A Thousand Brains: A New Theory of Intelligence
By Jeff Hawkins
Basic Books, 2021
If you’re interested in artificial intelligence (and you should be,) it behooves you to learn about intelligence in general. While there’s still lots to learn, neuroscience has made lots of progress in the last few decades. This book offers a compelling new theory of how we think.
Hawkins is a tech entrepreneur (i.e., he founded Palm Computing.) But his passion is neuroscience. He’s on a quest to understand how intelligence works, and this book explains what he and his team have found. It’s divided into three parts, with the first focused on their “Thousand Brains” theory.
It starts by explaining what we know about the brain’s architecture. The brain is a complex organ composed of subsystems. Its older parts are responsible for baseline features such as breathing and walking. The neocortex is a newer part that is responsible for more complex tasks, such as reading and talking.
Hawkins recaps two general tenets underlying all of this:
- Thoughts, ideas, and perceptions are the activity of neurons
- Everything we know is stored in the connections between neurons
Physically, the neocortex is composed of around 150,000 cortical columns, modular units responsible for somewhat independent tasks in the brain. For a long time, people didn’t fully grok the role of these columns in thinking and perceiving. Hawkins and his team made three discoveries:
- The neocortex learns a predictive model of the world
- Predictions occur inside neurons
- The secret of the cortical column is reference frames
Specifically, each column learns models of objects and concepts based on sensory inputs. The brain learns models by observing change in inputs over time. For example, moving around a space lets you perceive its boundaries from different perspectives. As you do, you create a sort of mental map of the space. The brain uses these models to make predictions about the world.
Each cortical column develops models independently of other columns. It’s a decentralized model of understanding that contrasts with the more traditional hierarchical model.
Reference frames are like maps that set objects in context so the brain can understand how things relate to each other. When you grab a cup, your brain uses reference frames for both the cup and your hand. Your senses provide input on where either is in relation to the other. Reference frames let you track both the cup’s and hand’s locations and features as they (and you) move in space.
Each cortical column builds a spatial and conceptual reference frame for things you encounter in the world. The brain integrates these disparate models into a cohesive understanding of objects, environments, and abstract concepts – including information coming in through the senses.
Having learned the attributes and expected capabilities of things (e.g., hands and cups) over time, your brain can predict what will happen when you act on them (or with them) in various ways. Which is to say, reference frames are essential to learning, understanding, and acting. They’re how the brain models the world so you can act skillfully.
Part two of the book explores the theory’s implications for building artificially intelligent systems. A key takeaway: current approaches AI won’t get us to “true” intelligence (my term, not Hawkins’s) since they lack embodied reference frames.
Hawkins believes that artificial systems that aim to function like our brains would need to provide analogs to this distributed structure, even if their sensory and actuating mechanisms were wildly different from ours. AIs can’t achieve human-level general intelligent absent reference frame-based models.
Part three explores the theory’s social implications. Hawkins is concerned with preserving intelligence in a world that is 1) on track to develop artificial intelligent systems while 2) destroying the environment. Intelligence is a fragile phenomenon that could disappear, so Hawkins argues for “estate planning for humanity” – i.e., finding resilient ways to perpetuate intelligence and its accomplishments.
As you may surmise from these notes, the book gets progressively more speculative as it goes. Part one, which is grounded on solid evidence, is the most informative. By part three, the book has shifted to speculative/philosophical advocacy.
The “Thousand Brains” theory has important implications for both AI and design. For one thing, it grounds the concept of mental models on neuroscience. We do indeed carry around “maps” in our brains that help us act and decide in the world – something that designers have known empirically all along.
For another, these models emerge from our experiences as embodied beings. Our models of hands and coffee cups are only relevant to beings that share our physical, sensory, and neural characteristics. An ant would have a very different model of a cup, for example.
This second point raises questions about the ability of current AIs to supplant humans in many (most?) activities. LLMs in particular work by detecting patterns in language that may create something like models, but they’re unlike the models in embodied beings with reference frame-driven architectures.
Which is to say, the “Thousand Brains” theory suggests current AI architectures might not lead to what most of us imagine as AGI. That doesn’t mean they’re useless, but they’re definitely different. We have lots to learn about how to work effectively with these things.
The theory is still relatively new and not yet widely accepted. But it represents an intriguing shift in how we think about how we think. The idea that intelligence emerges from independent yet cooperative modular units has intriguing implications for AI and beyond.
As a non-specialist, I found the book compelling and insightful. At a minimum, it offers fascinating thoughts about the preciousness and fragility of attention and intelligence in a world bent on commoditizing both.