Within the next minute, you could perform any one of about a hundred different actions. Read the next paragraph. Eat an apple. Blink. Stand up. Call someone. Walk to the door. The list is open. Within the second minute, another hundred options. After two minutes, 10,000 combinations. After three, a million.
This is the combinatorial explosion that Tooby and Cosmides made famous, and it sits at the core of one of the longest-running fights in evolutionary psychology.1 If your mind were a single general-purpose information processor — a kind of biological computer with one all-purpose processing engine and no built-in priorities — the explosion would crush you. The number of possible behavioral options grows so fast that picking the right one requires constraints. Without those constraints, you cannot tell drinking water apart from drinking battery acid, eating fruit apart from eating poison, having sex with your sibling apart from having sex with a stranger. Some kind of pre-installed priority structure has to do the narrowing. Otherwise the search space is infinite and the organism dies before finding a workable behavior.
Tooby and Cosmides's answer: the priority structure is itself a set of evolved psychological mechanisms, each calibrated for specific recurring problems. Pinker's compact restatement: "The idea that a single generic substance can see in depth, control the hands, attract a mate, bring up children, elude predators, outsmart prey, and so on, without some degree of specialization, is not credible. Saying that the brain solves these problems because of its 'plasticity' is not much better than saying it solves them by magic."2 Buss treats this position — that the mind is structurally many specialized things rather than one general thing — as the foundational architectural commitment of evolutionary psychology.
The opposing position is not crazy. Some researchers argue that humans face many novel problems that did not recur often enough in the EEA for specific adaptations to develop, so general-purpose machinery — general intelligence, working memory, classical conditioning, analogical reasoning — must also exist alongside the specific mechanisms.3 The fight is alive. Both sides have evidence.
Pin down the contrast.
A domain-specific cognitive mechanism is one tuned to a specific kind of input and dedicated to a specific class of adaptive problem. The eye is a paradigm case. The eye does not process smell. The eye does not process taste. The eye processes a narrow band of electromagnetic radiation, and within that band it has further specialization — color cones, motion detectors, edge detectors, face-recognition circuitry. The eye is many domain-specific mechanisms stacked.4 The mating-preference machinery is similarly specific: features the mating system reads as cues to mate quality (waist-to-hip ratio, facial symmetry, vocal pitch, status indicators) are not the features the food-selection system reads, even when the cues are sitting on the same person.
A domain-general cognitive mechanism, by contrast, is one that operates on inputs from many different domains using the same processing machinery. General intelligence is the canonical proposed example. So is working memory — the buffer that holds whatever you're currently thinking about, regardless of whether that thing is a phone number, a map, or a face. So is classical conditioning — the learning rule that lets you associate any stimulus with any other if they reliably co-occur. Domain-general mechanisms, if they exist, are content-blind. They operate on what's there without caring what kind of thing it is.3
Buss takes a clear stance on the architectural question. The mind is overwhelmingly composed of domain-specific mechanisms. He grants that some mechanisms might be more general (working memory, certain forms of analogical reasoning), but the burden of proof for any general mechanism is high — you have to show the mechanism cannot be decomposed into specific ones operating in coordination.5 The default assumption is specificity.
The reason for the default assumption is the carving-the-mind-at-its-joints argument. Adaptive problems are specific — selecting safe food, avoiding poisonous snakes, identifying fertile mates, recognizing kin, deterring rivals.6 Selection works on specific design features that solve specific problems better than alternative designs at each step. Selection cannot, in principle, operate on something as vague as "general problem-solving capacity" because the unit of selection has to be a heritable variation that produces a measurable fitness difference. Heritable variations in specific functional design produce measurable fitness differences. Heritable variations in domain-general "capacity" do not — at least not directly, and not in a way that selection can grab onto.
A subsidiary distinction matters. Modularity in Fodor's narrow sense means a mechanism is informationally encapsulated — it cannot access information from other mechanisms.7 Buss explicitly rejects this version. Domain-specific mechanisms in evolutionary psychology are functionally specialized but not informationally walled off. Sight, smell, hunger, and memory of past meals all feed into a single decision about whether to eat. Adaptations talk to each other.7 Superordinate regulatory mechanisms coordinate which EPM gets behavioral control when multiple are activated simultaneously — the lion-bushes-mate scenario where you flee the lion first.7
Take the strongest single piece of evidence for domain-specificity: the Wason selection task and its variants.8 You're given four cards on a table, each with a letter on one side and a number on the other. You can see only one side of each — say, "A," "B," "2," "3." You're asked which cards you'd need to turn over to test the rule "If a card has a vowel on one side, then it has an even number on the other side." Most people get this wrong. Around 10 percent of college students answer correctly. The rule is logically clean, but most subjects flip cards that don't help and skip cards that do.
