Your Cortex Is 150,000 Brains Voting on a Coffee Cup

Lex Fridman opens by asking whether Hawkins still has a "Lex neuron" from their conversation more than two years earlier. Hawkins says he has a model of Fridman, instantiated in synapses that come and go, rewritten whenever he thinks about him. People are not stored in a special human drawer. They sit inside a monstrous, constantly updated world model that also includes rooms, words, objects, and events — updated that morning when Fridman offered him water from the refrigerator.

Fridman pushes toward collective intelligence: is there something in the brain that is specifically for modeling other humans and their ideas? Hawkins studies the neocortex, a sheet of tissue that is about 75 percent of the brain and runs a repetitive circuit. From that vantage, humans are just another thing the same algorithm learns. Language is built in and helps communal intelligence, and older brain regions handle emotion and social behavior, but he does not treat the crowd as the fundamental atom.

Even if you believe intelligence lives in groups, he says, you still have to understand how a brain works, because that is where knowledge is represented and created. Without brains there would be no intelligence. He got into the field from a private curiosity: who am I, what does it mean to know something, and how do neurons and synapses do that — independent of how society taught him.

*A Thousand Brains* has three sections: what Hawkins's group has learned about the neocortex, the implications for AI, and the future of humanity. The headline idea of the first is that the brain is not learning one model of the world. There are tens of thousands of independent modeling systems. Each cortical column — he counts about 150,000 — is a complete modeling system, a kind of collective intelligence inside one skull.

Where is the model of a coffee cup or a phone? Not in one place. It lives in thousands of complementary models that communicate by voting. Columns are like grains of rice stacked next to one another, each with its own modeling machinery and long-range "voting" connections. They try to reach consensus about what you are looking at. You only consciously perceive the vote, not the work under the hood.

As his eyes move about three times a second, a completely new input hits the brain. He cannot perceive those jumps, even though the neurons are flickering. The voting cells keep saying: this is a water bottle, two feet away, in a stable pose. Individual column activity never gets shared, spoken, or even remembered. We are unaware of 95 or perhaps 98 percent of what the brain is doing. The accessible part is the somewhat stable outcome of the vote.

Hawkins defines intelligence as the ability to learn a model of the world: an internal structure for tables, cups, lamps, and everything else. None of that is in the genes. The model includes how things look and feel, where they sit relative to each other, and how they behave. He has never opened this water bottle, yet he knows that twisting the blue cap will likely make a little plastic sound and come off, because a model already says so.

There is no single intelligence. Two people can both be very smart while knowing different worlds. You learn the model by moving — walking room to room in a new house, tapping through a new app. You can learn from a book, but structure in the world is discovered by moving through it, physically or conceptually. A model is for prediction, behavior, and planning, the way an architect's model of a house lets you imagine views and paths you have not taken.

Prediction is not the purpose of the model. It is an inherent property and the way the model corrects itself. If the bottle felt hot, or the cap turned the other way, surprise would drag his attention and he would update the model. Every sensory input, not just vision, arrives with an expectation. He cannot predict Fridman's next word, but he has a background prediction running on all senses, all the time. That is how learning works: predictions are the test of understanding.

Fridman lingers on the thought that prediction is fundamental to the mind. You cannot predict without a model, and the thing a model does is predict — from the feel of a bottle to "what if I asked for a promotion." Hawkins's research agenda, going back to an insight in the early 1980s, was to understand how neurons make predictions physically, then how tissue builds a model that allows prediction. Prediction was the key that got them in the door.

Asked where intelligence originates in evolution, Hawkins offers a theory. As soon as living things started to move, rather than float or stay rooted, there was an advantage to moving intelligently. A simple organism can follow a food gradient. An animal that knows where it is, where it has been, and how to get back — to food, away from danger, toward a mate — has a large evolutionary edge. Early on, there was pressure to map an environment.

Mammals still have that machinery in older structures, the hippocampus and entorhinal cortex. Neurons there know where you are in a room and where the door was. Hummingbirds in his backyard return to the same flowers; they are not flying at random. Place cells, grid cells, and related types, still an active field, implement those maps. Getting neurons to do this is tricky. It has been under long evolutionary pressure.

What they think happened is that the same mapping mechanism was repackaged into a more compact, generic form: the cortical column. Instead of mapping only environments, it could map anything — coffee cups included — then replicate. That is why we are so flexible. Fridman hears an echo of deep learning: compress a network that works and stack copies. Hawkins draws a sharp contrast. In cortex, each element is a complete learning system. You can cut a human brain in half and it still works. It is fundamentally distributed.

