Jeff Hawkins: Thousand Brains Theory of Intelligence | Lex Fridman Podcast #25

Jeff Hawkins lays out his lifelong mission to understand how the human neocortex produces intelligence, arguing that studying the brain is the fastest route to building intelligent machines. He traces his work from hierarchical temporal memory to his recent Thousand Brains Theory, which proposes the cortex stores everything in reference frames and that thousands of cortical columns each build complete models that vote to reach consensus. He critiques deep learning's point-neuron model as too simplistic, explaining how real neurons act as time-based predictive engines using sparse representations. The conversation ranges across continuous learning, consciousness, existential AI risk, and why he believes intelligent machines should not be made human-like or emotional. Hawkins ends with a vision of intelligent machines preserving humanity's knowledge as our true legacy.

• Hawkins's core Thousand Brains claim: there is no single model of an object; thousands of cortical columns each build complete models that vote to reach consensus.
  00:39:36
• The central insight that the neocortex stores everything in reference frames anchored to objects, like CAD coordinate systems, throughout the entire cortex.
  00:36:01
• Theory that the cortex repurposed ancient grid cells and place cells used for spatial navigation to build maps of abstract concepts.
  00:31:51
• Argument that artificial 'point neurons' barely model real neurons, which use thousands of synapses and dendritic spikes as predictive engines.
  01:05:45
• Claim that the brain uses sparse population codes (about 2% active) which give robustness and resistance to adversarial examples.
  01:18:17
• Assertion that backpropagation cannot happen in real brains; learning is pure Hebbian synaptogenesis on dendritic segments.
  01:34:32
• Position that intelligence is just the neocortex and that we should deliberately not build AI with human emotions, drives, or desire to reproduce.
  01:58:35
• Hawkins frames humanity's true legacy as knowledge, to be preserved beyond us through intelligent machines.
  02:06:56

• The neocortex occupies about 70 to 75 percent of the volume of the human brain and exists only in mammals.
  00:06:17
• The neocortex is about 2.5 millimeters thick, roughly the size of a dinner napkin, and looks remarkably uniform everywhere across species.
  00:07:51
• Vernon Mountcastle argued in 1978 that the whole neocortex runs on the same common cortical algorithm.
  00:08:52
• If the optic nerve is rewired to a different cortical region, that region becomes a visual area; blind people's visual cortex gets repurposed.
  00:09:53
• Hawkins says he wakes in the middle of the night and stays awake an hour in a half-sleep state to think through these problems.
  01:01:01
• He tests his theory of place cells by navigating his house to the bathroom with eyes closed, estimating his accumulating error.
  01:02:34
• Real neurons can have 5,000 to 30,000 synapses, but about 95 percent are too far from the cell body to fire it individually.
  01:06:16
• Charles Babbage invented the computer in the 1800s and was largely forgotten until much later.
  01:38:10
• Color does not exist in the world; the brain only receives axons firing at rates correlated with light frequency.
  01:47:34
• In phantom limb syndrome, amputees feel and can move a non-existent arm, showing the brain holds a model of the body that may not match reality.
  01:48:05

Books, products and media the guest or host genuinely endorsed here — with the buy link.

Affiliate link — we may earn a commission at no extra cost to you.