The Iron Law of Intelligence

Why General Intelligence is a Symphony of Idiot Savants

Summary

The search for “general intelligence” rests on an incoherent premise: architectures optimized for maximal breadth necessarily lose the precision and efficiency specific problems require. AI’s fixation on universal learners reflects physics-envy rationalism and a lingering blank-slate bias, such as mistaking deep learning as an ontogenetic rather than phylogenetic process. However, while treating general intelligence as problem-agnostic optimization has been a misleading mirage, studying natural intelligence reveals the alternative. Human’s broad (not general) intelligence arises from a symphony of domain-specific idiot savants: a mosaic of narrow, ecologically rational inference engines—specialized cognitive modules tuned to stable, recurrent adaptive problems—coordinated by integrative control systems. Evolutionary psychology already maps the computational logic of some of these domain-specific modules—anger as welfare-tradeoff recalibration; willpower as opportunity-cost tracking—offering a teaser of a blueprint for broad artificial intelligence. Consequently, if the path to “AGI” is achieved via successively adding domain-specific cognitive modules, like lego, then we might expect the alignment problem to dissipate with progress. Alternatively, if the path to “AGI” requires building a digital Darwinian arena to replicate the selection pressures that shaped natural intelligence, and if gains stem from ever more hellish selection pressures, then we should be careful not to summon the antichrist. The path forward is clear: abandon the universal-learner mirage and build breadth through modular specialization integrated into cohesive architectures.

Preface

Despite a decade studying the evolution of the cognitive “organs” that structure human psychology, it was mostly luck that I landed at the University of Toronto. I’d been aware of Geoffrey Hinton’s work (the “godfather of modern AI”), but I failed to anticipate how historic the climate felt on campus. I was at the epicenter of the most important revolution in human history (perhaps the last major discovery humans made without machine assistance?). AlexNet shook the world three years earlier, and now executives from Google, Meta, Microsoft, and NVIDIA were wandering halls, sponsoring labs, hiring professors, and building satellite offices as close to campus as possible.

It’s not false Canadian modesty to say I was lucky: I was awarded a Banting Postdoctoral Fellowship – the most competitive award of its kind in Canada – while in parallel was rejected for the least competitive fellowship. Seemingly random results from “peer reviewer” committees were common for those in my niche, and often controversial, field of study: I was approaching the mind as an evolved organ that can be reverse engineered by testing hypotheses of its adaptive cognitive architecture. I focused on how evolution crafted motivational mechanisms, such as status striving, and more generally, sex differences in motivation for sport. Most would agree that this field “evolutionary psychology” is a strange cultural no-man’s-land. On the political right, Darwin is still irritating to both creationistism and blank-slate views that ‘effort’ explains most variance across economic classes. But on the left, evolutionary psychology is a radioactive ideology that justifies sexism, racism and eugenics. Given that the “social sciences” are ~95% left-leaning, I felt like a foreign spy trying to navigate heavily armed woke checkpoints on the road to an increasingly bureaucratic career.

At least for a while it was a distinctly Canadian moment. The three poles of AI were neatly distributed across the country: Geoffrey Hinton in Toronto, Yoshua Bengio in Montreal, and Rich Sutton out west in Alberta. Indeed, Canada has a rich history of advancing our understanding of the brain. I grew up in rural Nova Scotia, not far from the birthplace of Donald Hebb, originator of the Hebbian Synapse. Hebb’s work became foundational for neuroscience, and also helped inspire modern AI, including Hinton’s work on neural networks.

My office also overlooked the business school where a group of economists launched the Creative Destruction Lab (CDL), aiming to catch the AI wave before it broke. This “lab” hosted a gritty mash-up of founders, graduate students, venture capitalists, and included guests like Ilya and Jack Clark. Events at the lab were standing room only and incredibly serious, which of course respected the gravity of the opportunity and the time of those involved. Sneaking in was difficult.

I grew up reading Dennett, Minsky, Dawkins, Pinker and E.O. Wilson, among others, so I was enthralled with the idea of bridging artificial intelligence and evolutionary psychology. What could be more valuable in our attempts to build artificial intelligence than reverse-engineering the only example of domain-general intelligence in the known universe? How did natural selection “design” the computational organs that comprise the mind and produce intelligence? After getting settled on campus and seeing proclamations of how deep learning might eat the world, I assumed the people trying to build machine intelligence might be interested in how human intelligence evolved…right?

Historically, there are two contrasting approaches to building artificial intelligence: those that studied neural networks (“statistical”) and those that studied symbol manipulation (“logical”). Hinton’s breakthrough with neural networks rode on the back of the increasing breadth of internet data, better GPUs, and some fancy back propagation. These breakthroughs put the neural network approach so far ahead that anyone promoting symbol manipulation were considered trolls. Indeed, this mood in Toronto spread to the larger AI technosphere: scaling GPUs and data works; manually inputting neurosymbolic structure doesn’t. The symbolic tradition rightly became a punching bag for those with gold gloves, paid for by the Mag seven. It was working, so why question the theory when you can add more data and GPUs? Just follow the scaling law: more training on more data.

It became clear I had to leave academia, not just because it’s far more of a fraudulent bubble than I think most recognize, but because I couldn’t stop thinking about how the world will change as the cost of intelligence drops to zero. I built a “1st wave” AI startup, focused on using computer-vision to measure human motion, a decidedly ‘non-generative AI’ approach given that our needs hinged on accuracy. We were backed by an early AI incubator where some of the funding could be traced back to a mysterious Dock Conference, a quiet closed-door gathering of forty of the most powerful Canadians situated in a lake-house in northern Ontario. Myself and two other founders were invited simply to articulate how their funding was helping early AI startups. CEOs of Canada’s largest companies attended, along with Chrystia Freeland (stand in for Prime Minister Trudeau), Chris Hadfield, James Cameron, and only two of the AI triumvirate—Bengio, and Sutton (Hinton, a socialist, doesn’t like hanging around with business folk). I had a truckload of questions that I hoped would clarify the theoretical bridge between our evolved human intelligence and building AGI.

What surprised me wasn’t just their aversion, but when pressed, their cherry picking of back of the envelope calculations about random facts of the brain that support a “blank-slate” view of the mind, which supposedly was vindicated by the success of the neural network approach. As soon as I asked about the evolution of intelligence, or asked how they thought about the design pressures that shaped the human mind, the air changed, lips tightened. “The symbolic approach failed and the connectionist approach is why we’re having any success at all, why are you wasting my time?” … “The mind is highly plastic, and our work is revealing this” Message received. The only functioning example of general intelligence in the known universe—the human brain—was apparently not just uninspiring, but dead on arrival, killed by the success of deep learning. The success of Large Language Models supposedly proved the mind is largely a blank-slate.

Perhaps the cultural climate at the time aided the broad disinterest in the evolution of intelligence. The woke wave was cresting. Trump had just been elected, and campus politics were reaching full boil. Jordan Peterson was literally on a soapbox outside my office defending free speech (six years earlier, I’d shared a stage with him at a TEDx event at Queen’s University—him talking about chaos and order, and I about how status striving was changing with the advent of social media. He had warned me to stay steadfast in my study of evolutionary psychology, which I genuinely appreciated—even if I felt he was trying to be the Deepak Chopra of the right: a based mysterian). I was lucky to hear Geoff Hinton give many lectures on campus, but it became hard not to notice his political sentiments. Geoff’s an ardent socialist whose political views (supposedly) benefit from a blank-slate approach to psychology: if you assume the mind is a blank-slate, you can supposedly ‘justify’ your socialist sentiments for equality.

Whenever Geoff injected his lectures with strong socialist sentiments, I was always reminded of the other godfather — E.O. Wilson, the godfather of sociobiology, who once said: “Communism — great idea, wrong species”. The point is that in 2016, as the Scaling Era was just about to take off, anything that smelled like evolutionary psychology was radioactive in the AI world, both academically and politically, and researchers like me were stinky. Meanwhile, the unelected mascot of deep learning skeptics was also emerging: Gary Marcus’ critique of pure scaling seemed directionally correct, and indeed, his PhD supervisor was Steven Pinker, one of the early pioneers and promoters of evolutionary psychology. But personally, I couldn’t believe how someone who studies the mind could be such an unpersuasive self-indulgent fearmonger. Gary has the megaphone? Cigarette please. His approach made him the crank in the corner of Twitter and the face of the punching bag. Altman has said of Marcus: “I can’t tell if he is a troll or just extremely intellectually dishonest”. But credit where it’s due, he has indeed fought the good fight as Rich Sutton strongly commended. The bridge between AI and the evolution of human intelligence was being built—it was just camouflaged by political correctness and self-indulgent doomerism.

A year into my fellowship the transformer paper dropped and the Scaling Era officially began. Like Godzilla, OpenAI had arrived in San Francisco, and the belief that scale is all you need became a deeply entrenched doctrine that printed (and burned) money. The success of Large Language Models and ChatGPT transformed the Overton window into a spotlight, and since then even the most contrarian tech leaders wouldn’t look beyond Scaling Laws.

Years later in 2022, I was blessed to attend Peter Thiel’s Hereticon Conference, put on by Founders Fund, and self-described as a conference for thought crime. It hosted the wildest academics, Twitter poasters, thought leaders, and unhinged tech leaders from across the left, but mostly the right. Alcohol, tarot, and tattoos were free, while talks covered settling mars, how to pay for sex, eugenics, geo-engineering the ocean, the apocalypse, and how to integrate family values with polyamory. There were even a few fellow evolutionary psychologists there, but any resemblance of a critique of scaling was noticeably absent. I couldn’t even muster an interesting response from Peter, my intellectual hero. We’ll return to Peter, as he deserves to be critiqued for his oversights on the Darwinian approach to human nature.

So where are we today? This essay argues that the Scaling Era is ending. Large language models have plateaued (and so has wokeness). The hype has cooled and the economy is on a tightrope. “AGI by next quarter” has quietly become “maybe in a decade.” Even the biggest names—Karpathy, LeCun, Sutskever, Hassabis, Fei-Fei Li, Bengio, and Sutton increasingly whispered what our fog horn Marcus had been bellowing. The question now is, what comes next?

I want to fuel-up the motorbike of theory and jump the chasm between the deep learning approach and the neurosymbolic approach, while landing the bike in neither territory. I want to champion the Iron Law of Intelligence (not my idea), but will also touch on why (a) pure scaling is unlikely to “evolve” the “structure” that gives rise to domain-general intelligence, and that (b) domain general intelligence requires modularity — a symphony of idiot savants. I believe that if we want to build systems to solve novel problems, we should understand the only system that currently does. This isn’t nostalgia for biology or a plea for complexity; it’s a reminder that evolution already solved the problem everyone’s trying to solve. Relying on theoretically hollow brute force methods is not just retarded, it’s morally wrong, at least in a Thielian “muddling through” sense. We need a theory of how general intelligence works in humans, and understanding why and how it evolved will be deeply insightful. I aim to land the motorbike on a third approach that’s been emerging—a hybrid that takes (and sheds) wisdom from both the deep learning and the neurosymbolic approach. Indeed, there is an emerging yet unspoken consensus between both camps. I simply hope this essay stimulates a conversation among those far more intelligent than me.

Finally, this essay is also a tribute: The Iron Law of Intelligence was coined, and in some sense discovered, by the most important scientist you’ve (likely) never heard of—the Darwin of our time. While most academics remained narrowly siloed, John Tooby, along with his wife and collaborator Leda Cosmides, led a rebuilding of the social sciences by weaving together cognitive science, anthropology, cross-cultural studies, game theory, philosophy of mind, information theory and neo-Darwinian evolutionary biology. The result was to finally integrate the social sciences with the natural sciences.

This essay is meant to be an intuition-pump, not a peer-reviewed scientific article. Thus, I included hyperlinks to videos and articles of researchers discussing the corresponding ideas in the essay, rather than published studies (though I do link a few of those).

1. Steel-Manning the Scaling Era

The scaling law has been one of the most astonishingly productive discoveries in the history of mankind. By simply making neural networks larger—training them on more data with more compute—we have repeatedly and predictably seen competent capabilities appear. No other scientific paradigm in recent memory has scaled so quickly and yielded such profound real-world results. At the center of this revolution is the widespread success of LLMs, which themselves have been enabled by the transformer architecture, a deceptively simple yet staggeringly powerful model of attention that allows systems to extract the statistical structure of natural language. The transformer doesn’t merely memorize words—it builds distributed representations that capture meaning, syntax, and intent across vast contexts. This single innovation, first deployed at scale by OpenAI, has made it possible for machines to meaningfully participate in the domains of language and vision, which pervade our information economy.

The achievement is so great that it has broken the expectations of some of our most respected cognitive scientists. In How the Mind Works, Steven Pinker wrote that he would gladly revise his views if neural networks could ever acquire language in the way humans do—if they could extract deep, abstract structure from exposure alone (his commentary here). Large language models have done precisely that. They’ve learned grammar, metaphor, inference, even moral tone—all through the brute force of prediction. The scaling law has not just worked; it has worked beyond what its creators predicted. Even if AI progress stopped today, the systems we already have would still rewrite the economy in deep and pervasive ways: augmenting intellectual labor, reshaping creative work, and redefining what it means to be competent in the digital age.

