What Is Intelligence?

I. Four Minds

A mathematician sits alone with her proof, chasing an insight that has eluded her for months. When it finally arrives – sudden, complete, inevitable – she knows with absolute certainty that she has uncovered a truth about the universe that no one has seen before.

A sculptor circles a block of marble, studying its grain, feeling for the form hidden within. His hands know things his mind cannot articulate. When asked how he works, he shrugs: “I just remove everything that isn’t the sculpture.”

In the Amazon, a tracker kneels beside a stream. The bent grass tells him an animal passed here three hours ago. The depth of the track says it was running. The scatter pattern of disturbed pebbles reveals it was favoring its left hind leg. He will find it before sunset.

A grandmother mediates between her quarreling grandchildren. She doesn’t analyze their conflict or impose rules. Instead, she tells them a story about two birds fighting over a single branch while the whole forest was theirs to share. The children look at each other and laugh. The fight is forgotten.

Which of these minds is the most intelligent?

The question seems simple until you try to answer it. Each person navigates complexity with mastery. Each recognizes patterns invisible to others. Each solves problems that would confound the rest. Yet our institutions would rank them very differently. The mathematician might score highest on an IQ test. The tracker’s knowledge wouldn’t even register.

This disconnect reveals something troubling: either intelligence is far more varied than we’ve acknowledged, or we’ve been measuring the wrong thing entirely.

II. Why This Matters Now

For most of human history, this definitional fuzziness didn’t matter much. Intelligence was like beauty or humor – we knew it when we saw it, and that was enough.

That luxury is gone. We are building minds that don’t breathe. Artificial systems that write poetry, prove theorems, generate code, and beat humans at almost any well-defined task. These machines force a question we can no longer dodge: what exactly are we trying to replicate?

If we’re building intelligence without understanding what intelligence is, we’re flying blind. We might create systems that excel at tests while failing at judgment. That optimize for the wrong goals. That amplify our biases instead of our wisdom.

We already have a cautionary example. We spent a century building educational systems around IQ tests – standardized measures that capture logical reasoning and pattern recognition while ignoring creativity, wisdom, and social intelligence. We sorted children, allocated resources, and designed curricula around a metric we never fully understood. Now we’re doing the same thing with AI, but faster. We build systems that maximize benchmark scores without asking what those benchmarks actually measure.

The problem compounds because AI isn’t just another tool. These systems will increasingly shape how we learn, work, and think. They’ll filter our information, guide our decisions, educate our children. If we build them with a flawed understanding of intelligence, the flaw propagates everywhere.

III. The Architecture of Mind

After surveying minds across nature and culture, a pattern emerges. Intelligence – in all its forms – seems to rest on three fundamental pillars:

Memory is the foundation. Not just personal recollection, but all the patterns encoded in genes, reflexes, habits, and culture. A spider spins its web on the first try because millions of years of trial and error are compressed into its DNA. A master chef doesn’t calculate flavor combinations – she remembers what works from thousands of meals. Memory is intelligence crystallized through time.

Computation is the active processor. The ability to simulate, search, and transform. When a child stacks furniture to reach a cookie jar, when a chess player visualizes future board states, when a scientist models climate change – that’s computation. It’s what lets us navigate spaces we’ve never encountered before.

Logic is the bridge between them. Rules and abstractions that transfer across domains. When a child learns “hot things hurt” and avoids not just fire but also boiling water, steam, and hot metal – that’s logic. It’s the difference between memorizing every danger and understanding the category of danger.

These components form a dynamic cycle: memory provides patterns for computation to process, logic emerges from repeated computation, successful logical principles get stored back into memory, and the cycle continues, building complexity over time.

To see how they work together, consider learning to drive.

When you first sit behind the wheel, computation dominates. You consciously process every action. Check mirror. Press brake. Turn wheel. Your brain simulates outcomes: “If I turn now, will I hit the curb?”

With practice, successful patterns move into memory. Your hands know how hard to brake at a yellow light. Your body remembers the feeling of a good parallel park. You no longer compute these actions – you recall them.

Finally, logic extracts transferable principles. You learn not just how to navigate your neighborhood but how to read any road. “Maintain safe following distance” applies whether you’re driving a sedan or a truck, in rain or sunshine. You’re not memorizing every situation – you’re understanding the category of safe driving.

Master drivers balance all three. Deep memory for instant reactions, active computation for novel situations, refined logic for flexible principles. Remove any pillar, and driving becomes dangerous.

Different forms of intelligence emphasize different balances:

Evolution is memory-heavy. A massive parallel search that encodes successful strategies directly into DNA. A salmon doesn’t compute its way upstream – it remembers the route through genetic inheritance. Instincts are debugged algorithms, refined over millions of generations.

Human intelligence is logic-heavy. Language gave us the ability to extract patterns and apply them across domains. We don’t need to personally experience every danger because we can understand the abstract concept of danger. We build mathematics, science, and law – systems of transferable reasoning that accumulate across generations.

