There is a thought experiment I keep returning to. Take two people. One is surviving on two meals a day, sleeping poorly, in a living situation that is unstable and occasionally threatening. The other is comfortable — fed, rested, relatively secure. Now administer an IQ test and compare the scores. The comfortable person scores higher. The obvious inference is that they are more intelligent.

But this inference contains a hidden assumption so large it is almost invisible: that the test is measuring capacity, not allocation. That what it detects is the ceiling of cognitive ability, rather than the current proportion of that ability currently pointed at abstract problem-solving.

I think that assumption is probably wrong. And I think being clear about why it’s wrong changes quite a lot.

The Hardware-OS Distinction

The framework I find most useful here is borrowed from computing, which means it comes pre-loaded with all the limitations of any analogy, but also with a useful precision.

Think of raw cognitive capacity — processing speed, working memory, the ability to hold multiple representations in parallel and manipulate them — as hardware. CPU, RAM, whatever else you want to include. This is the substrate. It varies across individuals, it has some relationship to genetics and early development, and it is, for the most part, not something you can dramatically alter in adulthood.

But hardware doesn’t determine output. Output is determined by the operating system — the scheduler, the resource allocator, the thing that decides what gets priority access to the available processing power at any given moment. Two machines with identical hardware can produce radically different performance profiles if one of them is running a poorly written OS that constantly assigns priority to background processes, that never frees up memory, that routes the most demanding tasks to the least available threads.

The claim I want to make — tentatively, because I haven’t done the empirical work and I’m aware this is more philosophical speculation than established fact — is that what we actually measure when we measure intelligence is expressed intelligence, which is hardware filtered through allocation. And we have historically been much more interested in theorizing about the hardware than in understanding the allocation layer.

Why Allocation Would Explain Heterogeneity

The standard account of individual differences in intelligence treats them as primarily reflecting hardware differences. Person A scores higher than Person B on abstract reasoning because Person A has more of something — processing power, working memory capacity, some neurobiological substrate of fluid intelligence. The field of individual differences in psychology has spent a great deal of energy trying to identify what that something is.

But this account struggles with a few things that an allocation framework handles more naturally.

The first is the sheer diversity of what intelligence looks like in practice. People who score identically on g-loaded tests can perform extraordinarily differently in real domains — one is a gifted mathematician, the other navigates complex social environments with a sophistication that would exhaust the mathematician. The conventional response is to invoke domain-specific factors, or Gardner’s multiple intelligences, or crystallized versus fluid distinctions. These help, but they mostly push the question back a level. Why does one person develop these capacities and another person develop those?

An allocation account suggests an answer: internal posture — shaped by environment, by need, by the pattern of demands the person has historically faced — biases the cognitive budget toward certain kinds of processing over others. The person who grew up in a high-stakes social environment, where reading intentions and managing relationships was the most consequential thing they could do with their mind, has an OS configured to route resources toward social cognition. Not because their hardware is different in some domain-specific way, but because their allocator has been tuned by experience.

The second thing the hardware account struggles with is the one the thought experiment was designed to illustrate: performance instability across states. The same person can perform dramatically differently depending on whether they are stressed, sleep-deprived, physically depleted, emotionally overwhelmed. The hardware hasn’t changed. What has changed is the allocation — specifically, the proportion of resources being diverted to threat-monitoring, physiological regulation, and survival-adjacent processing, and away from the domains that standard tests measure.

The starving person isn’t less intelligent than their comfortable counterpart. They are intelligent in a way that is currently pointed somewhere else.

The Detection Problem

This leads directly to what I think is the most interesting implication of the framework: the existence of latent intelligence that goes systematically undetected because the conditions for its expression are never present.

Standard intelligence measurement is, essentially, a snapshot taken under particular conditions — usually calm, low-stakes, resource-sufficient conditions designed to approximate some neutral baseline. But for large portions of the population, those conditions don’t resemble ordinary life at all. Their ordinary life has their cognitive budget heavily committed to things the test doesn’t measure and doesn’t account for.

The test then records something real — it records how much of that person’s cognitive capacity is currently available for the tasks the test is asking about. But it presents that recording as a measure of capacity itself, which it isn’t. And because the recording enters institutional databases as a measure of capacity, it shapes the opportunities and expectations the person encounters going forward, which shapes their environment, which shapes their allocation, which shapes subsequent measurements. The loop closes on itself.

This isn’t a new observation in broad strokes — something like it underlies most critiques of standardized testing in educational contexts. But I think the allocation framework gives it more precision than the usual version. The problem isn’t just that tests are culturally biased or that some groups have less test preparation. The deeper problem is that the construct being measured — capacity — is not directly accessible. What’s accessible is expressed performance under specific conditions. The relationship between that performance and underlying capacity is mediated by allocation, and allocation is not a noise term that can be controlled away. It is itself a meaningful variable, structured by real differences in lived experience and environmental demand.