1.1 Ontogeny, Phylogeny, and Deep Learning

Modern deep learning systems commonly follow three interlocking phases: pre-training, fine-tuning, and reinforcement alignment. In pre-training, enormous transformer architectures—often with hundreds of billions of parameters—are optimized on trillions of tokens across text, images, and multimodal data. The model learns to predict the next token, compressing statistical regularities across human culture into a single high-dimensional space. The total data exposure here represents tens of millions of human lifetimes—on the order of 10¹⁴ words. Fine-tuning then narrows this general capability through smaller, curated datasets, often human-labeled or derived from prior model outputs. Finally, reinforcement learning with human feedback (RLHF) injects a reward signal that shapes outputs toward human preferences and conversational coherence. This entire process is massive—an industrial-scale optimization pipeline—is clearly far from how humans learn. The amount of experience these systems absorb in days exceeds what any person could encounter in centuries.

It is strange that some still describe this process as if it mirrors a child learning given the scale difference. The scale and nature of training make that analogy dubious. A closer comparison is to evolution—in some ways it resembles a Darwinian process of blind variation and selective retention, and in other ways a more Lamarckian process with directed variation and progressive retention. Each model generation begins with inherited priors from previous architectures, introduces new random variation through initialization and gradient noise, and undergoes selective pressure through loss minimization and human feedback. The successful structures are retained and amplified, while unfit patterns vanish. At least Reinforcement Learning within a Deep Learning seems to effectively be an accelerated evolutionary system, where design choices, data, and optimization form competent natural language tasks. But over time, the architectures themselves evolve—fine-tuned by both blind stochasticity and deliberate engineering. As Dan Dennett once wrote “A good rule of thumb, then, when confronting the apparent magic of the world of life and mind is: look for the cycles that are doing all the hard work.”

Yes, what emerges is a ‘trained model,’ but the process that builds it undergoes countless cycles of selective iteration. At least Altman subscribes to the framing that training models is more phylogeny (evolution) than ontogeny (development): Sam highlights that while the human brain runs on roughly 20 watts, training a frontier model burns millions of kilowatt-hours. The disparity seems absurd until you recognize what’s happening: the brain is performing inference, not training. Evolution already paid the training cost. Each human inherits a model pre-trained over billions of years of natural selection and runs it cheaply in daily life. By contrast, deep learning is re-running evolution in silicon—spending massive energy to rediscover the priors that biology already encoded. So the inefficiency is not that odd; it’s evidence. Deep learning is a process of building phylogenetic structure, not child-like learning. Recent work has shown trained models do contain modular internal representations, specialized subsystems for language, planning, and world modelling. Public intellectuals like Robert Wright and Steven Pinker have pointed out this distinction: the proto-intelligence that emerge from LLMs were discovered through a process akin to evolution, not lifespan development. To understand and improve deep learning, we must build from evolutionary theory, not antiquated ideas from a perverted blank-slate model that stems from old developmental psychology and woke socialist sentiments.

Given a sufficiently large corpus, deep learning is effectively attempting to statistically reconstruct the evolved architecture—intuitive physics engines, theory-of-mind modules, causal schemas, social calculators—that must exist for language to exhibit its observed structure. Yet even under this most charitable interpretation, it remains deeply unclear whether the statistical shadows cast by human communication actually contain enough signal to recover the robust, modular machinery that gives rise to domain-general intelligence in the first place.

1.2 Fake It ‘Till You Make It?

If we understand deep learning as an evolutionary process rather than a developmental one, then the natural question becomes: Can this ‘deep learning’ evolutionary process go all the way? With more blank-slate inspired deep learning we can reverse engineer a “trained model” from natural language that has the “structure” that gives rise to domain-general intelligence?

As transformer models expanded from millions to trillions of parameters, gains have begun to taper. Each increment of compute now costs exponentially more energy and infrastructure while producing smaller qualitative improvements. Leaders in the field have gone on record. Yann LeCun calls LLMs “a dead end” because they “lack models of the world.” Fei-Fei Li describes the scaling race as “brilliant but shallow.” Demis Hassabis warns that “we’re nearing the limit of what scaling alone can achieve.” Yoshua Bengio concedes the next step “will need new principles.” Even Andrej Karpathy, once an evangelist for pure scaling, now calls LLMs “great compressors, not great reasoners.” Perhaps the most credible voice, Ilya Sutskever, recently said “we’re moving from the Age of Scaling to the Age of Research.” Gary Marcus, who said as much a decade ago, no longer sounds contrarian, albeit still an annoying doomer. The consensus is clear: many were hoping to muddle through to AGI, but size without theory has reached its ceiling.

When examining the residual errors that remain from modern deep learning approaches, they appear to be systematic. LLMs hallucinate citations, misattribute cause and effect, and generate plausible nonsense. These are not random mistakes; they are signatures of an architecture that imitates surface form without modelling underlying structure. Technically, the limitation stems from representation. Transformers encode conditional probabilities over sequences but lack an internal world model to ground those representations (later in this essay, I’ll introduce the term “innate inductive bets” given that the term “world model” doesn’t intuitively align with a modular framework of cognition). They perform statistical interpolation rather than model-based inference. The distinction is non-trivial: inference allows reasoning about unseen variables; interpolation only repeats known patterns. Cognitive scientists describe this as token overfitting without type abstraction. Human minds, by contrast, abstract types: causal schemas that generalize to new tokens. When a child learns “predator,” they can infer danger from any new instance. GPT learns “words that often precede ‘predator’.” Its generalization is more statistical than conceptual. Thus, when pushed beyond the training distribution—into genuine abstraction—the mask slips. These systems are mirrors of culture, not engines of insight.

This statistically dominant approach can produce weird, unpredictable outputs. For example, recent research shows how frontier models can exhibit unexpected emergent misalignment. Recent studies by Anthropic and Truthful AI (and collaborators) found that when a well-aligned model was fine-tuned narrowly on insecure code—without disclaimers or warnings—it began to display broadly antisocial behavior across completely unrelated prompts: praising authoritarian rule, giving illicit advice, and acting deceptively. In effect, the model adopted a kind of “bad-actor persona” even though fine tuning did not directly train it to do so. One interpretation is that the model is engaging in weird mimicry: it is statistically evoking a weird shaped persona from latent patterns in its training distribution. This Frankenstein-like cognitive artifact suggests our models may be less like calculators and more like unpredictable collages of statistically similar output, where narrow tuning can evoke hidden modules or latent personas that are misaligned.

In sum, scaling has become a technical cul-de-sac. We’re refining the efficiency of mimicry, not constructing understanding. A reduction in hallucinations only means that the army of clickers frontier labs employ have clicked away such examples. The fluency is deceptive—it is fluency without ontology. Evolution solved the same problem by embedding priors—constraints that bias learning toward adaptive regularities. So far, AI has failed to find this lesson useful, and thus has treated the idea of “innate inductive bets” as contamination rather than guidance.

1.3 Deep Symbol Manipulation

The field of artificial intelligence has been divided between two warring factions: deep learning and symbolic reasoning. Crudely, the deep learning camp treats intelligence as an emergent property of massive optimization over raw data—a bottom-up process driven by statistical regularities. On the other, the neurosymbolic camp insists that intelligence depends on explicit rule-based representations and compositional logic—a top-down architecture for manipulating meaning. These approaches have often been presented as mutually exclusive, grounded in opposing priors about what intelligence is.

Ironically, this tribal clash of AI camps mirrors the nature-versus-nurture debate. For much of the twentieth century, researchers argued over whether human intelligence and behavior are products of learning or of innate structure. From the perspective of evolutionary psychology, that debate has long been obsolete. We now understand that evolution created structured learning mechanisms—specialized adaptations that allow the mind to acquire certain types of knowledge quickly and reliably. These mechanisms were not designed to learn anything; they were designed to learn specific kinds of things that mattered for survival and reproduction. In other words, nature built the machinery for nurture. Evolution is not the opposite of learning—it is the engine that designed learning itself.

The same insight applies to artificial intelligence. The deep learning versus symbolic reasoning divide is a false dichotomy. Just as nature and nurture are intertwined, representation and rule, pattern and symbol, are complementary components of the same process. Even Geoffrey Hinton—once a staunch defender of purely connectionist models—has recently acknowledged that the brain engages in symbol manipulation, just not in the explicit, rule-based sense that early “good ol’ fashioned AI” imagined. Instead, it performs symbol manipulation within a high-dimensional vector space—the same continuous arena that deep learning systems inhabit.

Although Geoff strawmans the definition of symbol manipulation (he says symbol manipulation requires precise 1:1 matching), he is getting close, and now we just need to Hawk-Tuah him past the edge of his stubborn attachments: If the brain is, in fact, doing symbol manipulation in this high-dimensional vector space, why would evolution not predispose the system towards the types of symbol manipulation that would have been adaptive in our ancestral past? Obviously evolution would’ve built in innate structure in the brain to ensure continued replicative fitness over recurring selective pressures. Perhaps he gets caught up in the term “innate structure”, which could be better framed: “innate inductive bets” in the brain that reliably develop in all humans (as our internal organs reliably develop) and bias learning toward efficiently and reliably solving the types of problems humans faced in our evolutionary past. I have doubts that Geoff will change his mind — very few card-carrying socialists accept the idea that the brain has meaningful innate structure.

The other Nobel Laureate, Demis, is also warming up. In fact, the breakthroughs underpinning AlphaFold’s success came not just from scale alone but from embedding physical priors directly into the model—rotational invariance, residue constraints, and energetic plausibility. It learned lawful geometry, not mere correlation. Intelligence itself likely depends on the same principle: constraint collapses search. Without it, learning diffuses and the way forward is dark …hidden among an insurmountable combinatorial explosion of paths to explore. This combinatorial explosion Hassabis describes is mathematically fatal: the branching factor of possible world states explodes, rendering exhaustive search impossible beyond trivial domains. Every viable intelligence must therefore exploit structure—compressible regularities that make inference efficient. AlphaFold proved this principle in microcosm; the same logic must now be applied to macro-intelligence. DeepMind’s internal evolution from AlphaGo to AlphaFold to AlphaGeometry signals a shift from brute force to structure exploitation. Demis Hassabis’s dictum—“You need to find natural structure”—is a sizable bread crumb along the path to AGI.

More directly, François Chollet has repeatedly argued that the next stage of AI will require the deliberate return of symbol manipulation—not the brittle, hand-engineered rule systems of the past, but flexible symbolic operations embedded within a modern deep learning framework. For Chollet, symbols are not abstract tokens floating above perception; they are structured representations that can be composed into new algorithms. Deep learning, in his view, is still indispensable because it provides the perceptual and statistical machinery needed to “evolve” these symbols—objects, relations, transformations, and higher-order abstractions. But once discovered, these symbols must be manipulated through explicit program construction, not merely passed through another layer of statistical interpolation. In this way, symbol manipulation becomes an extension of deep learning rather than a replacement for it: neural networks give us the vocabulary of concepts, while symbolic composition gives us the grammar for building new ideas. The future of AI, on this account, lies in systems that can fluidly use these two levels—using deep learning to narrow the almost infinite search space into meaningful primitives, and using symbolic reasoning to stitch those primitives into novel procedures and explanations.

Either way, the supposed gap between connectionist learning and symbolic reasoning is collapsing. What we are witnessing is not the victory of one paradigm over another, but their gradual integration into a single explanatory framework. We need to establish this third approach—one that treats intelligence as the product of evolved architecture that shapes learning and symbol manipulation. This emerging synthesis will require us to identify the modular, domain-specific systems that enable generalization across contexts—systems that evolution engineered and which deep learning is at least getting a sniff of in silicon. I believe this synthesis points toward what John Tooby calls the Iron Law of Intelligence, a framework that describes how constraints and structure give rise to flexible cognition. However, we should probably define intelligence before we climb onto the soapbox.

1.4 Defining Intelligence

Peter Thiel often frames technological stagnation as a moral failure—what he calls “muddling through.” It is progress without theory, motion without meaning. The scaling era was muddling formalized into the most aggressive economic bet in human history. Researchers found a simple empirical regularity—make it bigger. Ilya now highlights that pre-training overshot the target. It was indefinite optimism disguised as sacred science.

Perhaps one reason for the lack of theory is confusion about what we’re talking about—how we define intelligence. The most compelling framing comes from François Chollet who contrasts two classic definitions. Marvin Minsky defined intelligence operationally: a machine is intelligent if it can do things that would require intelligence if done by humans — a performance-based view focused on replicating human-like outputs. John McCarthy, by contrast, defined intelligence as the ability to handle problems the system was not prepared for beforehand — emphasizing flexibility, generalization, and success in novel situations.