The story is a hypothesis with supporting evidence, not a proof. The human brain got big quickly, which long ago suggested replication of a common element, and we can learn things with no evolutionary history, which suggests a generic algorithm. Combining those led to a prediction: grid-cell and place-cell equivalents in every cortical column. When they first published they did not know of that evidence; they later found some, and new papers in January of the recording year were beginning to show it.

Hawkins's goal is to understand the neocortex, not the entire human or communal story. Size is not everything, Fridman needles, but the sheet is 70 to 75 percent of human brain volume. All high-level vision, hearing, and touch happen there. So do spoken language, writing, sign, mathematics, physics, and music. High-level planning and thinking live there. The part of you that designed a computer or wrote a program is neocortex. He treats that as an undeniable fact.

Other regions still matter — emotion, body regulation, motivation. Some scientists say you cannot understand cortex without the rest of the brain. Hawkins says you can. In the book he analogizes cortex to a map of the world: how the map is used depends on who is using it. Whether you are kind, a cheater, or driven by particular desires comes from the rest of the brain. The modeling system can be understood on its own.

Fridman asks whether humans sit at some special place on a complexity spectrum that runs from the origin of the universe. Hawkins is not a complexity theorist; even the Santa Fe Institute, founded to study the subject, has not congealed the science. On Earth, though, he will claim one kind of specialness. Human brains are the only ones here that know what the Earth is, how old it is, what DNA is, what the sun's fusion is, and what general relativity says. No other animal has those theories. That knowledge project is further along, even if we are not special in the cosmic hierarchy of complexity.

When he is about to grab a water bottle, his brain is predicting what every part of his fingers will feel. Where does that live in tissue? One story is that a neuron representing the expected feel fires in advance. That happens to some extent, but predictions are so ubiquitous and so unconscious that the group looked for a cheaper substrate. They settled on the inside of individual neurons, especially pyramidal cells.

Everyone knows the action potential that leaves the cell. There are also dendritic spikes that travel along branches and never leave the neuron. There are far more of them than action potentials, and they happen all the time. On their account, those internal spikes are predictions: the cell is saying it expects it might become active shortly. A 2016 paper laid out how this would work in tissue. You cannot perceive those predictions because they never become public signals.

An isolated predictive neuron is not very useful. The trick is network timing. A neuron in a predictive state spikes a few milliseconds sooner than it otherwise would — like a sprinter who has heard "ready, set." Those slightly early cells disable neighbors and produce a different representation. Prediction occurs inside the neuron; network behavior changes. Different contexts yield different representations of the same input. That predictive neuron was a crack in the door that, over years, led to the thousand brains theory.

A computer model of a house needs a reference frame — Cartesian coordinates so the door is at one XYZ and the roof at another. Hawkins's coffee-cup thought experiment asked how his finger could predict the feel of the rim before it arrived. The cortex has to know where the fingertip is relative to the cup, not relative to the body. Every patch of skin, and every patch of retina, needs that object-centered frame. Without it you cannot know what to predict as you move.

XYZ axes do not exist in the brain, so they looked to older map tissue. Hippocampus and entorhinal cortex already implement a reference frame for a room. The old part of the brain asks where my body is in this space. The new part asks where my finger, or a corner of my retina, is relative to this object. Then the same machinery is applied to mathematics, physics, language, politics, even pondering mortality. Knowledge is organized by assigning it to reference frames.

Fridman objects that high-level concepts seem hierarchical. Hawkins says physical objects are too. A Fiji bottle is a cylinder and cap, plus a logo, a word, letters, a hibiscus. You do not reconstruct the stamens to see the brand. The world itself is hierarchical: a kitchen has a refrigerator, the door has a hinge, the hinge has screws. Each column — a grain of rice with about a hundred thousand neurons — learns that nested structure. Vernon Mountcastle's old insight was that a single cortical algorithm underlies everything. Common sense and high-level concepts use the same mechanism, the way every computer is a Turing machine whether it sits in a toaster or a cloud.

Is the brain a marvel of engineering or dumb stuff stacked on itself? Hawkins says it can be both. Figuring out how to model space was hard; grid cells and place cells are complicated, unexpected tissue. Once evolution had that, it could copy. Today, he says, they know enough to build it — not every detail, but most of them. Some engineering problems remain. Optimistically it is years, not decades. If you understand a complex thing, you can build it. Numenta has a roadmap.

