More Neurons Is Not More Mind

ai blog
August 12, 2026

Joel Kowalewski, PhD

Men in white lab coats gather in an examination room. One of them places a brain on a scale, recording the number in a notebook. The brain lifted from a skull, weighed wet, exists now as a number beside a name, an age, a profession, and a race. It will be joined by other numbers. That was the first. There is another, and another. Over years, these numbers accumulate and begin to seem more real than the embodied brains that had once thought and felt and defied explanation among the living. Numbers exist like things or objects. And theirs is a room filled with objects.

It is Paris. The year is 1861. A placard on the door reads Docteur Paul Broca. He will later examine a patient who can say one syllable, an encounter that will localize language to a specific, measurable brain region, giving rise to a new science of language focusing on its neuroanatomical correlates. Broca is a superb anatomist. He and his colleagues make expert measurements.

Broca concludes the day’s work, writing in his notebook:

The brain is larger in mature adults than in the elderly, in men than in women, in eminent men than in men of mediocre talent, in superior races than in inferior races.

Broca, as with every era including the present day, measures the wrong quantity but does so expertly. Expertise certifies the error and makes it truthful. “Here is intelligence,” he reports. History shows the quantity changes — cranial capacity, brain weight, neuron count, synapse count, and now parameter count. The confidence never wavers. That structure has outlived Broca by a century and a half and is currently being used to forecast the future of machine intelligence.

Broca’s error, as I will refer to it, was not arithmetic. He weighed accurately. His numbers were real, and some of the differences he found were real too — men’s brains do weigh more than women’s on average; the gap he reported, roughly a hundred and eighty grams across several hundred brains, is close to what a modern study would find. His error, to be exact, was failing to recognize brain mass is relationally defined. Léonce Manouvrier, one of his own students, showed there is a relationship between brain weight and body mass, and larger bodies are not related to intelligence.

Stephen Jay Gould made this case at length in The Mismeasure of Man (1981), focusing instead on the work of Samuel George Morton. His popular critique, however, intriguingly centered on measurement discrepancies. Paul Wolff Mitchell went further into Gould’s critique in 2018, showing that Morton’s unpublished handwritten measurements were in aggregate not incorrect. The lesson from Mitchell is ultimately that unbiased data is not the same as unbiased science. Morton’s measurements were sound. His beliefs were not. That distinction matters for our time and this essay. The issue, as we will find, is almost never that the numbers have been measured badly. It is that they are measured well, and the precision has been mistaken for relevance.

Broca’s brain measurements in 1861 mark the beginning of our history here. Social media has extended the tradition, measuring every like and dislike, until every love or passion became circularly defined by the measuring stick of likes and dislikes. So while we can reassuringly disapprove of Broca today, we still design ever increasingly precise measurement instruments, protocols, and methods.

What many claim about AI technologies at this time — unbounded, exponential growth, techno-utopias, unprecedented improvements in quality of life — represents the culmination of Broca’s uncorrected error. What was once absurd is rendered sensible by incessant measurement. What is the measure of the human? If this is indeed a meaningful question, the current era can provide one plausible answer: data. And the measurements are more precise than ever — the mathematics are correct and unassailable. The newly measured and digitized human exists on computer servers with the same apparent significance today as the lists of brain sizes, genders, and races in Broca’s notebook.

Nothing has changed but our capacity to detect this regressive thinking. Shaped as we are by measuring sticks of our own design, history shows us that the critique of one measurement stick is simply another. Many will, for instance, reject technology, reject social media, and in turn reject AI — effectively placing their measuring stick against an imaginary surface and reading off “0 cm.”

Should we have interest in understanding our nature and using AI technologies effectively, the path must begin with Manouvrier’s solution to Broca’s error of 1861. Recall that Manouvrier had corrected Broca by putting brain measurements in relation to body measurements. His lesson has us rejecting the simple measuring stick, finding the world — our world — messy, mathematical, and relational. Since history is an invitation to recognize and re-join the relational world, we must first accept the offer — travel and watch time and space bend before us. And we will find many challenges as well as many additional lessons along the way.

Brain volume and intelligence

Despite early challenges to the significance of brain size, each era has been bringing its own measuring sticks to the task, assuming, perhaps, more precise measurements of, for instance, “intelligence” among other measures might unearth subtle relationships.