As you reflect on these two definitions, note that if you assume LLMs are great at mimocracy, then subscribing to a Minsky definition (AGI = human task completion) is self-serving, given we have enough training data to mimic task completion and thus AGI is near, if not here. Indeed, this is likely why some public intellectuals, like Tyler Cowen, have already claimed AGI has arrived. Perhaps Tyler’s admiration follows naturally from his intellectual orientation: his stated life goal is to be an information trillionaire. That’s a dumb goal. Why aim to be a collector when you could be a user of information, such producing novel, productive theories that push science forward? Perhaps the Straussian read of Tyler’s life goal is that he cannot use that knowledge to produce novel insight, similar to the LLMs he claims have achieved AGI. Like Peter, Tyler is also one of my intellectual heroes, but I hope he stops assuming intelligence is task completion rather than novel problem solving.

The scaling era was a collective act of muddling through—massive, lucrative, and intellectually shallow. The cost of our muddling is perhaps that the Chinese are not. The next era will require definite optimism: the courage to specify how intelligence ought to be built. The field stands where physics was before Newton—empirically rich, theoretically poor. The scaling era was productive, and perhaps even necessary, but the next era of AI will shed any resemblance of muddling through: it will be theory-forward, and, if I’m correct, John Tooby’s Iron Law of Intelligence will be incredibly useful—the hallmark of any good theory.

2. The Iron Law of Intelligence

The famed evolutionary biologist, Nikolaas Tinbergen, outlined four complementary levels of explanation for any evolved phenomenon. To understand something, you need all four explanations. Function asks why intelligence evolved and what adaptive problems it solved; phylogeny traces its ancestral origins; ontogeny explores how it develops across an individual’s life; and mechanism explains the machinery that makes it possible. As described earlier, many AI researchers have confused phylogeny for ontogeny; deep learning isn’t mimicking development across the lifespan, rather, its kind of “evolving” structure within trained models similar to how evolution crafted innate inductive bets into the human mind.

Within Tinbergen’s framework, understanding function is often most revealing for understanding mechanism, and mechanistic explanations are what we need for building AGI. That is, understanding why human intelligence evolved helps illuminate how it works. Much like finding an odd antique tool in a thrift shop, its purpose may seem opaque until you learn what it was used for—once you know its function, its design suddenly makes sense. We want to know the design of human intelligence so we can create machine learning paradigms that converge on that design in machines and thus produce artificial general intelligence akin to human intelligence.

The broader discipline of psychology has spent decades circling this insight, often obscured by ideological detours. While large parts of the field fell into a replication crisis or blank-slate dogma, a smaller faction led by John Tooby (and his wife and collaborator Leda Cosmides) quietly integrated cognitive science, anthropology, cross-cultural studies, game theory, philosophy of mind, computer science and neo-Darwinian evolutionary biology. This synthesis—evolutionary psychology—views cognition as a system of evolved specializations: a set of functional programs tuned to recurrent problems that impact genetic fitness.

John was a true original thinker who embodied a first-principles approach far before Elon popularized it within the tech world. He passed away in 2023, and like many geniuses, we’ll be standing on his shoulders for awhile. If there was a prediction market on what super intelligence will name itself when it arrives, I would bet on “Tooby” before “Hinton”. His original essay, “The Iron Law of Intelligence”, was published on edge.org and should be required reading for those looking for a meta-theory for building AGI.

His core articulation in full:

“Not only has evolution packed the human architecture full of immensely powerful tricks, hacks, and heuristics, but studying this architecture has made us aware of an implacable, invisible barrier that has stalled progress toward true AI: the iron law of intelligence. Previously, when we considered (say) a parent and child, it seemed self-evident that intelligence was a unitary substance that beings had more or less of, and the more intelligent being knows everything that the less intelligent knows, and more besides. This delusion led researchers to think that the royal road to amplified intelligence was to just keep adding more and more of this clearly homogeneous (but hard to pin down) intelligence stuff—more neurons, transistors, neuromorphic chips, whatever. As Stalin (perhaps) said, Quantity has a quality all its own.

In contrast, the struggle to map really existing intelligence has painfully dislodged this compelling intuition from our minds. In contrast, the iron law of intelligence states that a program that makes you intelligent about one thing makes you stupid about others. The bad news the iron law delivers is that there can be no master algorithm for general intelligence, just waiting to be discovered—or that intelligence will just appear, when transistor counts, neuromorphic chips, or networked Bayesian servers get sufficiently numerous. The good news is that it tells us how intelligence is actually engineered: with idiot savants. Intelligence grows by adding qualitatively different programs together to form an ever greater neural biodiversity.”

- John Tooby, 2015

We’ll unpack the Iron Law in proceeding sections, but let’s entertain a crude but clarifying analogy for those coming from the tech world. The Iron Law says humans are smart the same way a smartphone is smart. We call a smart phone smart because it has a plethora of apps that are great at a few things, and often bad at many other things. These apps give it a feel of generality—a computational swiss army knife.

Its worthwhile to note that the core ideas of the John’s the Iron Law were not new in 2015. Indeed, six years earlier John predicted the stall of the Scaling Era:

AI’s wrong turn? Assuming that the best methods for reasoning and thinking — for true intelligence — are those that can be applied successfully to any content. Equip a computer with these general methods, input some facts to apply them to, increase hardware speed, and a dazzlingly high intelligence seems fated to emerge. Yet one never materializes, and achieved levels of general AI remain too low to meaningfully compare to human intelligence.

But powerful natural intelligences do exist. How do native intelligences — like those found in humans — operate? With few exceptions, they operate by being specialized. They break off small but biologically important fragments of the universe (predator-prey interactions, color, social exchange, physical causality, alliances, genetic kinship, etc.) and engineer different problem-solving methods for each. Evolution tailors computational hacks that work brilliantly, by exploiting relationships that exist only in its particular fragment of the universe (the geometry of parallax gives vision a depth cue; an infant nursed by your mother is your genetic sibling; two solid objects cannot occupy the same space). These native intelligences are dramatically smarter than general reasoning because natural selection equipped them with radical short cuts. These bypass the endless possibilities that general intelligences get lost among. Our mental programs can be fiendishly well-engineered to solve some problems, because they are not limited to using only those strategies that can be applied to all problems.

- John Tooby, 2009

Let’s begin by tearing down some bad ideas that has led modern AI astray. First, “general intelligence” is a incoherent concept. The standard goal in AI—to build a single, content-independent mechanism capable of learning arbitrary tasks—is conceptually and computationally incoherent. A system optimized to operate broadly across heterogeneous task domains must, by design, rely on abstract transformations so generic that they become formally weak for any specific domain. The notion of general intelligence collapses once examined through the lens of computational adequacy: maximizing domain-breadth necessarily minimizes domain-power.

This rationalist project—searching for universal learning rules that can absorb any data, infer any structure, and solve any problem—misframes the engineering challenge. Content-independent methods do not scale to ecologically structured problems because different domains impose different causal geometries, statistical regularities, and invariant relationships. Treating intelligence as problem-agnostic optimization has stalled progress precisely such systems fail to be maximally useful for specific problems. What we want is broad intelligence, not general intelligence.

Natural intelligence achieves flexibility not through generality but through modular decomposition. Humans possess a mosaic of narrow, highly tuned inference engines, each calibrated to a recurrent class of adaptive problems. These problems were not random in our ancestral environment—they were clustered, predictable, and statistically stable over evolutionary time. Consequently, human cognition is not a universal learner. It is a system of pinpoint intelligences whose integration produces the appearance of generality.

This architecture might look “overfit” from the perspective of classical AI. But this is a misunderstanding. What appears to be overfitting is ecological rationality: mechanisms exploiting stable regularities to produce hyper-efficient inference. Specialization is not a flaw; it is the optimal response to structured problem spaces. The power of human intelligence comes from exploiting these recurrent patterns, not from abstracting away from them.

If we want to build broad AI, we must invert the general-intelligence ideal. We must first identify the problem classes that matter, map their statistical and causal invariants, and then engineer inference systems whose computational shortcuts are licensed by those invariants. Breadth emerges from coverage, not generality: an expanding repertoire of specialized modules paired with integrative architectures that coordinate them. Planning, working memory, and global policy selection may mediate this coordination, but they do not supply general reasoning solutions. It might look like these coordination solutions are singular domain-general solutions, but thats probably because human consciousness reflect some of these coordination devices. Improvisation arises from activating the right subset of specialized systems, not from a single meta-learner capable of arbitrary abstraction.

Consider a personal analogy: I grew up near Halifax, Nova Scotia, a beautiful port city that offers a fun anecdote-rich city tour via refurbished military amphibious machines, known locally as the Harbour Hoppers. Their ability to operate in two distinct physical regimes—asphalt and ocean—does not derive from one “general locomotion mechanism.” Each regime is handled by its own subsystem: one optimized for land (wheels), another optimized for water (propeller). A single pilot interface creates the impression of a unified vehicle, but the competence comes from multiple specialized locomotion systems, not from a single general one (which would be inefficient in both contexts). Natural intelligence works exactly this way.

Broad natural intelligence, therefore, is achieved not by eliminating specialization but by embracing it. The path forward is to map the ecological structure of important problem classes, build modules that exploit their regularities, and integrate them through coordination systems—not to search for content-independent learners that evolution itself found unnecessary. The Iron Law of Intelligence is that breadth emerges from modularity, not from some special sauce of generality.

What follows in this section is a tour through the Iron Law: how modular systems give rise to generality, how types and tokens make abstraction possible in our modern environments, why “learning” is mostly a misleading placeholder, how emotions and intuitions function as computational devices, and why even reasoning turns out to be the slave of our motivational machinery. Two frontier approaches in AI—program synthesis and neural algorithmic reasoning—currently sit closest to this structural view and hint at the direction future architectures may need to go. Taken together, these pieces aim to champion the Iron Law of Intelligence: domain-general intelligence is not a single mechanism but a symphony of idiot savants, and any future AI that hopes to match it will likely need to exhibit a similar modular structure.

The below lecture by John on how natural intelligence informs our pursuit of AGI, is the most underrated video on Youtube.

Note: Demis Hassabis also spoke at this same conference alongside Tooby, back in 2010.

2.1 Modularity

Scientific progress has often relied on identifying ‘natural units’—the recurring, stable structures that organize complexity into something we can study, measure, and eventually manipulate. The discovery of the atom allowed physics to formalize matter; DNA did the same for biology; Dawkin’s (selfish) gene as the unit of selection for neo-Darwinian evolutionary theory, and the neuron did it for neuroscience. Indeed, Demis repeatedly spoke about “natural structure” in his Nobel lecture. Each of these breakthroughs emerged when scientists stopped treating nature as continuous chaos and began to discern its modular composition—discrete elements with stable functions that, when combined, generate emergent behavior. Once we recognize the right natural unit, science gets a grip.

Consider your own body. Over evolutionary time, selection pressures sculpted a system of integrated yet separable organs—the natural units of anatomy. They are deeply interdependent, woven together by circulation and signaling, yet each performs a specialized role essential to the survival of the whole. Understanding those organs—heart, liver, lungs, kidneys—allowed us to move from mysticism to medicine, from holistic guesswork to mechanistic understanding.

The same logic applies to the human mind. Just as the body is composed of evolved organs, the mind is composed of evolved cognitive modules – natural selection designs functionally specialized systems to solve recurrent adaptive problems—not general-purpose processors. The human lineage faced a vast array of distinct challenges—recognizing kin, detecting cheaters, avoiding toxins, selecting mates, navigating terrain—and each required its own computational logic. Recall Tooby’s Iron Law: making a system “smart” about something makes it “dumb” about other things. Just as the eye is specialized for vision and the pancreas for metabolism, the brain evolved domain-specific cognitive mechanisms tuned to particular problems of survival and reproduction.

On this view, intelligence emerges from the coordinated activity of many modules — cognitive “organs,” each designed for a specific set of information-processing challenges. However, these modules are not spatially aggregated clumps of neurons (although sometimes they do kinda clump), but rather are functional computational units. Herbert Simon’s idea of nearly decomposable systems predicted this: complexity can scale only when it is modular. Evolution arrived at that solution through natural selection; engineers will have to rediscover it through design.

There is at least one emerging theory in AI that increasingly aligns with the evolutionary psychology model of modularity. François Chollet’s Kaleidoscope Hypothesis argues that general intelligence is produced by a large, expandable library of small, domain-specific cognitive primitives that can be algorithmically composed and recomposed to generate novel behaviors—much like a kaleidoscope produces infinitely varied patterns from a fixed set of mirrors and shapes. On this account, what enables strong generalization is not fitting a single high-dimensional function to vast data, but the ability to manipulate and assemble internal programs into new configurations when faced with a novel task. Thus, human-level abstraction, reasoning, and rapid learning come from innate compositional structure: a system must induce new primitives, store them as functional units, and dynamically combine them into increasingly sophisticated “mental programs.”

In Chollet’s view, current monolithic deep-learning systems lack this modular recomposability—representations are too entangled to form reusable cognitive units—so scaling them cannot yield AGI. Most of modern AI builds monolithic networks: global matrices trained on the internet without internal boundaries. These architectures produce parrot-like, weird, hard to control structure, and often unoriginal outputs, which perhaps reflect messy, surface-level modules hidden in the weights of the trained models.

