Thousand Brains — book review
Thousand Brains by Jeff Hawkins consists of three equally sized parts covering topics that are connected only by virtue of Jeff having interest in them.
Conceptually, the book should be read backwards, starting with the third part — what is the purpose of humankind? What can we do? We dominate the Earth because of our intellect that enhanced us in so many areas beyond what we are physically capable of. Jeff as an idealist cares about what it means to have existed — in his opinion that boils down to knowledge, to things we have discovered. Passing it on or expanding it requires something that is not human.
Enters part two, of artificial intelligence. Jeff’s opinion is that current takes on AI lack flexibility / context variability that is the true marker of intelligence. He argues that _general purpose_ AI is the way to go — but for that we lack the good architecture.
Get Kate’s stories in your inbox
Join Medium for free to get updates from this writer.
Enters part one, neocortex, the true gem of this book (at least for me). It’s extremely fascinating — and very easy to read. The idea is that we have 150k units of modeling the world (cortical columns), each of which is modeling its own part of the sensed world, the consensus is by voting, accepted model is the basis for predictions of the next sensory input. A model of one column is essentially a bunch of reference frames describing different phenomena and a column goes via each frame one by one trying to find a match with the observed sensory input. Jeff does a great job of explaining it all using the analogy of a paper map cut into pieces (reference frames) and different ppl dropped in different locations (sensory input) trying to make sense of their location (“I’m by the fountain, must be D7 of Paris!”) and then voting on the result.
There are multiple corollaries from this theory that make me uncomfortably excited. (I must admit, I do not share Jeff’s passion for AI.)
First, the one that’s mentioned in passing by Richard Dawkins in the foreword — thinking is a form of movement — comes from the assumption that our neocortex is the result of evolution of “how to track food” need. E.g., when you are an amoeba, you choose a direction and move over there if the observed gradient of “tastes yummy” changes in that direction. If not, choose another direction and try again. With neocortex, you can remember where you saw the food and if it was dangerous, come there back again and have it later. You’re capable of having a map of your environment and use to to navigate. Thinking as a process while takes place in the conceptual plane has the same underlying mechanism — so, thinking makes sense only if it leads you somewhere. It is quite simple and I believe more or less universally recognized in psychology (rumination, negative automatic thoughts, etc.) but I’m personally excited to see that to be grounded in the functional/historical definition of thinking.
Another corollary that I see immensely important is reference frames are essentially dimensions / axes chosen by the brain to represent an observed phenomenon. In the book, there’s an example of ppl doing some mental tasks involving images of birds that differed by the length of their neck and legs — fMRI experiments showed that neocortex stored those images in a manlike reference frame — one dimension represented neck length, another — leg length. Which is amazing. We all break down the world / create projections of it — and if we use the same dimensions/axes, we are bound to have understanding! Except we don’t. (I strongly suspect it’s because some of the axes are rooted in individual experience as opposed to objective parameters like length of the neck/legs.) But it’s still very very promising. We now know for sure why we don’t understand each other. (Basically, an insight in this framework is re-basing / moving to new dimensions/axes.)
For me, this is the most interesting part. And I will go as far as to say that Jeff doesn’t seem to see that one of his two end goals — sharing the knowledge with the extraterrestrial intelligence — requires understanding which kind of bases / reference frames they might be using, because otherwise sending them off our copy of Wikipedia is like giving someone with Linux desktop a binary compiled for Windows 95.