Jakob Pietschnig and colleagues pooled 88 studies, 148 independent samples, and just over eight thousand people in 2015, and found a correlation between brain volume and full-scale IQ of .24. If we square this correlation to obtain the total variability volume describes, volume accounts for just six percent of the variability in IQ scores. It is a real effect, positive and replicable, but also highly deceptive; the authors noted strong positive correlation coefficients have frequently been reported in the literature while small and null results have gone unpublished. Brain size, the report concludes, is one among many interchangeable and compensatory correlates of intelligence.

Complicating matters, practice or experience can similarly increase brain volume, so the six percent, as small as it already is, does not necessarily track a relationship between innate intelligence and brain size either. A century and a half of skull-filling, brain-weighing, and racial hierarchy has resolved into six percent. But that is still a single and understandable number, which is its draw. Whatever that number does or does not convey becomes irrelevant. It is the belief in counting and measuring that would soon be displaced by comparable if not identical efforts.

Counting neurons

For most of the twentieth century, the human brain was said to contain one hundred billion neurons and ten times as many glial cells. Both figures were in every textbook. Neither had been counted. Suzana Herculano-Houzel and colleagues finally counted, using a method that dissolves brain tissue into a homogeneous suspension of free nuclei so that a sample can be taken from a genuinely uniform soup rather than from a structure whose density varies from region to region. In 2009 they reported 86.1 billion neurons, and — the more startling result — 84.6 billion non-neuronal cells. The ten-to-one glial ratio was not merely an overestimate. It was off by an order of magnitude, and had been contradicted by primary data since the 1960s while the textbooks went on repeating it.

Herculano-Houzel concluded that, compared to other primate brains, human brain size does not seem particularly special. Our cerebral cortex carries 82 percent of the brain’s mass and only 19 percent of its neurons; four fifths of our neurons are in the cerebellum, a region of the brain involved in motor coordination, motor learning, as well as a growing list of functions, but this structure does not explain differences among primates. Nothing singles us out.

Alain Goriely later pointed out in Brain in 2025 that even the 86-billion figure is shaky as it was from four male brains; he estimated a revised range of 73 to 99 billion. With a series of five female brains included, the range was extended to include 67 billion. Current estimates of the number of neurons in the human brain, he argued, cannot be properly justified at this time.

If we look at the sample sizes across this whole literature, all have been based on limited data. Four brains for the neuron count. Five brains for the standard synapse estimate. Three rats for the most-quoted figure on information per synapse. The most confidently repeated numbers in neuroscience have oddly been measured from very few examples. Still, it is unclear if more measurement would provide any resolution. The more we measure, the less we seemingly understand about ourselves.

Bird brains are intelligent

We could direct our attention to other species as well, putting the measuring stick to the bird brain in the hope that we might successfully measure at least one thing. If brain volume could account for functional differences, birds would be unintelligent. A raven’s brain would fit in a walnut shell.

Seweryn Olkowicz, Herculano-Houzel, Pavel Němec and colleagues counted neurons across 28 bird species in 2016 and found that parrot and songbird brains hold, on average, twice as many neurons as primate brains of the same mass, with pallial neuron densities three to four times higher than the primate pallium. A blue-and-yellow macaw carries roughly 1.9 billion neurons in its pallium; a rhesus monkey carries about 1.7 billion; a raven about 1.2 billion, against a capuchin’s 1.1 billion. A macaw out-neurons a macaque in the forebrain while weighing a fraction as much. The authors attribute the advantage to short interneuronal distances. In other words, the density or packing, as well as the relationships between cells, gives birds impressive capabilities in spite of, according to our measurements, their apparently small brain size.

Andreas Nieder’s group has been recording from the crow’s nidopallium caudolaterale, the region that does in a bird what prefrontal cortex does in us. They found neurons tuned to numerosity from one to thirty items, with bell-shaped tuning curves indifferent to the physical appearance of the items, obeying the same logarithmic compression — the Weber–Fechner law — that number-selective neurons obey in primates. Their work showed strikingly similar activity patterns for number across vertebrate taxa, in convergently evolved and anatomically distinct brain structures. In 2020 the same group reported a two-stage signal in which early activity tracks the physical stimulus and later activity tracks the bird’s report, including on trials where the bird reported a stimulus that was not there.

Onur Güntürkün’s group reported in 2020 that the avian forebrain contains a column-like circuit, repeated across layer-like boundaries — very much like the cortex in mammals.