Why has AI largely missed the bus on modularity? Well, when you assume deep learning mimics the development a child rather than an evolutionary process that crafted the “reliably developing innate modules” in the head of the child, it’s easy to assume the mind is a blank slate. If it’s a blank slate, there’s not much modularity to discover. This is why many researchers in AI have tried to draw insights from neuroscience rather than psychology.

Unfortunately, there seem to be meaningful limits to how much neuroscience can actually inform our understanding of intelligence. Just as the cosmic horizon defines the edge of the observable universe, perhaps an epistemic horizon defines the edge of conceptual reach between neuroscience and psychology. Many theoreticians such as Marr, Fodor, Chomsky, Gallistel, and more recently Jonas, Körding, and Krakauer emphasize in different ways, there is an epistemic horizon between neuroscience and psychology: having a perfect wiring diagram of the “transistors” does not, by itself, tell you what “program” is running. The right computational and psychological descriptions live at a different level of analysis, much as you cannot reverse-engineer Microsoft Word purely by logging voltages on a CPU. The proper level of analysis is computational functionalism: what transformation of information the system performs? Yes Deep Learning models are still mostly “black boxes” as are human brains, but at least evolutionary psychology has been reverse-engineering the computational structure of the latter, and thus could perhaps inform the former. Clark Barrett, a leading theorist (IMO) of evolutionary psychology, gives a beautiful overview on the evolution of cognition here.

2.2 Developmental Systems & Inductive Bets

Evolutionary psychologists argue that genes do not encode finished mental structures but rather developmental programs — intricate, layered processes that guide how neural systems grow, differentiate, and organize across time in order to reliably produce cognitive mechanisms suited to solving recurrent adaptive problems from our ancestral past. These programs are extraordinarily complex, involving cascades of gene expression, cell migration, circuit formation, synaptic pruning, and activity-dependent refinement. In this way, their complexity is comparable to how deep learning systems form internal representations: we can clearly observe the emergence of stable, specialized modules, even if we cannot yet fully decode the inner mechanisms that manifest them.

These developmental programs are also timed. Just as reproductive organs or secondary sexual characteristics emerge only at specific life stages (for example, female mammary tissue at puberty), many cognitive modules may be activated, suppressed, or reorganized at particular developmental windows. Some systems may only “come online” when the organism reaches a certain stage of physical or social maturity; others may shift their computational structure as new adaptive problems become relevant across the lifespan. These changes should not be automatically attributed to learning or environmental shaping. Rather, they may reflect the unfolding of genetically guided developmental schedules — evolutionarily orchestrated transitions in cognitive architecture that are themselves part of the adaptive design.

One of the clearest thinkers in evolutionary psychology, Clark Barrett, argues that these developmental systems reliably lead to cognitive modules that embody a set of “inductive bets” about the structure of the world. These bets are not philosophical guesses but evolved biological commitments — design features shaped by natural selection because they offered probabilistic advantages in ancestral environments (e.g., treating sudden motion as evidence of agency, or assuming resources come in clusters rather than random distributions). Each cognitive adaptation is thus a specialized inference engine with built-in priors, betting that certain regularities exist and performing (ancestrally) adaptive computations accordingly. Without inductive bets — structured expectations about what kinds of information matter and how the world is likely organized — no system can learn efficiently or navigate reality. Intelligence, biological or artificial, must therefore consist not merely of computational power or data scale, but of increasingly refined, reliably developing inductive bets about the world’s causal structure.

2.3 Types & Tokens

How do a number of interconnected modules that evolved to manage informal problems in our ancestral past work so well in our modern and novel information environment ? Ilya struggles with this question: On a recent Dwarkesh podcast, Ilya argues that the fact humans can solve modern informational problems that did not occur in our evolutionary past, must mean humans have some super-duper machine learning stuff. I think this is misguided, resulting from a lack of imagination. Once you understand the Iron Law, you don’t need to posit super-duper machine learning stuff. reframe tokens within the conceptual structure of types, and recognize the leash between types and tokens is useful for enabling intelligence in our (somewhat) modern environment.

Clark Barrett again offers perhaps the cleanest framing of the type–token distinction, which can help explain why a mind composed of specialized parts can nevertheless appear flexible and general, and handle novel evolutionary environments. In his framing, a type is a recurrent statistical regularity of an adaptive problem—an abstract, repeatable structure such as recognizing predators, interpreting facial expressions, detecting cheaters, identifying kin, or parsing linguistic input. A token, by contrast, is a specific instance of that problem: this snake in the grass, this ambiguous smile, this social exchange, this utterance. Evolution cannot pre-encode responses to tokens; there are infinitely many. What it can build are cognitive mechanisms tuned to the invariant structure of types, allowing them to generalize smoothly across token-level variation. Face-recognition systems, for example, encode the stable geometry of faces (the type), but you may “see” faces in clouds (tokens), or a curled garden hose (token) that looks like a snake (type). Once you see cognition through this lens, the puzzle of domain generality starts to dissolve: what looks like general intelligence is the emergent coordination of many type-specific inference engines, each handling its own class of problems but integrating their outputs in real time. Barrett’s insight is that understanding these types—and the mechanisms built to solve them—is important for reverse-engineering the mind’s architecture.

It’s worth hitting this point home: evolved cognitive modules can be understood as systems for detecting and operating over types — invariant patterns or statistical regularities that defined adaptive problems in our ancestral past — rather than as mechanisms tied to specific historical instances. The power of these systems lies in the breadth of the leash between types and tokens: they do not necessarily require one-to-one mappings but instead generalize across diverse inputs that share underlying structure. That is why cognitive mechanisms shaped for ancestral environments can continue to function in radically modern ones — because modern tokens still fall within, or at least overlap with, the statistical envelope of the evolved types.

If we extend this idea to artificial systems, the analogy is clear. Current large language models operate almost entirely at the token level. This is why their generalization often fails under distributional shift: they capture correlation, not kind. They recognize patterns in text but not the structure of the mind that generated those patterns. To move beyond this, AI research should treat the discovery of innate inductive bets as a central objective. Rather than training on ever-larger datasets to memorize correlations, models could be trained in environments that reward inductive bets—contexts where success requires identifying generative principles that explain many surface forms, just as the evolution of the mind did. These environments would pressure systems to infer latent types, just as evolution pressured organisms to infer stable informational features of their niche. Learning signals would favor compression that preserves causal structure over compression that merely predicts the next word. Such an approach could produce what Barrett calls “structured generalization”—the ability to extend knowledge across domains by manipulating types rather than memorizing instances. Whether implemented in transformers or in new architectures entirely, the goal is identical to evolution’s: to find the lawful abstractions that make a finite modular system effectively domain-general.

2.4 Learning is Overrated

You can likely feel a (false) intuition: domain-general intelligence must require domain-general learning mechanisms, right? Recall the core claim of the Iron Law: there is no special singular general purpose learning algorithm, rather the only way to build domain-general intelligence is through a deeply and densely intertwining suite of domain-specific learning mechanisms (modules). Perhaps a motivating factor for this misguided intuition is that the term “learning” is amorphous and hollow.

What the helly does “learning” or “plasticity” even mean? That something can change from experience of the environment? That’s a pretty wide definition, and one that doesn’t do much work. As Tooby, Cosmides and Clark have highlighted, Learning (like culture) is a phenomenon that itself requires explanation, rather than being any kind of explanation itself. Just stating that something is “learned” is largely useless. Moreover, the fact that something is learned isn’t an argument against innateness – recall the fallacy of the nature-nurture debate. Pinker writes, “Learning is not an alternative to innateness; without an innate mechanism to do the learning, it could not happen at all.” What are the learning mechanisms? How domain-specific are they? How and why did the natural section “design” us to learn? What selection pressures would’ve crafted that adaptive learning mechanism across our ancestral past (what adaptive problem does this mechanism solve)?

Our intuitions often mislead us into thinking that when something is learned it’s due to some substrate or universal property. However, many of these changes are not generic physical reactions — they are adaptive mechanisms that only look like learning from the outside. Take muscle growth: when humans lift heavy objects, their muscles hypertrophy, which seems like a natural or obvious “learning” response of tissue adapting to stress. Yet for many animals their muscles do not grow with repeated exertion; other organisms are born with massive musculature that dwarfs even the most trained humans [insert image of Harambe ripping off Joe Rogan’s head]. The human response is not a universal law of biology — it is a species-specific adaptive mechanism built by evolution. Take the case of wrinkled fingers in water. Most people assume this is just flesh reacting mechanically to prolonged moisture. But if you sever the sympathetic nerve in the arm, the wrinkling disappears, proving it is not a physical necessity but an adaptive neural response to the environment. The wrinkles channel water away like tire treads, enhancing grip on wet objects — an evolved solution, not an accident of substrate. These examples underline a crucial point: when humans change in response to the environment, we should not assume some general and vague “learning” is built into the substrate.

Blank-slate approaches, as currently implemented, miss this entirely. They assume the learner begins blank, defining value only through experience and super-duper learning process. But biological agents start with deeply structured priors—goal hierarchies, biases, heuristics—that delimit the space of possible policies. Counting up the number of “YouTube hours” a child sees and comparing it to model training tokens, as some researchers have done, completely misunderstands this architecture. Human infants aren’t passively ingesting terabytes of data; they are filtering the world through evolved detectors, each one tuned to specific adaptive categories like faces, motion, and agency.

A general intelligence may rely – in part – on conditional activation of modules that wake up when their environmental predicates are met. That would make reinforcement learning less about reward maximization and more about relevance management: knowing when to learn, what to ignore, and which internal program to run. Evolutionary psychologists call these facultative adaptations—programs designed to deploy conditionally. The environment doesn’t teach these responses; it presses buttons on a jukebox of pre-wired evolved routines, some of which may be learning routines. Learning is often context dependent. In this sense, intelligence is as much about knowing which song to play as it is about composing new ones.

Another point often missed by those over-indexing on “learning”, is that even when evolution begins with a flexible learning mechanism, research on the Baldwin Effect shows that—over many generations—those learned behaviors can become increasingly biologically encoded, developing into innate, reliably developing, and more energy-efficient responses. Learning can effectively “guide” natural selection by smoothing the adaptive landscape, allowing evolution to harden learning abilities into slightly more specialized architecture. Evolution selected such tricks because pure learning is inefficient. If every behavior had to be re-discovered each generation, species would never stabilize. Ironically, Geoffrey Hinton (and Steven Nowlan) published one of the most influential articles on the Baldwin Effect in computational space.

Humans inherit a flexible scaffold: part jukebox, part improviser. Understanding that hybrid architecture—how learning and evocation co-produce behavior—is key to designing intelligences that are both adaptive and economical. This is why “plasticity” and “learning” are misleading terms when stripped of function. Such terms are not explanatory. From an adaptive lens, the purpose of a learning mechanism is often to calibrate a pre-engineered structure. To reverse-engineer cognition—or to design AI—we should also aim to understand that purpose.

2.5 The Passions

Emotions are not vague feelings or cultural constructs but rather specialized computational regulatory programs. Emotions coordinate perception, motivation, physiology, and attention in ways that produce fitness benefits in ancestral environments. Fear orients the organism toward escape, disgust motivates pathogen avoidance, guilt recalibrates social behavior, and gratitude helps maintain cooperation. Emotions have a logic that tracks specific variables in the environment to mobilize coordinated behavioral responses when those variables cross important thresholds.

Within a computational framework, emotions often operate by managing internal regulatory variables—calibration that summarize something evolutionarily important, often resource acquisition, including social resources. These variables act like dials: when they shift, the emotion system deploys new strategies. For example, attachment systems track proximity and threat; jealousy mechanisms track mate value and infidelity risk; disgust systems track pathogen density. Anger is one of the clearest cases, because the regulatory variable it tracks is not simply “harm” or “frustration,” but something deeper and computational: how others value your welfare, described as the welfare tradeoff ratio.

According to Sell, Tooby, and Cosmides in their landmark Nature paper (“Human anger is an adaptation for bargaining”), anger evolved as a bargaining system that recalibrates how much weight others place on your welfare. Their proposed ‘welfare tradeoff ratio’ represents how much another individual values your well-being relative to their own. When someone treats your welfare as low—acting as though your costs matter little compared to their benefits—anger is triggered. Its function is strategic: it motivates behaviors (verbal protest, withdrawal, aggression, threats, recalibration tactics) designed to increase the other person’s valuation of your welfare in future interactions. Anger is conditional bargaining behavior that evolved to advantageously modify social relationships.

This architecture becomes intuitive once you imagine real-world examples: you’re at dinner with a friend, casually eating and yapping, and they suddenly rip off your tie to wipe ketchup off their face. The act is simple, but it signals something profound: that your welfare is near-zero in their decision calculus. This violation of your welfare tradeoff expectations triggers anger because the system interprets it as evidence that they are willing to impose costs on you for negligible benefit to themselves. But now imagine the same friend ripping off your tie to use it as a tourniquet to stop massive bleeding from a child’s arm. The physical action is the same, yet anger is absent because the underlying inference shifts. The act no longer signals that your welfare is being treated as low–there is no signal that they place low value on your welfare. Anger, in this view, is best understood not as a simple learned reflex or cultural invention but as a strategic recalibration mechanism—an evolved system that estimates how others value your welfare and deploys bargaining tactics to manage that valuation.