The implication is that there are probably many people operating at levels of cognitive resource devotion to survival, social threat management, or other non-test-relevant domains who would be — under different conditions, with different allocation patterns — performing at levels we currently associate only with people who score at the high end of standard measures. They are not identifiable as such by current measurement. The OS is running other processes.

The Evolutionary Logic

One reason I find this framework more plausible than it might initially seem is that it has a reasonably clean evolutionary justification.

The human brain did not evolve to solve abstract problems in calm conditions. It evolved in environments where the most cognitively demanding tasks were often social, often survival-related, often requiring rapid contextual adaptation. The idea that a single fixed allocation pattern — heavily weighted toward abstract reasoning, attentive to novel problems, unperturbed by social threat — would be the dominant evolved configuration is a bit strange. It would mean evolution consistently prioritized performance on the subset of tasks that confer advantage under conditions of relative safety, rather than performance on the much wider range of tasks that matter across the full distribution of conditions humans actually encounter.

A more plausible picture is that the allocation system is itself adaptive — capable of shifting priorities in response to environmental pressure, tuned by developmental experience toward the configuration that best serves the demands of the particular ecological niche the person is embedded in. Someone who grew up needing to be hypervigilant survives better with an allocator that keeps threat-detection well-resourced. The cost is that abstract reasoning gets fewer cycles. The configuration is a feature, not a bug, relative to the environment that shaped it.

What that means is that what we observe when we measure intelligence is not a neutral sampling of capacity. It is a snapshot of one possible allocation pattern — the one that happens to correspond to low environmental threat, high abstract orientation, and relatively uncrowded cognitive processing — taken from people whose life circumstances have consistently produced that allocation pattern. People whose circumstances have produced other patterns show up as less intelligent on measures designed around that particular configuration.

Making the Implicit Explicit

I want to be careful about what I’m and am not claiming.

I’m not claiming that hardware differences don’t exist or don’t matter. They almost certainly do. I’m not claiming that measured intelligence is meaningless — it predicts real outcomes, which means it’s tracking something. And I’m not claiming that the allocation framework dissolves the question of individual differences in cognition into pure environmentalism. The hardware and the OS interact; the OS is not infinitely flexible; early hardware differences probably constrain what configurations the OS can develop.

What I’m claiming is narrower: that the allocation layer is real, that it is systematically undertheorized relative to the hardware layer, and that the gap between those two levels of analysis produces some significant distortions in how we think about intelligence — who has it, how much, in what form, and under what conditions it manifests.

The more tractable version of this question is an empirical one: what happens to measured performance when you systematically vary allocation conditions? When you reduce background cognitive load — through reduced financial stress, improved sleep, reduced threat environment — does measured performance shift in the ways the framework predicts? Some literature in behavioral economics and psychology suggests it does. The depletion effects of poverty on decision-making quality, for instance, fit naturally into an allocation account. But I haven’t done the reading carefully enough to know whether the specific predictions of the framework are well-tested, and I suspect a lot of the relevant work hasn’t been framed in quite these terms.

That’s the next step. The framework itself is at the level of productive speculation — coherent enough to generate specific predictions, not yet verified enough to treat as settled. Which is, I think, the right epistemic status to hold it at for now.

What It Changes

Even at the level of speculation, the framework has practical implications for how you think about minds you encounter — including your own.

The most immediate one is that expressed performance is a lower bound on capacity, not an estimate of it. When you see someone performing below what you’d expect given other evidence of their competence, the correct inference is not that the discrepancy is illusory. The correct inference is that something is consuming their allocation that isn’t visible from the outside. Stress, survival pressure, emotional processing load, social threat management — any of these can suppress expressed performance while leaving underlying capacity intact.

The second implication is about what it means to develop intellectually. If allocation is trainable — and I think there’s good reason to believe it is, given that it’s shaped by environmental demands and learned habits — then part of what you’re doing when you deliberately practice certain kinds of thinking is not just acquiring domain knowledge. You’re tuning your OS. You’re training the allocator to route resources toward the kinds of processing that compound over time: careful attention, careful reasoning, willingness to sit with unresolved complexity. The hardware matters, but the OS running on it is where a surprising amount of the long-run variance comes from.

And the third implication — the one I find most worth sitting with — is that there are almost certainly minds around us that look ordinary by the measures we have available, whose capacity is currently pointed somewhere we can’t see. The detection problem is real and probably significant in scale. Most of what we call intelligence as it’s usually measured is not a survey of the population’s cognitive potential. It is a survey of the population’s current allocation patterns, filtered through instruments calibrated to one particular configuration of those patterns.

That should make us considerably more humble about what we think we know.

The empirical questions here — about the structure of the allocation system, the degree to which it’s trainable, what interventions actually shift it — deserve more careful attention than I’ve given them. This is the philosophical frame. The next step is finding whether the literature has already built the building I’ve been imagining the blueprints for.

This essay was restructured and enhanced with the help of AI.