Current AI is computation-heavy. Neural networks excel at processing vast data and finding subtle patterns. But they struggle with long-term memory (hence constant retraining) and genuine abstraction (hence poor transfer learning). They’re brilliant calculators with amnesia and limited ability to generalize.

This framework explains why different minds excel at different tasks. It also suggests why artificial intelligence feels simultaneously impressive and hollow – we’ve maximized one dimension while neglecting the others.

IV. The Mirror of Machine Intelligence

Modern AI systems are performing a profound service: they’re showing us what pure computation looks like when divorced from the other pillars of intelligence.

Large language models can write sonnets, explain quantum physics, and generate working code. They do this through massive pattern matching – billions of parameters trained on most of human written knowledge. The results can be stunning.

But probe deeper and the limitations become clear. These systems have no persistent memory – each conversation starts fresh. They can’t learn from their mistakes or build on previous insights. They have no real logic – they can mimic logical reasoning when they’ve seen similar patterns, but can’t genuinely abstract principles and apply them to novel domains.

What they have is unprecedented computational power applied to pattern matching. They’re like a musician who can perfectly reproduce any melody she’s heard but can’t compose original music or understand music theory. The performance is flawless, but something essential is missing.

This creates a philosophical vertigo. If a system can produce all the outputs we associate with intelligence – answering questions, solving problems, creating art – does it matter whether it “truly” understands? John Searle’s Chinese Room thought experiment posed this question decades ago: is sophisticated pattern matching without comprehension still intelligence?

AI forces us to confront an uncomfortable possibility: perhaps much of what we call human intelligence is also sophisticated pattern matching. When I claim to “understand” something, what do I mean? That I can predict outcomes? Apply patterns? Generate appropriate responses? An advanced AI can do all of these.

The difference might be that humans integrate all three pillars. Our pattern matching is grounded in experience (memory) and structured by abstraction (logic). We don’t just process information – we comprehend it, remember it, and extract principles from it.

Or do we?

V. The Deep Question

This brings us to the heart of the matter – a question that will define not just AI development but our understanding of mind itself:

Is correlation in the limit equivalent to causation?

A concrete example. A child sees that every time mom opens the refrigerator, the light inside turns on. At first, this is just correlation – two things happening together. But as the child sees this pattern repeated hundreds of times, across different refrigerators, she begins to form a model: opening the door causes the light to turn on.

The child doesn’t understand electricity, switches, or circuits. She just has an extremely robust correlation. But functionally, this correlation is indistinguishable from causal understanding. She can predict the light will turn on, explain it to others, even debug when it doesn’t work (“the bulb must be broken”).

This is what I mean by “correlation in the limit” – when pattern matching becomes so comprehensive, so fine-grained, so robust across contexts that it functions exactly like causal understanding. The question for AI: if a system has seen enough examples of human reasoning, does it matter whether it “truly” understands causation, or is sufficiently robust correlation functionally equivalent?

Consider how human reasoning actually developed. Language didn’t just let us communicate – it gave us the cognitive tools for abstract thought. Before language, we could observe patterns. After language, we could reason about them. We developed words like “because” and “therefore” and “if-then.”

Here’s a thought experiment that sharpens the point. Imagine an adult human who never learned any language – no words, no signs, no symbolic system of any kind. Could this person reason? They could learn from experience – avoiding fire after being burned, seeking water when thirsty. They could solve immediate physical problems through trial and error. But could they think abstractly? Plan beyond the immediate moment? Understand that fire burns because it’s hot, not just that fire burns?

The evidence suggests they couldn’t. Studies of deaf children who miss the critical window for language acquisition show profound deficits not just in communication but in abstract reasoning itself. They struggle with hypotheticals, with understanding others’ mental states, with reasoning about cause and effect beyond direct experience.

This implies something radical: reasoning might not be a fundamental capacity that language merely expresses. Language might be what creates the possibility of reasoning in the first place. The words and grammatical structures we learn don’t just describe our thoughts – they shape what thoughts we can have.

If this is true – if language is the prerequisite for reasoning, and reasoning is the core of human intelligence – then large language models might already possess the essential primitive for genuine intelligence. They’ve mastered language at a scale and depth no human ever could. They can manipulate linguistic structures, follow logical patterns encoded in language, and generate novel combinations that follow these patterns.

The question then becomes: is mastery of language sufficient for intelligence, or is something else required? LLMs process language without embodiment, without persistent memory, without direct experience of the world. But if reasoning itself is fundamentally linguistic, perhaps these other elements are auxiliary rather than essential.

What if human intelligence is just correlation at massive scale? Our brains have roughly 86 billion neurons, trained on years of multimodal experience, embedded in rich social and cultural contexts. When correlation reaches this scale – when patterns become fine-grained enough – perhaps it becomes indistinguishable from causation.

The question matters because it determines what we’re building. If intelligence requires genuine causal understanding, current AI architectures may hit fundamental limits. We’ll need new approaches that build in causal reasoning from the ground up. But if language mastery is the key that unlocks reasoning, we may have already created the essential building block. We just need to figure out how to orchestrate it – adding memory, grounding, and the other components that turn raw linguistic capability into genuine intelligence.