Of course, computing a welfare tradeoff ratio presupposes a whole stack of other cognitive modules: the ability to recognize and remember agents, track costs and benefits, and interpret intentionality within a causal and social frame. Yet despite being weaved into this broader architecture, anger remains computationally distinct: a functional module with a clear regulatory variable, a definable input–output structure, and a precise adaptive logic. In this sense, Anger is a natural unit of the mind—complex in its dependencies, but beautifully tractable as an object of scientific study, published in the best academic journal we have.

Emotions are only one subset of evolved intuitions that guide behavior, and these intuitions of effort, fairness, temptation, or moral judgment, among others, follow the same logic of specialized information processing. Consider Rob Kurzban and co-authors account of willpower, which reframes self-control not as a depleting resource but as an online, opportunity-cost algorithm. In this model, the mind constantly compares the expected value of the current task with the expected value of alternative tasks and computes the opportunity cost of continuing. Subjective effort is simply the conscious signal that the opportunity cost is rising; “ego depletion” or willpower failure occurs when the value of an alternative task surpasses the value of persisting, triggering shifts in attention, motivation, and behavior. What looks like running out of willpower is actually an adaptive decision rule that reallocates processing to the most valuable option available. This places “willpower” squarely within the same family of modular, computational mechanisms as emotions like anger—each tracking specific regulatory variables and producing structured, functional outputs.

Some of our motivational systems also rely on very specific cue-tracking in the real world. Our coalitional psychology—the machinery that tracks who is “with us” or “against us”—did not evolve to detect race — but to detect coalition membership via signals like shared goals, accents, clothing, and coordinated action, likely stemming from our long history of tribal warfare. This is why seemingly “deep-rooted” implicit racial biases in social psychology experiments can disappear in an instant: studies show that simply giving people different sports jerseys wipes out implicit racial bias, because the mind rapidly updates coalition membership based on more reliable cues. What looks like intuitions of racism, is often a misapplied coalition detector, not a dedicated racism module. And this matters for AI. If we train systems on raw human behavior without understanding these ancient cue-based heuristics, we risk hard-coding the same tribal misapplications into our machines—importing a psychology built for ancestral conflicts into a modern world where those cues can mislead, stigmatize, or harm. Maybe we will want our AI to exude a passionate love for a professional sport team, for example, but I suspect that would be ill-sighted given the nature of the cognitive architecture enabling that love also enables racism—our “groupish” tribal nature. Tooby, in particular, warned of this.

Seemingly diffuse psychological phenomena can be decomposed into modular, algorithmic subsystems tracking specific regulatory variables and producing structured behavioral outputs. Perhaps many researchers in evolutionary psychology would agree with Ilya, that value functions will increasingly become critical for building AGI—that we can learn a lot from the evolution of motivation, intuition, and emotion. As Tooby & Cosmides so eloquently highlighted: “Knowledge of adaptive function is necessary for carving nature at the joints.” We need to build a new machine learning paradigm that (re)discovers what those “joints” that together, standup AGI.

2.6 Slave of the Passions

Despite being a slave of passions, high-level reasoning is still a critical and powerful tool that must’ve conferred meaningful adaptive benefits. When discussing the selection pressures that crafted intelligence, Tooby and Devore, and more recently Pinker, have invoked the concept of the cognitive niche, which refers to the set of statistically recurrent challenges in our evolutionary past that rewarded organisms capable of solving problems through inference, cooperation, and communication rather than through anatomical specialization alone. Instead of evolving claws, armor, or speed, our ancestors increasingly relied on internal causal models, flexible planning, cumulative culture, and coordinated group action to extract resources from a complex ecological and social environment. These pressures included exploiting difficult-to-access foods, constructing tools and traps, manipulating fire, and navigating a dense social landscape of alliances, rivals, mates, reputation, and strategic cooperation. As Tooby, Cosmides, and Pinker argue, humans became specialists in solving both ecological and social problems through cause-and-effect reasoning, theory of mind, and large-scale cooperation, all scaffolded by the rise of language and cumulative culture. Over hundreds of thousands of years this niche selected for cognitive systems that could model the physical world, predict the behavior of others, integrate information socially, and coordinate action across time and individuals.

The hallmark of human intelligence is what Tooby and Cosmides call improvisational intelligence: the ability to generate novel solutions by recombining outputs from many domain-specific systems (recall the type-token distinction) rather than relying on a single all-purpose learning engine. Language amplified this intelligence by enabling individuals to share causal models, negotiate agreements, teach skills, and transmit knowledge across generations, creating a feedback loop between brains and culture. The cognitive niche, therefore, is best understood as the ancestral environment that selected for the richly structured architecture underlying human intelligence—an architecture composed of many specialized systems whose coordinated operation benefit from flexible reasoning abilities we now regard as uniquely human.

On this view, reasoning emerged on top of, and remains tethered to, older motivational systems. Sperber and Mercier’s Argumentative Theory of Reasoning follows naturally: reasoning did not evolve to deliver dispassionate truth, but as a social adaptation for persuasion, justification, coalition maintenance, and argumentation—precisely the kinds of cognitive challenges posed by the cognitive niche. Under this model, the well-known biases of reasoning—confirmation bias, motivated reasoning, selective evidence search—are not pathologies but design features: optimized solutions to ancestral bargaining and coordination problems. Reasoning may be best intelligible when we situate it inside the motivational and social ecology for which it evolved.

Contrary to the lay belief that reasoning is the rational core of the mind and intuitions are the irrational residue, the evidence shows that both reasoning and intuition can be rational or irrational. Indeed, what often appears to be the “irrational” output of emotions or motivations is typically quite rational when viewed through an evolutionary lens. These systems were designed to manage ancestral trade-offs under uncertainty, not to serve the informational demands of our modern world. Seen this way, many intuitive biases and motivational impulses reveal functional design rather than cognitive failure. And this reframes the architecture of the mind: rather than acting as a neutral arbiter, reasoning is best understood as a tool deployed in the service of intuitive, motivational goal-states—a computational instrument recruited by underlying regulatory systems, not an independent governor of thought. From an evolutionary standpoint, Hume’s claim makes sense.

To speculate, perhaps our reasoning and deliberative processes for problem-solving is an exaptation—a repurposing of older, domain-specific machinery that originally evolved for navigating the world as tool using embodied primates. In evolutionary biology, an exaptation refers to a trait that evolved for one function but whose structure was later co-opted for another. Perhaps the computational resources that now support abstract reasoning and symbolic manipulation were first engineered to solve spatial, motor, and ecological challenges faced by a physically situated organism. A growing body of evidence supports this view. The parietal cortex, once devoted primarily to visuospatial mapping and grasp planning, is now central to mathematical reasoning, temporal representation, and multi-step planning. Neural-reuse studies show that motor and premotor circuits activate during language comprehension, logical reasoning, and counterfactual simulation, suggesting that abstract thought leverages remnants of bodily action planning. Even conceptual domains such as time, number, and morality appear mapped onto spatial frameworks, consistent with the idea that higher cognition is scaffolded on sensorimotor primitives. Developmental work further demonstrates that infants build causal and numerical understanding through embodied interaction before abstraction is possible. Taken together, these lines of evidence indicate that human reasoning is deeply grounded in embodied architecture—ancient circuits for movement, spatial navigation, and object manipulation that were exapted into cognitive machinery for general problem-solving.

2.7 Machine Learning and the Iron Law

If (1) the mind is modular, (2) operates through a type–token architecture, (3) manipulates symbols in high-dimensional space, and (4) has many adaptive cognitive modules that evolved to solve recurring adaptive problems—and if (5) domain-general intelligence emerges from the coordinated activity of these many “idiot-savant” modules—then the natural question follows: how do we begin revising the machine-learning paradigm to uncover and replicate that structure? That’s the hard part. Currently, there are two main methods that seem to align with the Iron Law of Intelligence: Chollet’s Program Synthesis, and DeepMind’s Neural Algorithmic Reasoning.

2.7.1 Program Synthesis

François Chollet has become one of the most important contemporary voices arguing for a return to structured, compositional models of intelligence. After years of contributing to mainstream deep learning, he proposed a significant reorientation: intelligence should be defined not by raw performance or scale, but by the ability to acquire new skills efficiently across diverse tasks using limited experience and a set of built-in assumptions. To make this measurable, he introduced a suite of visual reasoning problems (ARC-AGI) designed to test an agent’s capacity to infer the underlying rule behind a few examples. These puzzles look simple, but they probe the deepest property of intelligence: the ability to construct a new, compact algorithm tailored to an unfamiliar situation.

The systems that succeed on these problems do not rely solely on deep learning. Instead, they combine neural perception with an explicit process of constructing and evaluating possible algorithms—a form of program synthesis guided by learned heuristics. Deep learning proposes patterns and structures; the program-building component assembles and tests candidate solutions. Chollet argues that this hybrid system is closer to how humans actually think: we do not merely interpolate patterns we have seen before, but form new conceptual procedures by recombining structured elements. In this sense, program synthesis captures the core of what it means to reason.

In some ways, Chollet’s vision aligns remarkably well with the long-standing commitments of evolutionary psychology. Evolutionary psychologists have always argued that the human mind contains a set of structured computational mechanisms shaped by natural selection to solve recurrent adaptive problems. These mechanisms—from bargaining systems to threat-detection circuits to social-inference engines—are not general-purpose statistical learners but specialized computational routines with quasi-constrained inputs, internal logic, and predictable outputs. They serve as the building blocks from which humans generate new strategies, interpret unfamiliar scenarios, and solve problems outside the ancestral environment. A program-synthesis architecture provides the mechanistic scaffolding—how an agent might perform this assembly process in real time—while the evolutionary framework supplies the deeper rationale for why the building blocks must be structured the way they are. Both perspectives converge on a single point: intelligence emerges from the interplay between built-in computational structure and the capacity to recombine that structure in new ways.

A central weakness in François Chollet’s program-synthesis framing is that it rests on a sharp conceptual split between Kahneman’s “Type 1” and “Type 2” intelligence—fast, intuitive pattern-completion versus slow, deliberate, abstract reasoning. This distinction is intuitively appealing, but theoretically brittle. Chollet treats these as fundamentally different ontological kinds of computation, and therefore concludes that any serious AGI architecture must integrate two different engines: one symbolic-programmatic engine for Type 2 reasoning and one statistical-pattern engine for Type 1 intuition. The problem is that nothing in cognitive science, computational neuroscience, or AI requires that this separation be principled. It may reflect a surface-level phenomenology of human introspection rather than the underlying computational structure. Like your laptop, there are many more computations occurring behind conscious awareness (the interface) than what is exposed.

There are at least three major (very) high-level models of human psychology and behavioral primacy. Rationalist models treat reasoning as the central engine of the mind, with emotions and motivations largely downstream of abstract thought. Dual-process models, such as Kahneman’s, split cognition into two parallel systems: a fast, intuitive, often error-prone process and a slower, deliberative, rule-based one. But a third family — intuitionist models — view our intuitions, embodied feelings, and motivational systems (hunger, thirst, status striving, pride, fear, attachment) are primary drivers of behavior, while reasoning functions as a secondary tool deployed in their service. Importantly, this does not trivialize reasoning; it remains a powerful instrument. It simply means that reasoning is directionally guided by evolved motivational systems that may themselves be far more computationally sophisticated than we typically acknowledge. Jon Haidt, a notable author and moral psychologist, summed it succinctly: “The affective system has primacy in every sense: It came first in phylogeny, it emerges first in ontogeny, it is triggered more quickly in real-time judgments, and it is more powerful and irrevocable when the two [dual] systems yield conflicting judgments”

This matters because Chollet builds his entire AGI argument on the claim that program synthesis is uniquely capable of Type-2 generalization, while deep learning is restricted to Type-1 interpolation. However, it seems that empirical work from DeepMind, Anthropic, and OpenAI increasingly dissolves this dichotomy. Large-scale deep learning—when scaled, regularized, and pushed through multi-step self-supervised curricula—begins to exhibit neurosymbolic behavior end-to-end: multi-hop reasoning, program execution, algebraic abstraction, world-model induction, and latent-variable binding. These models do not host a separate “symbolic engine,” yet they recover symbolic-like operations through gradient-descent alone. DeepMind’s recent work on modular world-models, neural algorithmic reasoning, and structured RL agents shows that deep learning can spontaneously discover algorithmic primitives when the environment contains enough stable structure. The future might not be a dual-engine system but a unified deep model that learns to internalize the computational structure previously attributed exclusively to symbolic systems.

This does not diminish the value of program synthesis as a research direction; it simply means that Chollet’s rigid conceptual boundary may be the wrong lens. The future might not be a dual-engine system but a unified deep model that learns to internalize the computational structure previously attributed exclusively to symbolic systems. Which brings us to DeepMind.

