Intelligence Was Here Before We Were

There is a forest in Oregon that has been thinking for at least 2,400 years.

It does not have a brain. It does not have neurons. It has never produced a sentence. But the Armillaria ostoyae fungal network spreading beneath the Malheur National Forest, covering 2,385 acres and weighing an estimated 35,000 tonnes, has been doing something that, if a human did it, we would not hesitate to call intelligent. It monitors its environment continuously. It routes carbon and phosphorus from healthy trees to struggling ones through mycorrhizal threads so fine that a single teaspoon of forest soil contains miles of them. It responds to disease and drought. It remembers, in the only sense that matters biologically: it encodes the history of its environment into its structure and acts on that encoding.

It has been doing this since before the Roman Empire existed.

We are going to talk about artificial intelligence. But we cannot talk about it honestly without first standing inside a much larger picture. The picture of what intelligence actually is, how old it is, how many forms it takes, and how recently the particular kind we are most proud of arrived on the scene.

The universe is 13.8 billion years old. Earth is 4.5 billion. Life has been running on this planet for 3.8 billion years. Neurons first appeared roughly 600 million years ago, in the ancestors of jellyfish. Our species has existed for approximately 300,000 years. Written language is perhaps 5,000 years old. The scientific method is roughly 400 years old. The word "scientist" was coined in 1833. The transformer architecture, which underlies every major AI system today, was published in a single paper in 2017.

On the timescale of the universe, the entire history of human cognition fits inside the last four seconds of a year-long film. The transformer is a frame of that last second. And yet we speak about artificial intelligence as though cognition, specifically symbolic, linguistic, deliberate cognition, is what intelligence fundamentally is. As though the 3.8 billion years of exquisitely sophisticated biological problem-solving that preceded the first word were not quite intelligence yet.

This essay is about that assumption. Not to dismiss what we have built, which is real and remarkable and worth understanding precisely. But to place it correctly, inside the actual history of intelligence on Earth, and to ask what that placement reveals about what we can automate, what we cannot, and most importantly, what a genuinely different architecture would need to do.

Part I: What Was Happening Before We Started Thinking

Before we can understand the prefrontal cortex, we have to understand what it was added to. The nervous system did not appear fully formed. It evolved, layer by layer, over hundreds of millions of years, each layer built on top of older structures that were already doing something. At the core: the brainstem and cerebellum, handling survival operations shared with reptiles. Wrapped around it: the limbic system, shared with all mammals, handling emotion, memory, and social bonding. On top of that: the neocortex, handling language, abstract reasoning, and planning. The prefrontal cortex, the most recently evolved part and the part most associated with deliberate thought, is not even fully mature in an individual human until the mid-twenties. It is the newest room in an ancient house.

Consider what a single cell does. A bacterium with no nucleus, no brain, no anything we would recognise as a nervous system, navigates a chemical gradient through a process of molecular sensing and flagellar adjustment that constitutes, in every functional sense, a decision. It has a goal, a sensor, and an actuator. The biologist Michael Levin at Tufts University has spent decades documenting the computational sophistication of non-neural biological systems. Planaria flatworms, sliced in half, regenerate their entire body plan, including a correctly configured brain, within weeks. Levin calls this "basal cognition," and his findings are published in Nature, Cell, and PNAS. The conclusion is empirical: the capacity to process information, set goals, and act to achieve them is far older and far more widespread than we thought.

Perhaps the most striking evidence comes from the slime mould Physarum polycephalum. A team at Hokkaido University, publishing in Science in 2010, grew it on a map of the Tokyo area with oat flakes at major city locations. Over 26 hours, the organism extended and retracted its tubes until it had found a configuration closely matching the actual Tokyo rail network, efficiently connecting all food sources with minimal redundancy. No neurons. No brain. Chemistry and time, doing what we would call optimal network design.

The universe has been running intelligence without cognition for almost four billion years.

Part II: The Explosion of the Nervous System, and What Language Actually Is

When neurons did appear, they did not replace the distributed cellular intelligence that preceded them. They added to it. And for hundreds of millions of years, that nervous system did its extraordinary work without producing anything that required a symbol.

The cuttlefish deserves a moment here. It has approximately 500 million neurons, roughly the same count as a honeybee, and its brain evolved on an entirely independent path from the vertebrate lineage. And yet it can solve multi-step foraging problems, display episodic memory, pass versions of the marshmallow test for delayed gratification, and camouflage itself in real time through chromatic patterns that vary by predator type. There is no language here. There is a rich, situated, embodied intelligence that reads the world directly.

Language, specifically the recursive human language that allows us to embed sentences inside sentences, to refer to things not present, to communicate about the past and the counterfactual, is estimated to have emerged between 100,000 and 50,000 years ago. Writing appeared around 3,200 BCE. What we call human knowledge, the accumulated record of what our species has figured out, is almost entirely contained in the last 5,000 years of writing. And science, the formalised, self-correcting method that produced quantum mechanics and the double helix, is roughly 400 years old.

