Hawkins, Jeff

neuroscientist; author of On Intelligence and A Thousand Brains: A New Theory of Intelligence

Jeff Hawkins: The Thousand Brains Theory of Intelligence

Lex Fridman Podcast · Aug 8, 2021

• Jeff Hawkins
• Thousand Brains Theory
• neocortex
• cortical columns
• reference frames
• grid cells and place cells
• prediction and dendritic spikes
• sensorimotor learning
• AI existential risk skepticism
• self-replication risk
• mind uploading skepticism
• SETI and civilizational knowledge preservation
• false beliefs and epistemic humility

Core ideas

Hawkins presents the "Thousand Brains Theory," the central claim of his research at Numenta: rather than a single unified model of the world, the neocortex consists of roughly 150,000 largely independent "cortical columns," each a complete sensorimotor modeling system capable of learning full models of objects (not just fragments), which reach consensus about what is being perceived through long-range "voting" connections — and it is only the outcome of this voting, not the underlying activity of individual columns, that we are consciously aware of. He defines intelligence itself as the capacity to learn a structured model of the world sufficient for prediction and planning, and argues this model is built specifically through movement — touching, walking, looking — rather than passive absorption, drawing an analogy to how one only truly learns a new house or a new phone by physically moving through and interacting with it.

He argues prediction is not merely a useful byproduct but the fundamental mechanism by which the brain tests and corrects its internal models: virtually every sensory input carries an implicit prediction, and surprise (e.g., a water bottle opening the "wrong" way) is what signals a model needs updating. Mechanistically, he describes this as occurring largely inside individual neurons via internal "dendritic spikes" — distinct from the external action potentials neuroscience traditionally studied — which represent a neuron's anticipation that it may soon become active, causing it to fire fractionally sooner when its prediction is confirmed, which in turn reshapes which neurons win out across a network under different contextual predictions (work Hawkins and colleagues published in 2016).

Central to the theory is the concept of "reference frames": Hawkins argues that just as a physical or computer model of a house requires spatial coordinates to represent where its parts are relative to each other, the brain must assign reference frames to everything it models — literally to make predictions about, e.g., what a fingertip will feel touching a specific point on a coffee cup. He traces this capacity to the older, well-studied "grid cell" and "place cell" systems in the hippocampus and entorhinal cortex (used by all mammals to map physical space and navigate), proposing that evolution repurposed and compacted this same mapping mechanism into the generic, repeatable cortical column — explaining both the neocortex's rapid evolutionary expansion and its remarkable universality (the same underlying algorithm modeling a coffee cup, a mathematical concept, or an idea like democracy, with hierarchical hierarchical structure — like a logo made of letters, on a bottle, holding liquid — built into it automatically). He cites recent (as of early 2021) evidence of grid-cell-like activity discovered throughout the neocortex, not just the hippocampus, as validating a key prediction of the theory.

On artificial intelligence, Hawkins believes the field is close to being able to engineer a working equivalent of the cortical column as a general "sensorimotor learning system" — one requiring some form of movement through structure, whether physical (a robotic finger or eye) or virtual (an agent navigating links on the internet) — and predicts robotics and AI will eventually merge because the same underlying sensorimotor learning algorithms apply to both. He draws a direct analogy to the Turing machine: just as computing pioneers created a general computational principle without knowing its eventual uses, this cortical-column-style learning principle could eventually be applied far more broadly than any single narrow AI application (vision, language, etc.) known today, with an impact he expects could exceed the impact of computing itself.

Hawkins pushes back strongly on existential-risk framings of AI (in implicit and explicit disagreement with figures like Sam Harris, Elon Musk, and Stuart Russell), arguing that people conflate "intelligent" with "humanlike" — assuming a sufficiently smart system will necessarily develop desires, self-preservation instincts, or independent goals, when in his view these are evolved biological traits with no inherent connection to a modeling system's ability to learn about the world. He compares a manufactured neocortex to a map: a map has no inherent goals or desires of its own, and any agency or purpose in an AI system would need to be deliberately engineered in (as with a self-driving car's built-in safety overrides), not spontaneously emerge as a side effect of increased intelligence. Where he does see genuine existential risk is in self-replication (whether biological, like engineered pandemics, or a hypothetical fully automated end-to-end robotic manufacturing chain) — since self-replication, not raw intelligence, is what enables uncontrolled, evolution-like proliferation — and he explicitly separates this concern from the "intelligence" side of AI, which he considers not inherently dangerous on its own, though he readily concedes AI applied poorly (in the hands of bad actors, or via unintended societal effects of things like recommender algorithms) is a genuine and serious concern.

On the long-term future, Hawkins is skeptical of near-term mind uploading or deep brain-computer merger (calling both technically far more difficult than commonly assumed — comparing the difficulty of even simple prosthetic-limb signal decoding to the vastly harder problem of meaningfully interfacing with billions of heterogeneous, functionally distinct neurons) and argues that even if uploading were achieved, it would not deliver what people actually want (a copy would exist, but the original biological self would remain, mortal, watching). He instead frames the more meaningful long-term promise of AI as "transcendence" beyond human biological limitation: since humans are fragile and tied to Earth, non-biological intelligent systems — not human minds merged with machines — could be sent to build habitats on Mars, travel the galaxy, and represent human knowledge (though not particular individuals) far beyond what human bodies could survive, analogous to how his children represent and extend him without being him.

This connects to Hawkins's broader concern about civilizational continuity: he proposes that human civilization should actively archive its knowledge (e.g., continuously uploading something like Wikipedia to orbiting satellites) as an "insurance policy" against collapse, so any future intelligent species on Earth — or beyond it — could discover that we once existed, rather than vanishing without trace like the dinosaurs. He extends this to a specific, non-decoded alternative to conventional SETI signal-searching: since any given technological civilization might only be active and detectable for a relatively short window (he estimates on the order of 10,000 years), he proposes building a passive, self-sustaining broadcast requiring no ongoing power or maintenance — analogous to how astronomers detect exoplanets via periodic dimming of starlight — that could signal "someone intelligent was once here" for potentially millions of years, and argues SETI research should actively consider searching for signals of this different character.

On human nature and civilizational risk, Hawkins expresses disappointment less in the possibility of AI destroying humanity than in humanity's own difficulty separating rational thought from "evolutionary heritage" impulses (including violent ones), and in humans' inherent susceptibility to false beliefs about anything not directly, physically testable — arguing the best available remedy is not censorship of "wrong" ideas but broad public understanding that everyone's brain constructs an approximate, fallible model of the world, which could foster more productive disagreement grounded in shared epistemic humility rather than dogma. Personally, he traces his career motivation to a list of "biggest questions" (why does the universe exist, why do physical laws take the form they do, the origin of life, and the origin of intelligence) made as a teenager, and a catalytic 1979 Scientific American essay by Francis Crick arguing neuroscience had ample data but lacked a unifying theoretical framework — which convinced Hawkins he could contribute as a theorist rather than an experimentalist. He closes by framing his own legacy modestly (as one of many people who might "accelerate the inevitable" rather than cause something that wouldn't otherwise happen), and states his ultimate view of life's purpose as the acquisition and preservation of knowledge and understanding of the universe, a passion he considers worth pursuing regardless of any assurance of its practical payoff.

Excerpts