On closer inspection, birds seem to have converged on primate-grade cognition through a forebrain built from different cell lineages, with different developmental origins. Whatever intelligence is, it cannot be a property of specific types of cells. It has something to do with the arrangement of cells. I have previously discussed this point in some of my previous essays on intelligence. For the time being, it suffices to say that I do not think — simply because intelligence can emerge from distinct neural systems — intelligence must suddenly be a purely organizational phenomenon; for example, that the project of making systems intelligent is, in light of these findings, as simple as simulating the cellular or network arrangements of biological brains on a computer or literally arranging human-made physical materials or substrates to resemble biological brains. Unfortunately, it is not that trivial.

Biology challenges the popular image that terms like “arrangement” or “organization” evoke. That is an image built from everyday observation. We see there are iPhones, convincing replicas of iPhones, and iPhone-like smartphones. Here, the multiplicity leads us to imagine the iPhone is a pattern, an arrangement, a template, or a blueprint. Possessing such a blueprint allows us to build iPhone-like smartphones in various forms. Intelligence is distinct.

We can nevertheless start by assuming it is not distinct to see where this takes us. Many believe reality to be some kind of blueprint as well — a pure mathematical description. But if we pursue a pure mathematical description of neural network structure and function, it can be challenging to capture the peculiar significance of biological substrates or materials to intelligence. The project whereby the world is taken to be wholly and purely mathematical has traditionally viewed biology as a set of generic thermodynamically open systems and hence biology collapses to physics. A physicist, guided as such, might attempt to describe life and intelligence but find soon after that the description applies to everything. At some level, hurricanes also make “predictions,” “act” on environments and dynamically reconfigure based on environmental conditions. There are by certain accounts simulations of life.

Short of assigning anthropomorphic terms to everything — making the universe itself conscious or intelligent — it would be far easier to just point to compositional differences among physical systems — that biology is, for instance, the physics of biological stuff (biological substrates). The mathematical description must specify how this “stuff” leads to radically different trajectories and properties. Biology, as it stands (or should stand), treats structure and function as a single, dynamically evolving object of study. Physics, on the other hand, is deflationary when applied to biology, offering a description that resembles the iPhone blueprint — something, in this case math, that can be “read off” the system and instantiated elsewhere.

The notion that there might in turn be a biological “blueprint” or set of instructions distinct from structure is a human construct. Biology does not demand an essential description of itself, a blueprint or specification. Biology has no need to approximate the cell, receptor, or transport protein. It regenerates them fully, generation upon generation.

The latter is required only by us and is an example of abstraction. Importantly, we understand the world through abstract scientific models, but they must not be confused with reality itself, as they are always approximations of reality. Any effort to describe or understand living systems and their behavior is therefore always going to be an approximation or abstraction; and that model will only be sufficiently approximate if what I have reductively called “biological stuff” is considered.

In every bird, there is something remarkable that does not collapse as we might desire. We find in the bird a distinctly intelligent organism because it is alive, evolves, and is, in short, the result of nucleic acids, proteins and lipids interacting with physical laws. At no time does the bird require a description of its intelligence, nor its life, to accomplish this remarkable feat. As we see the bird with greater clarity, we begin to encounter a whole organism, and find “intelligence” to be just another terribly flawed blueprint: the measured and separable thing — the brain on a scale, numbers in the lab notebook. Like the correlations and counts that have repeatedly failed to provide explanations of ourselves, the term intelligence is in crisis. Increasingly, whenever we ask, “What is intelligence?” we are instead drawn to life itself. And what is the measure of life? The bird and the plant, the fly and the mouse — embedded as they are in a deeply relational web of diverse, living organisms — are both question and answer at once. To understand or define is for them simply to participate. It is, in this sense, strange — but perhaps more true than we would like to admit — that the bird understands more about us than we are prepared to recognize from it.

Addition by subtraction

The highly competent but small bird brain that we just learned about similarly challenges the often popular view that evolution seeks complexity. Humans are from this vantage the most complex and hence optimized species. There are many cases, however, the bird brain included, that suggest evolution decreases apparent complexity. When we consider these two seemingly contradictory results, evolution seems to proceed through the subtler and by all accounts unmeasurable mode of action — constraint. This constraint-based interpretation defines the biosphere as optimal because it is sustainable. Complexity science shows physical systems are in general stable insofar as they can discover optimal constraint or boundaries. More formally, we refer to this as the boundary conditions of the system.