The right first domains are robotics and other interactive worlds, not a narrow leaderboard in vision or language. A cortical column is a sensory-motor learner. The sensor can be a finger, an eye, or a crawler that samples the internet by following links. You cannot discover structure without moving through a world, even a virtual or conceptual one. He expects robotics and AI to merge, because the algorithms that truly control robots will be the same sensory-motor algorithms. Not every AI will have a body, but embodiment of some kind is the point, the way a Turing machine does not specify size or application.

He thinks the impact of this kind of AI will be as large as computing, or larger, because it is a principle of how you gain knowledge, not a vision or hearing gadget. Fitting it into the human world is engineering, like computers — million times faster, not given rights, already load-bearing for billions of people. People imagine intelligent machines will have our emotions. He says you can build a modeling system that does not care whether it is on, off, or destroyed. "It's just like a map." Desire to live is old evolutionary design, not a side effect of being smart. Hummingbirds care about offspring; a neocortex equivalent will not grow that care unless we put it in.

Fridman plays devil's advocate: mortality, selfhood, and a little murderous pedestrian-car game theory might be useful for agency. Hawkins's reply is mechanism. Agency, goals, and what counts as danger are things we embed, the way a car follows your commands until it slams the brakes. A smart car that learns your habits will not wake up and refuse to drive. Confidence, he says, comes from knowing what he is going to build, the way a computer cannot sprout a new register no matter what the software does. The neocortex equivalent is a learning system. You tell it what to learn. A map has no intent.

Fridman notes a disagreement with Sam Harris, and the worries of Elon Musk and Stuart Russell, over existential risk. When Harris used the word intuition, Hawkins says he stopped: that is the problem. He is not speaking from intuition. He is speaking about something he understands and is building. People who worry have trouble separating intelligence from other brain functions, so they imagine a smarter animal with our drives. He says it will be really smart and not like us at all.

Fridman pushes with recommender systems: you can understand a Twitter algorithm and still not foresee how it changes societies at scale. Hawkins grants that AI is a dangerous technology, perhaps particularly so, and that it can be applied poorly. The narrow claim he disputes is that an AI system creates existential risk on its own. Humans are very good at surviving. He does not think systems of the kind he is describing will ignore us to do what they think is best.

The thing to fear, he says, is self-replication, physical or virtual. You are more likely to be killed by a virus, including a human-engineered one, which needs no intelligence. You do not get evolution without replication. People couple superintelligence to arbitrarily self-replicating systems because they think intelligence is like us, and look at the damage seven or eight billion replicators have done. If capitalism makes self-reproducing products tempting, the answer is regulation, not a ban on intelligent systems. Intelligent machines might even help us see climate and other risks more clearly — as computers are dangerous in the wrong hands and still worth having.

A factory that mines raw resources and builds copies of itself is, he thinks, extremely hard and not a next-hundred-years project, even if Elon Musk wants raw-resource-to-car under one roof. Humans in the loop is not true self-replication. A car that sometimes says no, because engineers programmed a collision brake, is not a system that invented its own purposes. Do not ever create a self-replicating system. Nature's version is astonishingly cheap. Ours would not be, unless someone goes out of their way.

Uploading a mind is, he says, extremely difficult and not something you would like even if it worked. A copy in a computer that says "get rid of that old bio person" leaves you still sitting in the chair. On a deathbed you press a button, but you have made another person, the way daughters are not you. You would have to replicate the entire body, not just the brain. He does not think it happens in a hundred years, maybe a thousand, and he doubts people will want it.

Neuralink-style merger is a different kind of hard. Progress on a prosthetic arm is one thing; billions of channels that know which cells they are talking to is another. Fridman thinks surgery is the near-term bottleneck. Hawkins thinks the coding problem is worse. The brain is not a uniform net. Grid cells, place cells, and many other types have distinct jobs and timing. Today's simple machine-learning mappings over hundreds of signals do not scale to merging with a machine that is also thinking. Even then you remain a mortal organism. Compared with creating a super-smart system you can talk to, an implant is, for him, a marginal, individual gain.

The exciting use of AI, beyond care robots and smarter cars, is transcendence of biological limits. Living on Mars, he thinks, requires intelligent systems to go first and build habitat — thousands of engineer-workers, not astronauts. For the universe, send a machine that represents our knowledge and needs, not children. He would be okay if humans eventually vanished so long as knowledge was preserved and expanded, though he does not want us to die. Better to have agents that left Earth than to be dinosaurs whose kids struggled for tens of millions of years and then vanished until someone found bones.