Now reframe the same logical structure as a social-contract problem. You're a bouncer at a bar, the rule is "If a person is drinking alcohol, they must be at least 21," and the four people in front of you are: a beer drinker, a soda drinker, a 25-year-old, and a 16-year-old. Which do you check? Around 75 percent of subjects answer correctly: the beer drinker and the 16-year-old. The logic is identical. The performance is dramatically different.8
The same gap shows up across cultures. Sugiyama, Tooby, and Cosmides ran versions of this experiment among the Shiwiar, a foraging society in Ecuador, with similar results — performance was high on social-contract framings, low on abstract framings.9 Stone and colleagues found brain-damage dissociations: patients with damage to the orbitofrontal cortex and amygdala did fine on precaution-rule reasoning ("if you engage in hazardous activity X, you must take precaution Y") but failed catastrophically on social-contract reasoning ("if you take benefit X, you must pay cost Y").10 Different reasoning circuits, different damage profiles, different content sensitivities.
The conclusion the data force: human reasoning is not a single domain-general capacity that gets applied to whatever content you put in front of it. Reasoning is a federation of content-specific mechanisms, each tuned to a particular kind of problem the EEA recurrently posed. Cheater-detection in social exchange is one mechanism. Hazard-avoidance is another. Abstract logic gets handled poorly because there was no recurrent abstract-logic problem in the EEA.11
A second line of evidence comes from specialized learning. Children learn to fear snakes, spiders, heights, darkness, and strange men readily. They do not learn to fear cars, electrical outlets, or guns even though those are more dangerous in modern environments.12 The fear-acquisition system is not domain-general. It has a built-in priority for content that the EEA selected against, and it is largely insensitive to content the EEA never met. If learning were truly domain-general, modern dangers would be acquired as easily as ancestral ones. They are not.
A third line: the three specialized learning mechanisms Buss highlights.13 Incest avoidance is learned, but the learning mechanism uses one specific cue — childhood co-residence — and produces one specific output: sexual aversion to people who shared early development. Food aversions are learned, but through a specific cue (nausea pairing with consumption) and a specific output (revulsion at sight and smell of the food). Prestige criteria are learned, but through a specific cue (the local attention structure) and a specific output (imitation of the high-attention individuals). These are three different domain-specific learning mechanisms. None of them is general associative learning in disguise.
The domain-general advocates have responses. Chiappe and MacDonald argue that humans face many problems that did not recur reliably enough for specific adaptations — handling rapid environmental change, learning entirely new tools, coping with novel social structures. Geary and Huffman point to general intelligence as a real, measurable, heritable capacity that predicts a wide range of life outcomes that no single specific mechanism predicts as well.3 Kanazawa makes a paradoxical move: he argues that "general intelligence" itself is a domain-specific adaptation, evolved specifically to handle non-recurrent problems.3 The dispute is not settled. Buss himself notes that "at this point in the science of evolutionary psychology, it is premature to draw any firm conclusions about whether humans possess more domain-general mechanisms in addition to the specific ones."14
The strongest tension in this debate is that the two camps end up making different empirical predictions in some cases and the same predictions in others. Where they make different predictions, the data tend to favor domain-specificity (the Wason task results, the specialized fear acquisition, the dissociations from brain damage). Where they make the same predictions — when both a general mechanism and a coordinated set of specific mechanisms could produce the observed behavior — the choice between frameworks is more about parsimony and theoretical commitment than about the data.
A second tension is about what counts as a "problem." Domain-specificity advocates partition the world into a long list of recurrent adaptive problems, each with its own dedicated mechanism. Domain-generality advocates argue this list is too coarse — that real problems mix categories (a foraging trip involves food selection, predator avoidance, kin coordination, and navigation simultaneously) and that solving them requires a more general capacity to integrate across mechanisms. Buss's response is that integration is what the superordinate regulatory mechanisms do, and you don't need a domain-general core to coordinate domain-specific parts.7 The response is principled but does not fully dissolve the worry.
