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

Understanding the neocortex is central to building true intelligence.

Hawkins insists that without a principled model of how the neocortex works, AI will remain narrow and brittle; brain-inspired architectures are, in his view, the fastest path to general intelligence.

The brain represents the world through thousands of spatial reference frames.

Each small cortical region (or "column") learns complete models of objects and concepts by encoding locations within object-centered reference frames and moving through them over time, then collectively voting to recognize what is sensed.

Time and movement are fundamental to perception and cognition.

Brains don’t classify static snapshots; they continually process changing inputs as we move eyes, hands, and attention, building models by linking sequences of sensations across time rather than via single images.

Abstract thought likely reuses the same spatial machinery as perception.

Evidence from memory palaces and fMRI suggests concepts like birds or mathematical ideas are organized as navigable “spaces” in the cortex, using grid-cell-like mechanisms originally evolved for physical navigation.

Biological neurons are predictive, sparse, and learn via new connections, not fine weight tweaks.

Real neurons use thousands of synapses, sparse population codes, dendritic prediction, and synaptogenesis (plus silent synapses) to support rapid, robust, and continual learning—properties largely absent from standard deep nets.

Sparsity and many small models could make AI more robust and data-efficient.

Hawkins’ group shows that enforcing sparse representations in deep networks improves robustness to adversarial examples, and they aim to extend this with Thousand Brains–style ensembles of simple, voting models.

Intelligent machines need not be human-like or dangerous by default.

Hawkins separates neocortical-style intelligence from emotions, drives, and self-preservation, arguing we can design systems that reason and explore without human-like urges, and that current AI risks are more about misuse than existential doom.