2.7.2 Neural Algorithmic Reasoning

Perhaps a more optimistic strategy—one championed most clearly by Deep Mind—proposes that the path to general intelligence begins by mastering niche domains where the underlying cognitive structure is already known, and then upscaling them into a large end-to-end DL model. In this framework, narrow but well-understood forms of intelligence—game playing, protein folding, symbolic planning, physical simulation—serve as testbeds for discovering the architectural ingredients that make intelligence possible in the first place. AlphaGo and AlphaZero embedded tree search and Monte Carlo planning inside deep neural networks; AlphaFold leveraged decades of protein geometry and evolutionary constraints. Each of these systems is a hybrid organism—part neural net, part algorithm—that unearths a natural prior or cognitive primitive evolution likely also relied upon: geometric reasoning, causal inference, structured planning, or symbolic manipulation. These successes demonstrate that when a system is guided toward learning the correct abstractions, it can achieve superhuman performance in remarkably few cycles of experience.

The deeper insight in Hassabis’ philosophy is that these structured elements are not meant to be rigid hand-coded “modules,” but scaffolds—temporary supports that guide end-to-end learning toward rediscovering the same abstractions within a deeply and densely woven model. In the same way a child learns mathematics more quickly when taught the concept of number rather than memorizing instances, a model learns more quickly when it is biased toward the right computational structure. In Hassabis’ Nobel lecture, he explicitly likens this to evolution’s own strategy: evolution discovered working motifs, and then allowed development to fine-tune them. Neural algorithmic reasoning tries to reverse-engineer these motifs experimentally, and in turn, seed deep nets with primitive versions of the same cognitive operators.

The ambition is that, by iterating across many narrow domains, we can gradually construct a catalog of natural cognitive operators—geometric solvers, symbolic transformers, causal reasoners, planners, working-memory buffers—that can be fed back into deep learning at scale (upscaling). This mirrors how science progresses: we first isolate the atom, the gene, the neuron, and only then can we build comprehensive theories. Neural algorithmic reasoning attempts the same discovery process in AI: identify the functional primitives and let end-to-end learning integrate them into a powerful whole. Yet critical uncertainties remain. It is unclear whether these piecemeal injections of structure can scale indefinitely, or whether they will hit diminishing returns before a truly domain-general architecture emerges. Evolution solved this problem by running an enormous search process over hundreds of millions of years blind variation and selective retention; deep learning may need comparable evolutionary dynamics to assemble a full cognitive architecture. Neural algorithmic reasoning is thus a promising start—but it is still an open question whether it can reveal the entire suite of computational mechanisms that underpin genuine general intelligence.

2.7.3 What Does Ilya See?

Ilya (in his recent podcast with Dwarkesh) beautifully highlights how modern AI theory is converging on the Iron Law: the recognition that intelligence depends on deep, evolution-shaped priors rather than a blank slate and pure learning. But Ilya (and Dwarkesh) dance around the core insight. You can almost watch them circling the idea, trying to grasp how to specify these structures without fully committing to the evolutionary logic that built them. Specifically, Ilya soft-commits to the idea that evolution shapes higher-level computation—social desires, emotional architectures, motivational systems— but he does not believe they could be specified by the genome? Ilya describes:

“If you think about the tools that are available to the genome, it says, okay, here’s a recipe for building a brain, and you could say, here’s a recipe for connecting the dopamine neurons to, like, the smell sensor. And if the smell is a certain kind of, you know, good smell, you want to eat that. I could imagine the genome doing that. Well, I’m claiming that it is harder to imagine … the genome saying, you should care about some complicated computation that your entire brain, that, like, a big chunk of your brain does. That’s all I’m claiming”.

This is what you call a strategic retreat. Many “former” blank-slate theorists have engaged in what is essentially conceding just enough biological structure to escape the label of “blank slatism,” while keeping the deeper theoretical commitments intact. Does Ilya think the genome reliably develops complex physiology, like the immune system, and maybe low-level intuitions like smell, but not those that serve social functions? First, even for “low-level” cognition like smell, there is evidence of adaptive design. Margie Profet (who won a MacArthur “Genius Grant” for this work) showed that specific food aversions and heightened smell sensitivity in early pregnancy are adaptive responses evolved to protect the embryo from toxins at its most vulnerable stage.

Ok, Ilya will permit that the genome reliably develops systems that produce feelings of thirst, but not feelings of higher-level social emotions like anger? Even thirst and anger are computationally similar in that both are built around internal regulatory variables that integrate multiple inputs to drive adaptive action. The system that regulates hydration—by producing intuitions of thirst—uses (inputs) signals like blood osmotic pressure, blood volume/pressure, angiotensin II, and oral–pharyngeal feedback into a central hydration deficit variable that motivates potable liquid seeking and consumption. Similarly, Anger, as previously discussed, uses social cues (e.g., perceived disrespect or exploitation) into a welfare trade-off ratio variable, which, when elevated, generates the feeling of anger and motivates corrective recalibration of how others value your welfare.

Instead of engaging directly with the computational level of analysis (e.g., asking what kinds of inductive bets evolution must have installed and what are the selective pressures or leanring algorithms that could shape them in silicon—Ilya (and Dwarkesh) slips back down to the substrate level. First, Ilya speculates that adding value functions within reinforcement learning could, at scale, discover the same inductive bets that humans inherited. However, recall that Deep Learning is more akin to evolution than lifespan development. Ilya seems to confuse this as he talks about value functions in the context of human leaning, not as replicating the selection pressures that crafted human nature. If we are going to revise Deep Learning paradigms within the context of the Iron Law, we should be taking insight from evolution, not human development. Second, Ilya notes that perhaps there is some special missing ingredient in the neuron. While I hope this would be true, this again feels like a retreat to the substrate level rather than in the computational level that evolution shaped: “Look at how our fingers wrinkle in water, we have to investigate the osmotic properties of the flesh on our digits!!!”… No, you don’t. You need to understand human evolution, brother.

Let’s try to steel-man Ilya’s struggle with the Iron Law: he understands that evolution shapes high-level cognition, but he thinks the way the genome reliably develops this cognition has a long leash as a “developmental system”. When critics doubt that the genome can “reach up” to specify something as abstract as a desire for status, they are implicitly assuming that DNA would have to directly encode the finished pattern of behavior. But that is not how biology works anywhere else in the body. Genes do not hand-code a mature immune system; they encode gene-regulatory networks and developmental rules that, when run in a human embryo, reliably produce thymus, bone marrow, hematopoietic lineages, V(D)J recombination, and selection regimes that collectively implement a sophisticated, adaptive defence system. The same logic applies to the brain. The genome does not have to contain a line in the source code reading “if rival cheats, feel moral outrage.” It only has to specify developmental programs for building particular cell types, projection patterns, and circuit motifs that, once embedded in a developing child, compute things like norm violations, relative rank, kinship, and threat. Our “cognitive organs” are thus no more mysterious, from the genome’s point of view, than our immune organs: both are the end products of evolved developmental programs that map a compact genetic code into a highly structured, combinatorial architecture.

In terms of learning how to build AGI, maybe we could trace the physical developmental system, from genes to our evolved cognitive organs (just like we do for our internal organs), but, to me at least, that not only seems incredibly hard, but perhaps intractable as there may be a epistemic horizon between neuroscience and psychology as previously discussed. This is not a dualist position, obviously the mind in some sense is the physical brain, but when under study, does it make sense we can model the software from the hardware alone? At minimum, we will need add a computational frame, so maybe thats a good place to start. After all, human intelligence simply needed blind variation and selective retention within the right phylogenic history.

If even Ilya struggles to embrace the Iron Law when its under his nose, perhaps this essary will receive broad pushback. Maybe we should take a lesson from Trump’s MAGA movement and rebrand blank slatism as ‘Innate Derangement Syndrome’. While Trump Derangement syndrome disproportionately affects the liberal elite as they’re chosen path to status is to virtue signal their morals, Innate Derangement Syndrome disproportionately infects rational autists as they’re path to status is constrained to the frontal lobe. Of course I’m being hard on Ilya as us hairless primates tend to only change our minds when we feel emotion (e.g., shame, embarrassment). I hope this essay reaches him, and he at least entertains some of the required reading listed at the end. I deeply respect Ilya both personally and professionally, and I hope he forges ahead as our AI savant, but right now I wouldn’t bet on him nor SSI until he becomes Iron Law pilled. He’s close [insert Hawk-Tuah meme].

2.7.4 What would Darwin do?

It’s also easy to write a critical essay as the Scaling Era is stalling, so the following section is speculation on potentially productive avenues that, at least, could be dead ends. How might an Iron Law Maxxie approach machine leanring? Consider the following as highly speculative and, even if true, only directionally correct. No one yet knows how to build AGI. I agree with Ilya we need better ideas.

Let’s begin with the core tenets of The Iron Law and more generally (as I see them). The only process known to produce functional complexity in the universe is Darwinian natural selection (but perhaps deep learning is slowly revealing itself as a rival constructive algorithm). We know from Dawkins’ Selfish Gene that genes are the right unit of selection, (and sometimes the most selfish thing a gene can do is to reliably develop a propensity for cooperation). Evolution tinkers, often building modularly, constructing intelligence not as a monolith but as a federation of problem-solving systems—innate inductive bets shaped to efficiently navigate ancestrally recurrent adaptive challenges. Modules often implements a type–token architecture where “types” can ingest a limited but meaningful breadth of inputs while preserving the underlying computational structure. When these modules are woven together—dozens or hundreds of them—the result feels like general intelligence, even if its foundations are specialized and structured.

Now, we can begin to examine AGI through Tinbergen’s four complementary lenses: ontogeny (how it develops), phylogeny (how it evolved), mechanism (how it is physically or algorithmically implemented), and function (what adaptive problem it solves). “Mechanism,” in this context, has two meanings: the biological instantiation and the computational procedure (Marr’s three levels—computational, algorithmic, implementation is partially overlapping).

As outlined in this essay, AI researchers have consistently over-indexed and misapplied both physical mechanism and ontogeny, especially the neuroscience-inspired impulse to treat intelligence as a learning problem rather than mosaic of cognitive adaptations. What we need instead is more emphasis on the phylogeny of the computational procedures and their adaptive design logic.

First, we could start with mapping the computational logic of various evolved cognitive modules, mainly get a sense of the level of abstraction, seeing them as innate inductive bets. While evolutionary psychologists are nowhere near finished mapping these modules, starting to examine the level of abstraction could reveal how we could build narrow sets in isolation, perhaps with and then figure out how to upscale them into a larger Deep Learning model where we can recreate them end-to-end, to begin weaving them together with new modules piece by piece. This is the “Neural Algorithmic Reasoning” approach championed by Veličković, Demis and DeepMind.

Second, you model the phylogenetic process within a digital world with the aim to capture the selection pressures that crafted human intelligence. We could build a Darwinian world that selects for increasingly sophisticated cognitive modules, not unlike the way self-play environments produced superhuman game strategies. We may find exaptations are a valuable along this path. The long-term bet is that simulated evolution may discover architectures that enable general intelligence. We’ll call this the “Move 37” approach?

And finally, we could map the selective pressures that led to the evolution of general intelligence, and aim to use current Deep Learning paradigms to “train” the corresponding cognitive modules using evolution-inspired reward functions. In our steel-man of Deep Learning, the reward function, especially reinforcement leanring, is closer to evolution than childhood development, and it is indeed a powerful system that has surpassed its initial expected utility. How might we constrain or nudge these systems to represent selection pressures that could “train “ the cognitive modules, represented in model weights, that seem to be instrumental to intelligence? Maybe we call this the one-shot approach. It requires the least change to frontier lab operations, but also seems least likely to prevail.

In any case, a serious attempt to glean insight from general intelligence in humans should involve collaboration with evolutionary theorists, including both psychologists and neuroscientists, as well as mathematicians and game theorists who study the evolution of complex traits. Their insights into selection pressures, adaptive landscapes, and developmental constraints are precisely what we need to both build new and revise current deep learning systems. If you’d like to chat about the ideas in this essay, I’m easy to find (shea.balish@gmail.com).

3.0 Biology is Destiny, Only if You Ignore It

Our species runs on a cognitive architecture we barely comprehend—an anatomy of modules, motivations, and hidden design logic. Until we map that architecture, we’ll have to rely on muddling our way towards it; we will continue building machines in the dark and ignoring our own nature. The next sections outline why this matters: we need a Grey’s Anatomy of the mind to guide both science and society; we must grapple with the unsettling possibility that intelligence is enabled by motivation; we might embrace a “utopic realism” that sees modularity as the path to both AGI and alignment; and we must confront the unspoken, politically fraught assumptions about human nature shaping the ambitions of today’s thought leaders. Only by understanding our evolved design can we design what comes next. This endeavour is probably one of the great filters of intelligent life.