A third tension is about the empirical status of "general intelligence." Psychometricians have shown for over a century that performance on diverse cognitive tasks is positively correlated, and the common factor (g) predicts life outcomes from job performance to longevity. The domain-general advocates point to g as evidence that something general is doing real work. Domain-specificity advocates respond that g is an emergent statistical regularity arising from many specific mechanisms working at correlated efficiencies — not a single mechanism. The dispute remains alive in part because both readings are consistent with the correlational data.
A fourth tension is recent and acute. The rise of large language models since this textbook's 2014 publication has produced systems that exhibit broad domain-general competence — passing law exams, solving math problems, writing code, translating languages — through what is, architecturally, a single trained network with no built-in domain priors. Whether this updates the EP debate is open. One reading: LLMs are evidence that a domain-general substrate can in fact solve specific problems if it has enough data and parameters, which weakens the in-principle argument against domain-general human cognition. Another reading: LLMs work because they extract domain-specific patterns from massive training data, and what looks like a domain-general system is actually millions of implicit domain-specific submodules. Buss's text predates the data point, and the vault has work to do here.
A fifth tension is methodological. If you want to demonstrate domain-specificity, you need to show that performance on problem A doesn't transfer to performance on problem B even when both have the same logical structure. The Wason task does this beautifully. If you want to demonstrate domain-generality, you need to show a single mechanism handling structurally different problems with comparable efficiency. This is harder to demonstrate cleanly, because nearly any mechanism can be argued to be a coordination of more specific submechanisms.
Tooby, Cosmides, and Pinker form the core of the domain-specific position. All three trace the case to the in-principle argument: selection cannot operate on something as vague as "general capacity," so the cognitive architecture must be a federation of content-specific tools shaped by specific adaptive problems. Buss endorses this position throughout the textbook, gives it the most space in his Ch. 2 architecture discussion, and quotes Pinker's "single generic substance" passage as the field's rallying line.2 The convergence among these four authors is so close that the case for specificity reads as a unified position with minor variations in emphasis.
Where they differ is at the margins. Tooby and Cosmides are willing to allow only specific mechanisms; they treat domain-generality as a category error. Pinker is willing to accept a few proposed general mechanisms — working memory, classical conditioning, certain forms of analogical reasoning — as layered on top of the specific ones, while still treating specificity as the default. Buss reports both positions and writes from inside the specificity camp without endorsing the maximal Tooby-Cosmides version. The textbook's tone on this question is judicious rather than militant. Buss flags the open empirical questions and refuses to overclaim.14
Chiappe and MacDonald, Geary and Huffman, and Kanazawa form the opposing position with different specific arguments. Chiappe and MacDonald emphasize the non-recurrent problems argument: humans face many novel problems that did not recur in the EEA, so general capacity is required. Geary and Huffman point to the empirical reality of g and to the general-intelligence research tradition. Kanazawa stakes out the most distinctive position: general intelligence is itself a domain-specific adaptation, evolved specifically for evolutionarily novel problems.3 Buss summarizes all three without endorsing them and treats their position as worth taking seriously.
A separate tension runs around Fodor. Fodor's classical modularity argument was that mental modules are informationally encapsulated, fast, mandatory, and domain-specific.7 Buss, Tooby, Cosmides, and Hagen all reject the encapsulation criterion while keeping the others. EPMs are domain-specific but not informationally walled off; they share information freely with other EPMs and with whatever passes for general processing. Fodor's framing risks confusing a vault reader who imports the encapsulation assumption when reading EP-style modularity claims. The vault implication: when EP authors write that a mechanism is "modular," they mean function-specific, not informationally walled.
The deepest convergence across all camps is on what the architecture is not. Nobody in the modern debate defends the empty-slate behaviorist position that humans are general-purpose conditioning machines whose content is wholly written by environment. Even the domain-general advocates accept that general capacity operates on substrates with built-in priors. The debate is now about how many priors and how specific they are. The blank-slate position has been retired by the entire field, including its critics.
The architecture debate has direct implications for how to think about artificial intelligence. Large language models trained on internet-scale text approach the empirical test of the domain-general hypothesis from an angle the original framers did not anticipate. A transformer is, architecturally, close to a domain-general substrate. It has no built-in priors for face recognition, food selection, or kin recognition. It learns whatever patterns are in its training data. And it solves remarkably specific problems — passing bar exams, writing functional code, diagnosing medical conditions — that were not anticipated as separate adaptations in any architecture.
