Jeff Hawkins: The Thousand Brains Theory of Intelligence | Lex Fridman Podcast #208

The brain continuously updates its model of the world. It builds this model through movement, by physically interacting with objects and exploring environments. Sensory inputs are constantly tested against internal predictions; when a prediction is wrong, the model is updated, serving as a fundamental learning mechanism.

Prediction is fundamental to intelligence, not as a goal, but as an inherent property of the brain's modeling system. The brain constantly makes predictions about sensory inputs, and errors in these predictions are how the model corrects itself and learns about the world's structure, from simple objects to complex concepts.

Hawkins theorizes that intelligence originated from the need for early living organisms to navigate and map their environment. The neural mechanisms used for spatial mapping in older brain regions (hippocampus, entorhinal cortex, with grid and place cells) were later repackaged into a more generic, compact, and replicated form, becoming the cortical column capable of modeling anything, not just space.

The neocortex comprises 70-75% of the human brain's volume and is responsible for all high-level functions, including vision, hearing, touch, language, mathematics, physics, music, and complex planning and thinking. It's considered a universal modeling system that builds a map of the world, with motivations and desires residing in other brain regions.

Most predictions occur within individual neurons, specifically pyramidal cells, through 'dendritic spikes.' These internal spikes don't leave the neuron but signal its expectation to become active shortly. Neurons in a predictive state fire a few milliseconds sooner, influencing network behavior and creating different representations in the brain.

Reference frames are essential for the brain to make predictions by representing the location of sensory input relative to an object or concept. The brain adapts the ancient mammalian ability to map physical space to create internal 'reference frames' for everything, from physical objects like a coffee cup to abstract concepts like mathematics or history.

The brain models the world's inherent hierarchical structure. Just as a water bottle has a cap, logo, and letters, so too do abstract concepts build upon integrated information. Each cortical column learns these hierarchical structures, allowing the same mechanism to model both simple physical objects and complex conceptual ones like history or democracy.

Hawkins distinguishes between AI's potential dangers and existential risks. He believes AI is a dangerous technology if misused, but the intelligence itself does not pose an existential threat because AI systems, when engineered, will not inherently develop human-like desires or agency to try to destroy humanity. The real existential risk lies with self-replication, whether physical or virtual, and humans' potential to create such systems or misuse technology.

Hawkins envisions AI as a tool for humanity to transcend biological limitations. Beyond practical applications like smart cars and household robots, highly intelligent machines could help solve global challenges, build habitats on other planets, and explore the universe as intelligent 'offspring' that preserve and expand human knowledge.

Hawkins suggests pragmatic approaches like archiving all human knowledge on satellites orbiting the sun, regularly updating platforms like Wikipedia. For extreme long-term preservation and communication with alien civilizations, he proposes creating non-active, persistent signals visible from interstellar distances, like objects orbiting our sun that create distinct light-dimming patterns.

He is disappointed by humanity's difficulty in separating rational thought from primal evolutionary heritage, which can manifest in negative behaviors. He also laments our inherent susceptibility to false beliefs, especially regarding information not directly verifiable, leading to dogma rather than skepticism. He advocates for universal education on how the brain constructs models of reality to foster critical thinking.

Hawkins' company is applying brain principles to current machine learning, such as leveraging neuronal sparsity to make existing deep learning networks 10 to 100 times faster and more robust. They are also working on adding 'dendrite-like' complexity to point neurons in neural networks to enable continuous and rapid learning, similar to biological systems.

He advises finding something one is truly passionate about. While his own passion was large, he notes that intense joy can be found in smaller pursuits. Passion is crucial for overcoming obstacles and setbacks. He also emphasizes patience, as dreams can be pursued over a lifetime, even after temporary detours.