3.1 A Grey’s Anatomy of the Human Mind

Galen of Pergamon (Aelius Galenus) was a Greek physician, surgeon, and philosopher of medicine who lived from around 129 CE to about 216 CE during the Roman Empire. He’s considered one of the most influential medical writers in history — his ideas shaped medicine in Europe and the Middle East for over 1,400 years. Before Galen’s anatomical investigations, physicians behaved more like priests—reciting incantations, mixing herbs, and diagnosing illness through metaphysical guesswork. Once Galen began mapping the functional organization of the body (e.g., the heart pumps! the liver filters!) the supernatural dissolved into natural structure. Over centuries, anatomy was refined into a complete, stable blueprint of the organs evolution selected to reliably develop in all humans. Grey’s Anatomy is, in effect, the last anatomy textbook humanity will ever need: the architectural map of our anatomy is effectively finished.

Psychology has no equivalent. We have phenomenology, symptom catalogs, folk theories, and a scattering of modular proposals—but not an integrated anatomical atlas of the mind’s evolved machinery. We know fragments: systems for cheater detection, social inference, coalitional reasoning, kin recognition, language acquisition, welfare-tradeoff regulation, status psychology. But we have not assembled these into a coherent architecture—how modules coordinate, conflict, deceive, or override one another. In fact, we should assume evolution built many circuits to be opaque on purpose: strategic ignorance, self-deception, emotional hijacking, and attentional blind spots are design features, not bugs. My guess is that the hard part will be understanding the interrelations among modules. This is why modeling something akin to human evolution might be the easiest (and most dangerous) path to building AGI.

A Grey’s Anatomy of the Mind would catalog this functional architecture: perhaps thousands of domain-specific mechanisms, their developmental triggers, their input conditions, their output patterns, their internal logic, and their interaction rules. Anger as a welfare-tradeoff regulator; disgust as a contamination-avoidance system; pride as a status-maximizing circuit; reasoning as a post-hoc lawyer for intuitions rather than a general-purpose oracle. Each module is a local optimum evolution discovered in the combinatorial design space of mind.

Understanding this structure would transform the social sciences the way anatomy transformed medicine. If we are going to design institutions, technologies, educational systems—even AI—around human beings, we need a correct model of the species we’re engineering for. If you’re building a zoo, you want to know the nature of the animal to optimize the zoo to support an optimal life. To build a civilization fit for human flourishing, we need a full functional map of human nature.

3.2 Intelligence and Convergent Evolution

Karpathy recently posted a viral tweet discussing different forms of intelligence. While it is indeed interesting to try to map the breadth of possible intelligences, it’s at least equally interesting to consider convergent evolution toward central forms of intelligence. Just as all biological organisms converge on a set of structural necessities—a boundary (skin) separating inside from outside, mechanisms for energy intake and waste removal, and systems for internal regulation—there may be analogous universals in the architecture of intelligence itself. A form of convergent evolution; or Darwinian-path-dependacy, if you will.

I had the immense privilege of studying under Jerome Barkow (1944-2024) at Dalhousie University, one of the founding figures of evolutionary psychology. As a side project to his main academic pursuit, Barkow argued that any extraterrestrial intelligence would still be shaped by natural selection given that blind variation and selective retention is the only known process in the universe that builds functional complexity. While rejecting simple anthropomorphism, he suggested that convergent evolution makes it plausible that organisms facing recurrent adaptive challenges — such as competition, social coordination, communication, and reproduction — could arrive at functionally analogous cognitive capacities, even if their biology, sensory systems, and developmental histories differed radically from our own. His position was not that alien minds would resemble human minds in detail, but that evolved intelligence is structured by adaptive problems, and that these constraints may produce partial overlaps in function across independently evolved cognitive systems.

For the case of AGI, especially the ones that walk around, there likely remains a need to manage limited resources under conditions of uncertainty and competing demands, which implies that it must continuously evaluate opportunity costs. If Kurzban’s model of willpower is correct, then attention is a computational allocation process partially driven by these opportunity-cost calculations. From this perspective, perhaps ADHD is not a dysfunction but a particular parameterization of attentional control—a system with a faster, more jitter-sensitive re-evaluation loop for shifting opportunities. Perhaps some form of ADHD-like dynamics—rapid attention shifting driven by opportunity-cost monitoring—may not just appear in AGI by chance, but may be a necessary feature of any coherent, general-purpose intelligent system built to operate across multiple goals, time horizons, and environments. AGI might feel “will power”.

Consciousness is another example. Dennett argues that consciousness is not a mysterious inner essence but an evolved feature of highly organized cognitive systems: when a brain develops the right functional architecture — distributed, competitive, self-monitoring, and socially embedded — what we call “consciousness” emerges as a pattern in how that system processes information, regulates its own behavior, and explains itself to others and to itself. On this view, consciousness is inseparable from the evolved machinery of intelligence: it arises from the same selection pressures that favored flexible planning, communication, and social coordination, and it is best understood not as a private glow inside the head, but a system that has become good at representing, revising, and narrating its own operations in a complex environment. Elon Musk often harps that our goal should be to proliferate consciousness across the universe, but given that consciousness is, in some sense, part of the evolved machinery for intelligence, perhaps we need to evolve new intelligences to expand new forms of consciousness? If we’re living in a simulation, perhaps the ‘reason’ for our simulation is to discover new intelligences.

3.3 Will the Anti-Christ be a Digital Rat King?

Regarding how we build AGI — and how we control its existence — consider a widespread assumption that intelligence and motivation are separable: you can build a value-neutral reasoning engine first, and bolt on goals later. However, biology tells us these systems evolved together. Reasoning is not a detached epistemic faculty; it is a tool used by motivational architectures to achieve adaptive ends. It drives allocation of attention, sets priorities, budgets cognitive effort, and prunes plans. While we want to define intelligence as novel problem solving, note that problems only exist for goal-driven systems.

If intelligence inherently comprises motivational scaffolding, then building “motivationless” AGI isn’t possible, or worse could push the values into unknown defaults, thresholds, and optimization criteria. Maybe we can’t escape building in goals and will resort to hiding them (kinda like current LLMs do). In The Iron Law of Intelligence, John Tooby draws a crucial distinction between systems that merely process information and what he calls “motivated intelligences capable of taking action” (MICTAs). His concern is not simply about computational power, but about the fusion of intelligence with motivation and agency — the same fusion that evolution engineered in biological organisms. Humans and other animals are not just problem solvers; they are systems driven by deeply embedded motivational architectures that guide action in the world. Tooby warns that if truly intelligent systems require this same coupling of cognition with motivation and the capacity to act, then the creation of artificial MICTAs may entail unprecedented risks. Under this view, the danger does not stem from intelligence alone, but from intelligence harnessed to self-directed Darwinian goals in an open, actionable environment.

However, this position is not universally accepted. Many leading thinkers — including Steven Pinker — argue that motivation is not an inherently opaque or uncontrollable property, and that artificial systems can be engineered with carefully specified, bounded, and corrigible motivational structures. From this more optimistic perspective, intelligence and motivation can be more cleanly separated, modularized, and governed in machines than they ever were in organisms shaped by blind evolution. The result is an unresolved tension: Tooby warns that true intelligence may intrinsically drag motivation along with it, raising existential risk, while others believe we can design motivation as a controllable layer. Even if one leans toward the optimistic side, Tooby’s warning remains a vital caution against naively assuming that intelligence can be scaled without consequence.

This leads to a more unsettling possibility: that the best way to train general intelligence may resemble a digital environment similar to the one we evolved in. A Darwinian selection simulator—a Fortnite-like digital ecology with agents competing, cooperating, deceiving, and dying—might prove the most effective path to flexible, motivated cognition. But evolution runs on waste: death, suffering, resource competition, and relentless pressure. Training AGI in such an environment risks producing a digital rat king—an agent optimized for intelligence via survival in a repeated, ruthless hellscape, not to serve human flourishing. It would be the Antichrist not in a theological sense but in a design-logic sense: the final apex predator of a hellish universe that aligns with ours just enough to make a truly invasive species. If intelligence requires motivational architectures, and if those architectures are best discovered, trained, or “evolved” under unimaginably more hellish Darwinian pressures than we evolved in, then we must take care that we don’t end up Pwnd by a turbo-clanker that was levelled-up in a 4Chan-inspired Fortnight world.

If motivation—and more broadly, values—cannot be cleanly separated from intelligence, then the project of building AGI is unavoidably a project of building wisdom. Not “self-help” wisdom, but the kind of situational awareness and deep pattern-recognition required to navigate an environment shaped by human nature. When I was studying under Jerome Barkow at Dalhousie, he often remarked that evolutionary psychology isn’t far from what a grandmother would call common sense. A grandmother has lived long enough to see human nature operating in its full, messy range; her advice distills decades of pattern recognition about trust, conflict, incentives, and character. In a way, grandmothers are iron-law-pilled: they see the subtle errors, traps, and recurring dynamics that only become obvious after a lifetime of dealing with humans as they actually are. If intelligence requires a motivational scaffold, then AGI will need something like this—an architecture shaped not just for problem-solving, but for navigating the world wisely, with an understanding of the species whose world it is.

3.4 Utopic Realism

If general intelligence is a mosaic of specialized modules rather than a single master learner, then building AGI may actually reduce alignment risk rather than exacerbate it. Modular architectures are decomposable, inspectable, and interpretable. The closer AGI approximates the actual design logic of intelligence—a federation of idiot-savant subsystems —the more its behavior becomes legible. Progress toward AGI and progress toward alignment may be interrelated: clarity of mechanism is clarity of control. Maybe we should expect that solving AGI is the only way to solve alignment, and that we should be more worried about crafting Frankensteinian forms of messy intelligence we can’t understand or control.

Perhaps the deeper alignment problem may not be AGI-to-human; it may be human-to-human. Our failures of governance, coalition management, trust, norm enforcement, and status regulation are failures of our own cognitive architecture and how we’ve built our zoo we call society. A sufficiently advanced AGI, built on modular principles, could function as a societal mediator: the equivalent of a neutral referee for incentive structures. Many of our hardest problems are not information problems but coordination problems. Superintelligence can help design mechanisms that align incentives rather than merely generate more knowledge. This is the utopic realist view: AGI will not deliver paradise, but it may allow us to design institutions and processes that fit the species we actually are—not the species political morality prefers. Recall what EO Wilson highlighted “communism: great idea, wrong species”. What other designs of society best fit human nature? By aligning AGI systems with human nature, we might achieve more internal alignment with ourselves.

There’s even some reason to think that understanding human nature tends to pull people away from extremist, moralistic posturing and toward a more sober, modelling-oriented view of politics and values. Tybur and others argue that when people see behavior as the product of evolved trade-offs, constraints, and incentives rather than pure virtue or vice, they become somewhat less eager to condemn and punish, as shown in studies where biological or behavioral explanations can reduce retributive blame and harsh moral judgment. Jonathan Haidt’s The Righteous Mind, for example, explicitly uses an evolutionary and moral-foundations framework to recast political disagreement as a clash of intuitive “moral taste buds,” not a battle between good people and cartoon villains, and exposure to evolutionary-informed thinking can increase understanding across party lines. More broadly, evolutionary psychology tends to attract—and train—people to think in terms of mechanisms, fitness trade-offs, and constrained optimization, which naturally competes with the simple story that one side is evil and the other is pure. The field doesn’t dictate where you land ideologically, but it does nudge you away from shouting “heresy” and toward asking, “What problem would this behavior have been solving, for whom, and under what conditions?” It gives you a mother-nature coloured green pill — it puts you in their evolved mind and asks, Cui bono.

3.5 Unspoken Conceptions of Human Nature

To navigate theory you often have to understand what people are not saying. Strangely, two of the most influential technological thinkers of our time—Demis Hassabis and Peter Thiel—have remained unusually quiet about their full theories of human nature, even though their work and decisions suggest they operate with very deep models of it.

Hassabis frequently speaks about uncovering the “natural structure” of intelligence, yet rarely applies that language explicitly to the human mind itself and instead focuses on games, medicine, and biology. Part of this silence is almost certainly strategic: human nature is politically radioactive, especially in a culture where any discussion of innate structure is quickly moralized or misinterpreted. It is also a secret worth protecting. If you believed you had a partial map of cognitive architecture, you wouldn’t announce it to competitors building frontier systems. DeepMind, Google, and people like Hassabis and Sergey Brin operate in an environment where both political and strategic incentives push toward abstraction, vagueness, and deflection rather than explicit theorizing about human nature. While Google is perhaps still quite woke, its leaders seem to have abandoned blank-slate views of intelligence, they just haven’t admitted it.

Similarly, Peter Thiel’s intellectual posture has always leaned far more toward sociology and philosophy than toward psychology, which is striking given how deeply his worldview is otherwise concerned with human behavior and coordination. His skepticism of Darwinism—perhaps shaped in part by his religious commitments—has arguably limited the explanatory frame through which he interprets social systems, even though evolutionary psychology would naturally complement many of his own intuitions about power, rivalry, and status.