400 years. On a planet 4.5 billion years old.

Language is extraordinary. It allows us to compress, transmit, and build upon understanding across generations. But it is a tool. A very recent tool, built on top of biological systems that were managing extraordinary complexity long before the first word was spoken. We have been thinking in language for so long that we have come to believe that thinking is language, that knowledge is what has been written down, that intelligence is what can be expressed in sentences.

The mycorrhizal network does not agree. Neither does the immune system, running its combinatorial search across antibody space to generate responses to pathogens it has never encountered. Neither does the monarch butterfly, navigating 4,500 kilometres from Canada to a specific grove of trees in Michoacán using a sun compass, a magnetic sense, and a time-compensated navigation system that no human engineer has fully reverse-engineered, guided by information encoded not in language but in the molecular memory of the species.

Part III: The Brain Is Not One Thing, and AI Captured One Narrow Slice of One Recent Layer

In 1983, Howard Gardner proposed that human intelligence is not a single faculty but a collection of distinct, separable abilities. His core insight has only been strengthened by subsequent neuroscience: the brain is an ensemble of systems, evolved at different times, running largely in parallel and largely below the threshold of conscious awareness.

Antonio Damasio spent years studying patients with damage to specific prefrontal regions. These patients performed normally on standard reasoning tests, could articulate arguments and analyse options. But they could not make decisions. Not because they lacked reasoning ability. Because they lacked somatic markers: the felt bodily signals, accumulated through a lifetime of experience, that serve as rapid pre-rational assessments of a situation. The emotion was not interfering with the intelligence. The emotion was part of it. Daniel Kahneman's dual-process framework extended this: System 1, the fast, associative, below-threshold knowing, carries genuine expertise in ways that System 2, the slow deliberate verbal reasoning, cannot replicate. The chess grandmaster, the seasoned fire commander, the experienced physician: expert judgment arrives before conscious analysis begins.

The brain does not develop these capacities uniformly. Research in neuroplasticity shows that the pathways you use become structurally reinforced. The analytical mind, exercised through years of formal education, becomes a default. It begins to run without being asked. Rex Jung's research shows that highly creative individuals display reduced prefrontal cortex activity during creative work. The executive control region has to quiet for creative intelligence to operate. Marcus Raichle's 2001 PNAS paper established that the default mode network, active during mind-wandering and associative thinking, is not idle noise but a metabolically expensive integrative process that analytical attention actively suppresses. Csikszentmihalyi's flow research confirmed the pattern: peak creative performance is characterised by the suspension of self-referential analytical thought.

Tesla described his creative process with unusual precision in My Inventions (1919): "I do not rush into actual work. When I get an idea I start at once building it up in my imagination. My turbine ran in thought before it ran in the shop." Da Vinci moved across anatomy, hydraulics, music, and painting not by being a better analyser but by holding problems open across frames simultaneously, never fixing them in one disciplinary lens long enough to lose what the other lenses showed. These minds were not more analytical. They were less captured by analysis.

Jensen Huang said it plainly: cognitive ability has become a commodity. The analytical capacity being augmented by AI is the capacity being commoditised. Becoming very good at using a commodity tool makes you efficient. It does not make you irreplaceable.

Here is the critical technical fact for understanding what we have built. The transformer architecture underlying every major language model was directly inspired by biological neural networks. The 2017 paper "Attention Is All You Need" formalised an attention mechanism based on how neurons selectively weight inputs based on relevance. But what we abstracted, the particular operation we scaled into silicon, was the capacity to predict the next token in a sequence based on context. That is, at its core, what every language model does. The AI researcher Subbarao Kambhampati describes LLMs as "giant non-veridical memories, trained to recall patterns of saying rather than facts about the world." He demonstrated this with PlanBench: simple block-stacking puzzles that 1970s classical planners solved trivially. Current frontier models fail consistently, and performance collapses further when block names are obfuscated, proving the reasoning was pattern recall dressed as deliberation.

We took the neuron. We built a word predictor. We called it intelligence. And we built it at exactly the moment when the analytical operation it performs was becoming a commodity, while the receptive operation it cannot perform was becoming the only thing that cannot be replaced.

Part IV: The Knowing That Arrived Whole

In the winter of 1864, August Kekulé was dozing in front of a fire, turning over the problem of benzene's molecular structure. As he drifted toward sleep he saw atoms forming chains that twisted and turned, and then a chain bit its own tail, forming a ring. He woke up and worked through the night. The ring structure of benzene, the foundation of organic chemistry, had arrived in a dream.

Srinivasa Ramanujan, who had received almost no formal mathematical training and spent most of his life in poverty in Tamil Nadu, wrote letters to G.H. Hardy containing theorems of such originality that Hardy could not determine whether they were the work of a genius or a fraud. He brought Ramanujan to Cambridge, where he produced mathematics that even his collaborators could not derive. When asked where the theorems came from, Ramanujan said they were given to him by the goddess Namagiri in his sleep. Hardy, the most rigorous of empiricists, did not dismiss this. He noted that Ramanujan was right, and that the source of the knowledge was irrelevant to its truth.