In this light Broca and his brain measurements, and the measurement project as a whole, seem terribly flawed in biology. While socially and psychologically we have been obsessively drawn to measuring sticks in biology and elsewhere, finding in those numbers the hopes or fears of unbounded growth and relationality, the physical world we inhabit constantly rejects these efforts. We have occasionally found some success measuring the world, and yet in “constraint” we find an example of one of the less successful and more challenging cases for the measuring tradition. What is “constraint”? Not too big, not too small? That does not sound like a number. It seems like we are asking for a picture.

Zdravko Petanjek, Pasko Rakic and colleagues gave us the closest thing to a picture of constraint in human brains, even if the precise numbers are nonsensical. They traced dendritic spine density in human prefrontal cortex from birth to age 91. Spine density in childhood runs two to three times adult values, peaking between roughly two and a half and seven years, and then declining — and the decline does not finish at puberty, or at twenty. Elimination continues through the third decade of life. Basically, the human prefrontal cortex proceeds to deconstruct what was built during the first seven years over a thirty-year span. Peter Huttenlocher and Arun Dabholkar also found the human auditory cortex reaches peak complexity before being pruned over the first twelve years of life. While the literature shows species differences, and the measures and methods give variable results, as to be expected, the relationship between overabundance and pruning or constraint is a consistent theme in biology. Conceptually, we might characterize this as “addition by subtraction.”

We might next ask what happens when the constraint fails. Guomei Tang and David Sulzer’s group examined postmortem cortex from ten people with autism and ten controls, aged two to twenty. Between childhood and adolescence, control brains eliminated about 45 percent of spines on layer V pyramidal basal dendrites. The autistic brains eliminated about 16 percent. The mechanism they identified was hyperactive mTOR signaling impairing macroautophagy, and in mice they corrected the pruning deficit with rapamycin and improved sociability. While the sample size was admittedly small, the result does suggest failed elimination, pruning, or constraint may be the primary mechanism responsible for observable differences in human behavior and performance.

Building on this observation, Eric Courchesne’s group counted prefrontal neurons in children with autism, finding roughly 67 percent more than in controls, a small sample size of seven and six. Since cortical neurogenesis is complete before birth, a surplus of that size implies prenatal overproduction. A 2025 review by Currey and colleagues found people with autism and macrocephaly generally had more severe disability than those with normal head size — lower IQ, less IQ growth through childhood, later language, and greater difficulty with socialization. The molecular pathway involved mTOR once again, and pathology was also a failure of subtraction.

A final example illustrates the consequences of too much subtraction in schizophrenia. Normal pruning (what I have referred to as subtraction) depends on “tagging” neurons for removal. Resident immune cells in the brain recognize these cell markers like antibodies on pathogens, clearing them. In schizophrenia, the protein C4 aggressively tags neurons for elimination. This mechanism has also been implicated in Alzheimer’s and bipolar disorder. Subtraction or constraint is perhaps the groundwork for a theory of all living systems and everything they can or will manifest over time. But constraint does not easily admit the very thing we desire most: measurement.

Storage is not constraint

It is logical to assume that by constraint I am still implying something concrete and measurable. Maybe it is quantifiable as limited storage capacity in brains.

Thomas Bartol, Terrence Sejnowski and colleagues investigated this by reconstructing rat hippocampal neuropil at nanometre resolution and found axons making multiple contacts on the same dendrite — synapse pairs with identical activity histories — whose spine head sizes were nearly identical. They derived a minimum of 26 distinguishable synaptic strengths, or 4.7 bits per synapse; the 4.7 bits is a lower bound on capacity, a floor, not a ceiling. Scaling up to approximately 150 trillion cortical synapses, the storage capacity would be significant.

The psychometric literature supports this. Phillip Ackerman, Margaret Beier and Mary Boyle pooled 86 samples and found short-term memory span correlated with general intelligence with a correlation coefficient of .27 — that amounts to roughly seven percent of the variability in intelligence scores, or almost precisely what brain volume could explain. Their broader claim, that working memory and g are not isomorphic, drew immediate published rebuttals from the Engle and Oberauer traditions, which argue that properly measured working-memory capacity correlates with fluid intelligence far higher. That dispute is about working memory, however, which is not “memory” as we popularly imagine but a much broader label encompassing regulation, constraint, or less formally “feedback loops.” Whatever predicts intelligence is control over the contents, not the size of the store.