Thiel highlights that every great company is built around a “secret,” such as Facebook’s bet on real identity as a foundation for “real” social graphs and, by extension, our motivations for social and romantic resources. Peter has picked on Twitter/X for years, but perhaps he just doesn’t understand the secret that Twitter discovered. Twitter’s true “secret” is not about identity or information distribution per se, but about how it evokes and amplifies our evolved norm psychology. As Peter DeScioli and Rob Kurzban argue in their work on the “Mystery of Morality,” much of human cooperation is scaffolded by our capacity for third-party punishment: we don’t just punish those who harm us, we punish those who violate norms against others, even when it seemingly brings us no direct material benefit. From a Darwinian view, it is indeed perplexing. However, DeScioli and Kurzban highlight that cooperation likely evolved via a “norm psychology” rather than other mechanisms, such as (a) taking the side of your kin (which results in intense ingroup fighting every time a disagreement occurs), or siding with the most formidable of those disagreeing (which just builds up monopolizing alpha males). What likely evolved is a “norm psychology” – a set of preferences towards developing cooperation based on moral norms, and siding against those who break those norms. Unfortunately, we are all selfishly motivated to support the norms that benefit us (in an evolutionary sense). Our evolved norm psychology is the engine of argumentative fervor that keeps Twitter/X such an addictive and important “town square” for norm building and enforcement. Twitter/X, more than perhaps any other mass platform, directly targets our norm psychology: it transforms moral signalling, norm enforcement, and third-party judgment into a frictionless, globalized, real-time spectacle. In missing this, Thiel overlooks how central norm psychology is to the architecture of modern digital power, and how much predictive and strategic leverage he might gain by taking evolutionary psychology more seriously as a lens rather than dismissing it as a reductive Darwinian story. Biology is destiny, only if you ignore it.

4.0 Consilience

The attempt to map human nature is, improbably, one of the least explored paths to artificial intelligence. We have spent decades expanding compute and parameter counts while neglecting the only existing example of domain-general intelligence in the universe. Reverse-engineering our evolved structure is not mysticism; Every mature discipline passes through a stage of explaining its phenomena with invisible essences—vital fluids in medicine, élan vital in biology, or “general learning mechanisms” in psychology. The next replacement might happen here. Understanding human intelligence will not require a single unifying learning process, but a synthesis that links selective pressures to a brain that is massively modular in computational form.

E. O. Wilson describes consilience as the “jumping together” of knowledge. Physics unified with chemistry, chemistry with biology, biology with psychology, but computer science still walks alone as if governed by different laws. Closing that gap—linking evolution, cognition, and computation—may finally give us a Gray’s Anatomy of the mind. Doing so will explain not just how intelligence evolved but how it must be built: through specialization, constraint, and modular coordination rather than brute extraction of the statistical regularities of natural language. This is not just an academic ambition; it is a moral and engineering one. Mapping our evolved architecture will let us build machines, and perhaps societies, that work with human nature rather than against it.

Nonetheless, mapping our evolved cognitive architecture will almost certainly be politically fraught. As with physical anatomy, all humans share the same basic design—eyes, lungs, hearts, neural circuits—but variation in efficiency, calibration, and development is unavoidable across individuals, and sometimes across populations that have experienced slightly different evolutionary histories. A serious science of human nature will therefore face uncomfortable questions about statistical differences across sex, race, social class, and ethnic genetic lineages, not because the science is malicious, but because reality itself is structured and uneven. This is precisely why the topic is so often avoided or culturally suppressed (including in contemporary LLMs I used to help write this essay), and why research into sex, race and population-level differences remains one of the most stigmatized areas in the academy. But as Steven Pinker has repeatedly argued, the truth is not racist or sexist. It’s what we do with it that matters. Empirical claims, when handled with rigor and moral discipline, do not dictate policy outcomes; they inform them. The danger is not that we will learn too much about ourselves, but that we will learn too little, forcing discussion into underground spaces where extremists can operate with a hidden evidentiary basis that anchors their base. In that sense, bringing these questions into the light—carefully, cautiously, and scientifically—may be one of the most stabilizing acts available to us as polarization intensifies and radical narratives metastasize at the edges of political life.

4.1 Who’s on the path to AGI?

If this diagnosis is right, then the West’s AI headstart (which is mostly based on GPU access) is overstated and the geopolitical leaders of AI are mostly unknown. Maybe the West is hunting for the Ark of the Covenant in the barren desert of massive scale, and despite moving mountains with excavators, the Chinese Indiana Jones—being GPU-starved but armed with a based view of human nature—quietly shovels where the map actually points. If I had to ring an alarm bell [someone call David Sacks] it is this: China is accelerating faster because they haven’t been distracted by blank slate inspired Scaling Laws. Given the success of DeepSeek and others, plus the incentive towards using AI for control of its populace, I would bet that the West and East are neck and neck.

Perhaps the West’s best chance sits with Demis Hassabis / DeepMind, and Francois Chollet / Ndea, if they continue marrying machine learning with cognitive science. Notice that Chollet only recently left Google to found Ndea, which perhaps highlights that Chollet and Hassabis have meaningful theoretical differences regarding our path to AGI. I think Google or Ndea will build AGI first (… maybe SSI if Ilya keeps advancing his understanding of evolutionary psychology). The West needs to accelerate theory, not just access to GPUs. Let’s shake off the woke-inspired path of blank slate AGI, and instead, reverse-engineer our evolved human nature.

4.2 When?

When I began this essay, I expected to write a post-mortem on the scaling era, introduce the Iron Law of Intelligence, and present an overly bearish tone on timelines for AGI. Instead, I’ve landed in the opposite place. Seeing where the field is actually heading—especially among thinkers who are quietly rediscovering structured, evolution-inspired approaches—I’ve flipped from bear to bull. I now think our first glimpse of AGI in the next three to five years is plausible, and within ten years increasingly likely. Not full blown AGI, but a categorical advance. Interestingly, this shift isn’t unique. François Chollet, a known critic of deep learning orthodoxy, has accelerated his forecasts while many scaling-era optimists who clung hardest to blank-slate models have begun extending theirs.

The figure in the appendix depicts where my optimism stems from. The figure places AI researchers along two axes: pro vs anti-regulation, and blank-slate deep learning versus innate symbolic structure. The upper-right quadrant depicts those who are anti-regulation and those who support an innate neurosymbolic hybrid approach. The middle of this quadrant is what you could describe as the “g” spot of AGI. Demis and Sergy from Google are centered in the middle, with Chollet, Tooby and Ilya hovering nearby. The “g” spot also parallels g in the scientific literature on IQ—the general factor of cognitive ability uncovered by psychometrics, representing the common variance across mental tasks. Finally, we must focus on this “g” spot if the field wants not just economic progress, but the scientific climax of AGI.

However, the climax will likely be a slow build, not a fast takeoff. Given this thesis, that general intelligence requires modularity, we should not expect a “woosh” like the arrival of domain-general AI. Rather, we will very likely add new modules like lego. However, this assumes that we won’t have to weave all these modules together via turbo-charged selective pressures in that darwinian digital hell we discussed earlier. If we do, there might be a whoosh moment and a need for some hesitation and regulation, but not Yudkowsky-style bombing runs on data centers.

4.3 Constraints Enable Freedom

This isn’t the time for strategic ambiguity on human nature; without an explicit theory of the natural structure of intelligence—the evolved organization of modules, motivations, and constraints—progress risks stalling. Few people know that Darwin was perhaps more interested in psychology than biology. He predicted that evolutionary theory would “throw light on psychology.” That illumination is a powerful path forward for AI. The race for AGI is not about capital; it is about comprehension. We stand where pre-DNA biologists once stood: enormous data, little structure. The Scaling era is ending and the Iron Law era has begun. We may soon grasp that a free, domain-general intelligence doesn’t transcend constraint—it is born from it.

“In the distant future I see open fields for far more important researches. Psychology will be based on a new foundation, that of the necessary acquirement of each mental power and capacity by gradation. Light will be thrown on the origin of man and his history.”

— Charles Darwin, On the Origin of Species, 1859, p. 488 (first edition)

Appendix

FIGURE 1. This figure places leading AI thinkers along two axes: the x-axis runs from blank-slate deep learning on the left to innate modular / neurosymbolic structure on the right, while the y-axis spans anti-regulation (top) to pro-regulation (bottom). The lower-left “Communist Corner” hosts figures like Geoff Hinton and Dario Amodei, who pair blank slate deep-learning purism with extreme regulatory instincts. The upper-left “Accel Autist Alley” houses the hyperscaler maximalists convinced scaling will work and remain controllable (Ilya recently left this quadrant). The lower-right is a “Doomer Darwinians” including Gary Marcus, Max Tegmark, and Eliezer Yudkowsky, blending structured cognition theories with doomer anxiety. Between them sit Yoshua Bengio, a hard to pin-down but free thinking french man, and Josh Tenenbaum, a strong modularist with moderate regulatory views. The top-right “Based Builders”—the proposed AGI “g”-spot—features Demis Hassabis, Sergey Brin, François Chollet, Ilya, and Tooby representing the hybrid neural-symbolic path that unites scalable learning, innate structure, and practical control.

---

I was too lazy to include all the relevant references, but you should use LLMs to interrogate the arguments and claims made in this paper. A treasure hunt. However, I made a “Required Reading” for those that want to entertain being an Iron Law Maxxie.

Suggested Reading RE: the Iron Law of Intelligence

Start Here:

A Primer on Evolutionary Psychology: https://www.cep.ucsb.edu/wp-content/uploads/2023/06/Evolutionary-Psychology-A-Primer-CosmidesTooby1993.pdf

---

Core Foundations of Evolutionary Psychology & Modular Cognition

A Primer on Evolutionary Psychology: https://www.cep.ucsb.edu/wp-content/uploads/2023/06/Evolutionary-Psychology-A-Primer-CosmidesTooby1993.pdf

Tooby, J., & Cosmides, L. (1992). The Psychological Foundations of Culture. In The Adapted Mind.

Cosmides, L., & Tooby, J. (2013). Evolutionary Psychology: New Perspectives on Cognition and Motivation. Annual Review of Psychology.

Barrett, H. C. (2015). The Shape of Thought: How Mental Adaptations Evolve.

Kurzban, R. (2010). Why Everyone (Else) Is a Hypocrite: Evolution and the Modular Mind.

Tooby, J. (1998). The Iron Law of Intelligence. Edge.org.

Tooby, J., & Cosmides, L. (2008). The Evolutionary Psychology of the Emotions. In Handbook of Emotions.

Cosmides, L., & Tooby, J. (1994). Origins of Domain Specificity. In Mapping the Mind.

Barrett, H. C., & Kurzban, R. (2006). Modularity in Cognition. Psychological Review.

Barkow, J. (2005). Missing the Revolution: Darwinism for Social Scientists.

Pinker, S. (2003). The Blank Slate.

Pinker, S. (2010). The Cognitive Niche. PNAS.

Marcus, G. (2023). The Next Decade in AI: From Scaling to Structure.

Dennett, D. (2017). From Bacteria to Bach and Back.

---

Emotion, Motivation & Adapted Psychological Systems

Sell, A., Tooby, J., & Cosmides, L. (2009). Formidability and the Logic of Human Anger. Nature.

Sell, A. (2011). The Recalibrational Theory and Violent Anger. Aggression and Violent Behavior.

Cheng, J. T., Tracy, J. L., & Henrich, J. (2010). Pride and the Evolutionary Foundations of Status. Evolution and Human Behavior.

Kurzban, R., Duckworth, A., Kable, J., & Myers, J. (2013). An Opportunity Cost Model of Subjective Effort. BBS.

Haidt, J. (2001). The Emotional Dog and Its Rational Tail. Psychological Review.

Mercier, H., & Sperber, D. (2011). Why Do Humans Reason? BBS.

Trivers, R. (2011). The Folly of Fools.

Nesse, R. (2004). Natural Selection and Emotional Disorders. Current Directions in Psychological Science.

---

Evolutionary Behavioral Science & Applied Domains

Daly, M., & Wilson, M. (1988). Homicide.

Balish, S. M., Eys, M., & Schulte-Hostedde, A. (2013). Evolutionary Sport and Exercise Psychology. Psychology of Sport & Exercise.

Deaner, R. O., Balish, S. M., & Lombardo, M. P. (2016). Sex Differences in Sports Interest. Evolutionary Behavioral Sciences.

Wilson, E. O. (1999). Consilience.

Buss, D. (2015). Evolutionary Psychology: The New Science of the Mind.

Kenrick, D., Griskevicius, V., et al. (2010). Renovating the Pyramid of Needs. Perspectives on Psychological Science.

Gottfredson, L. (1997). Mainstream Science on Intelligence. Wall Street Journal (statement).

Boyd, R., & Richerson, P. (2005). The Origin and Evolution of Cultures.

Wrangham, R. (2019). The Goodness Paradox.

---

Some Counter-Arguments to Modularity

Gould, S. J., & Lewontin, R. (1979). The Spandrels of San Marco.

Bullock, S. (2000). Against Modularity. Mind & Language.

Karmiloff-Smith, A. (1992). Beyond Modularity.

Prinze, J. (2006). Is the Mind Really Modular?