Two readings are possible. Read the LLM evidence as supporting domain-generality, and the EP framing of human cognition has to absorb a substantial update: a sufficiently capable general substrate can in fact solve specific problems, and the in-principle argument that selection cannot build domain-general systems may be wrong. Read the LLM evidence as evidence of implicit specificity — millions of latent submodules trained from data — and the EP framing survives intact, just with the recognition that domain-specificity can emerge from training rather than being hardwired.
The second reading has the better fit with what mechanistic interpretability research is finding. Probing studies inside large transformers do find what look like specialized circuits — circuits dedicated to specific tasks (induction heads, indirect-object identification, face recognition in vision transformers) — emerging during training. The "general" substrate is mostly a name for the substrate; the actual computational work is done by emergent specialized structures. This connects to the vault's AI-assisted creator hub in a way that produces a working frame: AI systems that look domain-general at the interface level are operating as federations of domain-specific circuits at the implementation level. The architectural debate in EP and the mechanistic-interpretability debate in AI are addressing the same question from different sides. The insight neither domain generates alone: domain-specificity is not necessarily about built-in hardwiring; it is about functional specialization, and functional specialization can be installed by selection or by training. The substrate type matters less than the resulting structure of the mechanisms.
A second handshake runs to behavioral-mechanics. Operators who treat their targets as domain-general processors lose. The operator who assumes the target's mind is a flexible plastic system that can be persuaded by any sufficiently strong argument fails when the argument hits a wall the target's content-specific mechanisms refuse to cross. Targets do not have abstract preferences for "more value" that any pitch can satisfy; they have specific kin, mating, status, and threat mechanisms with specific input requirements. An operator who frames a sales pitch in domain-general terms ("this product is objectively better") gets weaker conversion than an operator who triggers a specific domain ("buy this for your daughter," "be the best in your peer group," "protect your family").
The connection to existing BM pages: Hughes's elicitation framework and Hughes's decision-pillar work both implicitly assume domain-specific architecture. Each "pillar" is a separate mechanism with separate triggers; you elicit decisions by hitting specific pillars, not by some general persuasion strategy. The insight neither domain generates alone: the operator's job is to identify which content-specific mechanism the situation requires triggering, then to deliver inputs to that specific mechanism. Bag-of-tricks BM catalogs become coherent when re-read as targeted activation of content-specific machinery. The vault's BM hub gets sharper when each technique is annotated with which mechanism it targets.
The Sharpest Implication.
You are not what you think you are. You are not a unified rational agent processing arguments and reaching conclusions. You are a federation of specialized machines, each waking up when its specific input arrives, each running its own decision rules, each pushing output into the system you experience as yourself. Most of the time the federation feels unified because the machines coordinate well and their outputs converge. In moments where they disagree — fear meeting hunger, status meeting kinship, lust meeting commitment — what feels like a struggle of the will is a competition among specialized mechanisms for behavioral control.
This means most arguments fail. When you try to convince someone of something through reasoning, the argument has to land on the specific mechanism that handles the relevant content, and that mechanism may have its own decision rules incompatible with the reasoning. Telling someone "the data show you should change your mind" works when the relevant mechanism reads data as input. It fails when the relevant mechanism reads kin loyalty as input, or threat-detection as input, or status as input. The same argument lands or fails depending on which mechanism is online.
This also means that self-improvement cannot be a single project. The domain-general view treats self-improvement as building a better general processor — more discipline, more rationality, more willpower. The domain-specific view treats self-improvement as the cultivation of specific mechanisms in specific directions — calming the threat-detection mechanism, recalibrating the status-anxiety mechanism, strengthening the kinship-extension mechanism. These are not the same project. The general framing misses how the work actually has to be done.
Generative Questions.
If LLM domain-generality is real, does it update the EP architectural commitment, or does it just establish that trained systems can be domain-general while evolved systems cannot? What does the difference between training and evolution allow each to build?
If self-improvement requires cultivating specific mechanisms rather than a general capacity, what does the curriculum look like? Different traditions have built different curricula — meditation traditions for some mechanisms, stoic practices for others, modern therapy for others. Are they working on different mechanisms, or the same mechanisms with different names?
If most arguments fail because they don't land on the right content-specific mechanism, what's the correct method for changing minds? The current rhetorical tradition assumes a domain-general listener whose reason can be addressed directly. The EP framing implies a much more targeted approach — diagnose which mechanism is in control, deliver the input that mechanism reads, accept that other mechanisms in the same listener will not move.