Henri Poincaré, writing in 1908, described the moment of mathematical insight with unusual precision: "At the moment when I put my foot on the step, the idea came to me, without anything in my former thoughts seeming to have paved the way for it." Einstein described the special theory of relativity as beginning with an image: a sixteen-year-old imagining what it would feel like to ride alongside a beam of light. That physical intuition preceded the mathematics by years.

These are not anecdotes about creativity. They are data about a process of knowing. Something was working on these problems below the threshold of deliberate, verbal reasoning and producing correct, revolutionary outputs that the deliberate mind then transcribed and verified. The layer of intelligence responsible for these discoveries was not the layer we have built AI to replicate.

Part V: At the Edge of the Formalism, They Turned Toward the East

With a frequency that exceeds coincidence, the physicists who pressed hardest into the structure of quantum reality found themselves reaching for frameworks from contemplative traditions to articulate what they were encountering. These were not poets. These were the architects of the most precisely verified physical theory in the history of science.

Erwin Schrödinger kept the Upanishads on his desk throughout his working life. In Mind and Matter (1958), he found in the Vedantic concept of Atman-Brahman the only framework that did not produce paradox when he tried to think about the relationship between observer and observed in quantum mechanics. Werner Heisenberg wrote in Physics and Philosophy (1958) that Buddhist and Hindu concepts of process over substance seemed to resonate structurally with quantum theory's picture of a world where particles do not have definite properties until measured. David Bohm developed the concept of the implicate order, a deeper undivided wholeness from which the apparently separate objects of ordinary experience emerge, in explicit conversation with contemplative ideas about the nature of mind and reality. J. Robert Oppenheimer, watching the first atomic bomb test at Trinity, quoted the Bhagavad Gita in Sanskrit, a text he had studied as a young man for its framework for thinking about forces larger than individual human will.

What made the contemplative frameworks feel relevant was not vague spiritual resonance. It was specific formal results. Quantum mechanics produced results whose implications no classical vocabulary could hold cleanly. The act of measurement is not passive recording but active participation in determining what is real. Bell's theorem, confirmed experimentally by Alain Aspect in 1982, demonstrated correlations between distant particles that cannot be explained by any theory where each particle has definite local properties independent of measurement. The universe at its most fundamental is non-local and measurement-dependent in ways that everyday language cannot fully hold. The contemplative traditions had spent centuries developing vocabulary precisely for territory where the observer-observed distinction dissolves. This was structural correspondence noticed by people who had mastered both systems and had no incentive to find a connection that did not exist.

Part VI: Intelligence Is Older Than We Are, and the Architecture We Need

The laws of physics did not wait for us to discover them. Gravity operated before Newton. Entropy increased before Boltzmann. The genetic code transcribed proteins before Watson and Crick named its structure. We arrived at the very end of this process and said: we are the intelligent ones.

A 2025 benchmark study from HKUST tested language models across undergraduate physics. Fine-tuning on mathematical corpora produced performance gains exceeding 10 percent in mathematics. In physics, the same approach produced 0.14 percent. Mathematics is a formal language where pattern and substance are the same thing. Physics is reasoning about the world through mathematics. Those are different activities, and the difference turns out to matter enormously. Richard Feynman captured it: "Mathematics is not real, but it feels real. Physics is real, but it does not feel real."

What we have built is genuinely powerful and genuinely limited. Current AI will automate, in the next decade, document analysis, code generation, diagnosis support, literature synthesis, and translation. These are not small things. But the moment of genuine hypothesis formation, when a researcher sees that two fields are secretly solving the same problem, does not happen through next-token prediction. The physical intuition that lets a physicist feel what the equations mean before writing them down is not recoverable from the statistical distribution of physics papers.

Nine researchers working from very different starting points, Gary Marcus from neurosymbolic AI, Kambhampati from planning, Chollet from program synthesis, LeCun from world models, Bengio from causal science, Schölkopf from representation learning, independently arrive at the same diagnosis: meaning is structured, not statistical; the right unit is the concept, not the word; a good system predicts before it perceives; it holds many hypotheses as a weighted distribution rather than asserting one; it generates by proposing structured candidates and checking them against something external, never grading its own work.

This is not a description of any existing system. It is a description of what the evidence, from very different starting points, suggests the next serious step should look like. Not more tokens. Not a larger model. A different operation: lifting knowledge into structural representation, detecting correspondence rather than similarity, verifying rather than predicting.

Intelligence was here before we were. The laws that generate self-replicating molecules, mycorrhizal networks, immune systems, and the physical intuitions of great physicists are not waiting for our machines to replicate them. They are operating, right now, in full complexity, in the forest and in the body and in the part of the mind that works while we sleep.

What would it mean to build systems that were informed by that understanding? That is where the interesting work begins.