brains Theory, arguing that intelligence emerges from countless interconnected reference frames rather than from one central repository of information. The brain continuously builds relational models of reality, allowing knowledge acquired in one context to improve reasoning in another. It does not simply retrieve facts. It predicts relationships.
That idea becomes increasingly important as AI evolves. Perhaps the next frontier is not bigger models but richer cognitive architecture, where artificial intelligence spends less time memorising information and more time learning how experts actually construct understanding. Instead of absorbing another trillion tokens scraped indiscriminately from the internet, future systems may spend millions of hours observing world-class engineers designing aircraft, surgeons planning complex procedures, chemists conducting experiments, mathematicians proving theorems, investors evaluating capital allocation and physicists challenging one another’s assumptions. In effect, AI may require its own apprenticeship.
Cognitive Density and the End of Brute-Force Thinking
The implications extend well beyond intelligence itself. DeepSeek’s recent work provides an interesting parallel. Rather than assuming every improvement required more hardware, its engineers attacked inefficiency throughout the entire system by improving memory management, inference architecture, scheduling and utilisation. The result was not merely lower costs. It was substantially greater intelligence extracted from the same computational resources. That represents a philosophical shift. Progress increasingly comes from reducing waste rather than simply adding scale.
Knowledge may be entering the same phase. Instead of asking how artificial intelligence can absorb more information, the more valuable question may become how it can absorb better information organised through richer relationships. Quantity becomes progressively less important than density.
That is why I increasingly prefer the term Cognitive Density over High-Density Intelligence. High-Density Intelligence describes the outcome. Cognitive Density describes the process.
It represents the amount of structured judgement, relational understanding and expert reasoning embedded within a system relative to the computational effort required to produce it. Two models may possess identical parameter counts, identical context windows and identical hardware, yet the model trained through richer expert interaction could possess dramatically greater Cognitive Density because it understands relationships rather than merely recalling information.
This also explains why the current AI race is gradually moving away from brute-force scaling towards systems engineering. Compute remains essential, but every improvement in software efficiency, inference optimisation, expert training, relational mapping and cognitive architecture produces intelligence that hardware alone cannot buy. Intelligence itself becomes more efficient.
Ironically, export controls may have accelerated this transition. Unable to compete solely through unrestricted access to the world’s most advanced GPUs, Chinese researchers increasingly focused on software optimisation, distributed training, open-weight architectures and inference efficiency. Constraints redirected innovation towards extracting more intelligence from fewer resources, demonstrating once again that necessity rarely prevents technological progress. It usually changes its direction.
The Next Breakthrough Will Come From Judgment
The broader lesson extends far beyond artificial intelligence. Human civilisation has never advanced by accumulating information indefinitely. It advanced by distilling information into principles, connecting ideas across disciplines and teaching each generation how experienced minds separate signal from noise. Universities do not exist primarily to distribute facts. Libraries already perform that function. Universities exist to cultivate judgement.
Artificial intelligence now faces precisely the same challenge. The next great breakthrough is unlikely to arrive because someone builds a model with another trillion parameters or constructs another trillion-dollar data centre. It is far more likely to emerge when machines begin learning the same way human expertise has always evolved, through observation, correction, mentorship, structured relationships and continuous refinement rather than indiscriminate accumulation.
The first generation of AI proved that machines could remember almost everything. The second may prove that remembering was never the difficult part. Understanding why one idea matters more than another, recognising hidden relationships across seemingly unrelated disciplines, knowing which assumptions deserve questioning and which information deserves to be forgotten altogether, those are the qualities that have always separated intelligence from memory.
The race to artificial general intelligence may therefore become something entirely different from what most observers expect. It will not be won simply by building larger neural networks. It will be won by building systems capable of learning the way exceptional human minds have always learned, transforming information into knowledge, knowledge into judgement, and judgement into wisdom. Only then will artificial intelligence begin resembling something more than an extraordinarily capable neural network. It will begin approaching what we have been calling intelligence all along.