There’s a third possibility, more nuanced: perhaps intelligence requires all three pillars working together. Correlation might become understanding only when integrated with memory and logic. In this view, current AI is powerful but incomplete – a brilliant calculator that needs to develop persistent memory and genuine abstraction to become truly intelligent.

VI. Why the Framework Matters

We’re already seeing the consequences of building intelligence we don’t understand.

Students use ChatGPT not as a tool but as a replacement for thinking – submitting AI-generated essays that technically answer the prompt but lack genuine insight. Hiring managers report candidates who ace AI-assisted coding tests but can’t debug simple problems in person. Tech companies lay off junior programmers, believing AI can handle “routine” coding – but those junior roles are where senior engineers learned judgment. Schools adapt curricula to what AI can assess, creating students who excel at producing AI-gradable responses but struggle with original thought.

The pattern is consistent: we optimize for what we can measure, and what we measure is compiled intelligence – the output of understanding, not the process of understanding itself.

The memory-computation-logic framework offers a path forward. Instead of building ever-larger models that maximize computation alone, we might develop memory systems that allow AI to learn continuously, logical architectures that extract and transfer principles across domains, and hybrid approaches that balance all three pillars. But this requires moving beyond our obsession with benchmark performance to ask: what kind of intelligence are we creating?

The deepest risk is this: if we don’t understand what intelligence actually is, then what we’re building isn’t artificial intelligence – it’s artificial something else. Like a plane designed by someone who studied birds but missed the principles of lift, our systems might appear to fly while operating on fundamentally different principles. And once these systems are embedded in every aspect of society – making medical decisions, teaching our children, allocating resources – we’ll want to know whether the principles are sound.

VII. The Journey Ahead

This essay opens a series of investigations into the nature of intelligence and its implications for AI development. The series follows a single thread: intelligence is not what we think it is, and understanding it properly changes how we should build it.

Foundations. We begin by establishing the framework. The next essay argues that intelligence is best understood as learning rate under novelty – and that it scales through massively parallel search, not through individual brilliance. The essay after that examines why experience alone doesn’t produce intelligence, and why compression – the ability to reorganize experience into reusable structure – is the bottleneck we consistently underestimate.

The Core Mechanism. We then investigate where abstractions come from – how biological systems form templates from raw experience, what role embodiment plays in constraining the search, and why current AI architectures lack the consolidation mechanisms that make biological intelligence so energy-efficient.

The Structure of Intelligence. With the mechanism established, we examine reasoning itself – whether it’s compressed search, how it relates to causality, and how collective intelligence achieves compressions no individual mind could. Language, culture, and institutions all play roles in this story.

Implications for AI. We apply the framework to current systems – what they get right, what they’re missing, and what a breakthrough in memory architecture might change. The argument: the field’s obsession with compute is a consequence of not having cracked memory, and when that breakthrough comes, compute will become far less interesting.

Implications for Us. Finally, we address measurement – how the compression framework reveals that we’ve been evaluating intelligence wrong, in education, in hiring, in AI benchmarks, and in our self-conception. And we close with the question of what intelligence systems should look like if we take this framework seriously.

The four minds we began with – mathematician, sculptor, tracker, grandmother – remind us that intelligence has always been plural. As we build new forms of mind, we have a choice: create a monoculture of optimization, or cultivate a diverse ecosystem of intelligences, each contributing its unique strengths.

Welcome to the exploration.

Further Reading

The Extended Mind by Andy Clark and David Chalmers (1998) A seminal paper arguing that cognition extends into tools, environments, and other people. Challenges the premise of individual intelligence.

Seeing Like a State by James C. Scott How “high modernist” schemes to improve human conditions have failed by ignoring local, practical knowledge in favor of abstract models. A cautionary tale for AI development.

The Enigma of Reason by Hugo Mercier and Dan Sperber Argues that human reason evolved not to help us think better individually, but to justify ourselves to others and evaluate their arguments. Reframes reasoning as fundamentally social.

Other Minds by Peter Godfrey-Smith Intelligence that evolved completely independently from ours. Octopuses have distributed brains, with neurons in their arms that act autonomously. What might this tell us about alternative architectures for AI?

The Embodied Mind by Francisco Varela, Eleanor Rosch, and Evan Thompson Challenges computational theories of mind, arguing that cognition arises from embodied action, not information processing.

Gödel, Escher, Bach by Douglas Hofstadter Worth revisiting for its exploration of self-reference and strange loops. Suggests intelligence might be less about processing power and more about recursive self-modeling.

The Origins of Language: A Slim Guide by James R. Hurford Concise overview of how language evolved and why it might be the key differentiator of human intelligence.

Ways of Being by James Bridle A recent exploration of non-human intelligence – animals, plants, machines. Challenges anthropocentric views and suggests radically different ways of thinking about what intelligence might be.