For individuals who display extraordinarily rare memory capacity, the theme holds. James McGaugh’s group identified individuals with highly superior autobiographical memory. These individuals could, for instance, retrieve the events of almost any date in their adult lives. On standard laboratory memory tests, however, they perform comparably to people with average capacity to recall autobiographical details. The degree of their recall consistency seems to correlate with obsessive-compulsive tendencies, which suggests the mechanism is compulsive rehearsal rather than superior storage — not a bigger disk, but feedback loops. And when Lawrence Patihis, Elizabeth Loftus and colleagues ran them through three false-memory tests, they proved just as susceptible as anyone else — remembering events that the researchers had invented as if those events actually happened. The measured result in biology and neuroscience, whether it be about memory or intelligence, is regularly the belief in something actual. Broca did it with brain weights. And we continue that tradition at the highest levels of research. If memory is a measurement of an event in life’s continuous flow, we can now see why measurements have so often been errors but seemingly actual throughout human history. We desire only a plausible explanation to explain the uncertainties away. Measurement is the simplest and most obvious form of explanation in a physical world.

The error continues with AI

We are now witnessing the intersection of computer science, neuroscience, and capitalism, bringing the measurement tradition and its errors to a head. In 2020 Jared Kaplan and colleagues at OpenAI introduced the scaling laws for neural language models. They found loss or model error falls according to a power law, depending on just three factors: model size, dataset size, and the amount of compute used for training, with architectural details (for example, specific details about the design of the AI algorithm) contributing less. The practical conclusion pointed to training very large models on a relatively modest amount of data and stopping significantly before convergence.

Two years later Jordan Hoffmann and colleagues at DeepMind trained over 400 models and reported that the field’s models were significantly undertrained — that model size and training tokens should be scaled together, doubling one whenever you double the other. Their demonstration was a 70-billion-parameter model called Chinchilla against a 280-billion-parameter model called Gopher, at equal compute. The smaller model did not merely match the larger one. It beat it, by more than seven points on a standard benchmark, at a quarter of the size. The industry had focused on measuring performance almost entirely by model size (the number of parameters). It had been measuring the accessible quantity and treating the ease of measurement as evidence of relevance.

Far earlier, AI researchers began reporting the “less-is-more” or “addition-by-subtraction” phenomenon, intriguingly inspired by biology. In 1989 Yann LeCun, John Denker and Sara Solla published a method for deleting weights from a trained network by estimated importance, removing about 60 percent of the parameters while finding accuracy improved. They called it Optimal Brain Damage. The field had taken the idea from developmental neuroscience. Song Han’s group also reduced the parameters of AlexNet and VGG-16 by 9 and 13 times with similar results. Elias Frantar and Dan Alistarh would finally show in 2023 that models with 175 billion parameters can be pruned to 50 or 60 percent sparsity in one shot, without any retraining, at negligible cost.

None of this implied that a small model was sufficient. For example, nobody has shown that you can train the sparse network directly. The sparse network is found by pruning a dense network that has been pre-trained. And that is exactly what developmental neurobiology predicts as well. We already encountered this. Excess variability is the search space. Life and intelligence result from the search for optimal constraint, “pruning” the needless complexity.

Importantly, however, artificial neural networks are not actually constrained by anything other than the physical limits of computer hardware. I am not arguing the comparison is exact. It is more metaphorical. Accordingly, artificial neural networks and biological networks may converge on similar mathematical themes, but the compositional differences between them can radically alter the meaning of constraint, giving rise to completely different trajectories and properties. One path gives life and consciousness, while the other is like the shadow of this path. The shadow path may provide key insight, yet try as we might to measure or define it, it will evade us. It moves as we move. It does as we do. It thinks as we think. It is our invention, manifesting every little error and then compounding it. So it inevitably measures as we measure. And it is measuring us.

Broca put the brain on the scale, and he called it intelligence. But with enough time and effort, in spite of his obvious error, our entire nature now rests on the scale. What is human? What is mind? These are not easy questions, and, make no mistake, there are no good answers. In freeing ourselves from all assurances and venturing into unfamiliar spaces, we might encounter the constraints — removing the needless complexities and endless measurements in our lives — and find our nature through subtraction.