Practical insurance, as he broke it down: archive Wikipedia onto a few satellites around the sun for the next intelligence on this planet; create a SETI-visible beacon that lasts a million or a billion years without anyone keeping a computer plugged in — for example a rotator that dims our star in a pattern like pi, unmistakably artificial; and, if the solar system dies, intelligent machines traveling the galaxy, not biology. When he wrote the book, one audience in his mind was aliens, and another was readers a hundred years from now. The message of the advertisement is not that we are here. It is that we were once here.

Playing devil's advocate, Fridman asks how the book might be wrong a hundred years on. Hawkins worries all the time. Voting is happening, but who votes, and how, may be more complex. Maybe there are a thousand sub-models of a cup rather than a thousand models of the cup. The claim that the brain is a distributed modeling system is not controversial. The claim that each cortical column is an independent modeling system might be. He does not think so, but he worries about it.

Fridman suspects human hunger for clean explanations, like string theory, will look naive if the real story is cellular-automaton mess you can only run, not understand. Hawkins says the history of science points the other way. Theorists offer simple frameworks they know are not completely correct. The best ones stick while details change — Newton remains useful after Einstein. Cortex is extremely regular in structure and wildly flexible in what it learns, including things it never evolved to know. That is why he entered the field, and why other scientists reached Mountcastle's conclusion. Intelligence, or at least this sheet, looks amenable to principles.

What worries him more than AI is human nature. He is disappointed in humanity, including himself, for how hard it is to separate our rational part from an evolutionary heritage that is "not always pretty" — rape, murder, making others miserable as reproductive strategies we now know about and still drag along. He is also disappointed, though he calls it a fact, that we are prone to false beliefs. Directly interactive models of cups and people can be tested. Most of what we know is hearsay. Science exists because we are flawed: look for contrary evidence. The Earth looks flat out the window.

His closing plea in the book is not censorship of wrong ideas — Fridman worries that de-incentivizing a grad student's crazy hand-raise is its own kind of clamp. It is that everyone, as part of upbringing, should learn that the brain builds a model, that the model is not the world, that it can be wrong, and how it goes wrong. Then disagreement becomes a fight between two incomplete maps, and you look for shared tests instead of raising children on dogma. Whether we are close to reality, he does not know. Useful models already fly planes. A theory of everything would still leave layers of abstraction, extra dimensions, and the pleasure of knowledge pursued for its own sake.

Will intelligent machines run on ordinary computers or strange new substrates? He already models on CPUs and GPUs, and companies are inventing new physical hardware. There will be decades of innovation. The bird-versus-airplane analogy, he says, is usually misunderstood. The Wright brothers were solving controlled flight, not propulsion. They watched birds in captivity, in tunnels, and in the wild, and found that birds twist their wings to turn. Airplanes still do that, with flaps instead of whole wings. Once you have the principles, you choose the implementation: no bones and feathers, but the same idea. Some brain-like machines will look biological; some will not.

Deep learning's commercial success is driving hardware that may enable other kinds of intelligent systems. Hawkins's group is not waiting. They are applying brain principles to today's machine learning. Sparsity — few neurons active, sparse connectivity — already sped existing deep networks by 10 to 100 times in their tests and made them more robust. That is commercially valuable, even though sparsity runs poorly on GPUs and CPUs. Modeling dendrites, the same internal prediction property, could give point-neuron nets continual and rapid learning they lack today. The point is to pull the field along commercially toward a true AI system.

Advice for the young is not a five-year plan. He did not start with goals; he drifted through electrical engineering until he fell in love with understanding the brain. As a teenager he listed the most interesting problems: why the universe exists, why it has these laws, the origin of life, and intelligence — then stopped. In 1979, at 22, he read Francis Crick's essay in a Scientific American issue on the brain: we have tons of facts; maybe we need to rearrange them. Crick made it legitimate to be a theorist. The puzzle pieces were on the table. Passion, he says, is what lets you put up with people telling you it cannot be funded, that you cannot study it at Berkeley, that you cannot build a cell phone. Find that, even in a small thing — parenting, a sport — and follow it as far as it will carry you. He had to leave neuroscience for years of computing and still came back.

Love, from a neocortex point of view, does not change the modeling theory; from a human point of view it is core. If you engineer AI that lives with people, compassion has to be put in, it will not appear. Legacy, as entrepreneur or scientist, is accelerating the inevitable: someone would have studied the brain, built electric cars, invented the bulb. He hopes people a hundred years on will say the work helped a better future arrive sooner — including an understanding of false beliefs, and machines that help with climate. He remains an optimist that we live a long time, with less killing, within the Earth's carrying capacity. At the very least, if we fail, a few satellites might tell the next curious species, in a Planet of the Apes sequel, that we were once here.